A method and apparatus for generating multi-target molecules based on Monte Carlo tree search

Abstract

This invention discloses a method and apparatus for generating multi-target molecules based on Monte Carlo tree search, belonging to the field of target-based molecule generation. It includes: based on known protein sequences and molecular fragments, and user-defined multi-target attribute requirements, this invention employs a Monte Carlo tree construction method. Atom symbols that meet user requirements and have strong affinity are sequentially concatenated to the molecular fragments. The resulting complete molecule is then analyzed using chemical tools to obtain an evaluation value, which includes the user's multi-target attribute requirements. The constructed non-dominated molecules are then added to a global Pareto set for constructing the desired molecule. This invention uses a Monte Carlo tree search algorithm and proposes multi-target selection criteria, achieving molecule generation with user-specified multi-target attributes. It has the advantages of being simple, fast, effective, and having high search efficiency.

Description

Technical Field

[0001] This invention belongs to the field of target-based molecular generation, specifically relating to a method and apparatus for multi-target molecular generation based on Monte Carlo tree search. Background Technology

[0002] Target-based molecular generation is of great significance in drug development and other chemical research fields. In drug development, it can reveal the structure and interaction mechanisms between drugs and targets, providing data for drug molecule design and optimization, thus helping to improve the success rate of drug development and reduce costs. In other chemical research fields, understanding the interactions between molecules and targets can provide technical support for the laboratory preparation of more effective chemicals and catalysts.

[0003] Existing target-based molecular generation methods mainly fall into two categories. One is a multi-step generation method based on an autoregressive generation model. Patent document CN116779060A discloses a drug design method based on an autoregressive model, which includes: applying a word segmentation algorithm to biological text processing to train protein and ligand information in the dataset to obtain protein and ligand segmenters, and constructing an autoregressive model segmenter based on these; processing and transforming the original data in the dataset into a text format suitable for the autoregressive model, and then encoding it through the obtained segmenter to construct the training dataset required for the autoregressive model; training the autoregressive model using the training dataset to enable it to understand the SMILES representation of ligands and learn the interaction patterns between proteins and ligands; generating predicted ligand data using the trained autoregressive model, and performing post-processing using chemical information tools to obtain candidate ligands with specific chemical structures; evaluating and optimizing the obtained candidate ligands, and using chemical or bioinformatics tools for structure optimization and activity prediction to identify candidate molecules with the potential to become effective drugs.

[0004] However, this invention only involves the optimization of the structure and activity of molecules generated in the field of drug development, ignoring the multi-objective requirements that naturally exist in target-based molecule generation tasks, such as high drugability and low toxicity.

[0005] Another target-based molecular generation method is a one-step generation method based on a conditional generation model. Patent document CN111370074A discloses a method, apparatus, and computing device for generating molecular sequences, including: constructing a search tree for generating local sequences, where the initial local sequence is an empty character at the root node; extending multiple branches from the current local sequence, generating a complete molecule on each branch using the current molecular generation model, and calculating a score for each molecule; selecting the molecular sequence with the highest score, and if the score of this molecule is greater than the current target score, then this molecule is taken as a new target molecule, and its score is taken as the new target score; determining the next character of the current local sequence in the new target molecule, adding the next character to obtain a new local sequence; setting the new local sequence as the current local sequence, and repeating the above steps until no new target molecules are generated and the local sequences of the search tree are the same as the final target molecule.

[0006] However, this invention only involves improving the efficiency of molecular generation and lacks the ability to optimize multiple molecular properties at the same time, so it cannot meet the multi-objective needs of actual molecular generation. Summary of the Invention

[0007] The purpose of this invention is to provide a method and apparatus for generating multi-target molecules based on Monte Carlo tree search, which solves the problem that existing target-based molecule generation methods cannot simultaneously optimize multiple attribute targets.

[0008] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0009] In a first aspect, the present invention provides a method for generating multi-target molecules based on Monte Carlo tree search, comprising the following steps:

[0010] Step 1: Use the molecular fragment as the root node of the Monta Carlo tree and set a strong affinity threshold;

[0011] Step 2: Input the molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the first probability and its corresponding first-layer atomic symbols. Calculate the multi-objective selection criteria based on the first probability. Use the first-layer atomic symbols whose multi-objective selection criteria are greater than the strong affinity threshold as leaf nodes and splice them to the molecular fragments to obtain the first-layer molecular fragments.

[0012] Step 3: Input the first layer molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the second probability and its corresponding second layer atomic symbols. Calculate the multi-objective selection criterion based on the second probability. Use the second layer atomic symbols with the multi-objective selection criterion greater than the strong affinity threshold as leaf nodes and splice them to the first layer molecular fragments to obtain the second layer molecular fragments.

[0013] Step 4: Repeat steps 2 and 3 to update the evaluation values ​​of the molecules in reverse to obtain the evaluation values ​​of the leaf nodes, assisting in the multi-objective selection criteria, and iteratively update until the termination condition is met; add the non-dominated molecules to the global Pareto set based on the evaluation values ​​of the molecules to obtain the multi-objective molecules that meet the user's requirements.

[0014] This invention employs a Monte Carlo tree search algorithm. Based on known molecular fragments, target protein sequences, and user multi-target attributes, it first uses a pre-trained autoregressive model to obtain the first layer of atomic symbols and their probability distributions in the Monte Carlo tree. The molecular fragments are then used as the root node of the Monte Carlo tree. According to the probability distributions, the first layer of atomic symbols are concatenated to the molecular fragments to obtain the first layer of molecular fragments. These, along with the protein sequence, are then used in conjunction with a pre-trained autoregressive model to obtain the second layer of atomic symbols and their probability distributions. Finally, the second layer of atomic symbols are concatenated to the first layer of molecular fragments to obtain the second layer of molecular fragments.

[0015] After multiple iterations, the obtained complete molecules are characterized by a user-given objective function to obtain evaluation values. Following the branches of the Monte Carlo tree, the average of the evaluation values ​​of the complete molecules obtained from a given leaf node is used as the evaluation value of that leaf node. During the iterative process, to balance access to leaf nodes with high evaluation values ​​and those with strong affinity, this invention also proposes a multi-objective selection criterion. This process continues iteratively until a termination condition is met.

[0016] Based on the final evaluation value of the molecules, the non-dominated molecules are added to the global Pareto set to obtain multi-target molecules that meet the user's requirements.

[0017] Furthermore, the evaluation value of the molecule is obtained by the user providing objective functions based on the multi-objective attributes of the multi-objective molecule that meets the user's needs. Each objective function evaluates one objective attribute, resulting in multiple evaluation values ​​that represent different objective attributes. The average of the multiple evaluation values ​​is then calculated as the evaluation value of the molecule.

[0018] Furthermore, the multi-objective selection criteria are expressed by the formula:

[0019]

[0020] Where Up represents the multi-objective selection criterion, and W a N represents the evaluation value of leaf node a. a Let represent the number of times leaf node a is visited, c represent a constant balancing visit evaluation values ​​for leaf nodes with high affinity, P(a|C) represent the probability that the pre-trained autoregressive model selects leaf node a given the current input molecular fragments and protein sequences, N represents the total number of Monte Carlo tree search simulations, and N and N a The initial value is 1;

[0021] First item This indicates that the Monte Carlo tree search algorithm tends to visit leaf nodes with high evaluation values; the second term This indicates that the guided Monte Carlo tree search algorithm tends to visit leaf nodes with strong affinity.

[0022] Furthermore, the step of adding non-dominated molecules to the global Pareto set based on their evaluation values ​​includes:

[0023] The evaluation value of the newly generated molecule is compared with the evaluation values ​​of existing molecules in the global Pareto set. If the evaluation value of the newly generated molecule is higher than the evaluation value of existing molecules in the global Pareto set, the newly generated molecule is a non-dominated molecule. The newly generated molecule is added to the global Pareto set, and molecules dominated by the newly generated molecule are removed from the global Pareto set.

[0024] Furthermore, in a single reverse update, the molecule's evaluation value is updated in reverse to each leaf node that splices the molecule; in multiple reverse updates, the evaluation value of a certain leaf node is the average of the evaluation values ​​of molecules generated from that leaf node.

[0025] Furthermore, the auxiliary multi-objective selection criterion means that, based on the second term of the multi-objective selection criterion, the Monte Carlo tree search algorithm initially tends to visit leaf nodes with strong affinity; when a leaf node whose evaluation value has been updated is visited, the evaluation value of the leaf node is used to make the first term of the multi-objective selection criterion non-zero, and the auxiliary multi-objective selection criterion balances visiting leaf nodes with high evaluation values ​​and leaf nodes with strong affinity.

[0026] Furthermore, the termination conditions include: reaching a predefined maximum number of Monte Carlo tree generation levels, or a maximum number of iterations.

[0027] Secondly, in order to achieve the above-mentioned objectives, the present invention also provides a multi-target molecule generation device based on Monte Carlo tree search, including a search preparation module, a selection module, an expansion module, and an update module;

[0028] The search preparation module is used to set a strong affinity threshold by using molecular fragments as the root node of a Monte Carlo tree.

[0029] The selection module is used to input molecular fragments and protein sequences into a pre-trained autoregressive model to obtain a first probability and its corresponding first-layer atomic symbols. Based on the first probability, a multi-objective selection criterion is calculated. The first-layer atomic symbols with a multi-objective selection criterion greater than the strong affinity threshold are used as leaf nodes and spliced ​​to the molecular fragments to obtain the first-layer molecular fragments.

[0030] The extension module is used to input the first layer molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the second probability and its corresponding second layer atomic symbols. Based on the second probability, the multi-objective selection criterion is calculated. The second layer atomic symbols with the multi-objective selection criterion greater than the strong affinity threshold are used as leaf nodes and spliced ​​to the first layer molecular fragments to obtain the second layer molecular fragments.

[0031] The update module is used to repeat the selection module and the expansion module. It updates the evaluation value of the molecule in reverse to obtain the evaluation value of the leaf node, assists in the multi-objective selection criteria, and iterates until the termination condition is met. Based on the evaluation value of the molecule, the non-dominated molecules are added to the global Pareto set to obtain multi-objective molecules that meet the user's requirements.

[0032] Thirdly, to achieve the above-mentioned objectives, embodiments of the present invention also provide a multi-target molecule generation device based on Monte Carlo tree search, including a memory and a processor. The memory is used to store a computer program, and the processor is used to implement the multi-target molecule generation method based on Monte Carlo tree search provided in the first aspect of the present invention when the computer program is executed.

[0033] Fourthly, to achieve the above-mentioned objectives, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, wherein the computer program, when used with a computer, implements the multi-target molecule generation method based on Monte Carlo tree search provided in the first aspect of the present invention embodiments.

[0034] The beneficial effects of this invention are as follows:

[0035] (1) This invention uses the Monte Carlo tree search algorithm to store atomic symbols that can be spliced ​​to the target molecule fragments in each leaf node of the Monte Carlo tree. The leaf nodes also store probability and evaluation value data. The probability provided by the pre-trained autoregressive model represents the affinity between the current atomic symbol spliced ​​to the molecule fragment and the protein sequence. The generated molecule is passed through the objective function given by the user to obtain the evaluation value used to characterize the target attributes of the molecule. Finally, the molecule with the highest evaluation value is selected as the molecule that meets the user's multi-target attribute requirements. This invention combines molecular generation in the biological field with the Monte Carlo tree search algorithm in the computer field, which has the advantages of being simple, fast and effective.

[0036] (2) In order to improve search efficiency during the Monte Carlo tree search, this invention proposes a multi-target selection criterion to balance the access to leaf nodes with high evaluation values ​​and leaf nodes with strong affinity. Specifically, in the Monte Carlo tree search, in order to obtain molecules that meet the user's needs, the search initially tends to use leaf nodes with high access probability. After obtaining the molecules, the evaluation values ​​are updated to the leaf nodes in reverse. In subsequent search updates, as the evaluation values ​​in the leaf nodes increase, the first item in the multi-target selection criterion guides the access to leaf nodes with high evaluation values, ensuring that the target molecules formed have both strong affinity with the protein sequence and can meet the user's multi-target attribute requirements. Attached Figure Description

[0037] Figure 1 This is a flowchart of a multi-target molecule generation method based on Monte Carlo tree search provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram of the specific process of the multi-target molecule generation method based on Monte Carlo tree search provided in the embodiments of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of the multi-target molecule generation device based on Monte Carlo tree search provided in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram of the structure of a multi-target molecule generation device based on Monte Carlo tree search provided in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0042] The inventive concept of this invention is as follows: Based on known protein sequences and molecular fragments, as well as the user's multi-target attribute requirements, this invention adopts the method of constructing a Monte Carlo tree, sequentially splicing atomic symbols that meet the user's requirements and have strong affinity to the molecular fragments, and finally analyzing the properties of the complete molecule through chemical tools to obtain the molecule's evaluation value, which includes the user's requirements for the molecule's multi-target attributes. Then, the constructed non-dominated molecules are added to the global Pareto set to obtain multi-target targeted molecules that meet the user's requirements.

[0043] Monte Carlo Tree Search (MCS) is a heuristic search algorithm based on tree data structures that remains relatively effective even with a large search space. It primarily searches by approximating Nash equilibrium, combining the generality of stochastic simulation with the accuracy of tree search. Generally, the Monte Carlo Tree Search algorithm includes the following steps: (1) search; (2) expansion; (3) simulation; and (4) backpropagation.

[0044] Figure 1 This is a flowchart of a multi-target molecule generation method based on Monte Carlo tree search provided in an embodiment of the present invention. Figure 1 As shown, the embodiment provides a method for generating multi-target molecules based on Monte Carlo tree search, including the following steps:

[0045] S110 uses molecular fragments as the root node of a Monte Carlo tree to set a strong affinity threshold.

[0046] Before employing the Monte Carlo tree search algorithm, the molecular fragments are first used as the root nodes of the Monte Carlo tree, which are then used to sequentially attach atomic symbols to the molecular fragments during the subsequent algorithm search process.

[0047] In addition, to ensure that the resulting hierarchical molecular fragments obtained after splicing each atomic symbol to the molecular fragments have a strong affinity for the protein sequence, this embodiment also sets an additional strong affinity threshold.

[0048] S120: Input the molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the first probability and its corresponding first-layer atomic symbols. Calculate the multi-objective selection criteria based on the first probability. Use the first-layer atomic symbols with multi-objective selection criteria greater than the strong affinity threshold as leaf nodes and splice them to the molecular fragments to obtain the first-layer molecular fragments.

[0049] like Figure 2 As shown, in this embodiment, features are extracted from molecular fragments and protein sequences using a pre-trained graph neural network to obtain molecular embedding features and protein embedding features, respectively. These features are then input into a pre-trained autoregressive model, and the probability of the next atomic symbol is obtained through a softmax(·) normalization function. The probability of the next atomic symbol is a series of first probabilities used to evaluate the affinity between the atomic symbols spliced ​​onto the molecular fragments and the protein sequence. Following the order of the SMILE character matching rules, the first probability is the probability of the atomic symbol "C", the second probability is the probability of the atomic symbol "O", and so on, obtaining the first layer of atomic symbols and their corresponding first probabilities.

[0050] The multi-objective selection criterion Up is calculated based on the first probability. Leaf nodes with Up greater than the strong affinity threshold are selected for access, such as... Figure 2The selection section is shown. The probability obtained from the pre-trained autoregressive model corresponds to the affinity between the current atomic symbol spliced ​​into the molecular fragment and the protein sequence. The multi-objective selection criterion is expressed by the formula:

[0051]

[0052] Where Up represents the multi-objective selection criterion, and W a N represents the evaluation value of leaf node a. a Let represent the number of times leaf node a is visited, c represent a constant balancing visit evaluation values ​​for leaf nodes with high affinity, P(a|C) represent the probability that the pre-trained autoregressive model selects leaf node a given the current input molecular fragments and protein sequences, N represents the total number of Monte Carlo tree search simulations, and N and N a The initial value is 1; the first term This indicates that the Monte Carlo tree search algorithm tends to visit leaf nodes with high evaluation values; the second term This indicates that the guided Monte Carlo tree search algorithm tends to visit leaf nodes with strong affinity.

[0053] The first layer of molecular fragments is obtained by splicing the atomic symbols with strong affinity from the first layer of atomic symbols to the molecular fragments.

[0054] S130: Input the first layer molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the second probability and its corresponding second layer atomic symbols. Calculate the multi-objective selection criterion based on the second probability. Use the second layer atomic symbols with the multi-objective selection criterion greater than the strong affinity threshold as leaf nodes and splice them to the first layer molecular fragments to obtain the second layer molecular fragments.

[0055] The first-layer molecular fragments and protein sequences obtained from S120 are processed through a pre-trained autoregressive model to obtain second-layer atomic symbols and their corresponding second probabilities, which serve as extension information to guide the expansion of the Monte Carlo tree. Similarly, strong-affinity atomic symbols from the second-layer atomic symbols are concatenated to the first-layer molecular fragments to obtain the second-layer molecular fragments. Figure 2 As shown, when constructing leaf nodes from the root node downwards, the Monte Carlo leaf nodes are expanded in the manner described above, and the expanded nodes are child nodes.

[0056] S140, repeat S120 and S130, update the evaluation value of the molecule in reverse to obtain the evaluation value of the leaf node, assist the multi-objective selection criterion, and iterate until the termination condition is met; add the non-dominated molecules to the global Pareto set according to the evaluation value of the molecules to obtain the multi-objective molecules that meet the user's requirements.

[0057] During the repetition of S120 and S130, the splicing begins from the molecular fragment corresponding to the root node until the visited leaf node is the termination node, that is, the atomic symbol in the leaf node is a custom termination symbol. In this embodiment, the custom termination symbol is the $ symbol. Thus, the spliced ​​molecule is a complete molecule until the leaf node containing the termination symbol is the first leaf node with actual meaning.

[0058] Based on the user's multi-objective requirements for the target molecule, the user provides objective functions to evaluate the target attributes of the molecule. Each objective function evaluates one target attribute of the generated complete molecule and obtains a corresponding evaluation value. The average of the multiple evaluation values ​​obtained from the user's multi-objective attributes is then used to update each leaf node in the path leading to the molecule. During each update, the evaluation values ​​of each leaf node in the path are the same. When a leaf node is on a different path and a new average molecular evaluation value is obtained, the average of the evaluation values ​​at that leaf node is used as the evaluation value for that leaf node. For example, Figure 2 As shown, leaf nodes corresponding to high evaluation values ​​are Pareto nodes, indicating that the chemical properties of the molecule obtained after splicing the atomic symbols corresponding to this leaf node into molecular fragments better meet the user's needs. The evaluation value in each leaf node can be used to assist the multi-objective selection criterion. The Monte Carlo tree search algorithm initially tends to visit leaf nodes with strong affinity; when a leaf node whose evaluation value has been updated is visited, the evaluation value of the leaf node is used to make the first term of the multi-objective selection criterion non-zero, assisting the multi-objective selection criterion in balancing the access to leaf nodes with high evaluation values ​​and leaf nodes with strong affinity, thus ensuring a more efficient search. During the search process, the obtained molecules are stored in a pre-constructed global Pareto set. Based on the evaluation value corresponding to the molecule, each time a molecule is stored, non-dominated molecules are stored, and dominated molecules in the global Pareto set are removed.

[0059] like Figure 2 As shown, the process of selection, expansion, simulation, and updating is repeated until a termination condition is met. The termination condition includes: the atomic symbol in the visited leaf node is a predefined termination symbol; or a predefined maximum Monte Carlo tree generation layer or maximum number of iterations is reached. In this embodiment, the maximum Monte Carlo tree generation layer is set to 50 layers, and the maximum number of iterations is set to 1000. Based on the molecule's evaluation value R = (r1, ..., r... d Adding non-dominated molecules to the global Pareto set, where the non-dominated molecules are global Pareto molecules, yields multi-objective molecules that satisfy user needs, where R represents the set of user multi-objective attribute needs, r1, ..., r2. d It represents multiple different attributes.

[0060] Based on the same inventive concept, embodiments of the present invention also provide a multi-target molecule generation device 500 based on Monte Carlo tree search, such as... Figure 3 As shown, it includes a search preparation module 510, a selection module 520, an expansion module 530, and an update module 540;

[0061] Among them, the search preparation module 510 is used to set a strong affinity threshold by taking molecular fragments as the root node of the Monte Carlo tree.

[0062] The selection module 520 is used to input molecular fragments and protein sequences into a pre-trained autoregressive model to obtain the first probability and its corresponding first-layer atomic symbols. Based on the first probability, a multi-objective selection criterion is calculated. The first-layer atomic symbols with a multi-objective selection criterion greater than the strong affinity threshold are used as leaf nodes and spliced ​​to the molecular fragments to obtain the first-layer molecular fragments.

[0063] The extension module 530 is used to input the first layer molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the second probability and its corresponding second layer atomic symbols. Based on the second probability, the multi-objective selection criterion is calculated. The second layer atomic symbols with the multi-objective selection criterion greater than the strong affinity threshold are used as leaf nodes and spliced ​​to the first layer molecular fragments to obtain the second layer molecular fragments.

[0064] The update module 540 is used for repeating the selection module and the expansion module. It updates the evaluation value of the molecules in reverse to obtain the evaluation value of the leaf nodes, assists in the multi-objective selection criteria, and iterates until the termination condition is met. Based on the evaluation value of the molecules, the non-dominated molecules are added to the global Pareto set to obtain multi-objective molecules that meet the user's requirements.

[0065] For the multi-target molecule generation device based on Monte Carlo tree search provided in the embodiments of the present invention, since it basically corresponds to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this invention without creative effort.

[0066] Based on the same inventive concept, the embodiments also provide a multi-target molecule generation device based on Monte Carlo tree search, such as... Figure 4 As shown, it includes a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to implement the above-described multi-target molecule generation method based on Monte Carlo tree search when the computer program is executed.

[0067] The multi-target molecule generation device based on Monte Carlo tree search proposed in this invention can be a device such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, it is formed by the processor of any device with data processing capabilities loading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective... Figure 4 This is a schematic diagram of the structure of a multi-target molecule generation device based on Monte Carlo tree search provided in an embodiment of the present invention, except... Figure 4 In addition to the processor, memory, network interface, and non-volatile memory shown, the multi-target molecule generation device based on Monte Carlo tree search provided in this embodiment of the invention may also include other hardware depending on the actual function of any device with data processing capabilities, which will not be described in detail here.

[0068] Based on the same inventive concept, the embodiment also provides a computer-readable storage medium storing a computer program, which, when used by a computer, implements the above-described multi-target molecule generation method based on Monte Carlo tree search.

[0069] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of a wind turbine, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., mounted on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0070] It should be noted that the multi-target molecule generation device, multi-target molecule generation equipment, and computer-readable storage medium based on Monte Carlo tree search provided in the above embodiments all belong to the same concept as the multi-target molecule generation method embodiments based on Monte Carlo tree search. For details of their specific implementation process, please refer to the multi-target molecule generation method embodiments based on Monte Carlo tree search, which will not be repeated here.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for generating multi-target molecules based on Monte Carlo tree search, characterized in that, Includes the following steps: Step 1: Using the molecular fragment as the root node of the Monte Carlo tree, set a strong affinity threshold; Step 2: Input the molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the first probability and its corresponding first-layer atomic symbols. Calculate the multi-objective selection criteria based on the first probability. Use the first-layer atomic symbols whose multi-objective selection criteria are greater than the strong affinity threshold as leaf nodes and splice them to the molecular fragments to obtain the first-layer molecular fragments. Step 3: Input the first layer molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the second probability and its corresponding second layer atomic symbols. Calculate the multi-objective selection criterion based on the second probability. Use the second layer atomic symbols with the multi-objective selection criterion greater than the strong affinity threshold as leaf nodes and splice them to the first layer molecular fragments to obtain the second layer molecular fragments. Step 4: Repeat steps 2 and 3 to update the evaluation values ​​of the molecules in reverse to obtain the evaluation values ​​of the leaf nodes, assisting in the multi-objective selection criteria, and iteratively update until the termination condition is met; add the non-dominated molecules to the global Pareto set based on the evaluation values ​​of the molecules to obtain the multi-objective molecules that meet the user's requirements. The step of adding non-dominated molecules to the global Pareto set based on their evaluation values ​​includes: The evaluation value of the newly generated molecule is compared with the evaluation values ​​of existing molecules in the global Pareto set. If the evaluation value of the newly generated molecule is higher than the evaluation value of existing molecules in the global Pareto set, the newly generated molecule is a non-dominated molecule. The newly generated molecule is added to the global Pareto set, and molecules dominated by the newly generated molecule are removed from the global Pareto set.

2. The multi-target molecule generation method based on Monte Carlo tree search according to claim 1, characterized in that, The molecule's evaluation value is determined based on the multi-objective attributes of a multi-objective molecule that meets user needs. The user provides objective functions, each evaluating one objective attribute, resulting in multiple evaluation values ​​representing different objective attributes. The average of these evaluation values ​​is then used as the molecule's evaluation value. The multi-objective selection criteria are expressed by the following formula: in, Up Indicates the criteria for selecting multiple objectives. W a Leaf nodes The evaluation value, N a Leaf nodes The number of times the leaf nodes are visited, where c represents a constant balancing visit evaluation values ​​between leaf nodes with high affinity, and P(a|C) represents the leaf node selection by the pre-trained autoregressive model given the current input molecular fragments and protein sequences. The probability, N represents the total number of simulations for the Monte Carlo tree search, N and N a The initial value is 1; First item This indicates that the Monte Carlo tree search algorithm tends to visit leaf nodes with high evaluation values; the second term This indicates that the guided Monte Carlo tree search algorithm tends to visit leaf nodes with strong affinity.

3. The multi-target molecule generation method based on Monte Carlo tree search according to claim 2, characterized in that, In a single reverse update, the molecule's evaluation value is updated in reverse to each leaf node that splices the molecule; in multiple reverse updates, the evaluation value of a certain leaf node is the average of the evaluation values ​​of molecules generated from that leaf node.

4. The multi-target molecule generation method based on Monte Carlo tree search according to claim 2, characterized in that, The auxiliary multi-objective selection criterion means that, based on the second term of the multi-objective selection criterion, the Monte Carlo tree search algorithm initially tends to visit leaf nodes with strong affinity; when a leaf node whose evaluation value has been updated is visited, the evaluation value of the leaf node is used to make the first term of the multi-objective selection criterion non-zero, and the auxiliary multi-objective selection criterion balances visiting leaf nodes with high evaluation values ​​and leaf nodes with strong affinity.

5. The multi-target molecule generation method based on Monte Carlo tree search according to claim 1, characterized in that, The termination conditions include: reaching a predefined maximum number of Monte Carlo tree generation levels, or a maximum number of iterations.

6. A multi-target molecule generation device based on Monte Carlo tree search, characterized in that, This includes a search preparation module, a selection module, an expansion module, and an update module; The search preparation module is used to set a strong affinity threshold by using molecular fragments as the root node of the Monta Carlo tree. The selection module is used to input molecular fragments and protein sequences into a pre-trained autoregressive model to obtain a first probability and its corresponding first-layer atomic symbols. Based on the first probability, a multi-objective selection criterion is calculated. The first-layer atomic symbols with a multi-objective selection criterion greater than the strong affinity threshold are used as leaf nodes and spliced ​​to the molecular fragments to obtain the first-layer molecular fragments. The extension module is used to input the first layer molecular fragments and protein sequences into the pre-trained autoregressive model to obtain the second probability and its corresponding second layer atomic symbols. Based on the second probability, the multi-objective selection criterion is calculated. The second layer atomic symbols with the multi-objective selection criterion greater than the strong affinity threshold are used as leaf nodes and spliced ​​to the first layer molecular fragments to obtain the second layer molecular fragments. The update module is used to repeat the selection module and the expansion module. It updates the evaluation value of the molecule in reverse to obtain the evaluation value of the leaf node, assists in the multi-objective selection criteria, and iterates until the termination condition is met. Based on the evaluation value of the molecule, the non-dominated molecules are added to the global Pareto set to obtain multi-objective molecules that meet the user's requirements. The step of adding non-dominated molecules to the global Pareto set based on their evaluation values ​​includes: The evaluation value of the newly generated molecule is compared with the evaluation values ​​of existing molecules in the global Pareto set. If the evaluation value of the newly generated molecule is higher than the evaluation value of existing molecules in the global Pareto set, the newly generated molecule is a non-dominated molecule. The newly generated molecule is added to the global Pareto set, and molecules dominated by the newly generated molecule are removed from the global Pareto set.

7. A multi-target molecule generation device based on Monte Carlo tree search, comprising a memory and a processor, wherein the memory is used to store a computer program, characterized in that, The processor is configured to implement the multi-target molecule generation method based on Monte Carlo tree search as described in any one of claims 1-5 when executing the computer program.

8. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is used, it implements the multi-target molecule generation method based on Monte Carlo tree search as described in any one of claims 1-5.

Images