Chemical molecule information generation method and device, equipment and storage medium

Abstract

The present disclosure relates to a chemical molecule information generation method, device, equipment and storage medium. The method comprises: obtaining a pre-constructed chemical reaction knowledge graph; for each of at least one preset chemical conversion rule, extracting at least one target chemical conversion relationship corresponding to the preset chemical conversion rule from the chemical reaction knowledge graph; based on the at least one target chemical conversion relationship corresponding to the preset chemical conversion rule, constructing a reaction-constrained action space corresponding to the preset chemical conversion rule, wherein each action in the action space represents a chemical conversion operation between molecules; using a pre-trained generative flow network, generating a molecular structure and a synthesis path of a target molecule based on the action space corresponding to the at least one preset chemical conversion rule. Thus, the joint generation of the molecular structure and the synthesis path can be realized.

Description

Technical Field

[0001] This disclosure relates to the fields of machine learning and chemical information technology, and more specifically, to a method, apparatus, device, and storage medium for generating chemical molecular information. Background Technology

[0002] In recent years, deep learning-based molecular generation techniques have shown great potential in drug screening and materials design. Generative flow networks, as a novel probabilistic modeling method, can sample diverse candidate molecules based on reward distributions, offering unique advantages in solving structure optimization problems. However, existing molecular generation methods still have significant limitations in practical applications. For example, synthetic availability is often disconnected from the generation process. Traditional action-based generation methods fail to reflect the chemist's thought process of navigating chemical space, resulting in generated molecules with potential activity but lacking a clear synthetic pathway, significantly increasing the cost of subsequent experimental verification.

[0003] Therefore, how to achieve the joint generation of molecular structure and synthetic pathway is a technical problem worthy of attention. Summary of the Invention

[0004] The main purpose of this disclosure is to provide a method, apparatus, device and storage medium for generating chemical molecular information, so as to solve the technical problem of how to achieve the joint generation of molecular structure and synthetic pathway in the prior art, and to achieve the technical effect of joint generation of molecular structure and synthetic pathway.

[0005] To achieve the above objectives, a first aspect of this disclosure proposes a method for generating chemical molecular information, the method comprising: Obtain a pre-constructed chemical reaction knowledge graph, where nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules; For each of the at least one preset chemical transformation rules, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph; based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, construct the reaction-constrained action space corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules; A pre-trained generative flow network is used to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule.

[0006] In some possible implementations, the nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes; and From the chemical reaction knowledge graph, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, including: Knowledge graph embedding learning is performed on the chemical reaction knowledge graph to obtain multiple reaction template embeddings and multiple environmental constraint embeddings. The reaction template embedding represents the topological transformation logic contained in the reaction template node, and the environmental constraint embedding represents the reaction reagent node and the reaction condition node. The correlation between the reaction template embedding and the environmental constraint embedding is calculated, and multiple correlation degrees are obtained. Based on multiple correlation degrees, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule is determined.

[0007] In some possible implementations, based on at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed, including: From at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, determine at least one candidate chemical transformation relationship that matches the current molecular state in the generation process of the target molecule; For each candidate chemical transformation relationship in at least one candidate chemical transformation relationship: if the reaction template corresponding to the candidate chemical transformation relationship is a unimolecular reaction template, then the preset chemical transformation rule corresponding to the candidate chemical transformation relationship is instantiated as a unimolecular reaction action; if the reaction template corresponding to the candidate chemical transformation relationship is a multimolecular reaction template, then a common reactant matching the current molecular state and the candidate chemical transformation relationship is retrieved from the preset raw material library, and the combination of the preset chemical transformation rule and the common reactant is instantiated as a multimolecular reaction action; Based on the single-molecule and multi-molecule reaction actions corresponding to each candidate chemical transformation relationship, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed.

[0008] In some possible implementations, the molecular structure and synthetic pathway of the target molecule are generated based on the action space corresponding to at least one preset chemical transformation rule, including: Using a molecular encoder trained on a chemical reaction knowledge graph, the current molecular state during the generation process of the target molecule is mapped to a molecular state embedding in a vector space. Using a template encoder trained on a chemical reaction knowledge graph, each candidate reaction template is mapped to a reaction template embedding in a vector space; Based on the molecular state embedding and the reaction template embedding corresponding to each candidate reaction template, the probability of selecting each candidate reaction template is determined. For each candidate reaction template, if the candidate reaction template is a multi-molecule reaction template, the fusion feature is determined based on the molecular state embedding and the reaction template embedding. The matching degree between the fusion feature and the common reactant embedding vector is calculated, and the conditional probability of selecting the common reactant is determined based on the matching degree. For a single-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the single-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template; for a multi-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the multi-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting the common reactant. Based on the selection probabilities of single-molecule reaction actions and multi-molecule reaction actions, actions are selected from the action space corresponding to at least one preset chemical transformation rule, and the current molecular state is updated to generate the molecular structure and synthetic pathway of the target molecule.

[0009] In some possible implementations, the generating stream network is trained in the following manner: A loss function is constructed based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state. While keeping the parameters of the molecular encoder and the reaction template encoder pre-trained based on the chemical reaction knowledge graph frozen, the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix parameters in the decision-making process in the generative flow network are updated by minimizing the loss function, thus obtaining the trained generative flow network.

[0010] In some possible implementations, the reaction template includes a forward reaction template and a reverse reaction template, which represent chemical transformation processes with opposite reaction directions; and Before constructing the loss function based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state, the method also includes: For any state transition subprocess in the generation process of the target molecule: Determine the successor molecular state and the predecessor molecular state in this state transition subprocess; Based on the subsequent molecular state, candidate reverse reaction templates matching the subsequent molecular state are determined. Based on the matching relationship between the subsequent molecular state and the candidate reverse reaction template, the reverse selection probability corresponding to each candidate reverse reaction template is determined. From the candidate reverse reaction templates, the target reverse reaction template for converting the subsequent molecular state into the precursor molecular state is determined. Based on the reverse selection probability corresponding to the target reverse reaction template, the reverse state transition probability corresponding to the state transition subprocess is determined.

[0011] In some possible implementations, the chemical reaction knowledge graph is represented by a multi-relation heterogeneous graph, which contains multiple nodes and edges representing multiple semantic relations; The nodes include: reaction node, reactant node, product node, reagent node, reaction condition node, and reaction template node; Semantic relations include: participation relations, auxiliary relations, and mapping relations; Participation relationships characterize the compositional associations between reactant nodes, product nodes, and reaction nodes; Auxiliary relationships characterize the environmental constraint relationships between reaction reagent nodes, reaction condition nodes, and reaction nodes; The mapping relationship represents the rule attribution relationship between reaction nodes and reaction template nodes.

[0012] Secondly, embodiments of this disclosure provide an apparatus for generating chemical molecule information, the apparatus comprising: The acquisition unit is configured to: acquire a pre-constructed chemical reaction knowledge graph, wherein nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules; The extraction unit is configured to: extract at least one target chemical transformation relationship corresponding to each preset chemical transformation rule from a chemical reaction knowledge graph for each preset chemical transformation rule; and construct a reaction-constrained action space corresponding to the preset chemical transformation rule based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules; The generation unit is configured to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule, using a pre-trained generation flow network.

[0013] In some possible implementations, the nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes; and From the chemical reaction knowledge graph, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, including: Knowledge graph embedding learning is performed on the chemical reaction knowledge graph to obtain multiple reaction template embeddings and multiple environmental constraint embeddings. The reaction template embedding represents the topological transformation logic contained in the reaction template node, and the environmental constraint embedding represents the reaction reagent node and the reaction condition node. The correlation between the reaction template embedding and the environmental constraint embedding is calculated, and multiple correlation degrees are obtained. Based on multiple correlation degrees, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule is determined.

[0014] In some possible implementations, based on at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed, including: From at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, determine at least one candidate chemical transformation relationship that matches the current molecular state in the generation process of the target molecule; For each candidate chemical transformation relationship in at least one candidate chemical transformation relationship: if the reaction template corresponding to the candidate chemical transformation relationship is a unimolecular reaction template, then the preset chemical transformation rule corresponding to the candidate chemical transformation relationship is instantiated as a unimolecular reaction action; if the reaction template corresponding to the candidate chemical transformation relationship is a multimolecular reaction template, then a common reactant matching the current molecular state and the candidate chemical transformation relationship is retrieved from the preset raw material library, and the combination of the preset chemical transformation rule and the common reactant is instantiated as a multimolecular reaction action; Based on the single-molecule and multi-molecule reaction actions corresponding to each candidate chemical transformation relationship, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed.

[0015] In some possible implementations, the molecular structure and synthetic pathway of the target molecule are generated based on the action space corresponding to at least one preset chemical transformation rule, including: Using a molecular encoder trained on a chemical reaction knowledge graph, the current molecular state during the generation process of the target molecule is mapped to a molecular state embedding in a vector space. Using a template encoder trained on a chemical reaction knowledge graph, each candidate reaction template is mapped to a reaction template embedding in a vector space; Based on the molecular state embedding and the reaction template embedding corresponding to each candidate reaction template, the probability of selecting each candidate reaction template is determined. For each candidate reaction template, if the candidate reaction template is a multi-molecule reaction template, the fusion feature is determined based on the molecular state embedding and the reaction template embedding. The matching degree between the fusion feature and the common reactant embedding vector is calculated, and the conditional probability of selecting the common reactant is determined based on the matching degree. For a single-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the single-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template; for a multi-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the multi-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting the common reactant. Based on the selection probabilities of single-molecule reaction actions and multi-molecule reaction actions, actions are selected from the action space corresponding to at least one preset chemical transformation rule, and the current molecular state is updated to generate the molecular structure and synthetic pathway of the target molecule.

[0016] In some possible implementations, the generating stream network is trained in the following manner: A loss function is constructed based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state. While keeping the parameters of the molecular encoder and the reaction template encoder pre-trained based on the chemical reaction knowledge graph frozen, the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix parameters in the decision-making process in the generative flow network are updated by minimizing the loss function, thus obtaining the trained generative flow network.

[0017] In some possible implementations, the reaction template includes a forward reaction template and a reverse reaction template, which represent chemical transformation processes with opposite reaction directions; and Before constructing the loss function based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state, the method also includes: For any state transition subprocess in the generation process of the target molecule: Determine the successor molecular state and the predecessor molecular state in this state transition subprocess; Based on the subsequent molecular state, candidate reverse reaction templates matching the subsequent molecular state are determined. Based on the matching relationship between the subsequent molecular state and the candidate reverse reaction template, the reverse selection probability corresponding to each candidate reverse reaction template is determined. From the candidate reverse reaction templates, the target reverse reaction template for converting the subsequent molecular state into the precursor molecular state is determined. Based on the reverse selection probability corresponding to the target reverse reaction template, the reverse state transition probability corresponding to the state transition subprocess is determined.

[0018] In some possible implementations, the chemical reaction knowledge graph is represented by a multi-relation heterogeneous graph, which contains multiple nodes and edges representing multiple semantic relations; The nodes include: reaction node, reactant node, product node, reagent node, reaction condition node, and reaction template node; Semantic relations include: participation relations, auxiliary relations, and mapping relations; Participation relationships characterize the compositional associations between reactant nodes, product nodes, and reaction nodes; Auxiliary relationships characterize the environmental constraint relationships between reaction reagent nodes, reaction condition nodes, and reaction nodes; The mapping relationship represents the rule attribution relationship between reaction nodes and reaction template nodes.

[0019] Thirdly, embodiments of this disclosure provide an electronic device, including: Memory, used to store computer programs; A processor is configured to execute a computer program stored in the memory, wherein, when the computer program is executed, it implements the method of any embodiment of the method for generating chemical molecule information according to the first aspect of this disclosure.

[0020] Fourthly, embodiments of this disclosure provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any embodiment of the method for generating chemical molecule information as described in the first aspect above.

[0021] Fifthly, embodiments of this disclosure provide a computer program product including computer-readable code, which, when executed by a processor, implements the method of any embodiment of the method for generating chemical molecule information in the first aspect described above.

[0022] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects: In this disclosure, by acquiring a pre-constructed chemical reaction knowledge graph, and defining nodes in the knowledge graph as molecular entities and reaction templates, edges as relationships between nodes, and reaction templates as chemical transformation rules, a structured expression of the reaction knowledge contained in molecular entities, reaction templates, and relationships between nodes can be achieved. By extracting at least one target chemical transformation relationship corresponding to each of at least one preset chemical transformation rule from the chemical reaction knowledge graph, a correspondence can be established between the preset chemical transformation rules and the target chemical transformation relationships in the knowledge graph. By constructing a reaction-constrained action space corresponding to the preset chemical transformation rule based on the at least one target chemical transformation relationship, and defining each action in the action space as a chemical transformation operation between molecules, the selectable action range of the generative flow network can be limited to the range of chemical transformation operations corresponding to the preset chemical transformation rule. By employing a pre-trained generative flow network, based on the action space corresponding to at least one preset chemical transformation rule, the molecular structure and synthetic pathway of the target molecule can be generated. The molecular structure and synthetic pathway of the target molecule can be generated synchronously in the reaction-constrained action space, reducing the probability of the generation process deviating from the chemical transformation rules. Therefore, based on the relationships between molecular entities, reaction templates, and nodes in the chemical reaction knowledge graph, the target chemical transformation relationship corresponding to the preset chemical transformation rule can be converted into an action space constrained by the reaction. The pre-trained generation flow network can then generate the molecular structure and synthetic path of the target molecule based on the action space corresponding to at least one preset chemical transformation rule. This not only enables the joint generation of molecular structure and synthetic path but also improves the correspondence between the molecular structure and synthetic path of the target molecule, as well as the chemical rationality and generation effectiveness in the molecular generation process. Attached Figure Description

[0023] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of the disclosure and to make other features, objects, and advantages of the disclosure more apparent. The illustrative embodiments of the disclosure, along with their descriptions, are used to explain the disclosure and do not constitute an undue limitation thereof. In the drawings: Figure 1 A flowchart illustrating a method for generating chemical molecule information provided in this disclosure embodiment; Figure 2 A flowchart illustrating the process of constructing a reaction-constrained action space corresponding to a preset chemical transformation rule in a method for generating chemical molecule information provided in this embodiment of the present disclosure; Figure 3A A schematic diagram of the entity and relationship structure of a chemical reaction knowledge graph in a method for generating chemical molecular information provided in this embodiment of the present disclosure; Figure 3B A schematic diagram illustrating the decision-making process in a method for generating chemical molecular information provided in this embodiment of the disclosure; Figure 3C This is an example schematic diagram illustrating the joint generation of molecules and synthetic pathways in a method for generating chemical molecule information provided in this disclosure embodiment; Figure 4 A schematic diagram of a chemical molecule information generation device provided in an embodiment of this disclosure; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present disclosure, the technical solutions of the present disclosure will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present disclosure, and not all embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present disclosure.

[0025] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0026] In this disclosure, the terms "upper," "lower," "left," "right," "front," "rear," "top," "bottom," "inner," "outer," "middle," "vertical," "horizontal," "lateral," and "longitudinal" indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings. These terms are primarily for the purpose of better describing this disclosure and its embodiments, and are not intended to limit the indicated devices, elements, or components to having a specific orientation, or to be constructed and operated in a specific orientation.

[0027] Furthermore, in addition to indicating direction or positional relationship, some of the aforementioned terms may also have other meanings. For example, the term "above" may also be used in certain circumstances to indicate a dependency or connection. Those skilled in the art can understand the specific meaning of these terms in this disclosure according to the specific circumstances.

[0028] Furthermore, the terms "installation," "setup," "equipped with," "connection," "linked," and "socketing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral structure; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, or an internal connection between two devices, components, or parts. Those skilled in the art can understand the specific meaning of the above terms in this disclosure according to the specific circumstances.

[0029] The acquisition and processing of data in this disclosure comply with relevant laws, regulations, social ethics, and provisions concerning public interests. For example, the acquisition, storage, and processing of data in this disclosure were all carried out with the knowledge and consent of the users, and have obtained the full authorization of the relevant users.

[0030] Figure 1 This is a schematic flowchart illustrating a method for generating chemical molecular information according to an embodiment of this disclosure. This method can be applied to one or more electronic devices such as smartphones, laptops, desktop computers, portable computers, and servers. Furthermore, the execution entity of this method can be hardware or software. When the execution entity is hardware, it can be one or more of the aforementioned electronic devices. For example, a single electronic device can execute this method, or multiple electronic devices can cooperate with each other to execute this method. When the execution entity is software, this method can be implemented as multiple software programs or software modules, or as a single software program or software module. No specific limitations are made here.

[0031] like Figure 1 As shown, the method specifically includes: Step 101: Obtain a pre-constructed chemical reaction knowledge graph, where nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules.

[0032] In this embodiment, the chemical reaction knowledge graph can characterize the relationship structure between molecular entities and reaction templates, and between molecular entities themselves.

[0033] In some alternative implementations, chemical reaction knowledge graphs can be constructed based on reaction-related data.

[0034] In some optional implementations of this embodiment, the chemical reaction knowledge graph is represented by a multi-relation heterogeneous graph, which contains multiple nodes and edges representing multiple semantic relationships.

[0035] Among them, a multi-relation heterogeneous graph can be a graph structure that contains nodes of various types and edges with various semantics.

[0036] The nodes include the following types: reaction nodes, reactant nodes, product nodes, reagent nodes, reaction condition nodes, and reaction template nodes.

[0037] Reactant nodes can represent reactants that participate in the reaction.

[0038] Product nodes can represent the products generated by a reaction.

[0039] The reagent node can represent the reagents used in a chemical reaction, such as catalysts, solvents, oxidants, and reducing agents.

[0040] Reaction condition nodes can represent reaction temperature, reaction pressure, etc.

[0041] The reaction template node can represent a reaction template. As an example, the reaction template could be "alkene + water → alcohol".

[0042] The semantic relationships represented by edges include: participation relationship, auxiliary relationship, and mapping relationship.

[0043] Optionally, edges can also represent logical relationships such as participation, generation, requirement, influence, and application.

[0044] Participation relationships characterize the compositional associations between reactant nodes, product nodes, and reaction nodes.

[0045] The auxiliary relationship characterizes the environmental constraint association between the reaction reagent node, the reaction condition node, and the reaction node.

[0046] The mapping relationship represents the rule attribution relationship between reaction nodes and reaction template nodes.

[0047] Reaction nodes can represent reaction events or reaction instances. Reaction nodes can form compositional associations with reactant nodes and product nodes through participation relationships; reaction nodes can form environmental constraint associations with reagent nodes and reaction condition nodes through auxiliary relationships; and reaction nodes can form rule affiliation relationships with reaction template nodes through mapping relationships.

[0048] Reactant nodes represent reactants that participate in a reaction. Reactant nodes can form compositional relationships with product nodes and reaction nodes through participation relationships.

[0049] Product nodes can represent the products generated in a reaction. Product nodes can form compositional associations with reactant nodes and reaction nodes through participation relationships.

[0050] Participation relationships can characterize the compositional associations between reactant nodes, product nodes, and reaction nodes. As an example, participation relationships can include logical relationships such as "generate".

[0051] Auxiliary relationships can characterize the environmental constraint associations between reagent nodes, reaction condition nodes, and reaction nodes. As an example, auxiliary relationships can include logical relationships such as "requirements" and "influences".

[0052] Mapping relationships can characterize the rule attribution relationship between reaction nodes and reaction template nodes. As an example, mapping relationships can include logical relationships such as "applies to".

[0053] Compositional relationships can represent the compositional relationships between reactant nodes, product nodes, and reaction nodes. Compositional relationships can characterize which reactants participate in a reaction and which products they correspond to.

[0054] Environmental constraint correlations can represent the constraint relationships between reagent nodes, reaction condition nodes, and reaction nodes. They characterize the environmental constraints imposed on the reaction by reagents and reaction conditions.

[0055] Rule attribution relationships can represent the attribution relationship between reaction nodes and reaction template nodes. Rule attribution relationships can establish a correspondence between specific reaction nodes and the chemical transformation rules represented by the reaction template nodes.

[0056] It is understandable that using a multi-relation heterogeneous graph to represent a chemical reaction knowledge graph, and including multiple types of nodes and edges representing multiple semantic relationships, allows for a structured expression of different types of nodes and semantic relationships within the chemical reaction knowledge graph. Nodes, including reaction nodes, reactant nodes, product nodes, reagent nodes, reaction condition nodes, and reaction template nodes, enable a node-based expression of reaction composition, reaction environment, and reaction template. Semantic relationships, including participation relationships, auxiliary relationships, and mapping relationships, allow for the differentiation and expression of different types of relationships between nodes. Participation relationships characterize the compositional associations between reactant nodes, product nodes, and reaction nodes, allowing reaction composition information to be expressed within the chemical reaction knowledge graph. Auxiliary relationships characterize the environmental constraint associations between reagent nodes, reaction condition nodes, and reaction nodes, allowing reaction environment constraint information to be expressed within the chemical reaction knowledge graph. Mapping relationships characterize the rule attribution relationships between reaction nodes and reaction template nodes, enabling a corresponding expression between the chemical transformation rules represented by the reaction nodes and the reaction template nodes. Therefore, chemical reaction knowledge graphs simultaneously express reaction composition information, reaction environment constraint information, and rule attribution information in the form of multi-relationship isomer diagrams, which can further improve the chemical rationality and generation efficiency of molecular generation processes.

[0057] In some alternative approaches, chemical reaction data can be extracted from the Open Reaction Database (ORD) and a chemical reaction knowledge graph can be constructed based on the chemical reaction data.

[0058] Molecular entities can represent reactants, products, reagents, starting materials, common reactants, intermediates, or target molecules, etc.

[0059] A reaction template can represent a chemical transformation rule; that is, a reaction template can be an abstract rule description of a specific chemical transformation logic. As an example, a reaction template could be "alkene + water → alcohol".

[0060] In some optional implementations, the reaction template may include candidate reaction templates, forward reaction templates, reverse reaction templates, etc.

[0061] Since the reaction templates in the chemical reaction knowledge graph are extracted from reaction data, there are a large number of them. Therefore, a portion of these reaction templates can be selected as candidate reaction templates for the model (i.e., the generation flow network) to choose from. As an example, a candidate reaction template could be an "esterification reaction template," corresponding to the chemical transformation rule of "-COOH and -OH undergoing dehydration condensation to generate -COO-."

[0062] Forward and reverse reaction templates are relative terms. They represent chemical transformation processes with opposite reaction directions. For example, if the forward reaction template represents an esterification reaction, the corresponding reverse reaction template can represent an ester hydrolysis reaction; conversely, if the forward reaction template represents an ester hydrolysis reaction, the corresponding reverse reaction template can represent an esterification reaction.

[0063] Furthermore, the reaction template and the chemical transformation rule can have a one-to-one relationship.

[0064] Building blocks can be starting small molecules that can be used for molecular synthesis, such as basic chemical raw materials. Building blocks can be the starting point and basic unit of molecular generation processes.

[0065] Molecular generation can be described as the process of generating a target molecule. Specifically, molecular generation can start from an empty state or starting materials, and through a series of state transition steps, gradually generate intermediates or target molecules.

[0066] Common reactants can be other reagents or molecules that react with the current molecule during a chemical reaction. Common reactants are not the starting point of a synthetic pathway, but rather reactants added during the reaction process.

[0067] Step 102: For each preset chemical transformation rule in at least one preset chemical transformation rule, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph; based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, construct the reaction-constrained action space corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules.

[0068] In this embodiment, the preset chemical transformation rule can represent a predetermined chemical transformation rule.

[0069] Chemical transformation relationships can represent the logic of transformation from reactants to products. One chemical transformation rule can correspond to multiple instances of chemical transformation relationships. As an example, the chemical transformation rule "olefin + water → alcohol" can correspond to chemical transformation relationships such as "ethylene + H2O → ethanol", "propylene + H2O → propanol", and "butene + H2O → butanol".

[0070] The target chemical transformation relationship can be a chemical transformation relationship extracted from a chemical reaction knowledge graph that corresponds to the preset chemical transformation rule. One preset chemical transformation rule can correspond to at least one target chemical transformation relationship.

[0071] Each chemical transformation rule corresponds to an action space, and different chemical transformation rules correspond to different action spaces.

[0072] The reaction-constrained action space can be a set or range of actions constructed based on at least one target chemical transformation relationship corresponding to the preset chemical transformation rule. Each action in the action space represents an intermolecular chemical transformation operation, such as deprotecting a tert-butyloxycarbonyl (Boc)-protected amine to generate a free amine; or the operation of "-COOH dehydrating and condensing with -NH2 to generate -CONH-".

[0073] Furthermore, the action space can be the range of actions that the generative flow network can choose during molecular generation. The action space can be formed by combinations of candidate reaction templates, building blocks, or common reactants. In related technologies, tiling the action space fails to effectively encode the chemical synthesis logic, while this approach can limit the range of actions that the generative flow network can choose at each state transition step by using a reaction-constrained action space.

[0074] A chemical transformation operation is a specific action represented by an action. A chemical transformation operation can be expressed through a reaction template or a combination of a reaction template and common reactants, and after execution, the current molecular state can be updated to the intermediate or product of the reaction. As an example, a chemical transformation operation can represent the following operation: after obtaining a "Suzuki coupling reaction" and the corresponding "boronic acid building block," the halogenation site of the current molecule can be identified, and a new molecular topology after coupling can be generated. The molecular states during the generation process can include an empty state, intermediate product molecules, and the final target product molecule. After executing the chemical action, the current molecular state can be updated to the intermediate or product of the reaction, and a new molecular topology can be generated.

[0075] In some optional implementations, nodes and edges in the chemical reaction knowledge graph can be searched to extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule. Then, based on the current substructure state of the molecule, a set of matching candidate reaction templates is retrieved from the chemical transformation relationships; common reactants are retrieved based on compatibility associations within the chemical transformation relationships; and single-molecule reaction templates and combinations of multi-molecule reaction templates with corresponding common reactants are instantiated as a set of discrete actions in the current state. These discrete actions collectively constitute a reaction-constrained action space.

[0076] Step 103: Using a pre-trained generative flow network, based on the action space corresponding to at least one preset chemical transformation rule, the molecular structure and synthetic pathway of the target molecule are generated.

[0077] In this embodiment, the generative flow network can be used as a decision network constrained by chemical reaction rules (represented by the action space) at each state transition step. The generative flow network can introduce the reaction-constrained action space into the policy execution process, thereby jointly sampling the molecular structure of the target molecule and its corresponding synthetic path. Specifically, the generative flow network can select actions at each state transition step (also known as a state transition sub-process); the generative flow network can advance molecule generation through a forward policy execution process; the generative flow network can also include a reverse policy execution process to decompose complex molecular states into simpler substructures; the training process of the generative flow network can construct a trajectory balance loss function based on the forward and reverse probabilities of each state transition step and the reward value of the final state.

[0078] In some optional implementations of this embodiment, the generating stream network is trained in the following manner: The first step is to construct a loss function based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state.

[0079] The second step involves freezing the parameters of the molecular encoder and the reaction template encoder pre-trained based on the chemical reaction knowledge graph, and then updating the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix in the decision-making process in the generator flow network by minimizing the loss function, thus obtaining the trained generator flow network.

[0080] The process of generating a target molecule can be the process of generating its molecular structure and synthetic pathway. As an example, this process can start from an empty state or starting materials, and through a series of chemical actions, gradually update the current molecular state until the molecular structure and synthetic pathway of the target molecule are generated.

[0081] A state transition subprocess can be a subprocess in the generation of a target molecule that transitions from one molecular state to another, i.e., the process of state change in a generation flow network.

[0082] The probability of selecting a reaction action can be the probability of the generating flow network selecting a reaction action during the state transition subprocess, that is, the probability of selecting the complete action of the current state in the positive policy.

[0083] The reverse state transition probability can be the probability of returning from the successor molecule state to the predecessor molecule state, that is, the reverse probability in the execution of the reverse strategy or the probability of selecting the reverse reaction template.

[0084] The reward value corresponding to the terminating molecular state can be the value obtained after the generation process reaches the terminating molecular state.

[0085] The loss function can be a training objective constructed based on the probability of selecting a reaction action, the probability of reverse state transition, and the reward value corresponding to the termination molecular state.

[0086] The parameters of a molecular encoder can be the model parameters from the molecular encoder. A molecular encoder can be used to encode the current molecular state, thereby obtaining the embedding vector of the current molecular state, i.e., the molecular state embedding.

[0087] The parameters of a reaction template encoder can be the model parameters from the reaction template encoder. A reaction template encoder can be used to encode reaction templates to obtain the embedding vector of the reaction template, i.e., the reaction template embedding.

[0088] The parameters of the forward policy network can be learnable parameters of the forward policy network in the generative flow network. The forward policy network can be used for the synthetic evolution logic of molecules.

[0089] The parameters of the backward policy network can be learnable parameters of the backward policy network in the generating flow network. The backward policy network can be used for reverse backtracking analysis of the synthesis path.

[0090] The linear transformation matrix parameters can be trainable parameters used in the generative flow network to connect the molecular encoder and the policy decision module (including the forward policy network and the backward policy network).

[0091] In some alternative implementations, the positive probability of each state transition step (i.e., the probability of selecting a reaction action) can be used as the basis. And the reverse probability (i.e., the reverse state transition probability). and the reward value of the final state Constructing a trajectory balance loss function Here, Z is a globally learnable variable defined in the trajectory balancing loss function, representing the total flow in the generative flow network. The base of the logarithm can be, for example, e or other non-1 positive numbers. During training, the parameters of the molecular encoder and the reaction template encoder can be frozen without updating them. The goal is to minimize the loss function, updating the parameters of the forward policy network, the backward policy network, and the linear transformation matrix parameters in the decision-making process, thus obtaining the trained generative flow network.

[0092] It is understandable that by constructing a loss function based on the selection probability of reaction actions corresponding to each state transition subprocess in the target molecule generation process, the probability of reverse state transition, and the reward value corresponding to the terminating molecule state, the forward generation process, the reverse backtracking process, and the evaluation of the terminating molecule state can be incorporated into a unified training objective. By minimizing the loss function while keeping the parameters of the molecular encoder and the reaction template encoder pre-trained based on the chemical reaction knowledge graph frozen, the parameter perturbation of the pre-trained embedded representation during training can be reduced. By updating the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix parameters in the decision-making process in the generative flow network, the generative flow network can be trained on the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix parameters in the decision-making process while fixing the parameters of the molecular encoder and the reaction template encoder, thus obtaining the trained generative flow network.

[0093] In some application scenarios of the above-mentioned optional implementation methods, the reaction template includes a forward reaction template and a reverse reaction template, which represent chemical transformation processes with opposite reaction directions.

[0094] Based on this, before constructing the loss function based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state, the following steps can be performed for any state transition subprocess in the target molecule generation process: The first step is to determine the successor molecular state and the precursor molecular state in this state transition process.

[0095] The second step is to determine candidate reverse reaction templates that match the subsequent molecular state, based on the subsequent molecular state.

[0096] The third step is to determine the reverse selection probability of each candidate reverse reaction template based on the matching relationship between the subsequent molecular state and the candidate reverse reaction template.

[0097] The fourth step is to identify the target reverse reaction template from among the candidate reverse reaction templates to transform the subsequent molecular state into the precursor molecular state.

[0098] The fifth step is to determine the reverse state transition probability corresponding to the state transition subprocess based on the reverse selection probability corresponding to the target reverse reaction template.

[0099] The subsequent molecular state can be the molecular state after the transition in any state transition process. For example, the state of the intermediate molecule or the state of the product after the state transition.

[0100] The precursor molecular state can be the molecular state before the transfer in any state transfer process.

[0101] Here, the successor molecular state and the predecessor molecular state are relative terms. Each state transition subprocess can correspond to one successor molecular state and one predecessor molecular state.

[0102] A candidate reverse reaction template is a reverse reaction template that matches the subsequent molecular state. As an example, from all the reverse reaction templates, the molecule represented by the subsequent molecular state can be identified as the reactant reverse reaction template, thus obtaining a reverse reaction template that matches the subsequent molecular state, i.e., a candidate reverse reaction template.

[0103] The reverse selection probability can be the selection probability corresponding to the candidate reverse reaction template. As an example, the reverse selection probability can be obtained by calculating the matching degree (i.e., matching score) between the subsequent molecular state and the candidate reverse reaction template, and then performing mask correction and Softmax normalization.

[0104] The target reverse reaction template is a reverse reaction template determined from the candidate reverse reaction templates and used to transform the subsequent molecular state into the precursor molecular state. As an example, the target reverse reaction template for transforming the subsequent molecular state into the precursor molecular state can be randomly selected from the candidate reverse reaction templates according to the probability distribution of each candidate reverse reaction template.

[0105] The reverse state transition probability can be the reverse transition probability corresponding to the state transition subprocess. As an example, the reverse selection probability corresponding to the target reverse reaction template can be determined as the reverse state transition probability corresponding to that state transition subprocess.

[0106] It is understandable that by limiting the reaction template to include both forward and reverse reaction templates, and making these templates represent chemical transformation processes with opposite reaction directions, the templated expression of both forward generation and reverse backtracking processes can be achieved. By determining the subsequent and precursor molecular states in any state transition sub-process during the generation of the target molecule, the molecular state pairs required for calculating the reverse state transition probability can be determined. By identifying candidate reverse reaction templates that match the subsequent molecular states, the matching between the range of reverse reaction template selection and the subsequent molecular states can be achieved. By determining the reverse selection probability corresponding to each candidate reverse reaction template based on the matching relationship between the subsequent molecular states and the candidate reverse reaction templates, the probabilistic selection of the reverse reaction template can be realized. By identifying the target reverse reaction template from each candidate reverse reaction template to transform the successor molecular state into the predecessor molecular state, and determining the reverse state transition probability corresponding to the state transition subprocess based on the reverse selection probability corresponding to the target reverse reaction template, the reverse state transition probability can be provided for the construction of the loss function. This enables the modeling of the reverse probability from the successor molecular state to the predecessor molecular state in any state transition subprocess, thereby providing the reverse state transition probability for the construction of the loss function in the training process of the generative flow network.

[0107] Here, since each preset chemical transformation rule can correspond to an action space, at least one action space corresponding to at least one preset chemical transformation rule can be obtained.

[0108] The target molecule is the object generated by the generative flow network. The target molecule can be a candidate molecule with a target property distribution, the final target product molecule, or a molecule generated targeting a specific site.

[0109] Molecular structure can represent the structure of a target molecule. As an example, molecular structure can be represented as a two-dimensional molecular diagram.

[0110] A synthetic pathway is a generative route corresponding to the molecular structure of a target molecule. As an example, a synthetic pathway can be a logical chain that starts from an empty state and evolves through a series of ordered chemical actions to reach the target molecule. The synthetic pathway can record the reaction template number used at each step, the identifier (ID) of the building blocks consumed, and the structure of the intermediates generated.

[0111] In some alternative implementations, one can start with an empty molecule or a simple starting material (such as a benzene ring or alkane chain), set target information (such as protein structure, binding pocket features), calculate the probability distribution of each chemical transformation action in the current state through a forward policy network in a pre-trained generative flow network, filter out invalid actions according to preset rules, and select the next action from the action space corresponding to at least one preset chemical transformation rule using a sampling strategy (such as random sampling in GFlowNet or flow-based sampling). Apply the selected chemical transformation rule to generate a new molecular state. Repeat the above steps until the termination condition (such as molecular size limitation, synthesis of the target molecule) is met to obtain the molecular structure of the target molecule, and convert each chemical transformation rule in the generation trajectory into a standard chemical reaction representation (such as SMIRKS, SMARTS), forming the synthetic pathway of the target molecule.

[0112] In some optional implementations of this embodiment, the nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes.

[0113] Based on this, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule can be extracted from the chemical reaction knowledge graph in the following way: The first step is to perform knowledge graph embedding learning on the chemical reaction knowledge graph to obtain multiple reaction template embeddings and multiple environmental constraint embeddings. The reaction template embedding represents the topology transformation logic contained in the reaction template node, and the environmental constraint embedding represents the reaction reagent node and the reaction condition node.

[0114] The second step is to calculate the correlation between the reaction template embedding and the environmental constraint embedding, and obtain multiple correlation degrees.

[0115] The third step is to determine at least one target chemical transformation relationship corresponding to the preset chemical transformation rule based on multiple correlation degrees.

[0116] Knowledge graph embedding learning is the process of mapping nodes and relationships (represented by edges) in a chemical reaction knowledge graph into vector representations. For example, the TransR algorithm can be used to learn semantic vector representations between entity nodes in a unified latent space, resulting in molecular embeddings, reaction template embeddings, and environmental constraint embeddings.

[0117] Reaction template embedding can be a vector representation of a reaction template node after knowledge graph embedding learning. Reaction template embedding represents the topological transformation logic contained within a reaction template node, such as the topological transformation logic of a molecule from its initial state to its target state.

[0118] Environmental constraint embedding is a vector representation obtained after the reaction reagent node and reaction condition node have been learned through knowledge graph embedding. Here, since environmental constraint associations are formed between the reaction reagent node, reaction condition node, and reaction node, the environmental constraint embedding can include semantics related to the reaction environment.

[0119] Topological transformation logic refers to the transformation logic of molecular structures contained in reaction template nodes, such as changes in structural connectivity that occur during chemical transformations. Here, based on reaction template nodes and their corresponding reaction template embeddings, the topological transformation logic of molecules from the initial state to the target state can be extracted.

[0120] Association degree can represent the degree of association between reaction template embedding and environmental constraint embedding. As an example, association degree can be represented by a score. For example, the association score (i.e., association degree) between the reaction template embedding and the embeddings of the reaction reagent node and the reaction condition node can be calculated to extract chemical transformation relationships characterizing the suitability of the template under specific reaction conditions based on the association score.

[0121] In some alternative implementations, the correlation between the reaction template embedding and the environmental constraint embedding can be calculated based on the TransR knowledge graph embedding algorithm. The lower the correlation, the more reasonable the triple (reaction template, environmental constraint relationship, reaction reagent / condition); the higher the correlation, the less reasonable the triple.

[0122] In some optional implementations, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule can be determined by the value of the correlation degree. For example, from the chemical transformation relationships contained in the chemical reaction knowledge graph, the chemical transformation relationships with high correlation degree (e.g., correlation degree greater than a preset threshold), matching the current molecular state, and passing the chemical feasibility verification can be identified as at least one target chemical transformation relationship corresponding to the preset chemical transformation rule.

[0123] It is understandable that, since nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes, a node-based representation of reaction templates, reaction reagents, and reaction conditions can be achieved in the chemical reaction knowledge graph. By performing knowledge graph embedding learning on the chemical reaction knowledge graph, multiple reaction template embeddings and multiple environmental constraint embeddings are obtained. This allows the topological transformation logic, reaction reagent nodes, and reaction condition nodes contained in the reaction template nodes to be converted into computable vector representations. By calculating the correlation degree between the reaction template embeddings and the environmental constraint embeddings, multiple correlation degrees are obtained, which can quantify the degree of correlation between the topological transformation logic contained in the reaction template nodes and the environmental constraints represented by the reaction reagent nodes and reaction condition nodes. By determining at least one target chemical transformation relationship corresponding to the preset chemical transformation rule based on multiple correlation degrees, the topological transformation logic contained in the reaction template nodes and the environmental constraint embeddings can be used together to determine the target chemical transformation relationship. Therefore, it is possible to determine at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph based on the correlation degree between the reaction template embeddings and the environmental constraint embeddings, thereby improving the semantic correspondence between the target chemical transformation relationship and the topological transformation logic, reaction reagent nodes, and reaction condition nodes contained in the reaction template nodes.

[0124] In some optional implementations of this embodiment, the molecular structure and synthetic pathway of the target molecule can be generated based on the action space corresponding to at least one preset chemical transformation rule: The first step involves using a molecular encoder trained on a chemical reaction knowledge graph to map the current molecular state during the generation process of the target molecule into a molecular state embedding in a vector space.

[0125] The second step involves using a template encoder trained on a chemical reaction knowledge graph to map each candidate reaction template into a reaction template embedding in a vector space.

[0126] The third step is to determine the probability of selecting each candidate reaction template based on the molecular state embedding and the reaction template embedding corresponding to each candidate reaction template.

[0127] The fourth step involves determining the fusion features based on the molecular state embedding and reaction template embedding for each candidate reaction template. If the candidate reaction template is a multi-molecule reaction template, the matching degree between the fusion features and the common reactant embedding vector is calculated, and the conditional probability of selecting the common reactant is determined based on the matching degree.

[0128] Fifth, for a single-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the single-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template; for a multi-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the multi-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting the common reactant.

[0129] The sixth step involves selecting actions from the action space corresponding to at least one preset chemical transformation rule based on the selection probabilities of single-molecule reaction actions and multi-molecule reaction actions, and updating the current molecular state to generate the molecular structure and synthetic pathway of the target molecule.

[0130] A molecular encoder can be an encoder trained based on a chemical reaction knowledge graph. It can be used to map the current molecular state during the molecular generation process to a molecular state embedding in a vector space. The molecular state embedding can be a representation of the current molecular state in the vector space.

[0131] A template encoder can be another encoder trained based on a chemical reaction knowledge graph. The template encoder can be used to map reaction templates to reaction template embeddings in a vector space. A reaction template embedding can be a representation of a candidate reaction template in the vector space.

[0132] As an example, reaction template embedding can be matrix-dot productd with molecular state embedding to measure the matching score between the current molecular state and the candidate reaction template.

[0133] The vector space is the space where molecular state embedding and reaction template embedding reside.

[0134] Fusion features can be features obtained by fusing molecular state embedding and reaction template embedding.

[0135] A common reactant embedding vector can be a representation of a common reactant in a vector space. As an example, a common reactant can be a building block from a pre-defined raw material library. The building blocks can be used to form semantic embedding vectors through offline feature extraction.

[0136] Matching degree can be the degree of matching between fused features and common reactant embedding vectors.

[0137] The conditional probability of selecting common reactants can be the probability determined based on the degree of matching when the candidate reaction template is a multi-molecular reaction template.

[0138] The probability of selecting a single-molecule reaction action can be a probability determined for that single-molecule reaction action.

[0139] The probability of selecting a multimolecular reaction action can be a probability determined for that multimolecular reaction action.

[0140] Updating the current molecular state can mean changing the current molecular state after performing a selected action. For example, after performing an action (reaction) in the action space, the current molecular state can be updated to the intermediate or product of the reaction, and the action and the corresponding reactant information can be recorded in the synthetic pathway.

[0141] In some alternative implementations, molecular state embeddings can be computed. Reaction template embedding with each candidate reaction template Match score Where W1 represents the trainable matrix. W1 and between, and Both operations involve dot product. , Vector representation is used. This is combined with a validity mask generated by the substructure matching operator. Correct the match score, and apply the corrected match score. After performing the Softmax operation, the candidate reaction template is determined. probability Among them, the mask Indicates the current state s t Can the candidate reaction template r occur? i The sequence represents a series of reactions; if a reaction can occur, the mask is 1; otherwise, the mask is 0, and a match score is assigned. Setting it to negative infinity indicates that selection is prohibited.

[0142] If the selected candidate reaction template As a template for multimolecular reactions, then utilize Reaction template embedding with the selected reaction template The fusion features are used as query vectors, and their correlation with the pre-trained embeddings of common reactants in the initial raw material library (i.e., the preset raw material library) is calculated. The matching score (i.e., the matching degree). Matching score After performing the Softmax operation, the specific common reactants are selected. conditional probability Here, MLP stands for Multilayer Perceptron, which consists of multiple matrix multiplications and nonlinear activations and requires training. · represents a dot product. Combining the probabilities obtained above in a chain determines the complete action for the current state. Selection probability The probability of a unimolecular reaction is determined solely by the first probability. That is, the probability of selecting a unimolecular reaction is the probability of selecting the corresponding candidate reaction template; the probability of selecting a multimolecular reaction is the product of the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting a common reactant.

[0143] Then, based on the selection probability of single-molecule reaction actions and the selection probability of multi-molecule reaction actions, the single-molecule reaction action or multi-molecule reaction action with the highest selection probability can be selected from the action space corresponding to at least one preset chemical transformation rule, and the current molecular state can be updated to generate the molecular structure and synthesis route of the target molecule.

[0144] It is understandable that by using a molecular encoder trained on a chemical reaction knowledge graph, the current molecular state during the generation process of the target molecule can be mapped to a molecular state embedding in a vector space, thus achieving a vectorized expression of the current molecular state. Similarly, by using a template encoder trained on a chemical reaction knowledge graph, each candidate reaction template can be mapped to a reaction template embedding in a vector space, achieving a vectorized expression of each candidate reaction template. Based on the molecular state embedding and the corresponding reaction template embeddings of each candidate reaction template, the probability of selecting each candidate reaction template can be determined, enabling a matching calculation between candidate reaction template selection and the current molecular state. When the candidate reaction template is a multi-molecule reaction template, by determining the fusion feature based on the molecular state embedding and the reaction template embedding, calculating the matching degree between the fusion feature and the common reactant embedding vector, and determining the conditional probability of selecting the common reactant based on the matching degree, the association matching between the selection of the common reactant and the current molecular state and candidate reaction templates in multi-molecule reaction actions can be achieved. By separately determining the selection probability of single-molecule reaction actions and the selection probability of multi-molecule reaction actions, probabilistic selection of different types of reaction actions can be achieved. By selecting actions from the action space corresponding to at least one preset chemical transformation rule based on the selection probabilities of single-molecule and multi-molecule reaction actions, and updating the current molecular state, the state in the generation process of the target molecule can be progressively advanced. Therefore, action probabilities can be determined based on molecular state embedding, reaction template embedding, fusion features, and common reactant embedding vectors, and the molecular structure and synthetic pathway of the target molecule can be generated by selecting actions and updating the current molecular state.

[0145] It should be noted that, where there is no conflict, the technical features described in different alternative implementations can be included in the same embodiment. For the sake of brevity, they will not be elaborated here.

[0146] Based on the embodiments of this disclosure, by acquiring a pre-constructed chemical reaction knowledge graph, and making the nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, the edges in the chemical reaction knowledge graph represent the relationships between nodes, and the reaction templates represent chemical transformation rules, a structured expression of the reaction knowledge contained in the molecular entities, reaction templates, and relationships between nodes can be achieved. By extracting at least one target chemical transformation relationship corresponding to each of at least one preset chemical transformation rule from the chemical reaction knowledge graph, a correspondence can be established between the preset chemical transformation rules and the target chemical transformation relationships in the chemical reaction knowledge graph. By constructing a reaction-constrained action space corresponding to the preset chemical transformation rule based on at least one target chemical transformation relationship, and making each action in the action space represent a chemical transformation operation between molecules, the selectable action range of the generation flow network can be limited to the range of chemical transformation operations corresponding to the preset chemical transformation rule. By using a pre-trained generation flow network, based on the action space corresponding to at least one preset chemical transformation rule, the molecular structure and synthetic path of the target molecule can be generated. The molecular structure and synthetic path of the target molecule can be generated synchronously in the reaction-constrained action space, and the probability of the generation process deviating from the chemical transformation rule can be reduced. Therefore, based on the relationships between molecular entities, reaction templates, and nodes in the chemical reaction knowledge graph, the target chemical transformation relationship corresponding to the preset chemical transformation rule can be converted into an action space constrained by the reaction. The pre-trained generation flow network can then generate the molecular structure and synthetic path of the target molecule based on the action space corresponding to at least one preset chemical transformation rule. This not only enables the joint generation of molecular structure and synthetic path but also improves the correspondence between the molecular structure and synthetic path of the target molecule, as well as the chemical rationality and generation effectiveness in the molecular generation process.

[0147] Figure 2 This is a schematic flowchart illustrating the process of constructing a reaction-constrained action space corresponding to a preset chemical transformation rule in a method for generating chemical molecule information provided in this embodiment of the disclosure. Figure 2 As shown, the method specifically includes: Step 201: From at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, determine at least one candidate chemical transformation relationship that matches the current molecular state in the generation process of the target molecule.

[0148] In this embodiment, the process of generating the target molecule can be the process of generating the molecular structure and synthetic pathway of the target molecule. As an example, this generation process can start from an empty state or starting materials, and through a series of chemical actions, gradually update the current molecular state until the molecular structure and synthetic pathway of the target molecule are generated.

[0149] The current molecular state is the current state during the formation of the target molecule. For example, the current molecular state can be an empty state, an intermediate product molecule, or the final target product molecule. In some cases, the current molecular state can be represented as a two-dimensional molecular diagram structure.

[0150] Candidate chemical transformation relationships can be chemical transformation relationships that match the current molecular state, determined from at least one target chemical transformation relationship corresponding to the preset chemical transformation rule. For example, based on the current molecular state, at least one candidate chemical transformation relationship matching the current molecular state during the formation process of the target molecule can be determined by searching from at least one target chemical transformation relationship corresponding to the preset chemical transformation rule.

[0151] Step 202: For each candidate chemical transformation relationship in at least one candidate chemical transformation relationship: if the reaction template corresponding to the candidate chemical transformation relationship is a unimolecular reaction template, then the preset chemical transformation rule corresponding to the candidate chemical transformation relationship is instantiated as a unimolecular reaction action; if the reaction template corresponding to the candidate chemical transformation relationship is a multimolecular reaction template, then a common reactant matching the current molecular state and the candidate chemical transformation relationship is retrieved from the preset raw material library, and the combination of the preset chemical transformation rule and the common reactant is instantiated as a multimolecular reaction action.

[0152] In this embodiment, the unimolecular reaction template corresponds to an unimolecular reaction action. An unimolecular reaction action is a chemical transformation operation in molecule formation that involves only one molecule as a reactant. The unimolecular reaction template is a reaction template that describes an unimolecular reaction action.

[0153] Multimolecular reaction templates correspond to multimolecular reaction actions. A multimolecular reaction action refers to a chemical transformation operation in molecular formation that requires two or more molecules as reactants. A multimolecular reaction template is a reaction template that describes a multimolecular reaction action. The multimolecular reaction template needs to retrieve common reactants from a pre-defined raw material library and instantiate the combination of pre-defined chemical transformation rules and common reactants into a multimolecular reaction action.

[0154] The preset raw material library can be a set of raw materials used to retrieve common reactants.

[0155] Common reactants can be reactants that participate in the reaction together with the current molecular state during the chemical transformation process corresponding to the multi-molecular reaction template.

[0156] Step 203: Based on the unimolecular and multimolecular reaction actions corresponding to each candidate chemical transformation relationship, construct the reaction-constrained action space corresponding to the preset chemical transformation rule.

[0157] In this embodiment, the combination of monomolecular and multimolecular reaction actions corresponding to each candidate chemical transformation relationship can be used as the reaction-constrained action space corresponding to the preset chemical transformation rule.

[0158] It should be noted that, in addition to the contents described above, this embodiment may also include... Figure 1 The corresponding technical features described in the corresponding embodiments, thereby achieving Figure 1 For details on the technical effects of the method for generating chemical molecular information shown, please refer to [link / reference needed]. Figure 1 The relevant descriptions are presented concisely and will not be elaborated upon here.

[0159] Based on the embodiments of this disclosure, by determining at least one candidate chemical transformation relationship that matches the current molecular state in the formation process of the target molecule from at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, the action space constrained by the reaction and the current molecular state can be dynamically matched. When the reaction template corresponding to the candidate chemical transformation relationship is a monomolecular reaction template, the preset chemical transformation rule corresponding to the candidate chemical transformation relationship is instantiated as a monomolecular reaction action, thus realizing the conversion of the chemical transformation relationship corresponding to the monomolecular reaction template into a monomolecular reaction action. When the reaction template corresponding to the candidate chemical transformation relationship is a multimolecular reaction template, common reactants matching the current molecular state and the candidate chemical transformation relationship are retrieved from a preset raw material library, and the combination of the preset chemical transformation rule and the common reactants is instantiated as a multimolecular reaction action, thus realizing the conversion of the chemical transformation relationship corresponding to the multimolecular reaction template into a multimolecular reaction action. By constructing a reaction-constrained action space corresponding to the preset chemical transformation rule based on the single-molecule and multi-molecule reaction actions corresponding to each candidate chemical transformation relationship, the candidate chemical transformation relationship matching the current molecular state can be converted into a set of actions selectable by the generation flow network. This enables the instantiation of at least one target chemical transformation relationship corresponding to the preset chemical transformation rule into single-molecule and multi-molecule reaction actions, thereby ensuring that the reaction-constrained action space corresponding to the preset chemical transformation rule matches the current molecular state in the generation process of the target molecule.

[0160] The following describes the embodiments of this disclosure by way of example. However, it should be noted that the following content is only used to understand the technical solutions of the embodiments of this disclosure and does not constitute a limitation on the protection scope of the embodiments of this disclosure.

[0161] In recent years, deep learning-based molecular generation techniques have shown great potential in drug screening and materials design. Generative flow networks, as a novel probabilistic modeling method, can sample diverse candidate molecules based on reward distributions, offering unique advantages in solving structure optimization problems. However, existing molecular generation methods still have significant limitations in practical applications. First, synthetic availability is disconnected from the generation process. Traditional action-based generation methods fail to reflect the chemist's thought process of navigating chemical space, resulting in generated molecules with potential activity but lacking a clear synthetic path, significantly increasing the cost of subsequent experimental verification. Second, search efficiency is limited. Existing techniques often treat molecular generation as a stochastic exploration process lacking prior knowledge guidance. Its state representation is typically a molecular graph or SMILES (Simplified Molecular Input Line Entry System) string, with actions defined as low-level editing (atom addition, bond modification) or selecting a fixed template. This flattened action space fails to effectively encode chemical synthesis logic, causing the model to allocate a large number of sampling probabilities to chemically unstable intermediates, synthetically infeasible structures, or regions with low biological activity during the search process. Therefore, how to effectively integrate rich chemical reaction knowledge (such as reaction templates, raw material compatibility, etc.) into the sampling logic of the generation flow network to achieve the joint generation of molecular structure and synthetic path, and ensure the chemical rationality of the product and sampling efficiency, is a technical problem that urgently needs to be solved in this field.

[0162] In this embodiment, chemical reaction data is first extracted from the Open Reaction Database (ORD) to construct a system as follows: Figure 3A The diagram shows a chemical reaction knowledge graph. This knowledge graph contains six types of entity nodes: reaction nodes, reactant nodes, product nodes, reagent nodes, reaction condition nodes, and reaction template nodes. These nodes are linked through logical relationships such as "participation," "generation," "requirement," "influence," and "application." Furthermore, knowledge graph embedding techniques (such as the TransR model) are used to pre-train encoders on this chemical knowledge graph. These encoders map the entities to a unified low-dimensional latent space, obtaining feature vectors with structural semantic information. These feature vectors are used in subsequent generation decision-making processes to measure the matching degree between the current molecular state and candidate reaction actions.

[0163] This solution is based on the following: Figure 3A Based on the aforementioned chemical reaction knowledge graph, the technical solution adopted is as follows: The first step involves using a pre-trained knowledge graph encoder to perform offline feature extraction on all candidate reaction templates (the reaction templates in the knowledge graph are extracted from reaction data, and are numerous and varied; in the molecular generation system, refined reaction templates are used as candidate reaction templates for the model to select) and building blocks (small molecules that can be purchased on the market, such as those in the Enamine molecular library; candidate building blocks refer to small molecules for the model to select, and are a subset of building blocks), generating corresponding semantic embedding vectors, and storing them in the system's action knowledge base.

[0164] The candidate reaction templates here can be abstract rule descriptions of specific chemical transformation logics (such as Suzuki coupling, amidation reactions, Click reactions, etc.). The candidate building blocks can be inexpensive and readily available chemical reagents retrieved from the starting material library (i.e., the pre-set raw material library, a commercially available small molecule library, such as the Enamine molecular library, which stores building block molecules).

[0165] The second step involves inputting the current molecular state into a pre-trained molecular encoder and extracting features to obtain the molecular state embedding in a unified latent space.

[0166] The molecular state here can be an empty state at the beginning of the generation path, an intermediate product molecule in the middle of the generation path, or the final target product molecule, and is generally represented as a two-dimensional molecular graph structure. The embedding vector associated with this state is extracted by a pre-trained molecular encoder (such as a network composed of multiple graph convolutional layers or Transformer modules), and its dimension is consistent with the semantic embedding vector in the action knowledge base, thereby ensuring the comparability of different chemical entities in a unified semantic space.

[0167] The third step involves performing a search in candidate reaction templates and building blocks based on the current molecular state, and using a masking mechanism to block illegal actions that do not conform to chemical logic, thereby constructing a restricted action search space.

[0168] The chemical logic constraints here are primarily based on substructure matching constraints of specific reaction templates. This involves identifying functional groups in the current molecular structure to determine if it possesses the reaction sites required for chemical transformation with a specific template. Illegal actions refer to combinations of actions that are theoretically impossible or impractical in actual synthetic pathways. Introducing such constraints ensures that the retrieved reaction templates and building blocks, when input into a reaction execution program (such as the RDKit library or the OpenChem engine), consistently and uniquely output product molecules that conform to chemical valence rules. This avoids the generation flow network blindly searching in invalid or erroneous chemical spaces, significantly improving the efficiency of molecule generation and the reliability of synthetic pathways.

[0169] The fourth step involves performing two-stage hierarchical sampling in the restricted action search space using a policy network based on the molecular state embedding (also known as the state embedding vector) to determine the specific chemical transformation action for the current step.

[0170] The two-stage hierarchical sampling here refers to decomposing the complex molecular transformation decision into two sequential decision levels: the first level is the reaction template decision, which calculates the matching probability distribution of each candidate reaction template based on the current molecular state embedding, and selects a target reaction template accordingly; the second level is the common reactant decision, which further calculates and samples the required building block small molecules under the constraint of the determined target reaction template.

[0171] The hierarchical sampling mechanism here not only significantly reduces the search space for single-step decision-making but also ensures that the selected building blocks are chemically compatible with the selected templates. This structured sampling approach mimics the logical thinking of chemical experts in total synthesis design, enabling the generated molecules to maintain structural diversity while possessing extremely high synthetic route feasibility.

[0172] The fifth step is to execute the chemical action obtained from the sampling, update the current molecular state to the intermediate or product after the reaction, and record the action and the corresponding reactant information in the synthetic path trajectory.

[0173] The update here refers to simulating real chemical reaction processes using a cheminformatics execution engine (such as RDKit). For example, if the sampled chemical action is a "Suzuki coupling reaction" and the corresponding "boronic acid building block", the system will automatically calculate and generate the new molecular topology after coupling by identifying the halogenation sites of the current molecule.

[0174] The synthetic pathway trajectory here refers to a logical chain that starts from an empty state and evolves through a series of ordered chemical actions to the target molecule. This trajectory records in detail the reaction template number used, the building block ID consumed, and the intermediate structure generated at each step.

[0175] The sixth step involves calculating the trajectory balance loss based on the generated molecules and their synthesis pathways using a preset reward function, and then adjusting the learnable parameters in the policy network through a backpropagation algorithm to optimize the accuracy of subsequent sampling.

[0176] The reward function here is a multi-dimensional evaluation system designed to measure the quality of the generated molecule from different perspectives. For example, the reward function may include the molecule's quantitative drug similarity (QED, used to assess whether it is drug-like), synthetic availability score (SA, used to assess the ease of synthesis), and bioactivity values ​​for specific disease targets (such as a certain protein kinase or GPCR receptor) predicted based on a deep learning model.

[0177] The parameter adjustment here employs trajectory balance loss function optimization logic. Its core technical objective is to ensure that the probability of each synthetic path generated by the system is proportional to its final reward value. Through this training mechanism, the policy network can continuously learn and internalize the logic in the chemical knowledge graph, thereby automatically exploring novel molecular spaces that both conform to chemical laws and possess excellent pharmacological activity with a higher probability in subsequent generation tasks.

[0178] To further disclose the underlying algorithm implementation of the hierarchical sampling decision described in this invention, the following is in conjunction with the appendix. Figure 3B Provide a detailed explanation of the decision-making process.

[0179] like Figure 3B As shown in the "Trajectory Sampling" section, the system processes chemical reaction templates through a pre-trained template encoder to generate a reaction template embedding (also known as template embedding vector) matrix; simultaneously, it uses a molecular encoder to process the current molecular state. The system generates molecular state embeddings (also known as state embedding vectors). Furthermore, for each building block, the system uses the Morgan fingerprint algorithm to transform it into discrete binary features and maps them to a building block embedding vector matrix. All of these vectors are mapped to a unified latent feature space, providing a foundation for subsequent retrieval calculations.

[0180] The forward strategy here takes molecular state embedding, reaction template embedding matrix, and building block embedding vector matrix as input, and outputs the reaction template and building blocks selected by the strategy function, responsible for the molecular synthesis evolution logic; the backward strategy here takes molecular state embedding and reaction template embedding matrix as input, and outputs the reverse reaction template selected by the strategy function, responsible for the reverse backtracking analysis of the path; after the forward strategy outputs the corresponding action, the system calls and executes the reaction program according to the corresponding action, changing the molecular state from Advance to This process is repeated iteratively until the complete target molecule and its synthetic pathway are generated.

[0181] like Figure 3BAs shown in the "Detailed Forward / Backward Strategy" section, when executing forward decision-making, the system embeds the molecular state into the first MLP (Multilayer Perceptron) module. This module extracts the reactivity features of the current molecule and performs a matrix dot product operation with the reaction template embedding matrix to measure the matching score between the current state and each candidate reaction template. The system uses the Softmax function to convert the above matching score into a probability distribution and performs random sampling accordingly to determine the target reaction template. After determining the target reaction template, the system uses the Fusion module to interactively fuse the selected template features with the original molecular state embedding. The fused feature vector is input into the second MLP module and combined with the building block embedding vector matrix for secondary matching calculation. After secondary Softmax normalization and sampling, the system finally determines the target building block. When performing backward decision-making, the system also embeds the molecular state into the MLP module corresponding to the backward strategy, performs a matrix dot product operation between the output result and the reaction template embedding matrix to measure the matching score between the current state and each candidate reaction template; the system uses the Softmax function to convert the above matching score into a probability distribution, and performs random sampling accordingly to determine the target reverse reaction template.

[0182] To verify the effectiveness of this scheme, 20,000 molecules were randomly selected from the Enamine commercial chemical library as building blocks, and 105 representative and high-yield reaction templates were selected to construct the action space. Soluble epoxide hydrolase (sEH) was chosen as the target protein to generate the corresponding molecules. The scheme was compared with five existing molecule generation methods: GFlowNet, SyntheMol, RGFN, SynFlowNet, and SCENT. GFlowNet uses editing actions for molecule generation, while the other methods use fixed reaction templates. Experimental settings and pretreatment methods were kept consistent across all methods.

[0183] Experimental evaluation metrics include: sEH binding energy score (used to evaluate the binding strength between the molecule and the sEH target), LE score (ligand efficiency score, i.e., the value after normalizing the Vina docking binding energy using the number of non-hydrogen atoms in the molecule), QED score (used for molecular drug efficacy assessment), SA score (used for assessing the ease of molecular synthesis), and MedChem score (the proportion of generated molecules that pass industry standard filtering rules). The experimental data comparison results for each method are shown in Table 1. Table 1

[0184] As shown in Table 1, previous baseline models (such as RGFN and SCENT) tend to artificially increase the sEH score by increasing the mass of the generated molecule. This strategy leads to a sharp drop in the QED score (drug-likeness), causing the model to fall into a suboptimal state. In contrast, our proposed approach is located in the "elbow" region of the Pareto curve for multi-index optimization, achieving an efficient trade-off between various indices. Specifically, this approach sacrifices a very small amount of sEH score in exchange for a significant increase in indices such as QED, LE, and MedChem. This indicates that our approach does not rely on a simple search and exploitation strategy, but rather, guided by reaction knowledge, concentrates the sampling probability quality in a region of chemical space that is more pharmacologically realistic and synthetically feasible, thereby generating candidate drug molecules with superior overall performance.

[0185] Figure 3C The molecule generated using this method targeting the sEH site exhibits a Vina docking binding energy of -11.06, an SA score of 3.36, and a QED score of 0.78. Experimental results demonstrate that the molecule generated by this method possesses high target affinity while naturally carrying a synthetic pathway trajectory composed of the described reaction template and building blocks, proving its extremely high synthetic feasibility.

[0186] It should be noted that, in addition to the contents described above, this embodiment may also include the technical features described in the above embodiments, thereby achieving the technical effect of the method for generating chemical molecular information shown above. Please refer to the above description for details. For the sake of brevity, it will not be elaborated here.

[0187] Figure 4 This is a schematic diagram of a chemical molecule information generation device provided in an embodiment of the present disclosure. The chemical molecule information generation device includes: The acquisition unit 401 is configured to: acquire a pre-constructed chemical reaction knowledge graph, wherein the nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, the edges in the chemical reaction knowledge graph represent the relationships between nodes, and the reaction templates represent chemical transformation rules; Extraction unit 402 is configured to: extract at least one target chemical transformation relationship corresponding to each preset chemical transformation rule from a chemical reaction knowledge graph for each preset chemical transformation rule; and construct a reaction-constrained action space corresponding to the preset chemical transformation rule based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules; The generation unit 403 is configured to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule by using a pre-trained generation flow network.

[0188] In some possible implementations, the nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes; and From the chemical reaction knowledge graph, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, including: Knowledge graph embedding learning is performed on the chemical reaction knowledge graph to obtain multiple reaction template embeddings and multiple environmental constraint embeddings. The reaction template embedding represents the topological transformation logic contained in the reaction template node, and the environmental constraint embedding represents the reaction reagent node and the reaction condition node. The correlation between the reaction template embedding and the environmental constraint embedding is calculated, and multiple correlation degrees are obtained. Based on multiple correlation degrees, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule is determined.

[0189] In some possible implementations, based on at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed, including: From at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, determine at least one candidate chemical transformation relationship that matches the current molecular state in the generation process of the target molecule; For each candidate chemical transformation relationship in at least one candidate chemical transformation relationship: if the reaction template corresponding to the candidate chemical transformation relationship is a unimolecular reaction template, then the preset chemical transformation rule corresponding to the candidate chemical transformation relationship is instantiated as a unimolecular reaction action; if the reaction template corresponding to the candidate chemical transformation relationship is a multimolecular reaction template, then a common reactant matching the current molecular state and the candidate chemical transformation relationship is retrieved from the preset raw material library, and the combination of the preset chemical transformation rule and the common reactant is instantiated as a multimolecular reaction action; Based on the single-molecule and multi-molecule reaction actions corresponding to each candidate chemical transformation relationship, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed.

[0190] In some possible implementations, the molecular structure and synthetic pathway of the target molecule are generated based on the action space corresponding to at least one preset chemical transformation rule, including: Using a molecular encoder trained on a chemical reaction knowledge graph, the current molecular state during the generation process of the target molecule is mapped to a molecular state embedding in a vector space. Using a template encoder trained on a chemical reaction knowledge graph, each candidate reaction template is mapped to a reaction template embedding in a vector space; Based on the molecular state embedding and the reaction template embedding corresponding to each candidate reaction template, the probability of selecting each candidate reaction template is determined. For each candidate reaction template, if the candidate reaction template is a multi-molecule reaction template, the fusion feature is determined based on the molecular state embedding and the reaction template embedding. The matching degree between the fusion feature and the common reactant embedding vector is calculated, and the conditional probability of selecting the common reactant is determined based on the matching degree. For a single-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the single-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template; for a multi-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the multi-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting the common reactant. Based on the selection probabilities of single-molecule reaction actions and multi-molecule reaction actions, actions are selected from the action space corresponding to at least one preset chemical transformation rule, and the current molecular state is updated to generate the molecular structure and synthetic pathway of the target molecule.

[0191] In some possible implementations, the generating stream network is trained in the following manner: A loss function is constructed based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state. While keeping the parameters of the molecular encoder and the reaction template encoder pre-trained based on the chemical reaction knowledge graph frozen, the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix parameters in the decision-making process in the generative flow network are updated by minimizing the loss function, thus obtaining the trained generative flow network.

[0192] In some possible implementations, the reaction template includes a forward reaction template and a reverse reaction template, which represent chemical transformation processes with opposite reaction directions; and Before constructing the loss function based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state, the method also includes: For any state transition subprocess in the generation process of the target molecule: Determine the successor molecular state and the predecessor molecular state in this state transition subprocess; Based on the subsequent molecular state, candidate reverse reaction templates matching the subsequent molecular state are determined. Based on the matching relationship between the subsequent molecular state and the candidate reverse reaction template, the reverse selection probability corresponding to each candidate reverse reaction template is determined. From the candidate reverse reaction templates, the target reverse reaction template for converting the subsequent molecular state into the precursor molecular state is determined. Based on the reverse selection probability corresponding to the target reverse reaction template, the reverse state transition probability corresponding to the state transition subprocess is determined.

[0193] In some possible implementations, the chemical reaction knowledge graph is represented by a multi-relation heterogeneous graph, which contains multiple nodes and edges representing multiple semantic relations; The nodes include: reaction node, reactant node, product node, reagent node, reaction condition node, and reaction template node; Semantic relations include: participation relations, auxiliary relations, and mapping relations; Participation relationships characterize the compositional associations between reactant nodes, product nodes, and reaction nodes; Auxiliary relationships characterize the environmental constraint relationships between reaction reagent nodes, reaction condition nodes, and reaction nodes; The mapping relationship represents the rule attribution relationship between reaction nodes and reaction template nodes.

[0194] The chemical molecule information generation device provided in this embodiment can execute the corresponding steps of the chemical molecule information generation methods described above, thereby achieving the technical effects of the chemical molecule information generation methods described above. The chemical molecule information generation device and the chemical molecule information generation methods can refer to and cite each other in terms of specific implementation and technical effects. For the sake of brevity, they will not be elaborated here.

[0195] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present disclosure. Figure 5 The illustrated electronic device 500 includes at least one processor 501, a memory 502, at least one network interface 504, and a user interface 503. The various components in the electronic device 500 are coupled together via a bus system 505. It is understood that the bus system 505 is used to implement communication between these components. In addition to a data bus, the bus system 505 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 5 The general designated all buses as Bus System 505.

[0196] The user interface 503 may include a display, keyboard, or clicking device (e.g., mouse, trackball, touchpad, or touchscreen).

[0197] It is understood that the memory 502 in this embodiment of the present disclosure may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced Synchronous DRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), and Direct Rambus RAM (DRRAM). The memory 502 described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0198] In some implementations, memory 502 stores elements, executable units or data structures, or subsets thereof, or extended sets thereof: operating system 5021 and application program 5022.

[0199] The operating system 5021 includes various system programs, such as the framework layer, core library layer, and driver layer, used to implement various basic business functions and handle hardware-based tasks. The application program 5022 includes various applications, such as a media player and a browser, used to implement various application functions. The program implementing the method of this embodiment can be included in the application program 5022.

[0200] In this embodiment, by calling the program or instructions stored in memory 502, specifically the program or instructions stored in application program 5022, processor 501 executes the method steps provided in each method embodiment, including, for example: Obtain a pre-constructed chemical reaction knowledge graph, where nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules; For each of the at least one preset chemical transformation rules, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph; based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, construct the reaction-constrained action space corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules; A pre-trained generative flow network is used to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule.

[0201] The methods disclosed in the above embodiments of this disclosure can be applied to or implemented by processor 501. Processor 501 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware or by instructions in the form of software in processor 501. The processor 501 may be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this disclosure. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this disclosure can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software units in the decoding processor. The software units may be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory 502. Processor 501 reads the information in memory 502 and, in conjunction with its hardware, completes the steps of the above method.

[0202] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described above, or combinations thereof.

[0203] For software implementation, the techniques described herein can be implemented by units that perform the functions described above. The software code can be stored in memory and executed by a processor. The memory can be implemented within the processor or external to the processor.

[0204] The electronic device provided in this embodiment may be as follows: Figure 5 The electronic device shown can execute all the steps of the above-described methods for generating chemical molecule information, thereby achieving the technical effects of the above-described methods for generating chemical molecule information. For details, please refer to the above descriptions. For the sake of brevity, further details are omitted here.

[0205] This disclosure also provides a storage medium (computer-readable storage medium). This storage medium stores one or more programs. The storage medium may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as read-only memory, flash memory, hard disk, or solid-state drive; the memory may also include combinations of the above types of memory.

[0206] When one or more programs in the storage medium can be executed by one or more processors to implement the above-described method for generating chemical molecular information executed on the electronic device side.

[0207] The processor described above is used to execute a program for generating chemical molecule information stored in memory, to implement the following steps of a method for generating chemical molecule information executed on the electronic device side: Obtain a pre-constructed chemical reaction knowledge graph, where nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules; For each of the at least one preset chemical transformation rules, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph; based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, construct the reaction-constrained action space corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules; A pre-trained generative flow network is used to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule.

[0208] Furthermore, the computer program product provided in this disclosure embodiment may include computer-readable code that, when executed on a device, causes a processor in the device to implement the steps of a method for generating chemical molecule information that is executed on the electronic device side: Obtain a pre-constructed chemical reaction knowledge graph, where nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules; For each of the at least one preset chemical transformation rules, extract at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph; based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, construct the reaction-constrained action space corresponding to the preset chemical transformation rule, wherein each action in the action space represents a chemical transformation operation between molecules; A pre-trained generative flow network is used to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule.

[0209] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this disclosure.

[0210] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0211] It should be understood that the terminology used herein is for the purpose of describing particular exemplary embodiments only and is not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “described” as used herein may also include the plural forms. The terms “comprising,” “including,” “containing,” and “having” are inclusive and therefore indicate the presence of the stated features, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, elements, components, and / or combinations thereof. The method steps, processes, and operations described herein are not construed as requiring them to be performed in a particular order described or illustrated unless the order of performance is explicitly indicated. It should also be understood that additional or alternative steps may be used.

[0212] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for generating chemical molecular information, characterized in that, The method includes: Obtain a pre-constructed chemical reaction knowledge graph, wherein nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, edges in the chemical reaction knowledge graph represent the relationships between nodes, and reaction templates represent chemical transformation rules; For each of the at least one preset chemical transformation rules, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule is extracted from the chemical reaction knowledge graph; based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed, wherein each action in the action space represents a chemical transformation operation between molecules; A pre-trained generative flow network is used to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule. The nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes; and Extracting at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph includes: The chemical reaction knowledge graph is subjected to knowledge graph embedding learning to obtain multiple reaction template embeddings and multiple environmental constraint embeddings. The reaction template embedding represents the topology transformation logic contained in the reaction template node, and the environmental constraint embedding represents the reaction reagent node and the reaction condition node. The correlation between the reaction template embedding and the environmental constraint embedding is calculated, and multiple correlation degrees are obtained. Based on multiple correlation degrees, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule is determined; The process of generating the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule includes: Using a molecular encoder trained based on the chemical reaction knowledge graph, the current molecular state during the generation process of the target molecule is mapped to a molecular state embedding in a vector space. Using a template encoder trained based on the chemical reaction knowledge graph, each candidate reaction template is mapped to a reaction template embedding in the vector space; Based on the molecular state embedding and the reaction template embedding corresponding to each candidate reaction template, the probability of selecting each candidate reaction template is determined. For each candidate reaction template, if the candidate reaction template is a multi-molecule reaction template, the fusion feature is determined based on the molecular state embedding and the reaction template embedding. The matching degree between the fusion feature and the common reactant embedding vector is calculated, and the conditional probability of selecting the common reactant is determined based on the matching degree. For a single-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the single-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template; for a multi-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the multi-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting the common reactant. Based on the selection probability of single-molecule reaction actions and the selection probability of multi-molecule reaction actions, actions are selected from the action space corresponding to at least one preset chemical transformation rule, and the current molecular state is updated to generate the molecular structure and synthetic route of the target molecule.

2. The method according to claim 1, characterized in that, The construction of the reaction-constrained action space corresponding to the preset chemical transformation rule, based on at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, includes: From at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, determine at least one candidate chemical transformation relationship that matches the current molecular state in the generation process of the target molecule; For each candidate chemical transformation relationship in at least one candidate chemical transformation relationship: if the reaction template corresponding to the candidate chemical transformation relationship is a unimolecular reaction template, then the preset chemical transformation rule corresponding to the candidate chemical transformation relationship is instantiated as a unimolecular reaction action; if the reaction template corresponding to the candidate chemical transformation relationship is a multimolecular reaction template, then a common reactant matching the current molecular state and the candidate chemical transformation relationship is retrieved from the preset raw material library, and the combination of the preset chemical transformation rule and the common reactant is instantiated as a multimolecular reaction action; Based on the single-molecule and multi-molecule reaction actions corresponding to each candidate chemical transformation relationship, a reaction-constrained action space corresponding to the preset chemical transformation rule is constructed.

3. The method according to claim 1, characterized in that, The generated stream network is trained in the following manner: A loss function is constructed based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state. While keeping the parameters of the molecular encoder and the reaction template encoder pre-trained based on the chemical reaction knowledge graph frozen, the parameters of the forward policy network, the reverse policy network, and the linear transformation matrix parameters in the decision-making process in the generator flow network are updated by minimizing the loss function, thus obtaining the trained generator flow network.

4. The method according to claim 3, characterized in that, The reaction template includes a forward reaction template and a reverse reaction template, wherein the forward reaction template and the reverse reaction template represent chemical transformation processes with opposite reaction directions; and Before constructing the loss function based on the selection probability of the reaction action corresponding to each state transition subprocess in the target molecule generation process, the reverse state transition probability, and the reward value corresponding to the termination molecule state, the method further includes: For any state transition subprocess in the generation process of the target molecule: Determine the successor molecular state and the predecessor molecular state in this state transition subprocess; Based on the subsequent molecular state, a candidate reverse reaction template matching the subsequent molecular state is determined; Based on the matching relationship between the subsequent molecular state and the candidate reverse reaction template, the reverse selection probability corresponding to each candidate reverse reaction template is determined. From each candidate reverse reaction template, a target reverse reaction template for converting the subsequent molecular state into the precursor molecular state is determined; Based on the reverse selection probability corresponding to the target reverse reaction template, the reverse state transition probability corresponding to the state transition subprocess is determined.

5. The method according to any one of claims 1-4, characterized in that, The chemical reaction knowledge graph is represented by a multi-relation heterogeneous graph, which contains multiple nodes and edges representing multiple semantic relationships. The nodes include: reaction nodes, reactant nodes, product nodes, reaction reagent nodes, reaction condition nodes, and reaction template nodes; The semantic relationships include: participation relationships, auxiliary relationships, and mapping relationships; The participation relationship characterizes the compositional association between the reactant node, the product node, and the reaction node; The auxiliary relationship characterizes the environmental constraint association between the reaction reagent node, the reaction condition node, and the reaction node; The mapping relationship represents the rule attribution relationship between the reaction node and the reaction template node.

6. A device for generating chemical molecular information, characterized in that, The device includes: The acquisition unit is configured to: acquire a pre-constructed chemical reaction knowledge graph, wherein the nodes in the chemical reaction knowledge graph represent molecular entities and reaction templates, the edges in the chemical reaction knowledge graph represent the relationships between nodes, and the reaction templates represent chemical transformation rules; The extraction unit is configured to: extract at least one target chemical transformation relationship corresponding to each preset chemical transformation rule from the chemical reaction knowledge graph for each preset chemical transformation rule; and construct a reaction-constrained action space corresponding to the preset chemical transformation rule based on the at least one target chemical transformation relationship corresponding to the preset chemical transformation rule, wherein each action in the action space represents an intermolecular chemical transformation operation. The generation unit is configured to: use a pre-trained generation flow network to generate the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule; The nodes include reaction template nodes, reaction reagent nodes, and reaction condition nodes; and Extracting at least one target chemical transformation relationship corresponding to the preset chemical transformation rule from the chemical reaction knowledge graph includes: The chemical reaction knowledge graph is subjected to knowledge graph embedding learning to obtain multiple reaction template embeddings and multiple environmental constraint embeddings. The reaction template embedding represents the topology transformation logic contained in the reaction template node, and the environmental constraint embedding represents the reaction reagent node and the reaction condition node. The correlation between the reaction template embedding and the environmental constraint embedding is calculated, and multiple correlation degrees are obtained. Based on multiple correlation degrees, at least one target chemical transformation relationship corresponding to the preset chemical transformation rule is determined; The process of generating the molecular structure and synthetic pathway of the target molecule based on the action space corresponding to at least one preset chemical transformation rule includes: Using a molecular encoder trained based on the chemical reaction knowledge graph, the current molecular state during the generation process of the target molecule is mapped to a molecular state embedding in a vector space. Using a template encoder trained based on the chemical reaction knowledge graph, each candidate reaction template is mapped to a reaction template embedding in the vector space; Based on the molecular state embedding and the reaction template embedding corresponding to each candidate reaction template, the probability of selecting each candidate reaction template is determined. For each candidate reaction template, if the candidate reaction template is a multi-molecule reaction template, the fusion feature is determined based on the molecular state embedding and the reaction template embedding. The matching degree between the fusion feature and the common reactant embedding vector is calculated, and the conditional probability of selecting the common reactant is determined based on the matching degree. For a single-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the single-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template; for a multi-molecule reaction action in the action space corresponding to at least one preset chemical transformation rule, the selection probability of the multi-molecule reaction action is determined based on the probability of selecting the corresponding candidate reaction template and the conditional probability of selecting the common reactant. Based on the selection probability of single-molecule reaction actions and the selection probability of multi-molecule reaction actions, actions are selected from the action space corresponding to at least one preset chemical transformation rule, and the current molecular state is updated to generate the molecular structure and synthetic route of the target molecule.

7. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing a computer program stored in the memory, wherein when the computer program is executed, it implements the method of any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-5.

Images