A virtual screening method and screening device for large-scale compound library

Abstract

The present application proposes a virtual screening method and screening device for a large-scale compound library, including the following steps: obtaining a representative training set and obtaining multiple first docking molecules generated by a molecule generation model; setting a target protein, calculating the docking score of each first docking molecule with the target protein, combining each first docking molecule with the corresponding docking score to obtain a docking score training set, and the docking score training set is used to train the docking score prediction architecture to obtain a trained docking score prediction model; obtaining a set of optimized docking molecules generated by multiple optimized molecule generation models in parallel, selecting multiple optimized docking molecules with high target specificity scores from the set of optimized docking molecules as preliminary virtual screening results, and merging the preliminary virtual screening results with the first docking molecules to obtain the final virtual screening results. This scheme uses a drug-like database to select a representative training set of a super-large compound library so that virtual screening of a large-scale compound library can be performed using fewer computing resources.

Description

Technical Field

[0001] The present application relates to the field of drug design, and in particular to a virtual screening method and screening device for a large-scale compound library. Background Art

[0002] Ultra Large Virtual Screening (ULVS) refers to a method that uses computer technology and computer chemistry software to select potential lead compounds from a library of billions of compounds for biological activity experiments based on drug design theory. Many studies have discovered new active compounds targeting targets in this way, and have shown good potency and selectivity. At present, structure-based ULVS methods mainly include brute force docking and accelerated methods based on fragments and machine learning.

[0003] Brute force docking predicts the binding mode of each compound in the library with the target protein in a parallel computing manner. The computing resources required will inevitably increase linearly with the growth of the number of compound libraries. Taking the VirtualFlow platform as an example, it takes nearly 2 weeks to complete the screening of one billion molecules using 10,000 CPU cores at the same time. This is unaffordable for most medium-sized scientific research teams, and cannot be effectively expanded to larger compound libraries of tens of billions or even trillions.

[0004] Fragment-based ULVS uses the modular nature of ultra-large compound libraries to select representative chemical fragments that have the potential to bind to the target binding pocket as starting points. After that, based on the verified chemical reactions and molecular building block lists, the initial fragments are subjected to 1-3 cycles of expansion, docking, and selection to screen the ultra-large compound library in a hierarchical manner. The resources required for this type of method are within an acceptable range and are theoretically suitable for larger-scale virtual screening. However, the chemical reaction information and the list of molecular building blocks with corresponding functional groups required for this method are usually considered confidential information by compound suppliers and are difficult to obtain. In addition, when connecting molecular building blocks to fixed fragments in protein pockets, the docking pose filtering that needs to be considered and implemented is also relatively complex.

[0005] The ULVS acceleration method based on machine learning performs docking calculations on a small number of compounds in the database and uses this to train a neural network to predict the docking scores of the remaining molecules in the database. By iteratively docking and training the model, results close to brute force docking can be obtained when only 1%-5% of the entire space is docked. However, since its computational cost is linearly related to the total number of compounds, the performance of this type of algorithm is still not enough to handle larger-scale ULVS.

[0006] In summary, brute force docking and machine learning-based acceleration methods have large computing resource limitations when applied to larger-scale ULVS, while fragment-based ULVS methods require a lot of prior knowledge and manual processing and cannot be easily and quickly applied in the early drug development process. Summary of the invention

[0007] The embodiments of the present application provide a method and device for virtual screening of a large-scale compound library. The method generates a representative training set by using a drug-like data set, thereby reducing the number of samples. Multiple parallel generation instances are then constructed to virtually screen a library of tens of billions of compounds using fewer computing resources, thereby screening out specific active compounds for specific targets or cells.

[0008] In a first aspect, the present invention provides a method for virtual screening of a large-scale compound library, the method comprising:

[0009] Obtaining a representative training set, wherein the representative training set includes at least one type of drug small molecules;

[0010] Acquire a plurality of first docking molecules generated by a molecule generation model, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set;

[0011] Setting a target protein, calculating the docking score of each first docking molecule and the target protein, combining each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train a docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph;

[0012] Acquire the optimized docking molecule sets generated by multiple optimized molecule generation models in parallel, wherein the optimized molecule generation model is obtained by reinforcing iterative learning by introducing the docking score prediction model into the loss function of the molecule generation model, and multiple optimized docking molecules are obtained in each iterative learning, and all the optimized docking molecules obtained in each round of iteration are saved to obtain the optimized docking molecule set;

[0013] A plurality of optimized docking molecules with high target specificity scores are selected from the optimized docking molecule set as preliminary virtual screening results, and the preliminary virtual screening results are combined with the first docking molecules to obtain the final virtual screening results.

[0014] In a second aspect, the present application provides a virtual screening device for a large-scale compound library, comprising:

[0015] An acquisition module, used to acquire a representative training set, wherein the representative training set includes at least one type of drug small molecules;

[0016] A generation module, used for acquiring a trained molecule generation model to generate a plurality of first docking molecules, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set;

[0017] A docking score calculation module, which is used to set a target protein, calculate the docking score of each first docking molecule and the target protein, and combine each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train the docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph;

[0018] An optimization module is used to obtain in parallel an optimized docking molecule set generated by multiple optimized molecule generation models, wherein the optimized molecule generation model is obtained by introducing the docking score prediction model into the loss function of the molecule generation model through enhanced iterative learning, and multiple optimized docking molecules are obtained in each iterative learning, and all the optimized docking molecules obtained in each round of iteration are saved to obtain an optimized docking molecule set;

[0019] The screening module is used to select multiple optimized docking molecules with high target specificity scores from the optimized docking molecule set as preliminary virtual screening results, and merge the preliminary virtual screening results with the first docking molecules to obtain the final virtual screening results.

[0020] In a third aspect, an embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform a virtual screening method for a large-scale compound library.

[0021] In a fourth aspect, an embodiment of the present application provides a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes a program code for controlling a process to execute a process, wherein the process includes a virtual screening method for a large-scale compound library.

[0022] The main contributions and innovations of the present invention are as follows:

[0023] This scheme obtains a representative training set by obtaining a drug-like data set from a biologically active molecule library and screening it in a large-scale compound library based on the drug-like data set, so that the obtained representative training set has good drug-likeness and structural diversity, and the number is far less than the number of molecules in the large-scale compound library; this scheme introduces the trained docking score prediction model into the loss function of the trained molecule generation model to obtain an optimized molecule generation model, and generates molecules with good docking ability with the target protein through multiple parallel optimized molecule generation models, and this scheme also retains all molecules obtained in each round of iteration to avoid missing excellent molecules generated randomly at the beginning of the iteration; this scheme only needs to perform docking calculations on the first number of first docked molecules and the preliminary virtual screening results, and the calculation amount is only kept at the level of one hundred thousand. Compared with the traditional virtual screening docking calculations of millions, this scheme saves a lot of computing resources.

[0024] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0026] Figure 1 is a flow chart of a method for virtual screening of a large-scale compound library according to an embodiment of the present application;

[0027] Figure 2 is a logical schematic diagram for obtaining a representative training set according to an embodiment of the present application;

[0028] Figure 3 is a flow chart of using the representative training set to train a molecule generation framework according to an embodiment of the present application;

[0029] Figure 4 is a flow chart of training a docking score prediction framework using a docking score training set according to an embodiment of the present application;

[0030] Figure 5 is a schematic diagram of a process for selecting preliminary virtual screening results according to an embodiment of the present application;

[0031] Figure 6 According to an embodiment of the present application Figure 6 It is a bar graph of the docking scores of V-SYNTHES and REALSpace screening results finally obtained by the present invention;

[0032] Figure 7is a line graph of the final virtual screening results according to the embodiment of the present application and the enrichment of the V-SYNTHES data set;

[0033] Figure 8 is a structural block diagram of a virtual screening device for a large-scale compound library according to an embodiment of the present application;

[0034] Fig. 9 It is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0035] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with one or more embodiments of this specification. Instead, they are merely examples of devices and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0036] It should be noted that: in other embodiments, the steps of the corresponding method are not necessarily performed in the order shown and described in this specification. In some other embodiments, the steps included in the method may be more or less than those described in this specification. In addition, a single step described in this specification may be decomposed into multiple steps for description in other embodiments; and multiple steps described in this specification may be combined into a single step for description in other embodiments.

[0037] Embodiment 1

[0038] The present application embodiment provides a method for virtual screening of a large-scale compound library, which uses less computing resources to perform virtual screening on a compound library of tens of billions of compounds, thereby screening out specific active compounds for specific targets or cells. Specifically, referring to Figure 1 , the method comprising:

[0039] Obtaining a representative training set, wherein the representative training set includes at least one type of drug small molecules;

[0040] Acquire a plurality of first docking molecules generated by a molecule generation model, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set;

[0041] Setting a target protein, calculating the docking score of each first docking molecule and the target protein, combining each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train a docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph;

[0042] Acquire the optimized docking molecule sets generated by multiple optimized molecule generation models in parallel, wherein the optimized molecule generation model is obtained by reinforcing iterative learning by introducing the docking score prediction model into the loss function of the molecule generation model, and multiple optimized docking molecules are obtained in each iterative learning, and all the optimized docking molecules obtained in each round of iteration are saved to obtain the optimized docking molecule set;

[0043] A plurality of optimized docking molecules with high target specificity scores are selected from the optimized docking molecule set as preliminary virtual screening results, and the preliminary virtual screening results are combined with the first docking molecules to obtain the final virtual screening results.

[0044] In this scenario, we obtain a logical diagram representing the training set such as Figure 2 As shown, a drug-like data set is obtained from a bioactive molecule library, the drug-like data set includes drug-like molecules of various structures, and a large-scale compound library is queried to form an initial data set of multiple molecules with the highest similarity to each element in the drug-like data set, the elements in the initial data set are segmented to obtain multiple initial subsets, a clustering algorithm is used to cluster each initial subset into multiple subclasses, and at least one representative molecule is extracted from each subclass, and all representative molecules of all subclasses are integrated to obtain a representative training set.

[0045] In some embodiments, the present protocol uses ChEMBL33 as a library of bioactive molecules. ChEMBL is a large, open-access drug discovery database designed to collect medicinal chemistry data and knowledge in the drug research and development process.

[0046] In some embodiments, the present solution uses the REAL Space library as a large-scale compound library. The REAL Space library is a synthesizable chemical real space, in which the number of molecules has reached tens of billions.

[0047] This scheme obtains 2.37 million drug-like small molecules from the bioactive molecule library ChEMBL33 to form a drug-like data set, and uses each element in the drug-like data set as a query target to query in the large-scale compound library REAL Space to obtain the 1,000 molecules with the highest similarity to each element in the drug-like data set to form an initial data set. In other words, the number of elements in the initial data set obtained by this scheme should be 1,000×2.37 million, obtaining 2.37 billion molecules.

[0048] Furthermore, this scheme deduplicates the molecules in the initial data set when obtaining the initial data set. That is to say, in practical applications, the number of molecules in the initial data set is much smaller than 2.37 billion molecules. In this scheme, the number of molecules in the initial data set after deduplication is 850 million.

[0049] Furthermore, this scheme uses a similarity search tool to query multiple molecules with the highest similarity to each element in the drug-like data set in a large-scale compound library, wherein the similarity search tool used in this scheme is SpaceLight.

[0050] Specifically, this scheme uses the ChEMBL33 library to construct a drug-like dataset and query the REAL Space, because the molecules in the ChEMBL33 library have strong structural diversity, which can ensure that the results obtained by similarity query have good drug-likeness and structural diversity.

[0051] In some specific embodiments, the present solution divides the initial data set into 28 initial subsets, wherein the number of molecules in the initial subsets is manually set according to the application situation, or the molecules in the initial data set can be evenly distributed to each initial subset, which is not limited in the present solution.

[0052] Among the 28 initial subsets in this scheme, 27 initial subsets contain 30 million molecules, and the remaining 1 initial subset contains 17 million molecules. Experiments in this scheme show that such allocation will lead to better clustering effects.

[0053] In some specific embodiments, the present solution uses a clustering algorithm to cluster each subset into 300 small classes, and each small class contains 100,000 molecules.

[0054] Furthermore, this scheme uses the K-means algorithm to cluster each initial subset and uses the circles algorithm to extract at least one representative molecule in each subclass.

[0055] Specifically, when the circles algorithm is used to extract at least one representative molecule in each subclass, the threshold of the circles algorithm is set to 0.55.

[0056] In some specific embodiments, this solution integrates all representative molecules of all subcategories to obtain 24.29 million molecules constituting a representative training set.

[0057] Specifically, the representative data set obtained by this scheme provides sufficient data with high chemical space coverage for the training of the molecular generation model, enabling it to effectively infer and generate molecules in the target compound space of the non-training set.

[0058] In the step of "the molecular generation model is trained to obtain a molecular generation architecture using the representative training set", the molecular generation architecture is composed of an embedding layer, a long short-term memory network layer and a linear output layer connected in series in sequence.

[0059] In this solution, the dimension of the embedding layer is 1024, and the long short-term memory network layer is composed of 6 2048-dimensional LSTM (long short-term memory) layers.

[0060] Before the step of "the molecular generation model trains the molecular generation framework using the representative training set", the representative training set is cleaned and standardized.

[0061] Specifically, the process of cleaning and standardizing the representative training set includes: using the RDKit toolkit to extract the attributes of each molecule in the representative training set, removing non-drug abnormal molecules based on the attributes of each molecule, and removing molecules containing rare chemical elements in the SMILES string of the molecular structure.

[0062] Specifically, cleaning and standardizing the representative training set can ensure the consistency and reliability of the data.

[0063] Specifically, this solution uses the RDKit toolkit to extract the properties of each molecule, including the number of atoms and rings. Non-drug abnormal molecules refer to those with more than 70 atoms, more than 10 rings, and a longest carbon chain greater than or equal to 5. Non-drug abnormal molecules will cause obvious deviations in the molecular generation model, so non-drug abnormal molecules are removed from the representative data set.

[0064] Specifically, one or two characters representing rare chemical elements in the SMILES string of the molecular structure are rare words. By removing molecules containing rare words, the size of the vocabulary can be reduced, thereby speeding up the training of the molecular generation model.

[0065] In some embodiments, 23.48 million molecules remain in the representative training set after cleaning and normalization.

[0066] Furthermore, the representative training set is cleaned and standardized before data enhancement.

[0067] In this scheme, a plurality of different random SMILES are generated for each molecule in the representative training set for data augmentation.

[0068] Specifically, in this scheme, 200 different random SMILES are generated for each molecule in the representative training set for data enhancement. This scheme creates data sets corresponding to different iteration rounds through data enhancement, and this data enhancement method can increase the diversity of the representative training set and improve the generalization ability of the molecule generation model.

[0069] In this scheme, the flowchart of using the representative training set to train the molecular generation architecture is as follows Figure 3As shown, the molecules in the representative training set are segmented into word units, gradually converted into vector representations through the embedding layer, and then input into the long short-term memory network layer, and finally output by the linear output layer.

[0070] Furthermore, during the training of the molecular generation architecture, the batch size was set to 512, the initial learning rate was set to 1E-4, and the negative log-likelihood value was used as the loss function. The Adam optimization algorithm was used to minimize the loss function. The loss function formula of the molecular generation architecture is as follows:

[0071]

[0072] in, reflects the probability of sampling a SMILES sequence C from the molecule generation framework, Indicates that given the previously sampled word, Step sampling to word unit The minus sign is added to ensure that large positive values ​​correspond to lower probabilities. This scheme minimizes To maximize the sampling probability of the molecular sequences in the training set.

[0073] Specifically, this scheme saves the model weights of each iteration round during the training process, and ends the training when the iteration round reaches 200 or the learning rate drops to 1E-5.

[0074] In this scheme, a validation set is constructed, and after the training is completed, the validation set is used to validate the model of each iteration round, and the molecule generation architecture of the iteration round with the smallest average negative log-likelihood value is used as the molecule generation model, wherein the validation set is selected from the representative training set, and the molecules in the validation set do not participate in the training of the molecule generation architecture.

[0075] Exemplarily, if there are 23.48 million molecules in the representative training set, 100,000 molecules are randomly selected as the validation set.

[0076] Specifically, the validation set can evaluate the generalization ability of the molecular generation model.

[0077] In some specific embodiments, the present solution uses ICM-Pro software to calculate the docking score between each first docking molecule and the target protein.

[0078] In the step of "using the docking score training set to train the docking score prediction architecture to obtain a trained docking score prediction model", the docking score prediction architecture uses edge-based directed message encoding, and the hidden layer dimension of the encoder is 300.

[0079] Specifically, the flowchart of using the docking score training set to train the docking score prediction architecture is as follows: Figure 4 As shown, atomic features are represented as nodes and bond features are represented as edges as inputs of the docking score prediction model. In the message passing stage of the docking score prediction model, multiple iterations are performed within the molecule to pass information. The hidden state of the atom is obtained by aggregating the hidden state of the bond and used as the output feature representation of a molecule in the readout stage. In the docking score prediction architecture, a feedforward neural network is used to predict the predicted docking score of each molecule with the target protein.

[0080] Specifically, the message passing stage of the docking score prediction model was performed three times within the molecule to transfer information by using the chemprop library.

[0081] In this scheme, the docking score prediction architecture uses the L2 norm between the docking score of each first docking molecule and the target protein and the corresponding predicted docking score as the loss function, and uses ReLU as the activation function. The docking score prediction model is trained by the Adam optimization algorithm with 50 iterations and a batch size of 50, and the model of the round with the smallest loss function is saved as the docking score prediction model.

[0082] Exemplarily, G protein coupled receptor-cannabinoid receptor 2 (CB2) is used as a target protein as a test example, and 110,000 first docking molecules are selected. After calculating the docking score of each first docking molecule, a docking score set is obtained. 10,000 molecules in the docking score set are selected as a test set, and the remaining 10 molecules are used as a training set to train the docking score prediction framework.

[0083] In this scheme, the benefit of introducing the docking score prediction model into the loss function of the molecule generation model is that it can guide the optimization of the molecule generation model to generate molecules with stronger binding ability to specific targets.

[0084] In the step of "saving all optimized docking molecules obtained in each round of iteration to obtain an optimized docking molecule set", a memory is constructed to save all optimized docking molecules generated by each optimized molecule generation model in each round of iteration in the memory.

[0085] Specifically, although it is reasonable that as the number of iterations increases, the molecules obtained in each iteration will be better, since this scheme generates molecules randomly, excellent molecules may occasionally appear in a certain iteration. Therefore, this scheme saves all molecules in each iteration and uses target-specific scores for screening to avoid missing excellent molecules generated randomly at the beginning of the iteration.

[0086] In this scheme, each optimized molecule generation model generates 128 molecule samples in each round of iteration until the iteration stopping condition is met.

[0087] Specifically, the iteration stopping condition of each optimized molecule generation model in the present scheme is reaching a preset number of iterations or reaching a limit on the number of molecules, wherein the preset number of iterations is 15,000 and the limit on the number of molecules is 1 million. That is to say, each optimized molecule generation model in the present scheme stops iterating when it iterates 15,000 times or generates 1 million molecules, and all acquired molecules are stored in the memory.

[0088] In this protocol, the target specificity score, molecular weight, and diversity of each optimized docking molecule are calculated and combined in the form of a multi-parameter score, which is expressed as:

[0089]

[0090] Where MPO(C) is the multi-parameter score of molecule C, is the normalized target specificity score, is the scaling value of the molecular weight. For molecules with molecular weights between 200 and 600 Daltons, the scaling value is close to 1. For molecules with other molecular weights, the scaling value is close to 0. m is the number of molecules in the memory that have the same Murcko scaffold as molecule C. The upper limit of m is Set to 10 to control molecular diversity, , is the weight factor, , .

[0091] Furthermore, this scheme uses the sigmoid function to normalize the target specificity score to between [0,1]. The closer to 1, the better the target specificity score. The formula is as follows:

[0092]

[0093] in, , is the weight factor, , , d is the unnormalized target specificity score.

[0094] In this scheme, the target specificity score, molecular weight and diversity of each optimized docking molecule are calculated and combined to obtain a multi-parameter score, and a priori intelligent agent identical to the optimized molecule generation model is constructed. The multi-parameter score is combined with the negative log-likelihood of the prior intelligent agent to obtain the enhanced likelihood, and the square difference between the likelihood of the optimized molecule generation model and the enhanced likelihood is used as the loss function of the optimized molecule generation model.

[0095] Specifically, the formula for obtaining enhanced likelihood is expressed as:

[0096]

[0097] in, To enhance the likelihood, is the negative log-likelihood of the prior agent, is the multi-parameter score.

[0098] Specifically, the formula for the loss function of optimizing the molecular generation model is expressed as:

[0099]

[0100] in, To enhance the likelihood, Generate model likelihoods for optimized molecules.

[0101] Specifically, this scheme uses the Adam optimization algorithm to minimize the loss function of the optimized molecular generation model.

[0102] Before the step of "selecting multiple optimized docking molecules with high target specificity scores from the self-optimized docking molecule set as preliminary virtual screening results", the optimized docking molecule sets generated by each optimized molecule generation model are merged and duplicated.

[0103] In this scheme, the process diagram for selecting preliminary virtual screening results is as follows Figure 5 As shown, in this scheme, 100,000 molecules with the highest targeting scores were selected as the preliminary virtual screening results.

[0104] Specifically, this protocol uses the spatiallight tool to obtain the 100,000 molecules with the highest target specificity scores.

[0105] Exemplarily, this scheme constructed 10 parallel optimized molecule generation models, each of which generated 1 million molecules in all iterations, resulting in a total of 10 million molecules. Among the total of 10 million molecules, 100,000 molecules with the highest target specificity scores were selected as preliminary virtual screening results.

[0106] In this scheme, the purpose of merging the preliminary virtual screening results with the first docking molecule to obtain the final virtual screening results is that, from the perspective of computing resources, all docking calculations performed by the entire scheme need to be considered. Therefore, although the first docking molecule does not participate in the subsequent reinforcement learning, it still docks with the target protein, so the first number of first docking molecules must be combined to obtain the final virtual screening results.

[0107] Embodiment 2

[0108] From the experimental results in Table 1, we can see that the trained molecule generation model generates 94.96% of the molecules in the REAL Space library when idling, while the optimized molecule generation model generates 70.07% of the molecules in the REAL Space library. Although this ratio has decreased, it is still enough to support us in finding enough REAL Space molecules, because the number of REAL Space molecules actually used for subsequent analysis and testing is only 100,000. Therefore, we only need to select the first 150,000 molecules from the generated set for REAL Space search to meet the demand.

[0109]

[0110] Table 1 Proportion of target compound library molecules generated by the molecular generation model

[0111] The same protein target CB2 (PDBID: 5ZTY) and the same 22 small molecules as in the V-SYNTHES article were used to perform docking calculations on the same version of ICM-Pro (version 3.9-2b). The results showed that the mean deviation between our scores and the scores of the small molecules reported in the article was -0.90, which shows that we can reproduce the docking calculations in the original article very well. After that, the same settings were used to perform docking calculations on the CB2 target for 210,000 molecules in the final virtual screening results. Figure 6 The figure is a bar graph of the docking scores (the more negative, the better) of the REAL Space screening results finally obtained by V-SYNTHES and the present invention. The green bar graph in the figure represents the docking score distribution of the data set obtained by the present invention, and the purple bar graph represents the docking score distribution of the data set obtained by V-SYNTHES. The results show that compared with V-SYNTHES, the molecules in the data set of the present invention have more and stronger potential binding abilities, and the difference is significant. Further, in order to quantitatively compare the docking scores of the two data sets, we introduce the enrichment coefficient as shown in the following formula:

[0112]

[0113] in, represents the number of molecules with scores better than x in the data set obtained by the present invention, represents the number of molecules in the V-SYNTHES dataset that score better than x, Figure 7 is a line graph of the final virtual screening results of this scheme and the enrichment of the V-SYNTHES data set. The line in the graph is the ratio of the number of potential compounds in the final virtual screening results of this scheme to the number of potential compounds screened by V-SYNTHES. Figure 7As can be seen from the figure, at a threshold score of -40, the number of potential compounds in the final virtual screening results is equivalent to 74 times that of V-SYNTHES. Even at a threshold score of -30, the number of potential compounds in our dataset is close to 4 times that of V-SYNTHES. It is worth noting that V-SYNTHES obtained this dataset based on docking calculations of 1.5 million molecules, while we only performed docking calculations on 210,000 molecules. Therefore, it can be seen that with fewer docking calculations, we have achieved several to dozens of times of potential compound enrichment at different thresholds.

[0114] Embodiment 3

[0115] Based on the same idea, refer to Figure 8 , the present application also proposes a virtual screening device for a large-scale compound library, comprising:

[0116] An acquisition module, used to acquire a representative training set, wherein the representative training set includes at least one type of drug small molecules;

[0117] A generation module, used for obtaining a plurality of first docking molecules generated by a molecule generation model, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set;

[0118] A docking score calculation module, which is used to set a target protein, calculate the docking score of each first docking molecule and the target protein, and combine each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train the docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph;

[0119] An optimization module is used to obtain in parallel an optimized docking molecule set generated by multiple optimized molecule generation models, wherein the optimized molecule generation model is obtained by introducing the docking score prediction model into the loss function of the molecule generation model through enhanced iterative learning, and multiple optimized docking molecules are obtained in each iterative learning, and all the optimized docking molecules obtained in each round of iteration are saved to obtain an optimized docking molecule set;

[0120] The screening module is used to select multiple optimized docking molecules with high target specificity scores from the optimized docking molecule set as preliminary virtual screening results, and merge the preliminary virtual screening results with the first docking molecules to obtain the final virtual screening results.

[0121] Embodiment 4

[0122] This embodiment also provides an electronic device, referring to Fig. 9, comprises a memory 404 and a processor 402, wherein the memory 404 stores a computer program, and the processor 402 is configured to run the computer program to execute the steps in any of the above method embodiments.

[0123] Specifically, the processor 402 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0124] Among them, the memory 404 may include a large capacity memory 404 for data or instructions. For example, but not limitation, the memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid state drive (SSD), a flash memory, an optical disk, a magneto-optical disk, a tape, or a universal serial bus (USB) drive, or a combination of two or more of these. In appropriate cases, the memory 404 may include a removable or non-removable (or fixed) medium. In appropriate cases, the memory 404 may be inside or outside the data processing device. In a specific embodiment, the memory 404 is a non-volatile memory. In a specific embodiment, the memory 404 includes a read-only memory (ROM) and a random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (Programmable Read-Only Memory, PROM for short), an erasable PROM (Erasable Programmable Read-Only Memory, EPROM for short), an electrically erasable PROM (Electrically Erasable Programmable Read-Only Memory, EEPROM for short), an electrically alterable ROM (Electrically Alterable Read-Only Memory, EAROM for short) or a flash memory (FLASH) or a combination of two or more of these. In appropriate circumstances, the RAM may be a static random access memory (SRAM) or a dynamic random access memory (DRAM), wherein the DRAM may be a fast page mode dynamic random access memory 404 (FPMDRAM), an extended data output dynamic random access memory (EDODRAM), a synchronous dynamic random access memory (SDRAM), etc.

[0125] The memory 404 may be used to store or cache various data files required for processing and / or communication, as well as possible computer program instructions executed by the processor 402 .

[0126] The processor 402 reads and executes the computer program instructions stored in the memory 404 to implement any one of the virtual screening methods for large-scale compound libraries in the above embodiments.

[0127] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408 , wherein the transmission device 406 is connected to the processor 402 , and the input / output device 408 is connected to the processor 402 .

[0128] The transmission device 406 can be used to receive or send data via a network. Specific examples of the above-mentioned network may include a wired or wireless network provided by a communication provider of the electronic device. In one example, the transmission device includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 406 can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly.

[0129] The input and output device 408 is used to input or output information. In this embodiment, the input information may be a code table training set, a target protein, etc., and the output information may be a final screening result, etc.

[0130] Optionally, in this embodiment, the processor 402 may be configured to perform the following steps through a computer program:

[0131] S101, obtaining a representative training set, wherein the representative training set includes at least one type of drug small molecules;

[0132] S102, obtaining a plurality of first docking molecules generated by a molecule generation model, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set;

[0133] S103, setting a target protein, calculating a docking score between each first docking molecule and the target protein, combining each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train a docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph;

[0134] S104, obtaining in parallel a set of optimized docking molecules generated by multiple optimized molecule generation models, wherein the optimized molecule generation model is obtained by reinforcing iterative learning by introducing the docking score prediction model into the loss function of the molecule generation model, and multiple optimized docking molecules are obtained in each iterative learning, and all optimized docking molecules obtained in each round of iteration are saved to obtain an optimized docking molecule set;

[0135] S105, selecting a plurality of optimized docking molecules with high target specificity scores from the optimized docking molecule set as preliminary virtual screening results, and merging the preliminary virtual screening results with the first docking molecules to obtain a final virtual screening result.

[0136] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be described in detail here.

[0137] In general, various embodiments may be implemented in hardware or dedicated circuits, software, logic, or any combination thereof. Some aspects of the invention may be implemented in hardware, while other aspects may be implemented in firmware or software that may be executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flow charts, or using some other graphical representation, it should be understood that, as non-limiting examples, the boxes, devices, systems, techniques, or methods described herein may be implemented in hardware, software, firmware, dedicated circuits or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0138] Embodiments of the present invention may be implemented by computer software that is executable by a data processor of a mobile device, such as in a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets and / or macros may be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. A computer program product may include one or more computer executable components configured to perform an embodiment when the program is run. One or more computer executable components may be at least one software code or a portion thereof. In addition, at this point, it should be noted that, for example, Fig. 9 Any block of the logic flow in the program may represent program steps, or interconnected logic circuits, blocks and functions, or a combination of program steps and logic circuits, blocks and functions. The software may be stored on physical media such as memory chips or storage blocks implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as, for example, DVDs and their data variants, CDs, etc. Physical media are non-transitory media.

[0139] Those skilled in the art should understand that the technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A virtual screening method for a large-scale compound library, characterized in that: The following steps are involved: Obtaining a representative training set, wherein the representative training set includes at least one type of drug small molecules; Acquire a plurality of first docking molecules generated by a molecule generation model, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set; Setting a target protein, calculating the docking score of each first docking molecule and the target protein, combining each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train a docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph; Acquire the optimized docking molecule sets generated by multiple optimized molecule generation models in parallel, wherein the optimized molecule generation model is obtained by reinforcing iterative learning by introducing the docking score prediction model into the loss function of the molecule generation model, and multiple optimized docking molecules are obtained in each iterative learning, and all the optimized docking molecules obtained in each round of iteration are saved to obtain the optimized docking molecule set; A plurality of optimized docking molecules with high target specificity scores are selected from the optimized docking molecule set as preliminary virtual screening results, and the preliminary virtual screening results are combined with the first docking molecules to obtain the final virtual screening results.

2. The method for virtual screening of a large-scale compound library according to claim 1, characterized in that: In the step of "obtaining a representative training set", a drug-like data set is obtained from a bioactive molecule library, wherein the drug-like data set includes drug-like molecules of various structures, and a large-scale compound library is searched for multiple molecules with the highest similarity to each element in the drug-like data set to form an initial data set, the elements in the initial data set are segmented to obtain multiple initial subsets, a clustering algorithm is used to cluster each initial subset into multiple subclasses, and at least one representative molecule is extracted from each subclass, and all representative molecules of all subclasses are integrated to obtain a representative training set.

3. The method for virtual screening of a large-scale compound library according to claim 1, characterized in that: Before the step of "the molecular generation model uses the representative training set to train the molecular generation architecture", the representative training set is cleaned and standardized. The process of cleaning and standardizing the representative training set includes: using the RDKit toolkit to extract the attributes of each molecule in the representative training set, removing non-drug abnormal molecules based on the attributes of each molecule, and removing molecules containing rare chemical elements in the SMILES string of the molecular structure.

4. The method for virtual screening of a large-scale compound library according to claim 1, characterized in that: In the step of "the molecular generation model uses the representative training set to train the molecular generation architecture", a validation set is constructed, and after the training is completed, the validation set is used to validate the model of each iteration round, and the molecular generation architecture of the iteration round with the smallest average negative log-likelihood value is used as the molecular generation model, wherein the validation set is selected from the representative training set, and the molecules in the validation set do not participate in the training of the molecular generation architecture.

5. The method for virtual screening of a large-scale compound library according to claim 1, characterized in that: In the step of "representing each docked molecule in the form of a molecular graph", atomic features are represented as nodes and bond features are represented as edges as inputs to the docking score prediction model. In the message passing stage of the docking score prediction model, information is passed through multiple iterations within the molecule. The hidden states of atoms are obtained by aggregating the hidden states of bonds and used as the output feature representation of a molecule in the readout stage. In the docking score prediction architecture, a feedforward neural network is used to predict the predicted docking score of each molecule with the target protein.

6. The method for virtual screening of a large-scale compound library according to claim 1, characterized in that: In the step of "saving all optimized docking molecules obtained in each round of iteration to obtain an optimized docking molecule set", a memory is constructed to save all optimized docking molecules generated by each optimized molecule generation model in each round of iteration in the memory.

7. The method for virtual screening of a large-scale compound library according to claim 1, characterized in that: The target specificity score, molecular weight and diversity of each optimized docking molecule are calculated and combined to obtain a multi-parameter score, and a priori intelligent agent identical to the optimized molecule generation model is constructed. The multi-parameter score is combined with the negative log-likelihood of the prior intelligent agent to obtain the enhanced likelihood, and the square difference between the likelihood of the optimized molecule generation model and the enhanced likelihood is used as the loss function of the optimized molecule generation model.

8. A virtual screening device for a large-scale compound library, characterized in that: The following steps are involved: An acquisition module, used to acquire a representative training set, wherein the representative training set includes at least one type of drug small molecules; A generation module, used for acquiring a trained molecule generation model to generate a plurality of first docking molecules, wherein the molecule generation model is obtained by training a molecule generation architecture using the representative training set; A docking score calculation module, which is used to set a target protein, calculate the docking score of each first docking molecule and the target protein, and combine each first docking molecule with the corresponding docking score to obtain a docking score training set, wherein the docking score training set is used to train the docking score prediction framework to obtain a trained docking score prediction model, and each docking molecule is represented in the form of a molecular graph; An optimization module is used to obtain in parallel an optimized docking molecule set generated by multiple optimized molecule generation models, wherein the optimized molecule generation model is obtained by introducing the docking score prediction model into the loss function of the molecule generation model through enhanced iterative learning, and multiple optimized docking molecules are obtained in each iterative learning, and all the optimized docking molecules obtained in each round of iteration are saved to obtain an optimized docking molecule set; The screening module is used to select multiple optimized docking molecules with high target specificity scores from the optimized docking molecule set as preliminary virtual screening results, and merge the preliminary virtual screening results with the first docking molecules to obtain the final virtual screening results.

9. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor is configured to run the computer program to execute the method for virtual screening of a large-scale compound library according to any one of claims 1-7.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, wherein the computer program includes a program code for controlling a process to execute a process, wherein the process includes a virtual screening method for a large-scale compound library according to any one of claims 1 to 7.

Images