Compound design method, device, equipment and computer readable storage medium

Abstract

The present application discloses a compound design method, device, equipment and computer-readable storage medium, the method comprising obtaining a seed vector, which is a characteristic vector representation of a seed compound; performing a crossover operation and / or a mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector; and processing the derivative vector to obtain a derivative compound. Through the above-mentioned method, the present application can improve the diversity of the designed compound.

Description

Technical Field

[0001] The present application relates to the field of computational chemistry technology, and in particular to a compound design method, device, equipment and computer-readable storage medium. Background Art

[0002] In traditional drug research, scientists screen compound libraries, test compound targets one by one, and finally screen out lead compounds. Due to the high cost and failure rate of the process, computational chemists try to use computational modeling to predict the activity of compounds, simulate the binding of drugs in protein cavities with computers, and recommend and test a batch of molecules that may be active. However, this method is severely limited by the quality of virtual screening libraries. The number of molecules in existing compound libraries is generally in the hundreds of thousands, and the skeletons of the compounds have been extensively studied and screened by predecessors, making it difficult to find candidate compounds with new skeletons. Screening libraries with too few numbers and poor structural novelty can no longer meet the growing needs of research and development. Summary of the invention

[0003] The main technical problem solved by the present application is to provide a compound design method, device, equipment and computer-readable storage medium, which can improve the diversity of designed compounds.

[0004] In order to solve the above technical problems, a technical solution adopted in the present application is: to provide a compound design method, the method comprising obtaining a seed vector, which is a characteristic vector representation of a seed compound; performing crossover operation and / or mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector; and processing the derivative vector to obtain a derivative compound.

[0005] After the derivative vectors are processed to obtain the derivative compounds, the method further includes measuring the fitness of the derivative compounds based on the fitness function; and selecting candidate compounds from the derivative compounds according to the fitness.

[0006] Among them, candidate compounds are selected from the derivative compounds according to the size of fitness, including:

[0007] Step S1: selecting a target compound whose fitness meets a preset condition from the derivative compounds according to the size of the fitness;

[0008] Step S2: taking the derivative vector corresponding to the target compound as a seed vector, and continuing to perform a crossover operation and / or a mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector and measure the fitness of the derivative compound based on a fitness function;

[0009] Iterate steps S1 and S2 until the iteration termination condition is met, and then terminate the iteration cycle operation;

[0010] All derived compounds obtained were sorted in descending order according to fitness;

[0011] A predetermined proportion or a predetermined number of derivative compounds with better fitness are selected as candidate compounds.

[0012] The crossover operation of the seed vector based on the genetic algorithm includes: selecting two seed vectors from the seed vector set, selecting the exchange position of one of the seed vectors, and exchanging the value of the exchange position of the seed vector with the value of the corresponding position of another sub-vector.

[0013] The mutation operation on the seed vector based on the genetic algorithm includes: selecting a seed vector from a seed vector set, selecting a mutation position from the selected seed vector, and replacing the value at the mutation position with a new value.

[0014] Among them, processing the derivative vector to obtain the derivative compound includes: inputting the derivative vector into a molecular structure decoding model, decoding the derivative vector to obtain the derivative molecular structure, the molecular structure decoding model is a neural network model, and the derivative compound is obtained according to the derivative molecular structure.

[0015] Among them, before the derivative vector is input into the molecular structure decoding model and the derivative vector is decoded to obtain the molecular structure of the derivative compound through decoding, the method also includes: obtaining a sample matrix, which is a matrix representation of the sample compound; inputting the sample matrix into the encoding layer of the autoencoder and encoding to obtain a sample vector, which is a feature vector representation of the sample compound; inputting the sample vector into the decoding layer of the autoencoder and decoding to obtain a prediction matrix; calculating the loss between the prediction matrix and the sample matrix; iteratively updating the parameters of the autoencoder based on the loss until the loss is stable, and using the decoding layer and output layer of the updated trained autoencoder as the molecular structure decoding model, and the output layer is used to convert the compound represented by the matrix into a molecular structure representation.

[0016] Wherein, obtaining the seed vector includes: obtaining a SMILES string of a seed compound; performing one-hot encoding on the SMILES string of the seed compound to obtain a seed matrix, where the seed matrix is ​​a matrix representation of the seed compound; and encoding the seed matrix to obtain a seed vector.

[0017] The step of encoding the seed matrix to obtain the seed vector includes: inputting the seed matrix into a molecular structure encoding model, and performing encoding processing on the seed matrix to obtain the seed vector.

[0018] Among them, before the seed matrix is ​​input into the molecular structure encoding model and the seed matrix is ​​encoded to obtain the seed vector, the method also includes: obtaining a sample matrix, which is a matrix representation of the sample compound; inputting the sample matrix into the encoding layer of the autoencoder, encoding to obtain a sample vector, which is a feature vector representation of the sample compound; inputting the sample vector into the decoding layer of the autoencoder, decoding to obtain a prediction matrix; calculating the loss between the prediction matrix and the sample matrix; iteratively updating the parameters of the autoencoder based on the loss until the loss is stable, and using the encoding layer of the updated trained autoencoder as the molecular structure encoding model.

[0019] To solve the above technical problems, another technical solution adopted in the present application is: to provide a compound design device, which includes an acquisition module, an operation module and a decoding module, the acquisition module is used to obtain a seed vector, and the seed vector is a characteristic vector representation of a seed compound; the operation module is used to perform crossover operation and / or mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector; the decoding module is used to process the derivative vector to obtain a derivative compound.

[0020] To solve the above technical problems, another technical solution adopted in the present application is: to provide a compound design device, including a processor and a memory, wherein the memory stores instructions, and the processor is used to execute the instructions to implement any of the above compound design methods.

[0021] To solve the above technical problems, another technical solution adopted in the present application is: providing a computer-readable storage medium, which is used to store instructions / program data, and the instructions / program data can be executed to implement any of the above-mentioned compound design methods.

[0022] The beneficial effects of the present application are as follows: Different from the prior art, the compound design method provided by the present application is based on a genetic algorithm to develop and design compounds, which increases the explorable compound space, can obtain diversified compounds, and increases the selection space. Furthermore, during the operation, the complex chemical space is reduced to a one-dimensional vector, which enables the design algorithm to search the chemical space conveniently and efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a schematic diagram of a compound design method in an embodiment of the present application;

[0024] Figure 2 is a schematic diagram of a training process of a molecular structure model in an embodiment of the present application;

[0025] Figure 3 is a schematic flow chart of another compound design method in an embodiment of the present application;

[0026] Figure 4 is a schematic diagram of the structure of a compound design device in an embodiment of the present application;

[0027] Figure 5 is a schematic diagram of the structure of a compound design device in an embodiment of the present application;

[0028] Figure 6 It is a schematic diagram of the structure of a computer-readable storage medium in an embodiment of the present application. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and effect of the present application clearer and more specific, the present application is further described in detail below with reference to the accompanying drawings and examples.

[0030] In order to meet the needs of drug development for compound screening libraries, the inventors of this application have found that the molecular generation model based on deep learning can use a large-scale compound database to self-learn the writing rules of compounds, and represent the compounds as a dense continuous value vector, and then learn the structural characteristics of the compounds, generate compounds with new skeletons, and expand the searchable chemical space. On this basis, in order to generate molecules with certain characteristics, transfer learning or reinforcement learning methods can be used to guide model training, so that the chemical space generated by molecules is contracted to a specific area, and molecules that meet the conditions are sampled and generated in this area. For example, molecules with special functional groups can be generated. However, whether it is transfer learning or reinforcement learning, there is a problem that as the training proceeds, the diversity of the generated molecules gradually decreases, and the generated molecular skeletons gradually become single. Transfer learning is heavily dependent on the quality of small data sets. Too few samples and low diversity lead to premature convergence of the model, and the diversity of the generated compounds is poor. Overly complex function combinations in reinforcement learning make model training unstable and difficult to converge. If the scoring standard used by the scoring function is single, the model still converges prematurely, and the molecules obtained do not have diversity.

[0031] Based on this, the present application provides a compound design method, in which new compounds are developed and designed based on genetic algorithms, and a certain number of seed compounds are selected to simulate chromosomes in nature using the principle of simulating natural evolution in genetic algorithms to form an initial population. In each generation of evolution, the fitness of the entire population is evaluated, and a number of individuals are selected based on fitness to simulate natural selection in nature, and heredity and mutation produce the next generation of populations (i.e., derived compounds). Each generation repeats the cycle in this way to search for an optimal solution.

[0032] See also Figure 1 , Figure 1 is a schematic diagram of a compound design method in an embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment is not used as Figure 1The process sequence shown is limited. Figure 1 As shown, this embodiment includes:

[0033] S110: Obtain a seed vector.

[0034] The seed vector is a characteristic vector representation of the seed compound.

[0035] Based on the genetic algorithm, the first generation population of genetic evolution is constructed first, that is, the basic compound for compound design, i.e., the seed compound, needs to be obtained first. The seed compound can be any compound randomly selected from the compound database, which can be one or more. Different design requirements can also be used to perform specific screening on the seed compound, which is not limited here.

[0036] In the embodiments provided in the present application, the seed compound is also subjected to dimensionality reduction processing, and the complex chemical space is reduced to a one-dimensional vector. Specifically, the method of representing the compound by the molecular structural formula is converted into a method of representing the compound by a vector. Through dimensionality reduction processing, the design algorithm based on the genetic algorithm can be simplified to operations between vectors, which makes it more convenient and efficient to search the chemical space, and is more efficient.

[0037] S130: Perform a crossover operation and / or a mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector.

[0038] Through crossover and mutation operations between vectors, it is possible to simulate natural selection, inheritance, mutation, evolution, etc. in nature to generate new vectors (i.e., derivative vectors), that is, to generate new compounds to achieve compound diversification.

[0039] S150: Process the derivative vector to obtain a derivative compound.

[0040] After the operation between vectors is completed, the operation results are processed by dimensionality increase. Specifically, the method of using vectors to represent compounds is converted into the method of using molecular structures to represent compounds, and the specific structural formula of the compound is obtained, and then the derivative compound is determined.

[0041] In this embodiment, the development and design of compounds based on genetic algorithms increases the explorable compound space, can obtain diversified compounds, and increase the selection space. Furthermore, during the operation, the complex chemical space is reduced to a one-dimensional vector, which enables the design algorithm to search the chemical space conveniently and efficiently.

[0042] In one embodiment, the present application can use a neural network, take the chemical structure as input and output, and extract the vector output by the intermediate layer as a one-dimensional representation of the chemical structure. That is, the neural network model can be used to reduce or increase the dimensionality of the compound.

[0043] Among them, the autoencoder can be used to train a molecular structure encoding model and a molecular structure decoding model. The molecular structure encoding model can be used to reduce the dimensionality of the chemical structure and encode the chemical structure into a vector; the molecular structure decoding model can be used to increase the dimensionality of the vector and decode the vector into a chemical structure.

[0044] An autoencoder is a deep learning neural network that can be trained to make the input value and output value the same. It first compresses the input vector into a latent space, and then reconstructs and decodes the output so that the output is the same as the input. Specifically, the autoencoder mainly includes the encoding layer, the latent vector layer, and the decoding layer. The encoding layer contains several neurons, which can convert a large and sparse matrix into a dense one-dimensional vector composed of floating-point numbers (the vector in the latent vector layer). The decoding layer also contains several neurons, which can decode the dense one-dimensional vector into a large and sparse matrix.

[0045] In the training phase, a neural network is first established. This neural network can receive a large and sparse matrix. It is first converted into a vector of continuous values ​​through the embedding layer. These vectors are combined through multiple linear and nonlinear transformations to finally obtain a latent vector. This latent vector is decoded into a large and sparse matrix through multiple linear and nonlinear transformations. Since the parameters of these transformations are random or inaccurate, the decoded matrix and the original matrix are likely to be very different. Therefore, a certain metric is used to measure the difference between the decoded matrix and the original matrix, and the parameters in the neural network are updated by backpropagation according to the degree of difference. Then, the updated network is used to regenerate a new large and sparse matrix, and the difference between the decoded matrix and the original matrix is ​​calculated, and the parameters are updated again. Repeat multiple rounds until the difference gradually decreases and stabilizes (even if the difference is no longer reduced after the next cycle). After this training, a large and sparse vector is input, and finally an almost identical large and sparse matrix can be restored.

[0046] In one embodiment, the chemical structure can be one-hot encoded and converted into a matrix representation. Therefore, the above-mentioned neural network can be used to reduce and increase the dimensionality of the compound, and the above-mentioned training method can be used to train a molecular structure encoding model and a molecular structure decoding model.

[0047] See also Figure 2 , Figure 2 is a schematic diagram of a training process of a molecular structure model in an embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment does not Figure 2 The process sequence shown is limited. Figure 2 As shown, this embodiment includes:

[0048] S210: Obtain a sample matrix.

[0049] The sample matrix is ​​a matrix representation of sample compounds.

[0050] Among them, a compound library can be downloaded from the Internet, and effective compounds can be extracted from the compound library as sample compounds. The sample compounds can be screened to a certain extent, for example, when screening the sample compounds, chiral compounds can be removed, salt compounds can be removed, uncommon molecules can be deleted, molecules with too many heavy atoms can be removed, inorganic substances can be removed, etc. Different screening rules can be set according to different needs, and are not limited here.

[0051] After the sample compounds are selected, the selected sample compounds are converted into SMILES strings for representation. SMILES (Simplified molecular input line entry system) is a specification that uses ASCII strings to clearly describe molecular structures. Chemical structures can be written in the form of a SMILES string according to an existing set of regulations. For example, pyrimidine can be written as the SMILES string "c1ccncn1". The string can be thought of as a sentence consisting of several words. The above string of pyrimidine can be considered to consist of three words c, 1, and n. These words can be converted into vectors consisting of only 0s and 1s using one-hot encoding, and then the string can be converted into a matrix representation to obtain a sample matrix.

[0052] Take pyrimidine as an example. Its SMILES string is "c1ccncn1", which can be considered as consisting of three words c, 1, and n. These three words are disordered and discontinuous. These three words are considered as three states and represented by a vector consisting of 0 and 1. For example, if the first digit is c, the second digit is 1, and the third digit is n, then these three words can be represented as [1,0,0], [0,1,0], [0,0,1]. 1 means that the word is contained, and 0 means that the word is not contained. Then the structure of pyrimidine will be represented as a two-dimensional matrix [[1,0,0], [0,1,0], [1,0,0], [1,0,0], [0,0,1], [1,0,0], [0,0,1], [0,1,0]]. The so-called two-dimensional matrix means that the dimensions of the matrix can be understood as one dimension used to represent the vector length of each word and one dimension used to represent the length of each string. For example, in pyrimidine, the length of each word is 3, and the length of the entire pyrimidine string is 8. After encoding in this way, the pyrimidine structure is converted into a form that can be understood by computers.

[0053] After converting all compounds in the sample compound set into SMILES strings, special characters such as "$" and "#" can be added to the beginning and end of the SMILES string to indicate the beginning and end of the SMILES string, so as to distinguish and disconnect different strings, and the strings can also be deduplicated. The SMILES strings in the sample compound set can be uniformly encoded into an m*n matrix (m words, each word vector length is n). The longest SMILES string can be found, for example, its length is m. If the length of a SMILES string is less than m words, it is also represented as an m*n matrix, and all the missing elements are filled with 0. Similarly, find the longest word, for example, its length is n.

[0054] S230: Input the sample matrix into the encoding layer of the autoencoder, and encode to obtain a sample vector, wherein the sample vector is a feature vector representation of the sample compound.

[0055] S250: Input the sample vector into the decoding layer of the autoencoder, and decode to obtain a prediction matrix.

[0056] S270: Calculate the loss between the prediction matrix and the sample matrix.

[0057] S290: Iteratively update the parameters of the autoencoder based on the loss until the loss is stable, and obtain a molecular structure encoding model and a molecular structure decoding model.

[0058] The encoding layer of the updated trained autoencoder can be used as a molecular structure encoding model, and the decoding layer of the updated trained autoencoder can be used as a molecular structure decoding model.

[0059] In one embodiment, the autoencoder may further include an input layer, which may be used to convert a compound of a chemical structure formula into a matrix-form compound. The input layer and the encoding layer are then used together as a molecular structure encoding model. The molecular structure encoding model may take a compound of a molecular structure formula as input and output an encoded vector-form compound.

[0060] In one embodiment, the autoencoder may further include an output layer, which may be used to convert a matrix-type compound into a chemical structure-type compound. The specific conversion process is the inverse process of converting a chemical structure-type compound into a matrix-type compound. For details, please refer to the above description and will not be repeated here. The output layer and the decoding layer are then used together as a molecular structure decoding model. The molecular structure decoding model can take a vector-type compound as input and output a decoded molecular structure-type compound.

[0061] See also Figure 3 , Figure 3is a schematic diagram of another compound design method in the embodiment of the present application. It should be noted that if there are substantially the same results, this embodiment is not based on Figure 3 The process sequence shown is limited. Figure 3 As shown, this embodiment can combine the molecular structure encoding model, the molecular structure decoding model and the genetic algorithm to design compounds, specifically including:

[0062] S310: Obtain a seed vector.

[0063] Among them, a seed compound can be selected from the compound database to obtain the SMILES string of the seed compound; the SMILES string of the seed compound is one-hot encoded to obtain a seed matrix, which is a matrix representation of the seed compound; the seed matrix is ​​input into the molecular structure encoding model, the seed matrix is ​​encoded, and a seed vector is obtained. Please refer to the above description for details, which will not be repeated here.

[0064] S330: Perform a crossover operation on the seed vector based on a genetic algorithm to obtain a derivative vector.

[0065] Among them, the crossover operation can select two seed vectors from the seed vector set, select the exchange position of one of the seed vectors (can be one or more positions), and the seed vector and the exchange position can be selected randomly or by setting certain selection rules. The value of the selected exchange position of this seed vector is exchanged with the value of the corresponding position of another sub-vector. For example, there are two vectors [0.1, 0.2, 0.3] and [0.4, 0.5, 0.6]. Exchange the first position, then get two new vectors, [0.4, 0.2, 0.3] and [0.1, 0.5, 0.6]. For another example, exchange the first and third positions of the above two vectors, then get two new vectors [0.4, 0.2, 0.6] and [0.1, 0.5, 0.3].

[0066] S350: Perform mutation operation on the seed vector based on the genetic algorithm to obtain a derivative vector.

[0067] Among them, the mutation operation can select several seed vectors (the proportion of the vectors to be mutated can be pre-specified) from the seed vector set, and select mutation positions (can be one or more positions) from these seed vectors. The seed vectors and mutation positions can be selected randomly or by setting certain selection rules. The values ​​at these mutation positions are replaced with new values, which can be randomly replaced with any value or replaced with a set value. For example, there is a vector [0.1, 0.2, 0.3]. Select the first position and randomly replace this value with a value to obtain a new vector [0.5, 0.2, 0.3]. For another example, select the first and second positions and randomly replace the corresponding values ​​with new values ​​to obtain a new vector [0.2, 0.4, 0.3].

[0068] Both the crossover operation and the mutation operation are for generating new vectors (i.e., derivative vectors), deriving more vectors, and then deriving more compounds. The crossover operation and the mutation operation can simulate genetic evolution and can improve the diversity of compounds. The crossover operation and the mutation operation can be performed simultaneously, or in reverse order, or only one of the processes can be performed, that is, steps S330 and S350 are only for illustration, and one can be selectively executed, or the order can be reversed, which is not limited here.

[0069] S370: Process the derivative vector to obtain a derivative compound.

[0070] The derivative vector is input into the molecular structure decoding model, the derivative vector is decoded to obtain a derivative matrix, and then the derivative matrix is ​​transformed to obtain a derivative molecular structure, and then the derivative compound can be determined based on the derivative molecular structure. Please refer to the above description for details, which will not be repeated here.

[0071] S390: Measuring the fitness of the derivative compound based on the fitness function, and selecting a candidate compound from the derivative compound according to the fitness.

[0072] Fitness is a measure used to evaluate derivative compounds, such as whether the structure has good solubility, good activity, etc. In this way, the derivative compounds are linked to the criteria for judging their quality, that is, the fitness function is constructed.

[0073] Genetic algorithms simulate the process of evolution. As mentioned above, each derivative compound has an evaluation value, which represents the adaptability of the compound in the evolution process. For example, molecules with poor solubility and poor activity are more likely to be eliminated.

[0074] This evaluation criterion depends on the user's definition. The user can adaptively set the evaluation criterion (fitness function) according to the characteristic requirements of the compound to be designed. For example, the user hopes to obtain a compound with a large enough molecular weight. Then, thousands of derivative vectors are randomly generated, and these derivative vectors are converted according to the above, and a compound is obtained respectively, and then the molecular weight of these compounds is calculated. This molecular weight is the user-defined fitness. We sort these compounds in descending order according to the molecular weight, and select one or a batch of candidate compounds with a high ranking according to the parameters pre-defined by the user (the proportion or quantity selected each time).

[0075] In one embodiment, multiple rounds of crossover operations and mutation operations can be iteratively performed to obtain more derivative compounds, and then desired candidate compounds are selected from these derivative compounds. Specifically, the specific implementation method of selecting candidate compounds from derivative compounds according to the size of fitness may include: step S1: selecting a target compound whose fitness meets a preset condition from the derivative compounds according to the size of fitness; step S2: using the derivative vector corresponding to the target compound as a seed vector, and continuing to perform steps S330 and / or S350 to step S390 to measure the fitness of the derivative compound based on the fitness function. Iterate the above steps S1 and S2 until the iteration termination condition is met, and then terminate the iterative loop operation; arrange all the derived compounds obtained in descending order according to fitness; select a predetermined proportion or a predetermined number of derivative compounds with better fitness as candidate compounds.

[0076] Among them, the target compound that meets the preset conditions can be a fixed number (such as 10, 30, 50, etc.) of compounds with fitness ranking in the front row selected from the derivative compounds as the target compound; it can also be a fixed ratio (such as 1 / 10, 1 / 5, 1 / 3, etc.) of compounds selected from the derivative compounds according to fitness ranking from the front to the back as the target compound; it can also be a compound with fitness greater than a fixed threshold selected from the derivative compounds as the target compound. Among them, the number, ratio, and conditions of the selected target compound can be set as needed, and will not be repeated here.

[0077] For example, after completing an operation and obtaining the fitness of the derivative compounds, the derivative compounds can be sorted in descending order according to the fitness, the top-ranked target compounds can be selected, and the derivative vectors of these target compounds can be cross-operated and mutated to generate a batch of new one-dimensional derivative vectors. These new derivative vectors are input into the molecular structure decoding model, a new matrix is ​​decoded and converted into new derivative compounds, and the fitness of these derivative compounds is calculated. These derivative compounds are sorted in descending order according to the fitness, and the top-ranked target compounds are selected from them, and cross-operated and mutated to generate new one-dimensional derivative vectors. This iterative cycle is repeated, and all the generated derivative compounds are recorded. From these generated derivative compounds, the candidate compound with better fitness is found as the final result.

[0078] The number of iterations can depend on the set parameters and the characteristics of the data set itself. The iteration termination condition can be a pre-set number of iterations, which can range from dozens to hundreds, such as 200 to 400. The iteration termination condition can be a pre-set iteration duration, such as 8 hours, 12 hours, 24 hours, 48 ​​hours, etc.

[0079] Of course, iteration is not required and the desired candidate compounds can be obtained after one execution.

[0080] The solution provided by the above implementation method reduces the dimension of the complex chemical space into a one-dimensional vector by adopting a neural network algorithm, so that the design algorithm can search the chemical space conveniently and efficiently; it organically combines the chemical space and the genetic algorithm to overcome the problem that the generated compounds are gradually monotonous after reinforcement learning and transfer learning of the molecular generation model.

[0081] Below, the scheme of the present application will be described and illustrated through several specific experimental examples, but this should not bring too many limitations to the present application.

[0082] The latest ChEMBL28 database can be downloaded from the Internet, and the SMILES strings of the compounds are proposed. The sample compound structure must only contain atoms of hydrogen, carbon, nitrogen, oxygen, fluorine, sulfur, chlorine, and bromine. It does not contain chiral compounds, inorganic substances, and salt ions, and the number of heavy atoms is constrained to be within 70. These SMILES strings are converted into canonical forms. After deduplication, about 1.8 million SMILES are obtained. Use these SMILES to train a neural network. The implementation example is based on this neural network.

[0083] Experimental Example 1

[0084] Protein kinase B, also known as AKT, is a serine / threonine-specific protein kinase. It plays an important regulatory role in cell processes such as apoptosis, proliferation, and migration. AKT1 participates in the cell survival pathway through the apoptosis process, blocks apoptosis, and promotes cell survival. Clinical studies have found that AKT is overexpressed in a variety of human tumors such as gastric cancer and pancreatic cancer. AKT inhibitors can inhibit the activity of AKT and promote apoptosis of cancer cells.

[0085] Compound 1 is an AKT inhibitor in clinical research. By analyzing its interaction pattern, a pharmacophore model is established to evaluate the degree of match between the molecule and the pharmacophore, which is used as a fitness evaluation standard to find new molecules.

[0086] Specifically, a first predetermined amount of seed compounds is randomly selected from a compound database, and input into a molecular structure encoding model to obtain a seed vector; a crossover operation and a mutation operation are performed on the seed vector based on a genetic algorithm to obtain a plurality of derivative vectors; the derivative vectors are input into a molecular structure decoding model to obtain a plurality of derivative compounds; the matching degree between the derivative compound and the pharmacophore is evaluated using the above-mentioned pharmacophore model to obtain the fitness of the derivative compound; and a second predetermined amount of derivative vectors are selected from the derivative vectors according to the fitness as seed vectors to perform crossover operation and mutation operation, and this is iterated for 300 rounds to obtain a batch of new compounds, as follows:

[0087]

[0088] Experimental Example 2

[0089] Clinical studies have found that human isocitrate dehydrogenase 1 (IDH1) mutates in a variety of malignant tumors, such as glioma. The mutated IDH1 can convert α-ketoglutarate into 2-hydroxyglutarate. The latter is a carcinogen that accumulates in the body and promotes further deterioration of cancer. Clinical trials have shown that drugs that inhibit the activity of mutant IDH1 can effectively reduce the concentration of 2-hydroxyglutarate in the body and relieve cancer symptoms.

[0090] Compound 2 is the most promising inhibitor of mutant IDH1 currently under study. It is used as a template molecule, and the similarity of each generated molecule with the template molecule is calculated (using molecular fingerprint measurement) as the fitness of the molecule, and a group of similar molecules are searched from the potential space.

[0091] Specifically, a third predetermined amount of seed compounds are randomly selected from the compound database, input into the molecular structure encoding model, and obtain a seed vector; a crossover operation and a mutation operation are performed on the seed vector based on a genetic algorithm to obtain a plurality of derivative vectors; the derivative vectors are input into the molecular structure decoding model to obtain a plurality of derivative compounds; the similarity between each derivative compound and the template molecule is calculated respectively to obtain the fitness of the derivative compound; and a fourth predetermined amount of derivative vectors are selected from the derivative vectors according to the fitness as the seed vector for crossover operation and mutation operation, and this is iterated for 380 rounds to obtain a batch of new compounds, as follows:

[0092]

[0093] See also Figure 4 , Figure 4 Schematic diagram of the structure of the compound design device in the embodiment of the present application. In this embodiment, the compound design device 40 includes an acquisition module 41, a calculation module 42 and a decoding module 43.

[0094] Among them, the acquisition module 41 is used to obtain the seed vector, which is a characteristic vector representation of the seed compound; the operation module 42 is used to perform crossover operation and / or mutation operation on the seed vector based on the genetic algorithm to obtain a derivative vector; the decoding module 43 is used to process the derivative vector to obtain a derivative compound. In this way, the device develops and designs compounds based on the genetic algorithm, increases the explorable compound space, can obtain diversified compounds, and increases the selection space. Furthermore, during the operation, the complex chemical space is reduced to a one-dimensional vector, which enables the design algorithm to search the chemical space conveniently and efficiently. For the specific execution process, please refer to the description of the above embodiment, which will not be repeated again.

[0095] Furthermore, the compound design device 40 also includes a selection module (not shown in the figure), which is used to measure the fitness of the derivative compounds based on the fitness function; and select candidate compounds from the derivative compounds according to the size of the fitness.

[0096] Specifically, the selection module selects candidate compounds from the derivative compounds according to the size of the fitness, including: step S1: according to the size of the fitness, select the target compound whose fitness meets the preset conditions from the derivative compounds; step S2: use the derivative vector corresponding to the target compound as the seed vector, and continue to perform crossover operation and / or mutation operation on the seed vector based on the genetic algorithm to obtain the derivative vector to measure the fitness of the derivative compound based on the fitness function; iterate the steps S1 and S2 until the iteration termination condition is met, and end the iterative loop operation; arrange all the derived compounds obtained in descending order according to the fitness; select a predetermined proportion or a predetermined number of derivative compounds with better fitness as candidate compounds. In this way, more candidate compounds can be obtained, and it is easier to screen for better compounds. Please refer to the description of the above embodiment for the specific execution process, and will not be repeated again.

[0097] Further, the operation module 42 includes a cross operation submodule (not shown in the figure), which is used to select two seed vectors from the seed vector set, select the exchange position of one of the seed vectors, and exchange the value of the exchange position of the seed vector with the value of the corresponding position of another subvector to obtain a new derivative vector, thereby obtaining a derivative compound to enrich the derivative vector and improve the diversity of the derivative compound. For the specific execution process, please refer to the description of the above embodiment, which will not be repeated again.

[0098] Further, the operation module 42 includes a mutation operation submodule (not shown in the figure), which is used to select a seed vector from the seed vector set, select a mutation position from the selected seed vector, replace the value at the mutation position with a new value, obtain a new derivative vector, and then obtain a derivative compound to enrich the derivative vector and improve the diversity of the derivative compound. For the specific execution process, please refer to the description of the above embodiment, which will not be repeated again.

[0099] Furthermore, the decoding module 43 is used to input the derivative vector into the molecular structure decoding model, decode the derivative vector, and obtain the derivative molecular structure. The molecular structure decoding model is a neural network model; the derivative compound is obtained according to the derivative molecular structure.

[0100] Furthermore, the compound design device 40 also includes an encoding module (not shown in the figure), which is used to obtain the SMILES string of the seed compound; perform one-hot encoding on the SMILES string of the seed compound to obtain a seed matrix, which is a matrix representation of the seed compound; input the seed matrix into the molecular structure encoding model, encode the seed matrix, and obtain a seed vector.

[0101] Furthermore, the compound design device 40 also includes a model training module (not shown in the figure), which is used to obtain a sample matrix, which is a matrix representation of a sample compound; input the sample matrix into the encoding layer of the autoencoder, and encode to obtain a sample vector, which is a feature vector representation of the sample compound; input the sample vector into the decoding layer of the autoencoder, and decode to obtain a prediction matrix; calculate the loss between the prediction matrix and the sample matrix; iteratively update the parameters of the autoencoder based on the loss until the loss is stable, and use the decoding layer and output layer of the updated and trained autoencoder as a molecular structure decoding model, the output layer is used to convert the compound represented by the matrix into a molecular structure representation, and the encoding layer of the updated and trained autoencoder is used as a molecular structure encoding model.

[0102] The compound design device can be an independent server, a server cluster, or a module of a server, and can be used for model training, genetic algorithm, and further for designing compounds.

[0103] See also Figure 5 , Figure 5 1 is a schematic diagram of the structure of a compound design device in an embodiment of the present application. In this embodiment, the compound design device 10 includes a processor 11 and a memory 12.

[0104] The processor 11 may also be referred to as a CPU (Central Processing Unit). The processor 11 may be an integrated circuit chip having signal processing capabilities. The processor 11 may also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. A general-purpose processor may be a microprocessor or the processor 11 may also be any conventional processor, etc.

[0105] The compound design device 10 may further include a memory 12 for storing instructions and data required for the processor 11 to run.

[0106] The processor 11 is used to execute instructions to implement the method provided by any embodiment of the compound design method of the present application and any non-conflicting combination.

[0107] The compound design device can be a server, a desktop computer, a laptop computer, etc. It can be used for model training, genetic algorithms, and then used to design compounds.

[0108] See also Figure 6 , Figure 6It is a schematic diagram of the structure of a computer-readable storage medium in an embodiment of the present application. The computer-readable storage medium 20 of the embodiment of the present application stores instructions / program data 21, and when the instructions / program data 21 are executed, the method provided by any embodiment of the compound design method of the present application and any non-conflicting combination is realized. Among them, the instructions / program data 21 can form a program file and be stored in the above-mentioned storage medium 20 in the form of a software product, so that a computer device (which can be a personal computer, server, or network device, etc.) or a processor (processor) performs all or part of the steps of each embodiment method of the present application. The aforementioned storage medium 20 includes: various media that can store program codes, such as U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, or terminal devices such as computers, servers, mobile phones, and tablets.

[0109] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0110] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0111] The above description is only an implementation method of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A compound design method, characterized in that: include: Obtaining a seed vector, wherein the seed vector is a characteristic vector representation of a seed compound, wherein the seed vector is a one-dimensional vector obtained by performing dimensionality reduction processing on the seed compound through a neural network model; Performing a crossover operation and / or a mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector; Processing the derivative vector to obtain a derivative compound; The step of obtaining a seed vector comprises: Get the SMILES string of the seed compound; Performing one-hot encoding on the SMILES string of the seed compound to obtain a seed matrix, wherein the seed matrix is ​​a matrix representation of the seed compound; The seed matrix is ​​input into a molecular structure encoding model, and the seed matrix is ​​encoded to obtain the seed vector; wherein the molecular structure encoding model is obtained by training an autoencoder, and the molecular structure encoding model is used to perform dimensionality reduction processing on the chemical structure and encode the chemical structure into a vector.

2. The compound design method according to claim 1, characterized in that: After processing the derivative vector to obtain the derivative compound, the method further includes: Measuring the fitness of the derivative compounds respectively based on the fitness function; According to the fitness, candidate compounds are selected from the derivative compounds.

3. The compound design method according to claim 2, characterized in that: The step of selecting a candidate compound from the derivative compounds according to the fitness comprises: Step S1: selecting a target compound whose fitness meets a preset condition from the derivative compounds according to the size of the fitness; Step S2: using the derivative vector corresponding to the target compound as the seed vector, and continuing to perform the crossover operation and / or mutation operation on the seed vector based on the genetic algorithm to obtain the derivative vector to measure the fitness of the derivative compound based on the fitness function; Iterate the steps S1 and S2 until the iteration termination condition is met, and then terminate the iteration cycle operation; Arrange all the derived compounds obtained in descending order according to the fitness; A predetermined proportion or a predetermined number of derivative compounds with better fitness are selected as candidate compounds.

4. The compound design method according to claim 1, characterized in that: Performing a crossover operation on the seed vector based on a genetic algorithm includes: Two seed vectors are selected from the seed vector set, the exchange position of one of the seed vectors is selected, and the value of the exchange position of the seed vector is exchanged with the value of the corresponding position of another sub-vector.

5. The compound design method according to claim 1, characterized in that: Performing a mutation operation on the seed vector based on a genetic algorithm includes: A seed vector is selected from the seed vector set, a mutation position is selected from the selected seed vector, and a value at the mutation position is replaced with a new value.

6. The compound design method according to any one of claims 1 to 5, characterized in that: The step of processing the derivative vector to obtain a derivative compound comprises: Inputting the derivative vector into a molecular structure decoding model, decoding the derivative vector to obtain a derivative molecular structure, wherein the molecular structure decoding model is a neural network model; A derivative compound is obtained according to the derivative molecular structure.

7. The compound design method according to claim 6, characterized in that: Before inputting the derivative vector into a molecular structure decoding model and decoding the derivative vector to obtain a derivative molecular structure, the method further includes: Obtaining a sample matrix, wherein the sample matrix is ​​a matrix representation of sample compounds; Input the sample matrix into the encoding layer of the autoencoder, and encode to obtain a sample vector, wherein the sample vector is a feature vector representation of the sample compound; Inputting the sample vector into the decoding layer of the autoencoder and decoding to obtain a prediction matrix; Calculating the loss between the prediction matrix and the sample matrix; The parameters of the autoencoder are iteratively updated based on the loss until the loss is stable, and the decoding layer and output layer of the updated trained autoencoder are used as the molecular structure decoding model, wherein the output layer is used to convert the compound represented by the matrix into a molecular structure representation.

8. The compound design method according to any one of claims 1 to 5, characterized in that: Before inputting the seed matrix into a molecular structure coding model and performing coding processing on the seed matrix to obtain the seed vector, the method further includes: Obtaining a sample matrix, wherein the sample matrix is ​​a matrix representation of sample compounds; Input the sample matrix into the encoding layer of the autoencoder, and encode to obtain a sample vector, wherein the sample vector is a feature vector representation of the sample compound; Inputting the sample vector into the decoding layer of the autoencoder and decoding to obtain a prediction matrix; Calculating the loss between the prediction matrix and the sample matrix; The parameters of the autoencoder are iteratively updated based on the loss until the loss is stable, and the encoding layer of the updated trained autoencoder is used as the molecular structure encoding model.

9. A compound design device, characterized in that: include: An acquisition module, used for acquiring a seed vector, wherein the seed vector is a characteristic vector representation of a seed compound, wherein the seed vector is a one-dimensional vector obtained by performing dimensionality reduction processing on the seed compound through a neural network model; An operation module, used for performing a crossover operation and / or a mutation operation on the seed vector based on a genetic algorithm to obtain a derivative vector; A decoding module, used for processing the derivative vector to obtain a derivative compound; The compound design device also includes: The encoding module is used to obtain the SMILES string of the seed compound before the acquisition module obtains the seed vector; the SMILES string of the seed compound is one-hot encoded to obtain a seed matrix, and the seed matrix is ​​a matrix representation of the seed compound; the seed matrix is ​​input into a molecular structure encoding model, and the seed matrix is ​​encoded to obtain the seed vector; wherein the molecular structure encoding model is obtained by training an autoencoder, and the molecular structure encoding model is used to reduce the dimensionality of the chemical structure and encode the chemical structure into a vector.

10. A compound design device, characterized in that: The method comprises a memory and a processor, wherein the memory stores instructions, and the processor is used to execute the instructions to implement the compound design method according to any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium is used to store instructions / program data, and the instructions / program data can be executed to implement the compound design method according to any one of claims 1 to 8.

Images