Small molecule compound virtual screening method and system based on parallel Bayesian optimization

Abstract

The present application provides a method and system for virtual screening of small molecule compounds based on parallel Bayesian optimization, which is applied to multiple computing nodes that work independently and have a communication relationship. The computing nodes include: a master node and multiple slave nodes. The method includes: S1, each computing node uses a proxy model to predict the target properties of the target molecule library, obtains a prediction score and uncertainty, and calculates an EI score; S2, each computing node obtains small molecule compounds to be docked according to the number of collections in descending order of EI scores, performs molecular docking, and obtains a docking score; S3, the master node determines the number of collections of each computing node based on the docking score of each computing node; S4, the master node updates the proxy model based on the small molecule compounds to be docked and the corresponding docking score of each computing node, and repeats steps S1-S4 to perform multiple rounds of model training until a predetermined iteration round is reached. This method can avoid idle resources and save computing time.

Description

Technical Field

[0001] The present application belongs to the field of virtual screening technology in drug development, and in particular relates to a method and system for virtual screening of small molecule compounds based on parallel Bayesian optimization. Background Art

[0002] Drug development is a long and costly process. The discovery of lead compounds accounts for the largest and most time-consuming portion of the cost, meaning the discovery of lead compounds with specific medicinal properties followed by subsequent optimization and clinical trials. Virtual screening is the most commonly used method in lead compound discovery. Improving the efficiency of virtual screening and reducing its computational cost can significantly shorten the drug development cycle and reduce its cost. In practice, as drug synthesis technology matures, the number of synthesizable drug-like small molecules is rapidly increasing, and virtual screening is also limited by computational cost and time.

[0003] In order to solve the problem of unacceptable screening time caused by the increase in the size of the molecular library, parallel technology and artificial intelligence technology are introduced into the virtual screening process, and the docking scores of multiple molecules are calculated simultaneously, so as to solve the problems of large computing power requirements and long time consumption faced by virtual screening tasks by superimposing computing power. At present, the method based on parallel Bayesian optimization and the use of proxy models to fit the objective function have performed well in virtual screening, but the following problems still exist: 1) The black box evaluation time of each molecule is different and cannot be predicted. In the distributed parallel evaluation process, there are cases where distributed nodes have idle resources; 2) The evaluation process is parallelized, while the model prediction stage still uses single-machine serial execution, which does not fundamentally solve the problem of Bayesian optimization method being affected by the idle computing resources of a single node; 3) Since the purpose of Bayesian optimization is to obtain the global optimal solution, after the distributed parallel expansion, the acquisition function obtains the local optimum of local data, not the global optimum. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and system for virtual screening of small molecule compounds based on parallel Bayesian optimization, which can avoid idle computing resources and save computing time.

[0005] The present application provides a method for virtual screening of small molecule compounds based on parallel Bayesian optimization, which is applied to multiple computing nodes that work independently and have a communication relationship, wherein the computing nodes include: a master node and multiple slave nodes, and the method includes:

[0006] S1. Each computing node uses the proxy model to predict the target properties of the small molecule compounds in the corresponding target molecule library to obtain the prediction score and uncertainty of each small molecule compound, and calculate the EI score of each small molecule compound; wherein the target properties include affinity, toxicity, hydrophilicity and synthesizability;

[0007] S2. Each computing node obtains the small molecule compound to be docked according to the predetermined number of acquisitions in descending order of EI score, and performs molecular docking using docking software to obtain a docking score of the small molecule compound to be docked; wherein the docking score of the small molecule compound to be docked includes: the docking score of the small molecule compound to be docked and the target and the predicted docking score of the small molecule compound to be docked;

[0008] S3. The master node evaluates the ability of each computing node to screen small molecule compounds based on the docking score of the small molecule compound to be docked with the target, so as to determine the number of samples collected by each computing node;

[0009] S4. The master node updates the training data set of the proxy model based on the small molecule compounds to be docked and the corresponding docking scores of each computing node to update the proxy model, and repeats steps S1-S4 to perform multiple rounds of model training and molecular docking until a predetermined iterative round is reached to obtain the optimal value of the docking score between the small molecule compounds to be docked and the target of each computing node.

[0010] Furthermore, the target molecule library corresponding to each computing node is pre-built in the following way:

[0011] Obtain the name of the target associated with the disease, and search for structural information corresponding to the target in the PDB protein structure database;

[0012] selecting a small molecule compound library containing structural features of small molecule compounds according to the structural information;

[0013] Distributing the small molecule compound library equally to each computing node to construct a target molecule library corresponding to each computing node;

[0014] Among them, the targets include: protein targets and RNA targets; the structural features include: molecular formula, SMILE expression and atomic relative coordinates of small molecule compounds.

[0015] Furthermore, before each computing node uses the proxy model to predict the target properties of the small molecule compounds in the corresponding target molecule library, the method further includes:

[0016] Initialize the proxy model as follows:

[0017] Each computing node randomly collects small molecule compounds from the corresponding target molecule library, and performs molecular docking using the docking software to obtain docking scores between the randomly collected small molecule compounds and the target;

[0018] The master node constructs an initial training data set for the proxy model based on the small molecule compounds randomly collected by each computing node and the corresponding docking scores;

[0019] The master node initializes the proxy model using the structural features of the randomly collected small molecule compounds in the initial training data set as input variables and the docking scores corresponding to the randomly collected small molecule compounds as output variables;

[0020] Wherein, the proxy model is a deep graph message passing model.

[0021] Furthermore, the calculation of the EI score of each small molecule compound includes:

[0022] Each computing node obtains the mean prediction score based on the prediction score of each small molecule compound;

[0023] Each computing node calculates the EI score of each small molecule compound based on the predicted score mean and the uncertainty of each small molecule compound using the following formula:

[0024]

[0025] Where D N ={(G1, y1), ..., (G N ,y N )} is the docked small molecule compound and the corresponding docking score, θ is the parameter of the proxy model, y max is the maximum docking score of the docked small molecule compound, p(y|D N , θ) is the distribution of prediction scores of small molecule compounds predicted by the surrogate model.

[0026] Furthermore, the molecular docking is performed using docking software to obtain a docking score of the small molecule compound to be docked, including:

[0027] Each computational node uses the structural characteristics of the small molecule compound to be docked as an input variable, and uses the docking software to perform molecular docking in descending order of uncertainty of the small molecule compound to obtain the docking score between the small molecule compound to be docked and the target;

[0028] Each computing node monitors the docking process. When all the small molecule compounds to be docked on any computing node are docked, the docking process of other computing nodes will stop at the same time.

[0029] Other computing nodes determine the predicted score of the small molecule compound to be docked that has not completed docking when the docking process is stopped as the predicted docking score of the small molecule compound to be docked;

[0030] Wherein, the docking software is smina docking software.

[0031] Furthermore, the ability of each computing node to screen small molecule compounds is evaluated based on the docking score of the small molecule compound to be docked with the target of each computing node to determine the number of samples collected by each computing node, including:

[0032] The master node obtains the average docking score of each computing node based on the docking score of the small molecule compound to be docked and the target of each computing node;

[0033] The master node calculates the collection weight of each computing node using the following formula based on the average docking score of each computing node:

[0034]

[0035] Where, is the collection weight of the kth round of computing node i, is the set of docking scores between the small molecule compound to be docked and the target obtained in the k-1th round of calculation for node i,

[0036] The number of collections assigned to computing node i in the k-1th round;

[0037] The master node determines the number of collections for each computing node using the following formula based on the collection weight of each computing node:

[0038]

[0039] Where N is the set of computing nodes.

[0040] Furthermore, the updating of the training data set of the proxy model based on the small molecule compounds to be docked and the corresponding docking scores of each computing node to update the proxy model includes:

[0041] The master node incorporates the small molecule compounds to be docked and the corresponding docking scores of each computing node into the training data set of the previous round to update the training data set of the proxy model;

[0042] The master node uses the updated training data set of the proxy model to perform this round of training on the proxy model to update the proxy model.

[0043] The present application also provides a small molecule compound virtual screening system based on parallel Bayesian optimization, comprising: an agent model, a load balancing module, docking software, and a corresponding target molecule library arranged on each computing node, an adaptive acquisition module and an agent model update module arranged on a master node, wherein each computing node operates independently and has a communication relationship;

[0044] The proxy model is used to predict target properties of small molecule compounds in the corresponding target molecule library to obtain a prediction score and uncertainty for each small molecule compound; wherein the target properties include affinity, toxicity, hydrophilicity and synthesizability;

[0045] The load balancing module is used to calculate the EI score of each small molecule compound and obtain the small molecule compounds to be docked according to the predetermined number of collections in descending order of the EI score;

[0046] The docking software is used to perform molecular docking on the small molecule compound to be docked to obtain a docking score between the small molecule compound to be docked and the target; wherein the docking score of the small molecule compound to be docked includes: the docking score between the small molecule compound to be docked and the target and the predicted docking score of the small molecule compound to be docked;

[0047] The adaptive acquisition module is used to evaluate the ability of each computing node to screen small molecule compounds based on the docking score of the small molecule compound to be docked with the target of each computing node, so as to determine the number of samples to be collected by each computing node;

[0048] The proxy model updating module is used to update the training data set of the proxy model based on the small molecule compound to be docked and the corresponding docking score, so as to update the proxy model.

[0049] The virtual screening method and system for small molecule compounds based on parallel Bayesian optimization provided by the present application have the following technical effects: 1) each computing node uses the proxy model to respectively calculate the prediction score and uncertainty of the corresponding target molecule library, and uses the docking software to respectively calculate the docking score of the corresponding small molecule compound to be docked, and computing time can be saved by distributed computing; 2) each computing node selects the small molecule compound to be docked from the target molecule library by the EI score, which can make the accuracy of the small molecule compound to be docked higher; 3) the main node calculates the number of collections of each computing node by the docking score, and can match the appropriate number of collections for each computing node; 4) each computing node supervises the docking process with each other, which can avoid idle computing resources while obtaining sufficient training data sets to update the proxy model. In summary, the screening method and system of the present application can avoid idle computing resources and save computing time. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flowchart of a small molecule compound virtual screening method based on parallel Bayesian optimization provided in an embodiment of the present application is shown;

[0051] Figure 2 A schematic structural diagram of a small molecule compound virtual screening system based on parallel Bayesian optimization provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of this technical solution more clear, the following technical solution is further described in detail in conjunction with specific implementation methods. It should be understood that these descriptions are only exemplary and are not intended to limit the scope of this technical solution.

[0053] Example 1

[0054] See also Figure 1 , Figure 1 The flowchart of the small molecule compound virtual screening method based on parallel Bayesian optimization provided in the embodiment of the present application is shown. Figure 1 As shown, the method is applied to multiple computing nodes that work independently and have a communication relationship, the computing nodes include: a master node and multiple slave nodes, and the method includes:

[0055] S101. Construct a target molecule library.

[0056] In specific implementation, the target molecule library corresponding to each computing node is pre-built in the following way:

[0057] Step 1011: Obtain the name of a target associated with the disease, and search for structural information corresponding to the target in the PDB protein structure database.

[0058] Wherein, the targets include: protein targets and RNA targets.

[0059] Step 1012: Select a small molecule compound library containing structural features of small molecule compounds according to the structural information.

[0060] The structural features include: molecular formula, SMILE expression and atomic relative coordinates of small molecule compounds.

[0061] In this step, the existing conventional drug-like small molecule library can be directly used as the small molecule compound library, or corresponding drug-like small molecule compounds can be selected according to the structural information to construct a small molecule compound library. Here, the process of constructing the small molecule compound library is the process of storing the structural characteristics of the small molecule compounds.

[0062] Step 1013: Distribute the small molecule compound library equally to each computing node to construct a target molecule library corresponding to each computing node.

[0063] In this step, the small molecule compound library is evenly distributed to each computing node, so that each computing node can calculate the docking scores of multiple molecules at the same time, saving computing time by superimposing computing power.

[0064] S102: Initialize the proxy model.

[0065] In specific implementation, the proxy model is initialized in the following way:

[0066] Step 1021: Each computing node randomly collects small molecule compounds from the corresponding target molecule library, and performs molecular docking using docking software to obtain docking scores between the randomly collected small molecule compounds and the target.

[0067] In this step, since there is a communication relationship between the computing nodes, the slave node can send the docking scores of the randomly collected small molecule compounds and the target to the master node through the communication relationship.

[0068] Step 1022: The master node constructs an initial training data set for the proxy model based on the small molecule compounds randomly collected by each computing node and the corresponding docking scores.

[0069] Wherein, the proxy model is a deep graph message passing model.

[0070] In this step, the master node merges the docking scores of the randomly collected small molecule compounds and targets received from the slave nodes and the master node to obtain the randomly collected small molecule compounds and corresponding docking scores of each computing node to construct the initial training data set of the proxy model.

[0071] Step 1023: The master node initializes the proxy model using the structural features of the randomly collected small molecule compounds in the initial training data set as input variables and the docking scores corresponding to the randomly collected small molecule compounds as output variables.

[0072] In this step, the master node initializes the proxy model, obtains the parameters of the proxy model, and synchronizes the parameters of the proxy model to the slave node through the communication relationship.

[0073] S103. Each computing node uses the proxy model to predict the target properties of the small molecule compounds in the corresponding target molecule library to obtain the prediction score and uncertainty of each small molecule compound, and calculate the EI score of each small molecule compound.

[0074] Wherein, the target properties include: affinity, toxicity, hydrophilicity and synthesizability.

[0075] In specific implementation, the EI score of each small molecule compound is calculated in the following way:

[0076] Step 1031: Each computing node obtains a prediction score mean according to the prediction score of each small molecule compound.

[0077] Step 1032: Each computing node calculates the EI score of each small molecule compound based on the predicted score mean and the uncertainty of each small molecule compound using the following formula:

[0078]

[0079] Where, DN={(G1, y1), ..., (G N ,y N )} is the docked small molecule compound and the corresponding docking score, θ is the parameter of the proxy model, y max is the maximum docking score of the docked small molecule compound, p(y|D N , θ) is the distribution of prediction scores of small molecule compounds predicted by the surrogate model.

[0080] S104 , each computing node obtains the small molecule compound to be docked according to the descending order of EI score and the predetermined number of acquisitions, and performs molecular docking using docking software to obtain the docking score of the small molecule compound to be docked.

[0081] The docking score of the small molecule compound to be docked includes: the docking score between the small molecule compound to be docked and the target and the predicted docking score of the small molecule compound to be docked.

[0082] In this step, each computing node determines the small molecule compounds with higher EI scores as the small molecule compounds to be docked according to the predetermined number of collections in descending order of EI scores.

[0083] In specific implementation, the docking score of the small molecule compound to be docked is obtained by the following method:

[0084] Step 1041: Each computing node uses the structural features of the small molecule compound to be docked as an input variable, and performs molecular docking in descending order of uncertainty of the small molecule compound to be docked using the docking software to obtain a docking score between the small molecule compound to be docked and the target.

[0085] Wherein, the docking software is smina docking software.

[0086] In this step, each computing node performs molecular docking on small molecule compounds with higher uncertainty in descending order of uncertainty.

[0087] Step 1042: The computing nodes monitor each other's docking progress. When all the small molecule compounds to be docked on any computing node have completed docking, the docking processes of the other computing nodes will stop at the same time.

[0088] In this step, the purpose of stopping the docking process of other computing nodes at the same time is to prevent the computing resources of all computing nodes that have completed docking from being idle.

[0089] Step 1043: The other computing nodes determine the predicted score of the small molecule compound to be docked that has not been docked when the docking process is stopped as the predicted docking score of the small molecule compound to be docked.

[0090] In this step, the uncertainty of the small molecule compound to be docked that has not completed docking is low, that is, the prediction score of the small molecule compound to be docked obtained by the proxy model is more accurate. Therefore, it can be determined as the docking score predicted for the small molecule compound to be docked. In this way, while avoiding idle computing resources, a sufficient number of docking scores can still be obtained to update the training data set of the proxy model.

[0091] That is, after determining the docking score and the predicted docking score of the small molecule compound to be docked and the target, the slave node can send the docking score of the small molecule compound to be docked to the master node through the communication relationship.

[0092] S105. The master node evaluates the ability of each computing node to screen small molecule compounds based on the docking score of the small molecule compound to be docked with the target of each computing node, so as to determine the number of samples collected by each computing node.

[0093] In the specific implementation, the number of collections for each computing node is determined through the following steps:

[0094] Step 1051: The master node obtains an average docking score of each computing node based on the docking score of the small molecule compound to be docked and the target of each computing node.

[0095] In this step, in order to evaluate the ability of each computing node to screen small molecule compounds, each computing node only uses the docking scores of the small molecule compound to be docked and the target to calculate the mean docking score.

[0096] Step 1052: The master node calculates the collection weight of each computing node according to the average docking score of each computing node using the following formula:

[0097]

[0098] Where, is the collection weight of the kth round of computing node i, is the set of docking scores between the small molecule compound to be docked and the target obtained in the k-1th round of calculation for node i,

[0099] The number of collections assigned to computing node i in the k-1th round.

[0100] Step 1053: The master node determines the number of collections for each computing node using the following formula based on the collection weight of each computing node:

[0101]

[0102] Where N is the set of computing nodes.

[0103] S106: The master node updates the training data set of the proxy model based on the small molecule compounds to be docked and the corresponding docking scores of each computing node to update the proxy model.

[0104] In specific implementation, the proxy model is updated in the following way:

[0105] Step 1061: The master node incorporates the small molecule compounds to be docked and the corresponding docking scores of each computing node into the training data set of the previous round to update the training data set of the proxy model.

[0106] In this step, the master node merges the small molecule compounds to be docked and the corresponding docking scores received from the slave nodes and the master node into the training dataset of the previous round to update the training dataset of the proxy model.

[0107] Step 1062: The master node uses the updated training data set of the proxy model to perform this round of training on the proxy model to update the proxy model.

[0108] In this step, the master node updates the proxy model and obtains new parameters of the proxy model, and synchronizes the new parameters of the proxy model to the slave node through the communication relationship.

[0109] S107: Determine whether the predetermined number of iterations has been reached.

[0110] If the predetermined number of iterations has not been reached, steps S103 to S106 are repeated to perform multiple rounds of model training and molecular docking.

[0111] If the predetermined iteration round is reached, step S108 is executed to extract the optimal value of the docking score between the small molecule compound to be docked and the target of each computing node.

[0112] Example 2

[0113] See also Figure 2 , Figure 2The diagram shows the structure of the small molecule compound virtual screening system based on parallel Bayesian optimization provided in the embodiment of the present application. Figure 2 As shown, the system 200 includes: an agent model 201, a load balancing module 202, a docking software 203, and a corresponding target molecule library 204 arranged on each computing node, an adaptive acquisition module 205 and an agent model update module 206 arranged on the master node, and each computing node works independently and has a communication relationship;

[0114] The proxy model 201 is used to predict target properties of small molecule compounds in the corresponding target molecule library 204 to obtain a prediction score and uncertainty for each small molecule compound; wherein the target properties include affinity, toxicity, hydrophilicity, and synthesizability;

[0115] The load balancing module 202 is used to calculate the EI score of each small molecule compound and obtain the small molecule compounds to be docked according to the predetermined number of collections in descending order of the EI score;

[0116] The docking software 203 is used to perform molecular docking on the small molecule compound to be docked to obtain a docking score between the small molecule compound to be docked and the target; wherein the docking score of the small molecule compound to be docked includes: the docking score between the small molecule compound to be docked and the target and the predicted docking score of the small molecule compound to be docked;

[0117] The adaptive collection module 205 is used to evaluate the ability of each computing node to screen small molecule compounds based on the docking score of the small molecule compound to be docked with the target of each computing node, so as to determine the number of samples collected by each computing node;

[0118] The proxy model updating module 206 is configured to update the training data set of the proxy model 201 based on the small molecule compound to be docked and the corresponding docking score, so as to update the proxy model 201 .

[0119] The above content is only a preferred embodiment of the present invention. For ordinary technicians in this field, many changes can be made in the specific implementation methods and application scope based on the ideas of the present technical content. As long as these changes do not deviate from the concept of the present invention, they all fall within the scope of protection of this patent.

Claims

1. A virtual screening method for small molecule compounds based on parallel Bayesian optimization, characterized in that: The method is applied to a plurality of computing nodes that work independently and have a communication relationship, wherein the computing nodes include: a master node and a plurality of slave nodes, and the method includes: S1. Each computing node uses the proxy model to predict the target properties of the small molecule compounds in the corresponding target molecule library to obtain the prediction score and uncertainty of each small molecule compound, and calculate the EI score of each small molecule compound; wherein the target properties include affinity, toxicity, hydrophilicity and synthesizability; S2. Each computing node obtains the small molecule compound to be docked according to the predetermined number of acquisitions in descending order of EI score, and performs molecular docking using docking software to obtain a docking score of the small molecule compound to be docked; wherein the docking score of the small molecule compound to be docked includes: the docking score of the small molecule compound to be docked and the target and the predicted docking score of the small molecule compound to be docked; S3. The master node evaluates the ability of each computing node to screen small molecule compounds based on the docking score of the small molecule compound to be docked with the target, so as to determine the number of samples collected by each computing node; S4, the master node updates the training data set of the proxy model based on the small molecule compounds to be docked and the corresponding docking scores of each computing node to update the proxy model, and repeats steps S1-S4 to perform multiple rounds of model training and molecular docking until a predetermined iterative round is reached to obtain the optimal value of the docking score between the small molecule compounds to be docked and the target of each computing node; The calculation of the EI score of each small molecule compound includes: Each computing node obtains the mean prediction score based on the prediction score of each small molecule compound; Each computing node calculates the EI score of each small molecule compound based on the predicted score mean and the uncertainty of each small molecule compound using the following formula: Where D N ={(G1, y1), ..., (G N ,y N )} is the docked small molecule compound and the corresponding docking score, θ is the parameter of the proxy model, y max is the maximum docking score of the docked small molecule compound, p(y|D N , θ) is the distribution of prediction scores of small molecule compounds predicted by the surrogate model; The molecular docking is performed using docking software to obtain a docking score of the small molecule compound to be docked, including: Each computational node uses the structural characteristics of the small molecule compound to be docked as an input variable, and uses the docking software to perform molecular docking in descending order of uncertainty of the small molecule compound to obtain the docking score between the small molecule compound to be docked and the target; Each computing node monitors the docking process of each other. When all the small molecule compounds to be docked on any computing node are docked, the docking process of other computing nodes will stop at the same time. Other computing nodes determine the predicted score of the small molecule compound to be docked that has not completed docking when the docking process is stopped as the predicted docking score of the small molecule compound to be docked; Wherein, the docking software is smina docking software; The ability of each computing node to screen small molecule compounds is evaluated based on the docking score of the small molecule compound to be docked with the target of each computing node to determine the number of samples collected by each computing node, including: The master node obtains the average docking score of each computing node based on the docking score of the small molecule compound to be docked and the target of each computing node; The master node calculates the collection weight of each computing node using the following formula based on the average docking score of each computing node: Where, is the collection weight of the kth round of computing node i, is the set of docking scores between the small molecule compound to be docked and the target obtained in the k-1th round of calculation for node i, The number of collections assigned to computing node i in the k-1th round; The master node determines the number of collections for each computing node using the following formula based on the collection weight of each computing node: Where N is the set of computing nodes.

2. The method according to claim 1, wherein The target molecule library corresponding to each computing node is pre-built in the following way: Obtain the name of the target associated with the disease, and search for structural information corresponding to the target in the PDB protein structure database; selecting a small molecule compound library containing structural features of small molecule compounds according to the structural information; Distributing the small molecule compound library equally to each computing node to construct a target molecule library corresponding to each computing node; Among them, the targets include: protein targets and RNA targets; the structural features include: molecular formula, SMILE expression and atomic relative coordinates of small molecule compounds.

3. The method according to claim 1, wherein Before each computing node uses the proxy model to predict target properties of the small molecule compounds in the corresponding target molecule library, the method further includes: Initialize the proxy model as follows: Each computing node randomly collects small molecule compounds from the corresponding target molecule library, and performs molecular docking using the docking software to obtain docking scores between the randomly collected small molecule compounds and the target; The master node constructs an initial training data set for the proxy model based on the small molecule compounds randomly collected by each computing node and the corresponding docking scores; The master node initializes the proxy model using the structural features of the randomly collected small molecule compounds in the initial training data set as input variables and the docking scores corresponding to the randomly collected small molecule compounds as output variables; Wherein, the proxy model is a deep graph message passing model.

4. The method according to claim 1, wherein The updating of the training data set of the proxy model based on the small molecule compounds to be docked and the corresponding docking scores of each computing node to update the proxy model includes: The master node incorporates the small molecule compounds to be docked and the corresponding docking scores of each computing node into the training data set of the previous round to update the training data set of the proxy model; The master node uses the updated training data set of the proxy model to perform this round of training on the proxy model to update the proxy model.

5. A small molecule compound virtual screening system based on parallel Bayesian optimization, characterized in that: The system includes: a proxy model, a load balancing module, a docking software, and a corresponding target molecule library arranged on each computing node; an adaptive acquisition module and a proxy model update module arranged on the master node; each computing node works independently and has a communication relationship; The proxy model is used to predict target properties of small molecule compounds in the corresponding target molecule library to obtain a prediction score and uncertainty for each small molecule compound; wherein the target properties include affinity, toxicity, hydrophilicity and synthesizability; The load balancing module is used to calculate the EI score of each small molecule compound and obtain the small molecule compounds to be docked according to the predetermined number of collections in descending order of the EI score; The docking software is used to perform molecular docking on the small molecule compound to be docked to obtain a docking score between the small molecule compound to be docked and the target; wherein the docking score of the small molecule compound to be docked includes: the docking score between the small molecule compound to be docked and the target and the predicted docking score of the small molecule compound to be docked; The adaptive acquisition module is used to evaluate the ability of each computing node to screen small molecule compounds based on the docking score of the small molecule compound to be docked with the target of each computing node, so as to determine the number of samples to be collected by each computing node; The proxy model updating module is configured to update the training data set of the proxy model based on the small molecule compound to be docked and the corresponding docking score, so as to update the proxy model; The load balancing module is specifically used to calculate the EI score of each small molecule compound in the following manner: Each computing node obtains the mean prediction score based on the prediction score of each small molecule compound; Each computing node calculates the EI score of each small molecule compound based on the predicted score mean and the uncertainty of each small molecule compound using the following formula: Where D N ={(G1,y1),…,(G N ,y N )} is the docked small molecule compound and the corresponding docking score, θ is the parameter of the proxy model, y max is the maximum docking score of the docked small molecule compound, p(y|D N , θ) is the distribution of prediction scores of small molecule compounds predicted by the surrogate model; The docking software is specifically used to obtain the docking score of the small molecule compound to be docked by the following method: Each computational node uses the structural characteristics of the small molecule compound to be docked as an input variable and performs molecular docking in descending order of uncertainty of the small molecule compound to be docked to obtain the docking score between the small molecule compound to be docked and the target; Each computing node monitors the docking process of each other. When all the small molecule compounds to be docked on any computing node are docked, the docking process of other computing nodes will stop at the same time. Other computing nodes determine the predicted score of the small molecule compound to be docked that has not completed docking when the docking process is stopped as the predicted docking score of the small molecule compound to be docked; Wherein, the docking software is smina docking software; The adaptive collection module is specifically used to determine the number of collections of each computing node in the following manner: The master node obtains the average docking score of each computing node based on the docking score of the small molecule compound to be docked and the target of each computing node; The master node calculates the collection weight of each computing node using the following formula based on the average docking score of each computing node: Where, is the collection weight of the kth round of computing node i, is the set of docking scores between the small molecule compound to be docked and the target obtained in the k-1th round of calculation for node i, The number of collections assigned to computing node i in the k-1th round; The master node determines the number of collections for each computing node using the following formula based on the collection weight of each computing node: Where N is the set of computing nodes.

Images