Drug molecule optimization design method based on deep reinforcement learning and skeleton constraint

Abstract

The invention discloses a drug molecule optimization design method based on deep reinforcement learning and skeleton constraint, and belongs to the technical field of drug molecule optimization design, and the method comprises the steps: firstly building a molecule generation model based on deep learning, and training the molecule generation model in combination with data of an open source database; carrying out transfer learning on the pre-trained molecular generation model by utilizing a target attribute related compound database; on this basis, a strategy gradient algorithm is constructed, reinforcement learning based on target attributes is carried out in combination with a scoring model result, and skeleton constraints are added in a reinforcement learning sampling process to generate molecules with specific skeleton structures; and finally, comprehensively evaluating and screening the generated molecules by utilizing a multi-dimensional evaluation index, thereby identifying the patent medicine molecules with further research and development potential, rapidly obtaining molecular structures meeting target requirements, and greatly improving the efficiency of optimal design of the medicine molecules.

Description

Technical Field

[0001] The present invention relates to the technical field of drug molecule optimization design, and in particular to a drug molecule optimization design method based on deep reinforcement learning and skeleton constraints. Background Art

[0002] Drug development is a complex and lengthy process that usually includes the following stages: target discovery (identifying disease-related biomolecular targets), lead compound discovery (screening and identifying compounds that can bind to targets and produce biological activity), lead compound optimization (optimizing lead compounds through structural modification and chemical synthesis to improve their pharmacological and pharmacokinetic properties), preclinical studies (conducting in vitro and in vivo experiments to evaluate the safety and efficacy of candidate drugs), clinical trials (testing the safety and efficacy of drugs in human subjects), and drug approval (review and approval by regulatory agencies). Drug molecule design optimization is a crucial stage in the drug development process, which aims to improve the selectivity and affinity of drugs and the efficacy and safety of candidate drugs through systematic design and improvement in the lead compound discovery and optimization stages of drug development. Enhance the binding ability of drugs to targets by optimizing the molecular structure, while reducing interactions with non-targets to reduce side effects; improve the ADME characteristics of drugs, and improve the absorption, distribution, metabolism and excretion of drugs by adjusting the physicochemical properties of molecules; enhance the bioavailability of drugs, and improve the stability and effective concentration of drugs in the body through structural modification; reduce the toxicity of drugs, and reduce adverse reactions and toxicity through refined molecular design, thereby improving the safety of drugs.

[0003] Artificial intelligence-based molecular generation optimization technology is increasingly used in the field of drug design optimization. However, this technology faces a series of challenges in its implementation. The huge scale of chemical space far exceeds the order of magnitude of currently known compounds. This huge chemical space not only increases the complexity of the search, but also makes it more difficult to find molecules with specific functions. At the same time, molecular properties are extremely sensitive to small changes in structure, which requires molecular generation technology to have extremely high accuracy and robustness.

[0004] Traditional molecular generation methods, such as genetic algorithms, reinforcement learning, and generative adversarial networks, have made progress in some aspects, but they are insufficient in stability, reliability, and sample quality control. Genetic algorithms have high computational costs when dealing with large-scale search spaces, and their sensitivity to hyperparameters limits their application. Although reinforcement learning can optimize molecular structures through trial and error, the instability of its generation process limits its reliability in practical applications. Although generative adversarial networks have shown potential in generating innovative molecules, the uncontrollability of sample quality remains a problem that needs to be solved.

[0005] Given the limitations of existing technologies and the challenges of chemical space, it is of great significance to develop a method for optimizing the design of drug molecules based on deep reinforcement learning and skeleton constraints. This method should be able to integrate chemical knowledge and rules, optimize search strategies, and improve the generalization ability of the model to ensure that the generated molecules are both chemically reasonable and have the desired physicochemical properties. By improving the model architecture, training strategy, and evaluation mechanism, a more solid technical foundation can be provided for drug discovery. Summary of the invention

[0006] In response to the technical problems of low stability and reliability in the existing technology and difficulty in controlling the quality of generated samples, the present invention proposes a drug molecule optimization design method based on deep reinforcement learning and skeleton constraints to improve the accuracy and efficiency of drug discovery.

[0007] In a first aspect of the present invention, the present invention provides a method for optimizing the design of drug molecules based on deep reinforcement learning and skeleton constraints, comprising the following steps:

[0008] Step S1: constructing a molecular generation model based on a neural network;

[0009] Step S2: construct a training set, extract a screening data set from an open source database, and perform preprocessing;

[0010] Step S3: constructing a label library based on the training set and training the molecular generation model;

[0011] Step S4: construct a target attribute scoring function for reward calculation in molecular screening and reinforcement learning;

[0012] Step S5: Screening compounds with similar structures according to the molecules to be optimized as the migration data set, using the target attribute scoring function to eliminate molecules with low target attribute scores in the migration data set; using the migration data set for transfer learning to obtain an updated molecule generation model;

[0013] Step S6: construct a policy gradient algorithm to perform reinforcement learning on the molecular generation model after transfer learning to obtain an optimized molecular generation model;

[0014] Step S7: Sampling based on the molecular skeleton to be optimized according to the optimized molecular generation model to generate a molecular library to be further screened;

[0015] Step S8: Use the evaluation indicators to comprehensively evaluate and screen the molecules generated by the model, and guide the laboratory's next step of molecular optimization and chemical synthesis based on the preferred molecules.

[0016] Further, step S1 includes:

[0017] The molecular generation model includes a data input module, a generator, an optimizer and an output module;

[0018] The input module is used to input the chemical information of the compound into the generator, the generator generates a representation of the molecule based on the input data, the optimizer obtains the model that minimizes the loss function through iterative updating, and the output module outputs the result of the model output corresponding to the minimum loss function.

[0019] Further, step S2 includes:

[0020] The SMILES strings of druggable molecules were obtained from an open source database as a training set; the training set was preprocessed to remove duplicate molecules and invalid molecules, separate metal atoms bound to the molecules, and convert all molecules into a standardized SMILES form.

[0021] Further, step S3 includes:

[0022] Screening is performed based on the molecular composition and chemical expertise in the training set to determine a tag library; the tag library includes single characters, atomic types composed of two characters, and special environments;

[0023] Adopt one-hot encoding and use the tag library to encode each SMILES string;

[0024] Train neural networks for sequence modeling by maximizing the next token x in the target sequence t The goal is to maximize the probability of the correct label being assigned:

[0025]

[0026] During the training process, the model gradually grasps the rules and patterns of molecular generation by learning the sequence information of the compound structure; the model performance is evaluated using the validation set to ensure that the model can generate reasonable and diverse molecular structures, and the generation ability and stability of the model are optimized by adjusting the hyperparameters.

[0027] Further, step S4 includes:

[0028] Obtain molecules with target property test results and SMILES strings from open source databases as training data sets for property prediction models;

[0029] According to the optimization requirements of the target attributes, a target attribute prediction model is constructed, and a supervised learning regression model is used to predict the specific attributes of the molecule;

[0030] The model generates corresponding property prediction values ​​based on the input molecular chemical structure information; the molecular chemical structure information can be selectively input into SMILES strings, molecular fingerprints, molecular graph representations or 3D spatial position information according to the characteristics of the target attribute; these molecular information will be converted into processable feature representations in the model, and processed through a multi-layer neural network to ultimately output the predicted value of the target attribute.

[0031] Establish a target property scoring model, and score the target properties of the molecular compounds obtained by the target property prediction model according to the optimization target; the scoring method can be any of the following:

[0032] (1) When the target attribute is within a certain range, the score is high, and when it is outside the range, the score is 0;

[0033] (2) The higher the prediction result of the target attribute, the higher the score;

[0034] (3) The closer the target attribute is to the target value, the higher the score.

[0035] (4) Weight each attribute of multiple target attributes to generate a comprehensive score.

[0036] Furthermore, step S5 specifically includes:

[0037] According to the structure of the molecule to be optimized, a structural similarity algorithm is used to screen out compounds with similar structures from the database as the data set for transfer learning.

[0038] The screened compounds are evaluated using the target attribute scoring function. Based on the scoring results of the target attribute, molecules with low scores are eliminated and molecules with high scores are retained to ensure the quality of the transfer learning dataset.

[0039] The screened high-quality data set is input into the pre-trained molecular generation model, and the model parameters are adjusted through transfer learning so that the model can better generate molecules that meet the target property requirements.

[0040] Furthermore, step S6 further includes:

[0041] In the reinforcement learning framework, a policy gradient algorithm is constructed to optimize the molecule generation model; the policy gradient algorithm updates the model parameters by maximizing the reward function of the generated molecules, and the reward function is based on the feedback of the target attribute scoring function;

[0042] In the reinforcement learning process, the model gets reward values ​​based on the generated molecules, and the model parameters are continuously updated through the policy gradient algorithm to maximize the reward value.

[0043] After multiple iterations and training, reinforcement learning finally obtains an optimized molecular generation model that can better generate molecules that meet the target property requirements.

[0044] Further, step S7 includes:

[0045] Sampling is performed to obtain a sufficient number of generated molecular libraries based on the target skeleton for the next step of screening, and the quality of the generated molecular library is evaluated to ensure the diversity of the molecules in the generated molecular library and the conformity of the target attributes;

[0046] Multi-dimensional indicators are used to perform virtual screening of molecules in the generated molecular library, thresholds for specific evaluation indicators are set, molecules outside the threshold range are eliminated, and the remaining molecules are then comprehensively scored. The molecules with the highest comprehensive scores are selected to guide the molecular optimization and synthesis tasks of the medicinal chemistry laboratory.

[0047] Furthermore, multidimensional indicators include molecular synthesis feasibility, molecular docking, biological activity or drugability.

[0048] The present invention has the following beneficial effects:

[0049] The present invention discloses a drug molecule optimization design method based on deep reinforcement learning and skeleton constraints, belonging to the technical field of drug molecule optimization design. The present invention first establishes a molecule generation model based on deep learning, and trains the molecule generation model in combination with data from an open source database, and then uses a target attribute-related compound database to perform transfer learning on the pre-trained molecule generation model; on this basis, a policy gradient algorithm is constructed, and reinforcement learning based on the target attribute is performed in combination with the scoring model results, and skeleton constraints are added in the reinforcement learning sampling process to generate molecules with specific skeleton structures; finally, multi-dimensional evaluation indicators are used to comprehensively evaluate and screen the generated molecules, thereby identifying drug molecules with potential for further research and development, and being able to quickly obtain molecular structures that meet the target requirements, greatly improving the efficiency of drug molecule optimization design. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a brief flowchart of the drug molecule optimization design method based on deep reinforcement learning and skeleton constraints in the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present application to further clearly and completely describe the technical solutions in the embodiments of the present application. It should be noted that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of this application.

[0052] In order to make the invention purpose, technical scheme and advantages of this application clearer, the technical scheme of the present invention is described clearly and completely in conjunction with the accompanying drawings and specific implementation schemes. It should be pointed out that for ordinary technicians in this field, several changes and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of this application.

[0053] The most important step in new drug development is the optimization of lead compounds based on drugability, but traditional methods have long cycles, high costs, and may not necessarily produce the desired results. The development of artificial intelligence technology has promoted the development of drug research and development, and a series of cutting-edge technologies such as deep learning have effectively improved the efficiency of drug research and development.

[0054] The present invention proposes compound optimization based on the skeleton of candidate compounds, which has high generation capacity and single or multi-target molecular property optimization capabilities, can explore known and unfamiliar areas of chemical space, and provides a better choice for molecular optimization.

[0055] In the first embodiment of the present invention, the present invention discloses a drug molecule optimization design method based on deep reinforcement learning and skeleton constraints, comprising the steps of:

[0056] S1: Building a molecular generation model based on neural network;

[0057] A neural network architecture suitable for the molecule generation task is selected, such as a recurrent neural network (RNN) architecture. In this embodiment, a standard RNN, a long short-term memory network (LSTM) or a gated recurrent unit (GRU) can be selected, and an appropriate activation function, such as ReLU or tanh, can be selected.

[0058] The model input is defined as a SMILES string, and one-hot encoding is used to encode each SMILES string using a tag library.

[0059] Determine the number of RNN layers, the number of nodes in each layer, and other parameter settings based on the complexity of the task.

[0060] Generate new molecules using the trained model through sampling methods.

[0061] S2: Build training sets, extract screening data sets from open source databases, and perform preprocessing;

[0062] Obtain SMILES strings of druggable molecules from an open source database as a training set;

[0063] The training set was preprocessed to remove duplicate molecules and invalid molecules, separate the metal atoms bound to the molecules, and convert all molecules into standardized SMILES form.

[0064] S3: Build a label library based on the training set and train the molecular generation model;

[0065] Based on the inspection of the molecular composition in the training set and the screening of chemical expertise, a library of tokens is determined. The library of tokens consists of single characters, atom types consisting of two characters, and special environments represented by square brackets (such as [nH]).

[0066] Adopt one-hot encoding and use the tag library to encode each SMILES string;

[0067] Train neural networks for sequence modeling by maximizing the next token x in the target sequence t The goal is to maximize the probability of the correct label being assigned:

[0068]

[0069] During the training process, the model gradually grasps the rules and patterns of molecular generation by learning the sequence information of the compound structure; the model performance is evaluated using the validation set to ensure that the model can generate reasonable and diverse molecular structures, and the generation ability and stability of the model are optimized by adjusting the hyperparameters.

[0070] S4: Construct target attribute scoring functions for reward calculation in molecular screening and reinforcement learning.

[0071] Obtain molecules with target attribute laboratory test results and SMILES strings from open source databases as attribute prediction model training data sets, such as PubChem database, ChEMBL database, Tox21 database, etc.;

[0072] According to the optimization requirements of the target attribute, a target attribute prediction model is constructed, and a supervised learning regression model is used to predict the specific attributes of the molecule. By using the chemical structure of the known molecules in the training data and their corresponding target attributes, the regression model can learn and establish the mapping relationship between the structure and the attribute, thereby predicting the attributes of the newly generated molecules. In this embodiment, the target attributes include solubility, toxicity or pharmacokinetic properties.

[0073] The model generates corresponding property prediction values ​​based on the input molecular chemical structure information. The molecular chemical structure information can be selectively input into SMILES strings, molecular fingerprints, molecular graph representations, or 3D spatial position information according to the characteristics of the target attribute. These molecular information will be converted into feature representations that the model can process, and then processed by a multi-layer neural network to finally output the predicted value of the target attribute.

[0074] Establish a target property scoring model, and score the target properties of the molecular compounds obtained by the target property prediction model according to the optimization target; the scoring method can be selected from any of the following: when the target property is within a certain range, it is scored high, and when it is out of the range, it is scored 0; the higher the target property prediction result, the higher the score; the closer the target property is to the target value, the higher the score. When multiple target properties need to be optimized, each property can be weighted to generate a comprehensive score to consider multi-dimensional optimization goals.

[0075] S5: Based on the structure of the molecule to be optimized, compounds with similar structures are screened out as transfer learning datasets; molecules with low scores are screened out using the target attribute scoring function, and then the screened transfer dataset is used for transfer learning to update the molecule generation model;

[0076] According to the structure of the molecule to be optimized, a structural similarity algorithm is used to screen out compounds with similar structures from the database as a data set for transfer learning. In this embodiment, similarity calculation based on molecular fingerprint or graph representation is used.

[0077] The screened compounds are evaluated using the target attribute scoring function. Based on the scoring results of the target attribute, molecules with low scores are eliminated and molecules with high scores are retained to ensure the quality of the transfer learning dataset.

[0078] Use the transfer learning dataset to perform transfer learning and obtain an updated molecular generation model;

[0079] Transfer learning aims to use the desired optimized molecular properties as conditions to control the generation process so that the generated molecules meet specific property requirements.

[0080] Select a compound database that is highly relevant to the target attribute as a transfer learning database. Depending on the molecular optimization goal, this compound library can be a set of compounds with high molecular attributes or high specific bioactivity supported by laboratory data, or a set of compounds with similar structures to the generated skeleton.

[0081] The dataset is cleaned and preprocessed, and then input into the pre-trained molecular generation model for transfer learning. The specific process of transfer learning includes freezing part of the layers of the trained molecular generation model to keep the original weights and parameters; retraining the remaining layers of the model to better adapt to the new data. The frozen part of the layer is the underlying feature extraction part of the model, which contains the understanding of the basic structure of the compound. In order to optimize the specific properties of the generated drug molecules, transfer learning freezes part of the layers of the trained molecular generation model, and then uses smaller training data to train the molecular generation model to achieve the purpose of optimizing the drug molecule generation model.

[0082] This solution uses an iterative training approach to continuously adjust model parameters to achieve optimal performance on a database of compounds related to target properties. During the training process, a validation set is used for real-time evaluation to prevent overfitting of the model.

[0083] In an embodiment of the present invention, a pre-trained target property prediction model is used to predict the target properties of the output results of the molecular generation model after transfer learning, and then the target properties of the output molecules are scored according to the target scoring model. The training is continuously iterated until the target properties of the output molecules are consistent with the expected target properties, and the molecular generation model after the transfer learning is obtained.

[0084] S6: Construct a policy gradient algorithm to perform reinforcement learning on the molecular generation model after transfer learning to obtain an optimized molecular generation model;

[0085] Define the environment and goal of reinforcement learning, use the molecular generation model based on skeleton constraints as a carrier, generate molecular structures as inputs to the molecular property prediction model, and use the molecular property scoring model to evaluate the quality of each generated molecule, i.e. as a reward function. The goal is to continuously optimize the generated molecules to achieve the best performance in terms of molecular properties.

[0086] Through literature research and chemical laboratory research, the optimized molecular skeleton and optimized molecular properties are determined; skeleton constraints are added to the molecular model after transfer learning, and the molecular generation model based on skeleton constraints is input into the pre-trained molecular property prediction model, and the predicted molecular properties are compared with the optimized molecular properties. The molecular generation model based on skeleton constraints is iteratively trained by minimizing the difference between the predicted molecular properties and the expected optimized molecular properties, and the molecular generation model optimized by the policy gradient algorithm is obtained.

[0087] In an embodiment of the present application, optimizing molecular properties includes one or more of drug activity, toxicity, solubility or pharmacokinetic properties.

[0088] According to the policy gradient algorithm, the gradient of the policy network is calculated and these gradients are used to update the network parameters. By repeatedly generating samples, calculating rewards and updating policies, the weights and biases of the network are gradually adjusted by maximizing the expected value of the reward function until the performance of the policy network on the target attribute reaches a satisfactory level.

[0089] S7: Sampling is performed according to the optimized molecular generation model to generate a molecular library to be screened, and multi-dimensional evaluation indicators are used to comprehensively evaluate and screen the generated molecules, and the laboratory is guided for further optimization and chemical synthesis based on the preferred molecules.

[0090] The sample is sufficient to obtain a generated molecular library based on the target skeleton for the next step of screening. The generated molecular library is quality evaluated to ensure the diversity of the molecules in the generated molecular library and the conformity of the target attributes.

[0091] Multi-dimensional indicators are used to perform virtual screening of molecules in the generated molecular library, thresholds for specific evaluation indicators are set, molecules outside the threshold range are eliminated, and the remaining molecules are then comprehensively scored. The molecules with the highest comprehensive scores are selected to guide the molecular optimization and synthesis tasks of the medicinal chemistry laboratory.

[0092] Multidimensional indicators include molecular synthesis feasibility, molecular docking, biological activity or drugability.

[0093] In Example 2 of the present invention, the present invention takes the active skeleton of PDE4 inhibitor as an example and provides a method for optimizing and generating drug molecules based on the active skeleton of PDE4 inhibitor, which is used to solve the technical problem of optimizing the drugability of the target skeleton. This example is described in detail below.

[0094] S1. Select water solubility and biological activity as the target properties, and build a neural network for molecule generation, which will serve as the basic model for molecule generation;

[0095] S2. Obtain the zinc database and screen the database. On the basis of controlling the molecular diversity and the drug-forming properties of the molecules, screen the molecules predicted to have water solubility within the drug-forming range. Based on the screening database, determine the marker library and pre-treat the molecules.

[0096] S3, use the screened zinc database to pre-train a neural network, which will serve as the basic model for molecule generation;

[0097] S4. Based on these two target attributes, a target attribute prediction model is constructed. Water solubility is predicted using data from a public database, and PDE4 activity is modeled and predicted using laboratory activity test data. The model generates corresponding water solubility and activity prediction values ​​based on the input molecular chemical structure information (such as a SMILES string or a molecular fingerprint). In the scoring stage, after each molecule is predicted based on the two target attributes, the higher the attribute, the higher the score, and then the weighted sum is calculated based on the weight distribution. The specific weight distribution is: PDE4 activity accounts for 0.7, and water solubility accounts for 0.3;

[0098] S5. Based on the active skeleton of the molecule to be optimized, compounds with similar structures are screened in the molecular database as the data set for transfer learning. The target attribute scoring algorithm is called to score the molecules in the data set and the molecules with low scores are eliminated. The eliminated data set is input into the pre-trained molecular generation model, and the model parameters are further adjusted through transfer learning so that the model can generate molecules with good water solubility and PDE4 inhibitory activity;

[0099] S6. Based on the molecular generation model optimized by transfer learning, a policy gradient algorithm is constructed to further optimize the water solubility and PDE4 activity of the generated molecules. The reinforcement learning process is that each time a new molecule is generated, the model obtains a corresponding reward based on the molecular properties it generates; through multiple iterations and training, the model uses the policy gradient algorithm to continuously adjust its parameters so that the generated molecules gradually optimize the target properties. After each round of iteration, the molecules generated by the model will be evaluated based on the water solubility and PDE4 activity scores, and the generation strategy will be updated. After multiple rounds of reinforcement learning training, the final molecular generation model can better generate molecules that meet the target properties. In this embodiment, high PDE4 activity and good water solubility are selected as the molecular target properties;

[0100] S7. After reinforcement learning is completed and the molecular generation model is optimized, the molecules are sampled based on the model and the skeleton constraints. A large number of molecules with different structures are generated by sampling to form a molecular library to be further screened; the molecules generated by reinforcement learning are comprehensively evaluated and screened using multi-dimensional indicators (including molecular synthesis feasibility scores, molecular docking, biological activity prediction or drugability evaluation) to identify preferred molecules with PDE4 inhibitory activity that have the potential for further research and development.

[0101] The present invention discloses a drug molecule optimization design method based on deep reinforcement learning and skeleton constraints, belonging to the technical field of drug molecule optimization design. The present invention first establishes a molecule generation model based on deep learning, and trains the molecule generation model in combination with data from an open source database, and then uses a target attribute-related compound database to perform transfer learning on the pre-trained molecule generation model; on this basis, a policy gradient algorithm is constructed, and reinforcement learning based on the target attribute is performed in combination with the scoring model results, and skeleton constraints are added in the reinforcement learning sampling process to generate molecules with specific skeleton structures; finally, multi-dimensional evaluation indicators are used to comprehensively evaluate and screen the generated molecules, thereby identifying drug molecules with potential for further research and development, and being able to quickly obtain molecular structures that meet the target requirements, greatly improving the efficiency of drug molecule optimization design.

Claims

1. A drug molecule optimization design method based on deep reinforcement learning and skeleton constraints, characterized in that: include: Step S1: constructing a molecular generation model based on a neural network; Step S2: construct a training set, extract a screening data set from an open source database, and perform preprocessing; Step S3: constructing a label library based on the training set and training the molecular generation model; Step S4: construct a target attribute scoring function for reward calculation in molecular screening and reinforcement learning; Step S5: Screening compounds with similar structures according to the molecules to be optimized as the migration data set, and using the target attribute scoring function to eliminate molecules with low target attribute scores in the migration data set; Use the migration dataset to perform transfer learning and obtain an updated molecular generation model; Step S6: construct a policy gradient algorithm to perform reinforcement learning on the molecular generation model after transfer learning to obtain an optimized molecular generation model; Step S7: Sampling based on the molecular skeleton to be optimized according to the optimized molecular generation model to generate a molecular library to be further screened; Step S8: Use the evaluation indicators to comprehensively evaluate and screen the molecules generated by the model, and guide the laboratory's next step of molecular optimization and chemical synthesis based on the preferred molecules.

2. The method according to claim 1, characterized in that: The step S1 comprises: The molecular generation model includes a data input module, a generator, an optimizer and an output module; The input module is used to input the chemical information of the compound into the generator, the generator generates a representation of the molecule based on the input data, the optimizer obtains the model that minimizes the loss function through iterative updating, and the output module outputs the result of the model output corresponding to the minimum loss function.

3. The method according to claim 2, characterized in that The step S2 comprises: The SMILES strings of druggable molecules were obtained from an open source database as a training set; the training set was preprocessed to remove duplicate molecules and invalid molecules, separate metal atoms bound to the molecules, and convert all molecules into a standardized SMILES form.

4. The method according to claim 3, characterized in that The step S3 comprises: Screening is performed based on the molecular composition and chemical expertise in the training set to determine a tag library; the tag library includes single characters, atomic types composed of two characters, and special environments; Adopt one-hot encoding and use the tag library to encode each SMILES string; Train neural networks for sequence modeling by maximizing the next token x in the target sequence t The goal is to maximize the probability of the correct label being assigned: During the training process, the model gradually grasps the rules and patterns of molecular generation by learning the sequence information of the compound structure; the model performance is evaluated using the validation set to ensure that the model can generate reasonable and diverse molecular structures, and the generation ability and stability of the model are optimized by adjusting the hyperparameters.

5. The method according to claim 1, characterized in that The step S4 comprises: Obtain molecules with target property test results and SMILES strings from open source databases as training data sets for property prediction models; According to the optimization requirements of the target attributes, a target attribute prediction model is constructed, and a supervised learning regression model is used to predict the specific attributes of the molecule; The model generates corresponding property prediction values ​​based on the input molecular chemical structure information; the molecular chemical structure information can be selectively input into SMILES strings, molecular fingerprints, molecular graph representations or 3D spatial position information according to the characteristics of the target attribute; these molecular information will be converted into processable feature representations in the model, and processed through a multi-layer neural network to ultimately output the predicted value of the target attribute.

6. The method according to claim 5, characterized in that The target property scoring according to the predicted target property value in step S4 includes: A target property scoring model is established, and the target properties of the molecular compounds obtained by the target property prediction model are scored according to the optimization target. The scoring method can be any of the following: (1) When the target attribute is within a certain range, the score is high, and when it is outside the range, the score is 0; (2) The higher the target attribute prediction result, the higher the score; (3) The closer the target attribute is to the target value, the higher the score; (4) Weight each attribute of multiple target attributes to generate a comprehensive score.

7. The method according to claim 1, characterized in that The step S5 further comprises: According to the structure of the molecule to be optimized, a structural similarity algorithm is used to screen out compounds with similar structures from the database as the data set for transfer learning; The screened compounds are evaluated using the target attribute scoring function, and molecules with low scores are eliminated based on the scoring results of the target attributes, while molecules with high scores are retained to ensure the quality of the transfer learning dataset; The screened high-quality data set is input into the pre-trained molecular generation model, and the model parameters are adjusted through transfer learning so that the model can better generate molecules that meet the target property requirements.

8. The method according to claim 1, characterized in that The step S6 comprises: In the reinforcement learning framework, a policy gradient algorithm is constructed to optimize the molecule generation model; the policy gradient algorithm updates the model parameters by maximizing the reward function of the generated molecules, and the reward function is based on the feedback of the target attribute scoring function; In the reinforcement learning process, the model gets reward values ​​based on the generated molecules, and the model parameters are continuously updated through the policy gradient algorithm to maximize the reward value. After multiple iterations and training, reinforcement learning finally obtains an optimized molecular generation model that can better generate molecules that meet the target property requirements.

9. The method according to claim 1, characterized in that: The step S7 comprises: Sampling is performed to obtain a sufficient number of generated molecular libraries based on the target skeleton for the next step of screening, and the quality of the generated molecular library is evaluated to ensure the diversity of the molecules in the generated molecular library and the conformity of the target attributes; Multi-dimensional indicators are used to perform virtual screening of molecules in the generated molecular library, thresholds for specific evaluation indicators are set, molecules outside the threshold range are eliminated, and the remaining molecules are then comprehensively scored. The molecules with the highest comprehensive scores are selected to guide the molecular optimization and synthesis tasks of the medicinal chemistry laboratory.

10. The method according to claim 9, characterized in that The multidimensional indicators include molecular synthesis feasibility, molecular docking, biological activity or drugability.

Images