Methods, apparatus, and media for performing retro-synthetic analysis of a compound molecule
By combining Markov logic networks and reinforcement learning algorithms, the retrosynthetic analysis of compound molecules is automated, solving the problem of low efficiency in existing technologies and realizing efficient synthetic route design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA TELECOM CORP LTD
- Filing Date
- 2021-09-03
- Publication Date
- 2026-07-24
AI Technical Summary
Current retrosynthetic analysis methods are inefficient and difficult to automate, making it difficult to design synthetic routes for complex molecules and relying on human experience.
A method combining Markov logic networks and reinforcement learning algorithms is adopted. The trained Markov logic network predicts candidate retrosynthetic pathways of compounds, and the Q-learning model is used to optimize the selection of synthetic pathways. The training and reward mechanism are combined with actual reaction data.
It enables automated retrosynthetic analysis of compound molecules, improving the efficiency and accuracy of synthetic route design and reducing human intervention.
Smart Images
Figure CN115762652B_ABST
Abstract
Description
Technical Field
[0001] This disclosure generally relates to methods, apparatus, and media for retrosynthetic analysis of compound molecules. Background Technology
[0002] Retrosynthetic analysis has long been the most fundamental and commonly used method for designing organic synthetic routes. It is a reversible logical thinking method that begins by analyzing the chemical structure of the target molecule. Based on the characteristics of the interatomic connections (chemical bonds) within the molecule, it comprehensively applies knowledge of organic chemical reaction methods and mechanisms to select appropriate chemical bonds for cleavage, transforming the target molecule into smaller intermediates. These intermediates are then used as new target molecules, cleaved into even smaller intermediates, and so on, until readily available starting materials are found. However, the retrosynthetic process is entirely manual, performed by experienced chemists, resulting in low efficiency. For complex molecules, chemists often struggle to find suitable retrosynthetic routes. Summary of the Invention
[0003] A brief overview of this disclosure is given below to provide a basic understanding of some aspects of it. However, it should be understood that this overview is not an exhaustive summary of this disclosure. It is not intended to identify key or essential parts of this disclosure, nor is it intended to limit the scope of this disclosure. Its purpose is merely to present certain concepts of this disclosure in a simplified form as a prelude to the more detailed description that follows.
[0004] According to one aspect of this disclosure, a method for retrosynthetic analysis of compound molecules is provided, comprising: determining the probabilistic graph weights of a single-step reaction in a Markov logic network based on actual reaction learning data, thereby obtaining a trained Markov logic network; receiving the encoded molecular structure of a target compound; inputting the encoded molecular structure into the trained Markov logic network and predicting multiple candidate retrosynthetic pathways of the target compound through the Markov logic network; inputting the multiple candidate retrosynthetic pathways into a reinforcement learning model, training the reinforcement learning model based on existing actual reaction processes, such that pathways among the multiple candidate retrosynthetic pathways that match actual synthesis pathways receive rewards, and updating the transition matrix in the reinforcement learning model based on the obtained rewards; and using the reinforcement learning model, determining one or more selected retrosynthetic pathways among the multiple candidate retrosynthetic pathways using the finally obtained transition matrix.
[0005] According to another aspect of this disclosure, an apparatus for retrosynthetic analysis of compound molecules is provided, comprising: a memory having instructions stored thereon; and a processor configured to execute the instructions stored in the memory to perform the method as described above.
[0006] According to another aspect of this disclosure, a computer-readable storage medium is provided, comprising computer-executable instructions that, when executed by one or more processors, cause the one or more processors to perform the method described above according to the present disclosure. Attached Figure Description
[0007] The accompanying drawings, which form part of this specification, illustrate embodiments of this disclosure and, together with the specification, serve to explain the principles of this disclosure.
[0008] This disclosure will be more clearly understood with reference to the accompanying drawings and the following detailed description, wherein:
[0009] Figure 1 An example of an existing single-step reaction is shown.
[0010] Figure 2 Examples of two candidate synthetic routes for exemplary compounds are shown.
[0011] Figure 3 An exemplary target compound is shown.
[0012] Figure 4 A candidate synthetic route for an exemplary target compound is shown.
[0013] Figure 5 Another candidate synthetic route for the exemplary target compound is shown.
[0014] Figure 6 This is a flowchart illustrating an example process of the inverse synthetic analysis method according to this disclosure.
[0015] Figure 7 An exemplary configuration of a computing device that can implement embodiments of the present disclosure is shown. Detailed Implementation
[0016] The following detailed description is based on the accompanying drawings and provides various exemplary embodiments of the present disclosure to aid in a comprehensive understanding. Various details are included in the following description to aid understanding; however, these details are considered exemplary only and not intended to limit the present disclosure, which is defined by the appended claims and their equivalents. The words and phrases used in the following description are intended only to provide a clear and consistent understanding of the present disclosure. Additionally, descriptions of well-known structures, functions, and configurations may have been omitted for clarity and brevity. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of the present disclosure.
[0017] The technology of using AI to perform retrosynthetic analysis of compounds is a major advancement in the pharmaceutical and chemical industries.
[0018] The following describes some background technologies and concepts involved in this technology.
[0019] 1. Retrosynthetic analysis of compounds
[0020] Retrosynthetic analysis is a method in organic synthetic chemistry for determining efficient synthetic routes to obtain target compounds in multi-step synthesis. This is accomplished by rationally cleaving the target molecule into precursors with simple structures. After analysis, the same steps are repeated to convert each precursor into readily available or commercially available compounds. Elias James Klee proposed this concept. One or more synthetic routes can be obtained through this retrosynthetic process, but the key is to perform a rational retrosynthetic analysis.
[0021] (1) Transformation-based strategy
[0022] The key to transformation-based strategies is to derive simpler molecules from more complex target molecules. However, even after various transformations, compounds with complex structures do not always lead to reduction, thus requiring multiple steps to build it up.
[0023] (2) Strategies based on structural objectives
[0024] After focusing on structures that can be used as intermediates or starting materials, strategies for binding with target molecules can be considered. Searching in both directions of the inverse configuration (inverse configuration and actual configuration) becomes possible.
[0025] (3) Topology strategy
[0026] Identify one or more independent or related fractures. Topological strategies may lead to the identification of critical substructures or the use of rearrangement reactions.
[0027] (4) Stereochemical strategy
[0028] Many target molecules have different stereocenters that need to be controlled. Stereochemical strategies simplify the target by deleting, moving, or reducing the elements at these stereocenters.
[0029] 2. Markov Logical Network
[0030] In recent years, scholars both domestically and internationally have proposed important methods such as statistical relation learning and probabilistic graphical models to effectively handle complexity and uncertainty problems, attracting considerable attention. Simply put, statistical relation learning obtains likelihood models from relational data by integrating relational / logical representations, probabilistic reasoning, uncertainty handling, machine learning, and data mining. Currently, relational / logical representations are still mainly based on first-order predicate logic, used to concisely express domain knowledge and effectively handle logical problems; while probabilistic graphical models are a generalized method for representing and processing uncertain knowledge, mainly encompassing Bayesian networks, Hidden Markov Models, Markov Decision Processes, and neural networks. A natural idea is to integrate statistical relation learning (especially relational / logical representations) and probabilistic graphical models. In 2004, Richardson and Domingos of the University of Washington first proposed Markov Logic Networks and demonstrated the possibility of Markov Logic Networks as a unified framework for statistical relation learning. The main reasons are as follows: Firstly, from a probabilistic and statistical perspective, Markov logic networks not only provide a concise and effective means of describing large Markov networks, but also have the ability to flexibly incorporate modular knowledge domains into Markov networks. Secondly, from the perspective of first-order predicate logic, Markov logic networks add excellent uncertainty handling capabilities to first-order predicate logic and can tolerate incompleteness and contradictions in the knowledge domain. Therefore, many important tasks in statistical relation learning (such as aggregation classification, relation prediction, social networks, and object recognition) can be unified into the Markov logic network system. Thus, the international artificial intelligence community generally recognizes Markov logic networks as a relatively perfect method for representing and handling the complexity and uncertainty of first-order predicate logic and probabilistic graphical models, possessing significant research value and broad application prospects, and has become a research hotspot in artificial intelligence, machine learning, and data mining.
[0031] 3. Reinforcement Learning Algorithms
[0032] Q-learning (reinforcement learning) is a model-independent reinforcement learning algorithm based on Markov decision processes.
[0033] A standard Markov decision process can be represented by a quintuple.<S,A,P,R,γ> It means that, among them:
[0034] S is a discrete bounded state space;
[0035] A is a discrete action space;
[0036] P is the state transition probability function, which represents the probability that the agent will transition to state s' after choosing action a in state s.
[0037] R is the reward function, used to calculate the immediate reward value obtained by the agent after choosing action a from the current state s and transitioning to the next state s'. R is determined by the current state and the chosen action, reflecting the Markov property.
[0038] γ is a discount factor used to determine the relative proportion of delayed returns (future returns) to immediate returns. The larger the γ, the more important the delayed returns are.
[0039] The goal of the Markov decision problem is to find a strategy that maximizes the long-run cumulative value of its reward function (i.e., the long-run cumulative reward).
[0040] Q-learning, through continuous exploration of the environment and accumulation of historical experience, allows the agent to strengthen itself through trial and error, ultimately achieving the goal of autonomously selecting the optimal action. That is, regardless of the current state, it can provide the optimal path to the target state. In this algorithm, the environment and actions influence each other; the choice of action affects the state of the environment, and the environment can also influence the choice of action by providing feedback on the merits of actions through a reinforcement reward function R.
[0041] To automatically predict candidate synthetic routes for compounds, the encoded molecular structure needs to be input into the prediction model. Therefore, the compound molecule must first be encoded.
[0042] Molecular structure encoding
[0043] First, in some embodiments, OpenCV can be used to perform image recognition on molecular structure images. OpenCV is a computer vision and machine learning software library capable of implementing many general-purpose algorithms in image processing and computer vision. Then, based on the recognition information obtained from image recognition of molecular structure images using OpenCV, the recognized compound molecules can be encoded. For example, the molecular formula of the compound can be encoded in, for example, the V3000 format to form the encoded data of a mol file in the form of a chemical matrix (i.e., the encoded molecular structure).
[0044] .mol files are a commonly used format for storing molecular structures. Software such as ChemDraw can be used to convert .mol files into visual molecular structures.
[0045] The following is an example of encoded data in .mol file format.
[0046] Untitled Document-1
[0047] ChemDraw06182015012D
[0048] 0 0 0 0 0 0 V3000
[0049] M V30 BEGIN CTAB
[0050] M V30 COUNTS 13 14 0 0 0
[0051] M V30 BEGIN ATOM
[0052] M V30 1 C -2.475000 -0.000000 0.000000 0
[0053] M V30 2 C -2.062500 -0.714471 0.000000 0
[0054] M V30 3 C -1.237500 -0.714471 0.000000 0
[0055] M V30 4 C -0.825000 -0.000000 0.000000 0
[0056] M V30 5 C -1.237500 0.714471 0.000000 0
[0057] M V30 6 C -2.062500 0.714471 0.000000 0
[0058] M V30 7 C 0.000000 -0.000000 0.000000 0
[0059] M V30 8 C 0.412500 -0.714471 0.000000 0
[0060] M V30 9 C 1.237500 -0.714471 0.000000 0
[0061] M V30 10 C 1.650000 -0.000000 0.000000 0
[0062] M V30 11 C 1.237500 0.714471 0.000000 0
[0063] M V30 12 C 0.412500 0.714471 0.000000 0
[0064] M V30 13 C 2.475000 -0.000000 0.000000 0
[0065] M V30 END ATOM
[0066] M V30 BEGIN BOND
[0067] M V30 1 2 1 2
[0068] M V30 2 1 2 3
[0069] M V30 3 2 3 4
[0070] M V30 4 1 4 5
[0071] M V30 5 2 5 6
[0072] M V30 6 1 6 1
[0073] M V30 7 1 4 7
[0074] M V30 8 2 7 8
[0075] M V30 9 1 8 9
[0076] M V30 10 2 9 10
[0077] M V30 11 1 10 11
[0078] M V30 12 2 11 12
[0079] M V30 13 1 12 7
[0080] M V30 14 1 10 13
[0081] M V30 END BOND
[0082] M V30 END CTAB
[0083] M END
[0084] A .mol file is generated based on the character relationships between characters and feature points in a molecular structure image, as well as the connection relationships between feature points. In addition, a .mol file may include other content such as an attribute matrix.
[0085] Note that using mol files to store molecular structures is just one example; molecular structures can also be stored in other encoding formats.
[0086] This disclosure applies Markov logic networks to molecular retrosynthetic analysis. By training the network with the relationships between the logic networks, the molecular structure of the precursor molecule of the target compound is inferred. Traditional retrosynthetic methods rely solely on human experience. The retrosynthetic analysis method based on Markov logic networks and AI algorithms disclosed in this disclosure allows for machine-based retrosynthetic analysis. This significantly reduces manual labor and accelerates the development of the pharmaceutical industry.
[0087] AI-assisted retrosynthesis is a key application area in the field of pharmaceutical artificial intelligence. This technology can rapidly advance the design of synthetic routes for complex structures such as drugs of varying sizes and natural products, and has become a typical application scenario for AI in the pharmaceutical field.
[0088] Markov Logic Network Training
[0089] Markov logic networks can be viewed as a combination of first-order predicates (i.e., first-order logic or first-order predicate logic) and probability. The first-order logic knowledge base can be seen as establishing a series of hard rules on a set of possible worlds; if a world violates even one of these rules, its probability of existence is zero. The basic idea of Markov logic networks is to relax these hard rules; that is, when a world violates even one of these rules, the probability of that world existing decreases, but it is not completely impossible. The fewer rules a world violates, the greater its probability of existence.
[0090] First-order predicate logic can be expressed in the form of conjunctive normal form, thus forming a probabilistic graphical network. This disclosure uses, for example, a Tuffy Markov logic network framework for computation, and obtains good results.
[0091] The first-order predicate conjunction normal form is expressed as: Where P and Q represent propositions.
[0092] To predict candidate retrosynthetic routes for target compounds, this disclosure determines the probabilistic graph weights of single-step reactions in a Markov logic network based on actual reaction learning data, thereby obtaining a trained Markov logic network.
[0093] Determining the probabilistic graphical weights of single-step responses in a Markov logic network requires weight inference. This inference process is divided into two parts: "evidence" and "query." "Evidence" is an objective description of the facts, a set of graphically descriptive data; "query" is a definition based on the characteristics of the molecular structure.
[0094] Figure 1 An example of an existing single-step reaction is shown. Using this single-step reaction as an example, the probabilistic graphical weights for this single-step reaction can be inferred as follows.
[0095] (1) Evidence detection
[0096] First, detect the characteristics of the reaction, such as Figure 1 The structure indicated by "1" in the diagram. Extract the features of this structure.
[0097] (2) Query confirmation
[0098] The rules determine that "1" can be decomposed into: Figure 1 The features of "2" and "3" in the query confirmation are then used. A large amount of actual reactions can be collected to form actual reaction learning data, which can then be used as rules.
[0099] (3) Determine the probability
[0100] From "1", we can deduce "2" and "3", and from this we can deduce the meaning of "1". Figure 1 The single-step response shown is assigned a weighted probability (probabilistic graphical weight). In one embodiment, the probabilistic graphical weight of a single-step response in a Markov logic network can be determined based on the ratio (or frequency) in which the single-step response occurs in actual response learning data, such that the larger the ratio (i.e., the more common or reliable the single-step response is), the larger the probabilistic graphical weight.
[0101] By repeating the processes (1), (2), and (3) above, the trained Markov Logic Network (MLN) can be obtained.
[0102] Using a trained Markov logic network, multiple candidate (i.e., possible) retrosynthetic routes for a target compound can be predicted.
[0103] Q Learning and Training Process
[0104] When multiple candidate inverse synthesis paths exist, Q-learning can be used to infer one or more ideal candidate inverse synthesis paths.
[0105] like Figure 2 As shown, structure "1" has two possible inferences: it can be synthesized using compounds "2" and "3", or it can be synthesized using compounds "4" and "5". Q-learning is used to make decisions during the synthetic route selection process.
[0106] This disclosure utilizes real-world reaction learning data containing over 50,000 existing chemical reactions to train a reinforcement learning model (Q-learning model). During training, rewards are given based on the results. If a candidate retrosynthetic pathway aligns with existing reactions, it receives a higher score reward.
[0107] The technology disclosed herein mainly includes the following two steps.
[0108] 1. Step 1: Single-step reaction calibration learning to build the MLN network.
[0109] Based on the current chemical reaction (actual reaction learning data), the weights of the probability graph are determined, thus obtaining a trained Markov logic network. This Markov logic network can then predict multiple candidate synthetic pathways for the compound to be synthesized.
[0110] 2. Second step: Q learning and training
[0111] When multiple candidate synthesis paths exist, exhaustive search is performed using computational power to train the reinforcement learning model. During training, paths that match actual experimental results (i.e., actual chemical reactions) receive higher rewards. For example, Figure 4 and Figure 5 As shown Figure 3 The diagram shows two candidate synthetic routes for the target compound. Among them, Figure 4 The path matches the actual situation, resulting in a higher reward. It should be noted that... Figure 5 The target compound shown in the upper left corner is... Figure 4 The target compound shown at the top left is the same, in which Figure 4 A relatively simple set of symbols was used.
[0112] By using rewards for specific paths, the transition matrix in the reinforcement learning model can be continuously updated, thereby training the model. Ultimately, the trained transition matrix can be used to determine one or more desirable inverse synthesis paths.
[0113] The method of this disclosure is described in more detail below.
[0114] Figure 6 This is a flowchart illustrating an example process of the inverse synthetic analysis method according to this disclosure.
[0115] like Figure 6As shown, in one embodiment, a method for retrosynthetic analysis of compound molecules is provided, comprising: determining the probability graph weights of a single-step reaction of a Markov logic network based on actual reaction learning data, thereby obtaining a trained Markov logic network (101); receiving the encoded molecular structure of a target compound (102); inputting the encoded molecular structure into the trained Markov logic network, and predicting multiple candidate retrosynthetic paths of the target compound through the Markov logic network (103); inputting the multiple candidate retrosynthetic paths into a reinforcement learning model, training the reinforcement learning model based on existing actual reaction processes, such that paths among the multiple candidate retrosynthetic paths that match actual synthesis paths receive rewards, and updating the transition matrix in the reinforcement learning model according to the obtained rewards (104); and using the reinforcement learning model, determining one or more selected retrosynthetic paths among the multiple candidate retrosynthetic paths using the finally obtained transition matrix (105).
[0116] In one embodiment, the step (104) of inputting multiple candidate inverse synthesis paths into a reinforcement learning model and training the reinforcement learning model based on existing actual response processes may include: determining whether each single-step response in each candidate inverse synthesis path is consistent with an existing actual single-step response; and if a specific single-step response in a candidate inverse synthesis path is consistent with an existing actual single-step response, then assigning a single-step response reward value to the candidate inverse synthesis path.
[0117] In one embodiment, the step of assigning a single-step reaction reward value to a candidate retrosynthetic route may further include: determining the single-step reaction reward value based on the reactant (i.e., starting material or precursor) cost and the chemical reaction cost of a specific single-step reaction, such that the lower the reactant cost or chemical reaction cost, the higher the single-step reaction reward value. This encourages the selection of retrosynthetic routes with lower reactant costs or chemical reaction costs. Reactant costs include, for example, the price of the reactants themselves. Chemical reaction costs include, for example, the resources required to carry out the chemical reaction.
[0118] In one embodiment, the step (104) of inputting multiple candidate inverse synthesis paths into a reinforcement learning model and training the model based on existing actual reaction processes may include: assigning a negative step reward value to each single-step reaction in the candidate inverse synthesis paths. This encourages the selection of inverse synthesis paths with fewer total reaction steps in a single-step reaction.
[0119] In one embodiment, the step of assigning a single-step reaction reward value to a candidate retrosynthetic route may include determining the single-step reaction reward value based on the properties of the reactants in a particular single-step reaction. In one embodiment, the properties may include one or more of the reactants' chemical stability, storage requirements, chemical toxicity, environmental hazards, and ease of commercial access. This can encourage the selection of retrosynthetic routes involving reactants with favorable properties (e.g., high chemical stability).
[0120] In one embodiment, the step of assigning a single-step reaction bonus value to a candidate retrosynthetic route may include assigning an additional bonus value to the candidate retrosynthetic route if a specific reactant in a particular single-step reaction is the same as a reactant in another single-step reaction within the candidate retrosynthetic route. This encourages the selection of retrosynthetic routes involving a smaller total number of reactants. Clearly, if the total number of reactants (i.e., the variety of reactants) is smaller, it is easier to obtain all the desired types of reactants.
[0121] In one embodiment, the step (101) of determining the probabilistic graphical weights of single-step reactions in a Markov logic network may include: determining the probabilistic graphical weight of the single-step reaction in the Markov logic network based on the rate at which the single-step reaction occurs in actual reaction learning data, such that the larger the rate, the larger the probabilistic graphical weight. Thus, inverse synthesis paths consisting of more frequently occurring single-step reactions (i.e., more common single-step reactions) can be preferentially predicted.
[0122] Figure 7 An exemplary configuration of a computing device 1200 capable of implementing embodiments of the present disclosure is shown.
[0123] Computing device 1200 is an example of a hardware device capable of applying the above aspects of this disclosure. Computing device 1200 can be any machine configured to perform processing and / or computation. Computing device 1200 can be, but is not limited to, a workstation, server, desktop computer, laptop computer, tablet computer, personal data assistant (PDA), smartphone, in-vehicle computer, or a combination thereof.
[0124] like Figure 7As shown, computing device 1200 may include one or more components that can be connected to or communicate with bus 1202 via one or more interfaces. Bus 1202 may include, but is not limited to, Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus. Computing device 1200 may include, for example, one or more processors 1204, one or more input devices 1206, and one or more output devices 1208. The one or more processors 1204 may be any type of processor and may include, but is not limited to, one or more general-purpose processors or dedicated processors (such as dedicated processing chips). Processor 1204 may, for example, be configured to implement the method described above for retrosynthetic analysis of compound molecules. Input device 1206 may be any type of input device capable of inputting information to the computing device and may include, but is not limited to, a mouse, keyboard, touchscreen, microphone, and / or remote controller. The output device 1208 can be any type of device capable of presenting information, and may include, but is not limited to, a monitor, speaker, video / audio output terminal, vibrator and / or printer.
[0125] The computing device 1200 may also include or be connected to a non-transitory storage device 1214, which may be any non-transitory storage device capable of storing data, and may include, but is not limited to, disk drives, optical storage devices, solid-state storage, floppy disks, flexible disks, hard disks, magnetic tapes or any other magnetic media, compressed disks or any other optical media, cache memory and / or any other storage chip or module, and / or any other medium from which a computer may read data, instructions and / or code. The computing device 1200 may also include random access memory (RAM) 1210 and read-only memory (ROM) 1212. ROM 1212 may store executable programs, utilities, or processes in a non-volatile manner. RAM 1210 provides volatile data storage and stores instructions related to the operation of the computing device 1200. The computing device 1200 may also include a network / bus interface 1216 coupled to a data link 1218. Network / bus interface 1216 can be any kind of device or system capable of enabling communication with external devices and / or networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication devices and / or chipsets (such as Bluetooth). TM Equipment, 802.11 equipment, WiFi equipment, WiMax equipment, cellular communication facilities, etc.
[0126] This disclosure can be implemented as any combination of apparatus, system, integrated circuit, and computer program on a non-transitory computer-readable medium. One or more processors can be implemented as integrated circuits (ICs), application-specific integrated circuits (ASICs), or large-scale integrated circuits (LSIs), system LSIs, super LSIs, or ultra LSI components that perform some or all of the functions described in this disclosure.
[0127] This disclosure includes the use of software, application programs, computer programs, or algorithms. Software, application programs, computer programs, or algorithms may be stored on a non-transitory computer-readable medium to cause a computer, such as one or more processors, to perform the steps described above and in the accompanying drawings. For example, one or more memories may store the software or algorithm in executable instructions, and one or more processors may be associated with executing a set of instructions of the software or algorithm to provide various functionalities according to embodiments described in this disclosure.
[0128] Software and computer programs (also referred to as programs, software applications, applications, components, or code) include machine instructions for programmable processors and can be implemented in high-level procedural languages, object-oriented programming languages, functional programming languages, logic programming languages, assembly languages, or machine languages. The term "computer-readable medium" means any computer program product, apparatus, or device used to provide machine instructions or data to a programmable data processor, such as magnetic disks, optical disks, solid-state storage devices, memories, and programmable logic devices (PLDs), including computer-readable media that receive machine instructions as computer-readable signals.
[0129] For example, computer-readable media may include dynamic random access memory (DRAM), random access memory (RAM), read-only memory (ROM), electrically erasable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage devices, magnetic disk storage devices or other magnetic storage devices, or any other medium that can be used to carry or store required computer-readable program code in the form of instructions or data structures, and that can be accessed by a general-purpose or special-purpose computer or a general-purpose or special-purpose processor. As used herein, a disk or disc includes compact discs (CD), laser discs, optical discs, digital versatile discs (DVD), floppy disks, and Blu-ray discs, wherein a disk typically copies data magnetically, while a disc copies data optically using a laser. Combinations of the above are also included within the scope of computer-readable media.
[0130] The subject matter of this disclosure is provided as examples of apparatus, systems, methods, and programs for performing the features described herein. However, other features or variations are contemplated in addition to those described above. It is anticipated that the components and functions of this disclosure can be implemented using any emerging techniques that may replace any of the above-described implementations.
[0131] Furthermore, the above description provides examples and does not limit the scope, applicability, or configuration set forth in the claims. Changes may be made to the function and arrangement of the elements discussed without departing from the spirit and scope of this disclosure. Various processes or components may be appropriately omitted, substituted, or added in various embodiments. For example, features described with respect to certain embodiments may be combined in other embodiments.
[0132] Furthermore, in the description of this disclosure, the terms “first,” “second,” “third,” etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or order.
[0133] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring the operations to be performed in the specific order shown or in sequential order, or requiring the execution of all illustrated operations to achieve the desired result. In some cases, multitasking and parallel processing can be advantageous.
Claims
1. A method for retrosynthetic analysis of compound molecules, comprising: Based on actual reaction learning data, the probabilistic graphical weights of single-step reactions of the Markov logic network are determined, thereby obtaining the trained Markov logic network. Receive the encoded molecular structure of the target compound; The encoded molecular structure is input into a trained Markov logic network, and the Markov logic network is used to predict multiple candidate retrosynthetic pathways for the target compound. The multiple candidate inverse synthesis paths are input into the reinforcement learning model, and the reinforcement learning model is trained based on the existing actual reaction process. The paths among the multiple candidate inverse synthesis paths that match the actual synthesis paths are rewarded, and the transition matrix in the reinforcement learning model is updated based on the obtained rewards. as well as Using a reinforcement learning model, the final obtained transition matrix is used to determine one or more selected inverse synthesis paths from the multiple candidate inverse synthesis paths.
2. The method according to claim 1, wherein, Training reinforcement learning models includes: Determine whether each single-step reaction in each candidate retrosynthetic pathway is consistent with an existing actual single-step reaction; and If a specific single-step reaction in a candidate retrosynthetic pathway is consistent with an existing actual single-step reaction, then the candidate retrosynthetic pathway is given a single-step reaction reward value.
3. The method according to claim 2, wherein, The single-step reaction reward value assigned to the candidate retrosynthetic pathway includes: Based on the reactant cost and chemical reaction cost of the specific single-step reaction, a single-step reaction reward value is determined such that the lower the reactant cost or chemical reaction cost, the greater the single-step reaction reward value.
4. The method according to claim 2, wherein, Training reinforcement learning models includes: For each single-step reaction in a candidate retrosynthetic path, a negative step reward value is assigned to the candidate retrosynthetic path.
5. The method according to claim 2, wherein, The single-step reaction reward value assigned to the candidate retrosynthetic pathway includes: The reward value for a single-step reaction is determined based on the properties of the reactants in that specific single-step reaction.
6. The method according to claim 5, wherein, The properties include one or more of the following: chemical stability of the reactant, storage requirements, chemical toxicity, environmental hazards, and ease of commercial availability.
7. The method according to claim 2, wherein, The single-step reaction reward value assigned to the candidate retrosynthetic pathway includes: If a specific reactant in a particular single-step reaction is the same as a reactant in another single-step reaction in the candidate retrosynthetic pathway, then the candidate retrosynthetic pathway is given an additional bonus value.
8. The method according to claim 1, wherein, The probabilistic graphical weights for determining the single-step responses of a Markov logic network include: Based on the ratio of the single-step response to appear in the actual response learning data, the probability graph weight of the single-step response of the Markov logic network is determined such that the larger the ratio, the larger the probability graph weight.
9. An apparatus for retrosynthetic analysis of compound molecules, comprising: A memory that stores instructions; as well as The processor is configured to execute instructions stored in the memory to perform the method according to any one of claims 1 to 8.
10. A computer-readable storage medium comprising computer-executable instructions, which, when executed by one or more processors, cause the one or more processors to perform the method according to any one of claims 1 to 8.