Small molecule drug optimization method based on evolutionary computation and multi-granularity surrogate model approximation

Abstract

The present invention discloses a small drug molecule optimization method based on evolutionary computation and multi-granularity proxy model approximation, comprising: 1. modeling the multi-attribute optimization problem of small drug molecules as an expensive constrained multi-objective optimization problem; 2. uniformly generating an initial molecular population in a discrete chemical space and performing an expensive optimization target evaluation and constraint evaluation on the initial molecular population; 3. encoding the evaluated small drug molecules into a continuous latent space to obtain a continuous vector representation of the molecules; 4. using the continuous vector representation of the molecules, the corresponding target values ​​and the constraint values ​​as training data to train a multi-granularity proxy model; 5. executing a proxy model-assisted evolutionary algorithm to select high-quality small drug molecules for decoding into a discrete chemical space, and performing an expensive optimization target evaluation and constraint evaluation on the initial molecular population; 6. repeating steps 3-5 to output the optimal drug molecule. The present invention aims to efficiently optimize multiple attributes of small drug molecules using a relatively low evaluation cost, thereby providing technical support for drug research and development.

Description

Technical Field

[0001] The present invention relates to multiple interdisciplinary fields such as computer science, artificial intelligence, and bioinformatics. Specifically, the present invention proposes a drug small molecule optimization method based on evolutionary computation and multi-granularity agent model approximation. Background Art

[0002] The basic process of new drug research and development includes four stages: drug discovery, preclinical research, clinical research, and registration and approval. It usually takes more than ten years and hundreds of millions of dollars for a new drug to go from discovery to approval for market use. In the first stage of new drug research and development, namely the drug discovery stage, it is necessary to first discover biological target macromolecules that have pharmacological functions and can be acted upon by drugs; secondly, high-throughput screening and other technologies are used to obtain lead compounds with certain biological activity; then, the lead compounds are optimized to obtain lead compounds with high biological activity, pharmacodynamic properties, and safety; finally, the lead compounds are evaluated and optimized to obtain preclinical candidate compounds with high drugability. It is worth mentioning that many new drug research and development projects often fail due to the poor properties of preclinical candidate compounds. Therefore, optimizing two types of small molecule drugs, lead compounds and lead compounds, has become a key link in new drug research and development.

[0003] In order to optimize the hit compounds and lead compounds, it is necessary to explore all possible chemical small molecules (10 23 ~10 60The vast space of small molecule properties (comprising multiple entities) is used to discover small drug molecules that meet the constraints and exhibit superior properties. Traditional small molecule optimization methods primarily rely on experimental screening, which is time-consuming and labor-intensive. In recent years, with the rapid development of artificial intelligence optimization technology and the massive accumulation of biological big data, small molecule optimization has become more cost-effective and efficient. Although existing research has made significant progress in optimizing single-attribute small molecule properties, actual drug development often requires small molecules to possess multiple excellent properties simultaneously. Optimizing only a single attribute is insufficient to meet the actual needs of drug development. The difficulty of optimizing multiple attributes of small molecule drugs lies in the interdependence of different attributes. Optimizing a single attribute of a small molecule may cause changes in, or even deterioration of, other attributes. Furthermore, small molecule optimization often involves complex constraints, such as structural similarity constraints, non-toxicity constraints, and Lipinski's five rules, making the problem of optimizing small molecule properties even more difficult to solve. Evolutionary algorithms, a class of meta-heuristic intelligent optimization algorithms inspired by various natural phenomena and the behavior of biological populations, possess strong adaptability, stability, concurrency, and self-organization. They are not easily trapped in local optima and can be combined with various constraint-handling strategies, making them suitable for solving constrained multi-attribute optimization problems for small drug molecules. Currently, several studies have used evolutionary algorithms to solve these constrained multi-attribute optimization problems for small drug molecules, achieving promising results. However, practical small drug molecule optimization often involves expensive attributes, and evaluating these attributes can be time-consuming and resource-intensive. Clearly, these small drug molecule optimization problems pose significant challenges for evolutionary algorithms, which require multiple iterations. Summary of the Invention

[0004] In order to solve the above-mentioned problems, the present invention proposes a drug small molecule optimization method based on evolutionary computation and multi-granularity proxy model approximation, in order to dynamically adjust the approximate granularity of the proxy model according to the constraint violation, thereby achieving higher proxy model approximation accuracy using fewer computing resources, thereby reducing the cost of evaluating drug small molecules and further optimizing drug small molecules.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The present invention is characterized in that a small molecule drug optimization method based on evolutionary computation and multi-granularity agent model approximation is performed according to the following steps:

[0007] Step 1: Obtain the kth drug molecule to be optimized And use RDKit tool to identify the kth drug small molecule Conduct biological attribute evaluation and obtain Molecular drug properties , synthetic accessibility and molecular toxicity ; Thus, using formula (1) to construct The multi-attribute optimization goal And its constraints:

[0008] (1)

[0009] In formula (1), represents the search space, and Respectively represent The violation of the molecular weight constraint and the violation of the number of hydrogen bond donors constraint, and Represents small drug molecules The molecular weight and number of hydrogen bond donors, and represent the maximum values ​​of molecular weight and number of hydrogen bond donors, respectively;

[0010] Molecular drug properties , synthetic accessibility and molecular toxicity Recorded as Any j-th optimization goal ;j=1,2,3;

[0011] Will and Recorded as Any i-th constraint of ;i=1,2;

[0012] Step 2: Define and initialize the current iteration number to Iter=0; and define the Iter generation drug small molecule population ,in, represent The kth drug molecule in the population, N represents the size of the population;

[0013] Generate SMILES strings uniformly in the discrete chemical space for initialization Each drug small molecule in , thus calculated according to formula (1) The violation degree of the multi-attribute optimization objectives and constraints of each drug molecule in ;

[0014] Step 3: Train the variational autoencoder to Encoding from discrete chemical space to continuous latent space yields The optimal latent vector ;

[0015] Step 4: Build two proxy models and use The optimal potential vector of each drug small molecule and its corresponding multi-attribute optimization target value and constraint violation degree are used as training data to train each proxy model separately to obtain a trained proxy model;

[0016] Step 5: Use the trained zth agent model to assist the evolutionary algorithm to iteratively optimize the drug molecule properties, thereby obtaining the final drug small molecule population as the final optimization result.

[0017] The present invention also features a small molecule drug optimization method based on evolutionary computation and multi-granularity agent model approximation in that step three is performed as follows:

[0018] Step 3.1, build a variational autoencoder, including: encoder and decoder, and Process and obtain The reconstructed word vector ;

[0019] Step 3.2: Use formula (3) to calculate the total loss of the variational autoencoder in the Iter generation :

[0020] = (3)

[0021] In formula (3), represents the kl divergence loss of the Iter generation, Represents the reconstruction loss of the Iter generation, 、 Represents the weights corresponding to the two losses; and:

[0022] (4)

[0023] (5)

[0024] In formula (4) and formula (5), represents the logarithmic function;

[0025] Step 3.3: Use back propagation and gradient descent to train the variational autoencoder and calculate To update the network parameters until Until convergence, the variational autoencoder with the optimal parameters of the Iter generation is obtained;

[0026] Step 3.4: The kth drug molecule Input the optimal parameters of the Iter generation variational self-encoder and get The optimal latent vector ,in, express The optimal potential feature of the rth dimension in , thus obtaining The optimal latent vector .

[0027] Furthermore, the step 3.1 includes:

[0028] Step 3.1.1. Randomly initialize the neural network weights in the encoder and decoder.

[0029] Step 3.1.2: For the kth drug molecule Perform word embedding to obtain the kth word vector of the Iter generation ,in, express The nth word vector in , Represents the total number of word vectors;

[0030] Step 3.1.3, Input into the encoder and get The mean and variance , thus obtaining the kth drug molecule of the Iter generation according to formula (2) The latent vector ,in, express The r-th dimension feature in ;

[0031] (2)

[0032] In formula (2), e represents the exponential function, esp is the random noise sampled from the standard normal distribution;

[0033] Step 3.1.4, the decoder is autoregressive Process and obtain the kth drug molecule The reconstructed word vector ={ },in, express The mth word vector in .

[0034] Furthermore, the step 4 includes:

[0035] Step 4.1, use formula (6) to calculate the z-th dimension Correlation matrix The kth row in Column Elements , thus obtaining the kth drug small molecule in the zth agent model and The correlation vector :

[0036] (6)

[0037] In formula (6), is the r-th dimension parameter of the z-th proxy model of the Iter-th generation; Indicates the Small molecule drugs The optimal latent vector The optimal latent feature of the rth dimension in; z=1,2;

[0038] Step 4.2: Construct training data for the zth agent model , including: training data set of the first proxy model and the coarse-grained training dataset of the second proxy model ;

[0039] Step 4.3: Use formula (8) to get the mean of the zth surrogate model of the Iter generation :

[0040] (8)

[0041] In formula (8), represents an N-dimensional all-1 vector, and T represents transpose;

[0042] Step 4.4: Use formula (9) to solve the parameter set of the zth agent model of the Iter generation { }, and get the trained z-th agent model:

[0043] (9)

[0044] In formula (9), Indicates the training data of the z-th surrogate model The true objective optimization value or true constraint violation degree of ; and ∈{ , , }.

[0045] Furthermore, the step 4.2 includes:

[0046] Step 4.2.1. Construct a training dataset for training the first agent model ,in, Represents the kth drug small molecule The sum of the three optimization objectives;

[0047] Step 4.2.2, if The molecular weight constraint violation degree and hydrogen bond donor number constraint violation degree are both greater than 0, that is, any i-th constraint is not satisfied, then the k-th drug small molecule of the Iter generation is obtained using formula (7): Constraint violation of aggregate value , thereby obtaining the coarse-grained training data of the second proxy model :

[0048] (7)

[0049] In formula (7), respectively The maximum and minimum values ​​on the i-th constraint;

[0050] like Among the molecular weight constraint violation and hydrogen bond donor number constraint violation, only one constraint violation is greater than 0, that is, there is only one unsatisfied constraint, which is denoted as , then construct the fine-grained training data of the second proxy model ,in, Represents the kth drug small molecule of the Iter generation The constraint violation degree of the unsatisfied constraints.

[0051] Furthermore, the step five includes:

[0052] Step 5.1, The corresponding optimal latent vector set As the parent molecule;

[0053] Step 5.2: Based on the optimization target value and constraint violation degree of the drug small molecule, perform crossover mutation operation on the optimal potential vector in the parent molecule to generate the daughter molecule ;

[0054] Step 5.3, use The corresponding trained z-th agent model pair The multi-attribute optimization objectives and constraint violation degrees of each drug small molecule are predicted, and then the environmental selection method of the non-dominated sorting genetic algorithm NSGA-II is used to select the parent molecule according to the prediction results. and daughter molecules After screening N high-quality drug small molecules as new parent molecules, return to step 5.2 and execute sequentially Second, get the The parent molecule in the second cycle is used as the Iter+1 generation drug small molecule population ;

[0055] Step 5.4, Input the trained decoder for processing and obtain The optimal latent vector of each drug small molecule in is decoded into a SMILES string;

[0056] Step 5.5: Use RDKit to The decoded SMILES string is evaluated to obtain its true multi-attribute optimization objective and constraint violation degree;

[0057] Step 5.6: After assigning Iter+1 to Iter, return to step 3 and execute sequentially until the maximum number of iterations is reached, thereby outputting the last generation of drug small molecule population as the final optimization result.

[0058] The electronic device of the present invention includes a memory and a processor, and is characterized in that the memory is used to store a program that supports the processor to execute the drug small molecule optimization method, and the processor is configured to execute the program stored in the memory.

[0059] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the drug small molecule optimization method when the computer program is run by a processor.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The multi-granularity surrogate model approximation method proposed in this invention dynamically adjusts the granularity of the surrogate model approximation based on the position of the evolving molecular population in the adaptive landscape. If the evolving molecular population violates all constraints, a surrogate model is trained for each target and a coarse-grained model is trained for all constraints. If the evolving molecular population only violates some constraints, a surrogate model is trained for each target and a fine-grained model is trained for each unsatisfied constraint. If the evolving molecular population does not violate any constraints, only one surrogate model is trained for each target. Based on this strategy, the proposed method can dynamically adjust the training model strategy based on the constraint violations of the drug small molecule, thereby improving the approximation accuracy of the surrogate model.

[0062] 2. The present invention designs a special model updating strategy, which dynamically adjusts the comparison criteria of the advantages and disadvantages of drug molecules according to the distance between the molecular population and the feasible region. When the molecular population is far away from the feasible region, the constraint dominance principle is used to update the training data and the molecular population, thereby improving the degree of constraint satisfaction of the drug molecules; when the molecular population is close to the feasible region, the multi-objective method is used to update the training data and the molecular population, thereby balancing the optimization of drug molecule properties and constraint satisfaction; when the molecular population completely enters the feasible region, the training data and the molecular population are updated only based on the multi-attribute target value, thereby maximizing the optimization of molecular properties. By dynamically adjusting the update method of the training data and the molecular population, the proposed strategy can reduce the impact caused by the approximate error of the proxy model and avoid the drug molecules from falling into the local optimum. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flow chart of the method of the present invention;

[0064] Figure 2 Schematic diagram of the training of the multi-granularity proxy model in the present invention. DETAILED DESCRIPTION

[0065] In this embodiment, a small molecule drug optimization method based on evolutionary computation and multi-granularity agent model approximation is described. Figure 1 As shown, it is applied to optimize the structure of small drug molecules and is carried out in the following steps:

[0066] Step 1: Obtain the kth drug molecule to be optimized And use RDKit tool to identify the kth drug small molecule Conduct biological attribute evaluation and obtain Molecular drug properties , synthetic accessibility and molecular toxicity ; Thus, using formula (1) to construct The multi-attribute optimization goal And its constraints:

[0067] (1)

[0068] In formula (1), represents the search space, and Respectively represent The violation of the molecular weight constraint and the violation of the number of hydrogen bond donors constraint, and Represents small drug molecules The molecular weight and number of hydrogen bond donors, and represent the maximum molecular weight and the number of hydrogen bond donors, respectively.

[0069] Molecular drug properties , synthetic accessibility and molecular toxicity Recorded as Any j-th optimization goal ;j=1,2,3;

[0070] Will and Recorded as Any i-th constraint of ;i=1,2.

[0071] Step 2: Generate an initial population of drug small molecules uniformly in the discrete chemical space and evaluate the violation of the multi-attribute optimization objective and its constraints;

[0072] Step 2.1. Define and initialize the current iteration number as Iter=0; and define the Iter generation drug small molecule population ,in, represent The kth drug molecule in the population, N represents the size of the population;

[0073] Generate SMILES strings uniformly in the discrete chemical space for initialization Each drug small molecule in

[0074] Step 2.2: According to formula (1), we can get The violation degree of the multi-attribute optimization objective and its constraints for each drug molecule in .

[0075] Step 3: Train the variational autoencoder. Encoding from discrete chemical space to continuous latent space yields Continuous vector representation of ;

[0076] Step 3.1, build a variational autoencoder, including: encoder and decoder, and Process and obtain The reconstructed word vector ;

[0077] Step 3.1.1. Randomly initialize the neural network weights in the encoder and decoder.

[0078] Step 3.1.2: For the kth drug molecule Perform word embedding to obtain the kth word vector of the Iter generation ,in, express The nth word vector in , Represents the total number of word vectors;

[0079] Step 3.1.3, Input into the encoder and get The mean and variance , thus obtaining the kth drug molecule of the Iter generation according to formula (2) The latent vector ,in, express The r-th dimension feature in ;

[0080] (2)

[0081] In formula (2), e represents the exponential function, and esp is the random noise sampled from the standard normal distribution.

[0082] Step 3.1.4, the decoder is autoregressive Process and obtain the kth drug molecule The reconstructed word vector ={ },in, express The mth word vector in .

[0083] Step 3.2: Use formula (3) to calculate the total loss of the variational autoencoder in the Iter generation :

[0084] = (3)

[0085] In formula (3), represents the kl divergence loss of the Iter generation, represents the reconstruction loss of the Iter generation, 、 Represents the weights corresponding to the two losses; and:

[0086] (4)

[0087] (5)

[0088] In formula (4) and formula (5), Represents a logarithmic function.

[0089] Step 3.3: Use back propagation and gradient descent to train the variational autoencoder and calculate To update the network parameters until Until convergence, the variational autoencoder with the optimal parameters of the Iter generation is obtained;

[0090] Step 3.4: The kth drug molecule Input the optimal parameters of the Iter generation variational self-encoder and get The optimal latent vector ,in, express The optimal potential feature of the rth dimension in . Thus we get The optimal latent vector .

[0091] Step 4: Figure 2 As shown, build two proxy models and use The optimal potential vector of each drug small molecule and its corresponding multi-attribute optimization target value and constraint violation degree are used as training data. Each proxy model is trained separately to obtain the hyperparameters of each model:

[0092] Step 4.1, use formula (6) to calculate the z-th dimension Correlation matrix The kth row in Column Elements :

[0093] (6)

[0094] In formula (6), is the r-th dimension parameter of the z-th proxy model of the Iter-th generation; Indicates the Small molecule drugs The optimal latent vector The optimal latent feature of the rth dimension in; z=1,2;

[0095] Using formula (6), we can get the relationship between the kth drug small molecule and the zth agent model. The correlation vector .

[0096] Step 4.2: Construct training data for the zth agent model , including: training data set of the first proxy model and the coarse-grained training dataset of the second proxy model ;

[0097] Step 4.2.1. Construct a training dataset for training the first agent model ,in, Represents the kth drug small molecule The weighted summation of the three optimization objectives is performed. Since the dimensions of the three optimization objectives are different, the target value of each optimization objective is first normalized to the interval [0, 1] and then the summation operation is performed.

[0098] Step 4.2.2, if The molecular weight constraint violation degree and hydrogen bond donor number constraint violation degree are both greater than 0, that is, any i-th constraint is not satisfied, then the k-th drug small molecule of the Iter generation is obtained using formula (7): Constraint violation of aggregate value , thereby obtaining the coarse-grained training data of the second proxy model :

[0099] (7)

[0100] In formula (7), respectively The maximum and minimum values ​​on the i-th constraint.

[0101] like Among the molecular weight constraint violation and hydrogen bond donor number constraint violation, only one constraint violation is greater than 0, that is, there is only one unsatisfied constraint, which is recorded as , then construct the fine-grained training data of the second proxy model ,in, Represents the kth drug small molecule of the Iter generation The constraint violation degree of the unsatisfied constraints.

[0102] Step 4.4: Use formula (8) to get the mean of the zth proxy model of the Iter generation :

[0103] (8)

[0104] In formula (8), represents an N-dimensional all-one vector, and T represents transpose.

[0105] Step 4.5: Use formula (9) to solve the parameter set of the zth agent model of the Iter generation { }, and get the trained z-th agent model:

[0106] (9)

[0107] In formula (9), Indicates the training data of the z-th agent model The true objective optimization value or true constraint violation degree of ; and ∈{ , , }.

[0108] Step 5: Use the surrogate model trained in step 4 to assist the evolutionary algorithm in iteratively optimizing the molecular properties of the drug;

[0109] Step 5.1, The corresponding optimal latent vector set As the parent molecule;

[0110] Step 5.2: Based on the optimization target value and constraint violation degree of the drug small molecule, perform crossover mutation operation on the optimal potential vector in the parent molecule to generate the daughter molecule ;

[0111] Step 5.3, use The corresponding trained z-th agent model pair The multi-attribute optimization objectives and constraint violation degrees of each drug small molecule are predicted, and then the environmental selection method of the non-dominated sorting genetic algorithm NSGA-II is used to select the parent molecule according to the prediction results. and daughter molecules After screening N high-quality drug small molecules as new parent molecules, return to step 5.2 and execute sequentially Second, get the The parent molecule in the second cycle is used as the Iter+1 generation drug small molecule population .

[0112] In order to balance the optimization of molecular population properties and the satisfaction of constraints, when screening high-quality drug small molecules, the proposed method dynamically adjusts the method of comparing the advantages and disadvantages of molecules according to the position of the molecular population in the search space. Specifically, when the molecular population is far from the feasible region, the proposed method uses the constraint dominance principle to compare the advantages and disadvantages of different molecules, as shown in Equation (10):

[0113] (10)

[0114] In formula (10), and For two small drug molecules, Representative drug small molecules The degree of constraint violation, Representative molecules The i-th constraint value of Representative molecules The jth target value of The degree of violation Equal to the numerator The degree of violation , if we consider that the molecule Better than molecules , then the molecule All target values ​​of and is better than at least one target value. The target value of . If the numerator The degree of violation Smaller than the molecule The degree of violation , then the molecule can be directly considered Better than molecules .

[0115] When the molecular population is close to the feasible region boundary, the proposed method first calculates the constraint violation values ​​of all molecules, and then treats the constraint violation values ​​as optimization targets together with the original three optimization targets, thereby comparing the advantages and disadvantages of different molecules based on the four optimization target values. When the molecular population completely enters the feasible region, the proposed method only compares the advantages and disadvantages of molecules based on the three optimization target values.

[0116] Step 5.4, Input the trained decoder for processing and obtain The optimal latent vector of each drug small molecule in is decoded into a SMILES string;

[0117] Step 5.5: Use RDKit to The decoded SMILES string is evaluated to obtain its true multi-attribute optimization objective and constraint violation degree;

[0118] Step 6: After assigning Iter+1 to Iter, return to step 3 and execute sequentially until the maximum number of iterations is reached, thereby outputting the last generation of drug small molecule population as the final optimization result.

[0119] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0120] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.

Claims

1. A small molecule drug optimization method based on evolutionary computation and multi-granularity agent model approximation, characterized in that: The steps are as follows: Step 1: Obtain the kth drug molecule to be optimized And use RDKit tool to identify the kth drug small molecule Conduct biological attribute evaluation and obtain Molecular drug properties , synthetic accessibility and molecular toxicity ; Thus, using formula (1) to construct The multi-attribute optimization goal And its constraints: (1) In formula (1), represents the search space, and Respectively represent The violation of the molecular weight constraint and the violation of the number of hydrogen bond donors constraint, and Represents small drug molecules The molecular weight and number of hydrogen bond donors, and represent the maximum values ​​of molecular weight and number of hydrogen bond donors, respectively; Molecular drug properties , synthetic accessibility and molecular toxicity Recorded as Any j-th optimization goal ;j=1,2,3; Will and Recorded as Any i-th constraint of ;i=1,2; Step 2: Define and initialize the current iteration number to Iter=0; and define the Iter generation drug small molecule population ,in, represent The kth drug molecule in the population, N represents the size of the population; Generate SMILES strings uniformly in the discrete chemical space for initialization Each drug small molecule in , thus calculated according to formula (1) The violation degree of the multi-attribute optimization objectives and constraints of each drug molecule in ; Step 3: Train the variational autoencoder to Encoding from discrete chemical space to continuous latent space yields The optimal latent vector ; Step 4: Build two proxy models and use The optimal potential vector of each drug small molecule and its corresponding multi-attribute optimization target value and constraint violation degree are used as training data to train each proxy model separately to obtain a trained proxy model; Step 4.1, use formula (6) to calculate the z-th dimension Correlation matrix The kth row in Column Elements , thus obtaining the kth drug small molecule in the zth agent model and The correlation vector : (6) In formula (6), is the r-th dimension parameter of the z-th proxy model of the Iter-th generation; Indicates the Small molecule drugs The optimal latent vector The optimal latent feature of the rth dimension in; z=1,2; Step 4.2: Construct training data for the zth agent model , including: training data set of the first proxy model and the coarse-grained training dataset of the second proxy model ; Step 4.2.

1. Construct a training dataset for training the first agent model ,in, Represents the kth drug small molecule The sum of the three optimization objectives; Step 4.2.2, if The molecular weight constraint violation degree and hydrogen bond donor number constraint violation degree are both greater than 0, that is, any i-th constraint is not satisfied, then the k-th drug small molecule of the Iter generation is obtained using formula (7): Constraint violation of aggregate value , thereby obtaining the coarse-grained training data of the second proxy model : (7) In formula (7), respectively The maximum and minimum values ​​on the i-th constraint; like Among the molecular weight constraint violation and hydrogen bond donor number constraint violation, only one constraint violation is greater than 0, that is, there is only one unsatisfied constraint, which is denoted as , then construct the fine-grained training data of the second proxy model ,in, Represents the kth drug small molecule of the Iter generation The constraint violation degree of the unsatisfied constraints; Step 4.3: Use formula (8) to get the mean of the zth surrogate model of the Iter generation : (8) In formula (8), represents an N-dimensional all-1 vector, and T represents transpose; Step 4.4: Use formula (9) to solve the parameter set of the zth agent model of the Iter generation { }, and get the trained z-th agent model: (9) In formula (9), Indicates the training data of the z-th agent model The true objective optimization value or true constraint violation degree of ; and ∈{ , , }; Step 5: Use the trained zth agent model to assist the evolutionary algorithm to iteratively optimize the drug molecule properties, thereby obtaining the final drug small molecule population as the final optimization result.

2. The method for optimizing small drug molecules based on evolutionary computation and multi-granularity agent model approximation according to claim 1, characterized in that: Step three is to proceed as follows: Step 3.1, build a variational autoencoder, including: encoder and decoder, and Process and obtain The reconstructed word vector ; Step 3.2: Use formula (3) to calculate the total loss of the variational autoencoder in the Iter generation : = (3) In formula (3), represents the kl divergence loss of the Iter generation, Represents the reconstruction loss of the Iter generation, 、 Represents the weights corresponding to the two losses; and: (4) (5) In formula (4) and formula (5), represents the logarithmic function; express The mean of express The variance of represents the kth drug small molecule The k-th word vector of the Iter-th generation after word embedding; Step 3.3: Use back propagation and gradient descent to train the variational autoencoder and calculate To update the network parameters until Until convergence, the variational autoencoder with the optimal parameters of the Iter generation is obtained; Step 3.4: Add the kth drug molecule Input the optimal parameters of the Iter generation variational self-encoder and get The optimal latent vector ,in, express The optimal potential feature of the rth dimension in , thus obtaining The optimal latent vector .

3. The method for optimizing small drug molecules based on evolutionary computation and multi-granularity agent model approximation according to claim 2, characterized in that: The step 3.1 includes: Step 3.1.

1. Randomly initialize the neural network weights in the encoder and decoder. Step 3.1.2: For the kth drug molecule Perform word embedding to obtain the kth word vector of the Iter generation ,in, express The nth word vector in , Represents the total number of word vectors; Step 3.1.3, Input into the encoder and get The mean and variance , thus obtaining the kth drug molecule of the Iter generation according to formula (2) The latent vector ,in, express The r-th dimension feature in ; (2) In formula (2), e represents the exponential function, esp is the random noise sampled from the standard normal distribution; Step 3.1.4, the decoder is autoregressive Process and obtain the kth drug molecule The reconstructed word vector ={ },in, express The mth word vector in .

4. The method for optimizing small drug molecules based on evolutionary computation and multi-granularity agent model approximation according to claim 3, characterized in that: The step five includes: Step 5.1, The corresponding optimal latent vector set As the parent molecule; Step 5.2: Based on the optimization target value and constraint violation degree of the drug small molecule, perform crossover mutation operation on the optimal potential vector in the parent molecule to generate the daughter molecule ; Step 5.3, use The corresponding trained z-th agent model pair The multi-attribute optimization objectives and constraint violation degrees of each drug small molecule are predicted, and then the environmental selection method of the non-dominated sorting genetic algorithm NSGA-II is used to select the parent molecule according to the prediction results. and daughter molecules After screening N high-quality drug small molecules as new parent molecules, return to step 5.2 and execute sequentially Second, get the The parent molecule in the second cycle is used as the Iter+1 generation drug small molecule population ; Step 5.4, Input the trained decoder for processing and obtain The optimal latent vector of each drug small molecule in is decoded into a SMILES string; Step 5.5: Use RDKit to The decoded SMILES string is evaluated to obtain its true multi-attribute optimization objective and constraint violation degree; Step 5.6: After assigning Iter+1 to Iter, return to step 3 and execute sequentially until the maximum number of iterations is reached, thereby outputting the last generation of drug small molecule population as the final optimization result.

5. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the drug small molecule optimization method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the small molecule drug optimization method according to any one of claims 1 to 4 are executed.

Images