Molecular virtual screening method based on parallel depth map Bayesian optimization framework

Abstract

The invention is suitable for the technical field of virtual screening, and provides a molecular virtual screening method based on a parallel depth map Bayesian optimization framework. According to the method, a depth map Bayesian optimization framework is used for performing a virtual screening task; a parallel screening strategy is designed under a complex condition, that is, a parallel depth map Bayesian optimization framework is used for parallel operation on a plurality of computing nodes. In order to further reduce the time cost, the agent model is trained by using an incremental learning method, and meanwhile, the problem of disastrous forgetting is relieved by using an EMA strategy and an OCS sampling method. Compared with a serial depth map Bayesian optimization framework, the parallel depth map Bayesian optimization framework provided by the invention has the advantages that when the same screening task is carried out, the time cost is reduced by 75.1%, the training time of an agent model is reduced from 13.97 h to 4.49 h, and meanwhile, the screening performance is basically kept stable.

Description

Technical Field

[0001] The present invention belongs to the technical field of virtual screening, and in particular relates to a molecular virtual screening method based on a parallel deep graph Bayesian optimization framework. Background Art

[0002] Virtual Screening (VS) was originally introduced to solve the problems of low efficiency and high cost of compound screening in the drug development process. Traditional drug discovery relies on a large number of experimental screenings, which require testing of thousands of compounds. In contrast, virtual screening uses computer simulation technology to significantly speed up the screening process and reduce costs. Therefore, it plays a vital role in auxiliary screening, cost control and efficiency improvement in drug development, providing strong support for new drug development.

[0003] Incremental learning is a machine learning method that allows the model to continuously learn as it receives new data while retaining what it has learned to avoid forgetting. This method is suitable for scenarios where data is constantly changing, such as recommendation systems, real-time monitoring, and autonomous driving. It can be updated without retraining the entire model, saving resources and improving adaptability. Common methods include preventing the model from overfitting new data through regularization, retaining some old data to maintain the model's memory, or using generative models to simulate the distribution of old data to maintain knowledge consistency.

[0004] Catastrophic forgetting refers to the situation where a machine learning model loses its memory of previous tasks when learning new tasks incrementally or continuously, resulting in the overwriting of learned knowledge. Especially in neural networks, when updating new task data, parameter adjustments may make it impossible for the model to accurately complete previous tasks. This phenomenon is very unfavorable in new scenarios that require long-term learning, because it is difficult for the model to maintain stable performance on multiple tasks.

[0005] In recent years, with the emergence of virtual compound libraries, the number of small molecule compounds has exploded, and screening libraries have become larger and more diverse, which undoubtedly increases the cost of virtual screening. Traditional molecular docking virtual screening methods require calculating the positions and conformations of a large number of molecules one by one. For large-scale virtual screening, the demand for computing resources is extremely high. Especially when the number of molecules is huge or the docking pattern is complex, the screening process becomes extremely time-consuming and inefficient, resulting in a significant increase in the time cost of the entire screening process. In some complex screening tasks, it may take decades to complete using traditional methods, or even the screening task cannot be completed at all. For this reason, the present invention proposes a molecular virtual screening method based on a parallel deep graph Bayesian optimization framework. Summary of the invention

[0006] The object of the present invention is to provide a molecular virtual screening method based on a parallel deep graph Bayesian optimization framework, aiming to solve the problems raised in the above-mentioned background technology.

[0007] The object of the present invention is achieved through the following technical solutions: A molecular virtual screening method based on a parallel deep graph Bayesian optimization framework includes the following steps: Environment setting: Deploy the parallel deep graph Bayesian optimization framework on different GPUs of different servers, and each GPU is regarded as an independent computing node; allocate an agent model to each computing node, and the parameters of each agent model are different; Data allocation: Initially, randomly and evenly divide the total database into several equal parts equal to the number of agent models, each part corresponding to a computing node, and each computing node branch runs independently on its corresponding computing node; Execute the virtual screening task, specifically including: Randomly select a part of the molecules from the database of each computing node for black-box evaluation, and use the molecular true values obtained from the evaluation as the initial training data; use the initial training data to train the agent model of each computing node. After the training is completed, the agent model predicts the affinity of the remaining molecules in the database, and calculates the predicted mean and variance of each data sample through the dropout technology; the acquisition function selects the molecule with the best properties for black-box evaluation according to the predicted mean and variance of the agent model; after each computing node branch completes the black-box evaluation, it uses the data evaluated by other computing nodes together with its own data as a new training data set; in each iteration, the training process of the agent model is as follows: based on the new training data set, combined with a part of the old task data selected from the core set, re-train on the basis of the agent model trained in the previous iteration; during the training process, use the EMA strategy to alleviate the problem of catastrophic forgetting; the replay strategy maintains the core set to store the parameters of the agent model in the previous iteration number, that is, the old task data. When the model learns a new task, use the OCS sampling method to select the most suitable data sample from the old task data and put it into the core set; Repeat the above steps until the preset number of iterations is reached; after reaching the preset number of iterations, screen out the top k molecules with the best performance in the total database according to the true value data obtained from all black-box evaluations.

[0008] Furthermore, the specific formula of the EMA is as follows: ; where represents the parameters saved by the agent model in the current t-th iteration; is a parameter with a value from 0 to 1; is the parameter obtained after training the surrogate model in the t-th iteration; is the parameter saved by the surrogate model in the previous iteration.

[0009] Furthermore, the OCS sampling method selects the most suitable data samples from three aspects and puts them into the core set, specifically as follows: Representativeness: Evaluate its representativeness by calculating the similarity between the gradient of the data sample and the average gradient of the old task dataset. The specific formula is as follows: ; where is the similarity score between the gradient of the data sample and the average gradient of the old task dataset; is the gradient of the data sample ; , indicating that the t-th iteration belongs to the n-th data sample of the current iteration dataset . Each data point consists of its smile formula x and its affinity true value y, respectively represent the smile formula and its affinity true value of the n-th data sample in the t-th iteration; is the average gradient of the old task dataset; is the transpose of the average gradient of the old task dataset; Diversity: Evaluate its diversity by calculating the gradient difference between the data sample and other data samples. The specific formula is as follows: ; where is the diversity score of the data sample ; represents the data samples in the dataset except ; is the number of data samples in the dataset ; p represents the subscript of other data points except the n-th data point; n represents the n-th data point; represents the gradient of the p-th data point in the t-th iteration; represents the transpose of the gradient of the p-th data point in the t-th iteration; Similarity: Evaluate its similarity by calculating the cosine similarity score between the average gradient of the data sample and the data sample randomly selected from the core set. The specific formula is as follows: ; where represents the similarity score of the data sample ; is the set of data samples randomly sampled from the core set C; Denotes a set of data samples randomly sampled from the core set The average gradient; Denotes a set of data samples randomly sampled from the core set The transpose of the average gradient; Finally, the formula for selecting the most suitable data samples from the old task data and putting them into the core set is as follows: ; where u is the index of the obtained data sample.

[0010] Furthermore, the acquisition function adopts the EI acquisition function, and the specific form of the EI acquisition function is as follows: ; where Denotes the current best point; Denotes the expected improvement value that is expected to exceed the current best point at the current point x; Denotes the predicted mean of the surrogate model for the objective function at point x; Denotes the normalized distance of the predicted mean relative to the current optimal value; Denotes the predicted standard deviation at point x; Denotes the cumulative distribution function of the standard normal distribution; Denotes the probability density function of the standard normal distribution.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention uses a deep graph Bayesian optimization framework to perform virtual screening tasks; a parallel screening strategy is designed in complex situations, that is, a parallel deep graph Bayesian optimization framework is used to run in parallel on multiple computing nodes. In order to further reduce the time cost, the present invention uses an incremental learning method to train the surrogate model, and at the same time uses the EMA strategy and the OCS sampling method to alleviate the catastrophic forgetting problem. When performing the same screening task, the parallel deep graph Bayesian optimization framework proposed by the present invention reduces the time cost by 75.1% compared with the serial deep graph Bayesian optimization framework, and the training time of the surrogate model is reduced from 13.97 h to 4.49 h, while the screening performance remains basically stable. Brief Description of the Drawings

[0012] Figure 1 Is a flowchart of a molecular virtual screening method based on a deep graph Bayesian optimization framework.

[0013] Figure 2 Is an overview diagram of a molecular virtual screening method based on a deep graph Bayesian optimization framework.

[0014] Figure 3It is a flowchart of a molecular virtual screening method based on a parallel deep graph Bayesian optimization framework.

[0015] Figure 4 It is a core set sample selection method.

[0016] Figure 5 The top three molecules with the best properties are screened out by the parallel deep graph Bayesian optimization framework proposed in the present invention.

[0017] Figure 6 It is the total running time of the framework screening task.

[0018] Figure 7 It is the running time of the surrogate model training part of the framework.

[0019] Figure 8 It is a line graph of the successful matching quantity of the top 1000 molecules with the best properties screened out by the framework. Specific implementation manners

[0020] For a clearer understanding of the technical features, objectives, and beneficial effects of the present invention, the technical solution of the present invention will be described in detail below, but it should not be construed as a limitation on the implementable scope of the present invention.

[0021] The following describes the specific implementation of the present invention in detail with reference to specific embodiments.

[0022] An embodiment of the present invention provides a molecular virtual screening method based on a deep graph Bayesian optimization framework, and its flowchart is as Figure 1 shown ( Figure 1 focusing on describing the iterative optimization process based on the surrogate model and the data acquisition function), Figure 2 which is an overview diagram of the method ( Figure 2 fully showing the key steps in the machine learning process).

[0023] The deep graph Bayesian optimization framework is used to perform the virtual screening task. The framework is an iterative optimization process, mainly including three parts: The surrogate model, which is used to predict the affinity between molecules and target proteins, approximately evaluate the objective. After initial training, the surrogate model is used to predict the affinity between all molecules and target proteins in the dataset. In the present invention, Chemprop is used as the surrogate model of the deep graph Bayesian optimization framework, which can capture the complex relationships in the molecular graph and provide high-precision property predictions. The mean and variance of each candidate data sample are obtained through dropout, that is, the prediction and uncertainty of the model.

[0024] The acquisition function collects a set of candidate data samples with the best properties through the prediction information (i.e., mean and variance) of the surrogate model. The present invention uses the Expected Improvement (EI) acquisition function, which can effectively balance the relationship between exploration and exploitation. The Expected Improvement function characterizes the expectation of improving the current optimal solution at the candidate point.

[0025] Black-box evaluation: Use the molecular docking tool smina for black-box evaluation. Smina is an improved and extended version of AutoDock Vina, providing faster docking speed and more flexible functions.

[0026] The steps for performing virtual screening tasks are as follows: Randomly select 1% of the molecules from the database, use the molecular docking tool smina for black-box evaluation to obtain the true values (true affinity values) of the molecules, and use the obtained true value data as the initial training data to train the surrogate model (Chemprop). After training, use the surrogate model to predict the affinity of the entire dataset (except for the molecules selected for initial training before), and obtain the prediction mean and variance of each data sample through the dropout technique, that is, the uncertainty information of each data sample.

[0027] Use the EI acquisition function to collect a set of candidate data samples with the best properties (such as the top 1% molecules in terms of score) according to the prediction mean and variance of the surrogate model; the formula of the EI acquisition function can be defined as: ; where represents the current candidate point, represents the current best point, represents the predicted mean of the surrogate model for the objective function at point x; represents the mathematical expectation; represents the expected improvement value that is expected to exceed the current best point at the current point x. The specific form is: ; where represents the current best point; represents the normalized distance of the predicted mean relative to the current optimal value; represents the predicted standard deviation at point x; represents the cumulative distribution function of the standard normal distribution; represents the probability density function of the standard normal distribution.

[0028] The collected candidate data sample sets are evaluated using the molecular docking tool smina in a black-box manner to obtain the true values of each data sample. The newly evaluated true-value data samples are added to the training set, the training set is updated, and the surrogate model is retrained. Then, the next iteration is entered, and the above steps are repeated. As the number of iterations increases, the performance of the surrogate model also becomes stronger. Finally, after reaching the preset number of iterations, the top k optimal molecules in the entire database are selected based on the true-value data obtained from all black-box evaluations.

[0029] An embodiment of the present invention provides a molecular virtual screening method based on a parallel deep graph Bayesian optimization framework, aiming to solve the problem of efficient screening in complex scenarios, especially when the compound library is large, a single device cannot complete the screening task, or the time cost is too high. The present invention parallelizes and extends the deep graph Bayesian optimization framework, enabling it to be flexibly deployed on multiple GPUs of multiple servers to perform the screening task in parallel, thereby significantly reducing the time cost while maintaining the screening performance.

[0030] During the parallelization and extension process, the present invention solves two key problems: First, to avoid the redundant evaluation problem, that is, multiple computing nodes evaluating the same candidate solution wastes resources, the present invention adopts a partitioning strategy. Specifically, the total database is randomly and evenly divided into several small databases according to the number of computing nodes, ensuring that the data samples of each computing node are not repeated, avoiding multiple computing nodes evaluating the same candidate solution, and thus avoiding redundant evaluation. Second, for the problem that the evaluation durations of different samples may be different during the screening task execution of the parallel deep graph Bayesian optimization framework, the existing BO framework needs to wait for all samples in the previous round to be evaluated before starting the sample selection and evaluation in the next round, which limits the parallel efficiency. For this reason, the present invention adopts a parallel scheme in which each computing node is assigned a surrogate model. Each surrogate model branch runs independently on its own computing node, and then shares the training data after black-box evaluation. Without waiting for all samples to be evaluated before starting the next round, each computing node can directly use the data that has been black-box evaluated by other computing nodes to continue running, thereby avoiding waiting time and improving the parallel efficiency.

[0031] The flowchart of the molecular virtual screening method based on the parallel deep graph Bayesian optimization framework is as Figure 3 shown and is described as follows: First, deploy the parallel deep graph Bayesian optimization framework on different GPUs of different servers. Each GPU is regarded as a computing node to enable the model to run simultaneously on multiple computing nodes, effectively reducing the time cost. In scenarios with a large amount of data, each computing node is assigned an agent model, and the parameters of each agent model are different. Initially, the total database is randomly and evenly divided into several equal parts equal to the number of agent models (i.e., the number of node branches). Each computing node branch runs independently on each computing node (which can be different GPUs of the same server or different servers). First, randomly select a part of the molecules from the database for black-box evaluation, and the obtained molecular true values are used as the initial training data. Subsequently, the agent model is trained using the initial training data. After training, the agent model is used to predict the affinity of the molecules in the database (except for the part of the molecules selected for initial training before), and the predicted mean and variance of each data sample are obtained through the dropout technique. The acquisition function selects the molecule with the best properties for black-box evaluation based on the predicted mean and variance of the agent model. After each computing node branch completes the black-box evaluation, the data that has been evaluated by other computing nodes and its own data are used as new training data and added to the initial training set. Then, the agent model is retrained using the new training set, entering the next iteration, and repeating the above steps until the preset number of iterations is reached. After reaching the preset number of iterations, according to the true value data obtained from all black-box evaluations, the top k molecules with the best performance in the total database are selected.

[0032] To significantly reduce the time cost of training the agent model, we introduce an incremental learning method. The agent model is trained using the incremental learning method, and each time only new data is used to train on the basis of the agent model trained in the previous iteration. Although this method reduces the time cost, the catastrophic forgetting problem has become a major challenge, that is, the model will forget the knowledge of old tasks when learning new tasks. To address this issue, we design two strategies, EMA (Exponential Moving Average) and OCS (Optimized Core Set Sampling), to effectively alleviate catastrophic forgetting.

[0033] First, the EMA strategy is introduced. The specific formula of EMA is as follows: ; where represents the parameters saved by the agent model in the current t-th iteration; is the parameters saved by the agent model in the previous iteration; are the parameters obtained after training the agent model in the t-th iteration; is a parameter with a value from 0 to 1, which controls the ratio of the current model parameters to the model parameters in the previous iteration.

[0034] The model parameters saved this time consist of two parts: one is the parameters of the surrogate model obtained from the current training, representing new task knowledge; the other is the parameters of the surrogate model in the previous iteration, representing old task knowledge. In this way, the surrogate model in each iteration can take both new and old knowledge into account, thus alleviating the forgetting rate of old knowledge.

[0035] Based on the EMA strategy, we introduce the replay strategy to further alleviate the catastrophic forgetting problem. The replay strategy is one of the commonly used methods to alleviate the forgetting problem. Similar to the human learning process, when learning new knowledge, it is necessary to review the old knowledge to alleviate forgetting. The same is true for machine learning models. The replay strategy maintains a core set to store old task data. When the model learns a new task, a part of the old task data is put in for further training to alleviate the catastrophic forgetting problem. Most previous works generally use reservoir sampling to randomly select data samples into the core set. However, the potential of each data sample and its impact on the model performance are different, and the most suitable data needs to be selected to be put into the core set. Therefore, which data to select into the core set is a key issue. The present invention improves the OCS sampling method and selects the most suitable data samples from three aspects to be put into the core set to maximize the model performance. Figure 4 The schematic diagram of the core set sample selection method is shown.

[0036] 1) Representativeness: The selected data samples need to be representative of the old task dataset, so that the model can review as much old knowledge as possible from a small part of the core set data. Therefore, the gradients of the selected data samples need to be as close as possible to the average gradient of the old task dataset.

[0037] Therefore, the representativeness is evaluated by calculating the similarity between the gradient of the data sample and the average gradient of the old task dataset. The specific formula is as follows: ; where is the gradient of the data sample ; , indicating the nth data sample in the dataset of the current iteration at the tth iteration , and each data point consists of its smile formula x and its affinity true value y. respectively represent the smile formula and its affinity true value of the nth data sample at the tth iteration; is the average gradient of the old task dataset; is the transpose of the average gradient of the old task dataset; is the similarity score between the gradient of the data sample and the average gradient of the old task dataset; the cosine similarity score is used, and the range is 0 to 2, which represents the similarity degree between the gradient of the current data sample and the average gradient of the old task dataset. The higher the score, the more similar the gradient of the current data sample is to the average gradient of the old task dataset. Then it can be considered that the current data sample has a good representativeness for the samples of the old task dataset. The data samples sampled through the formula can better represent the overall data characteristics, thereby improving the accuracy and efficiency of the model.

[0038] 2) Diversity: The selected data samples need to be diverse, so that the model can review as comprehensive old knowledge as possible from a small part of the core set data. If the gradient of a data sample is more different from the gradients of other data samples, then its diversity in the gradient space is higher.

[0039] Therefore, the diversity of a certain data sample is evaluated by calculating the gradient difference between it and other data samples. The specific formula is as follows: ; where represents the data samples in the dataset except ; is the number of data samples in the dataset ; p represents the subscripts of other data points except the nth data point; n represents the nth data point; represents the gradient of the pth data point at the tth iteration; represents the transpose of the gradient of the pth data point at the tth iteration. The formula measures the diversity between a certain data sample and other data samples in , is the diversity score of the data sample

[0040] ; the score range is 0 to 2, and a higher score indicates higher diversity of the data point. When the differences between the training data are large, the features and labels of the data samples have a large variation range. In this case, selecting samples with high diversity can reduce redundancy, ensure that the selected sample set has good representativeness and generality, and thus improve the generalization ability of the model.

[0041] 3) Similarity: The selected data samples need to have a certain similarity with the past tasks, because the goal of incremental learning is to train on the data samples representing the current task and prevent the performance degradation of the previous tasks. ; Among them is a set of data sample collections randomly sampled from the core set C; that is, the data samples of past tasks, representing old knowledge. Represents the data sample Similarity score of; Represents a set of data samples randomly sampled from the core set Average gradient of; Represents a set of data samples randomly sampled from the core set Transpose of the average gradient. The formula measures a certain data sample That is, the cosine similarity score between the current task and the data samples randomly selected from the core set C That is, the average gradient of past tasks. Selecting data samples with high scores can ensure the performance of the current task while preventing the degradation of the performance of previous tasks and improving the anti-forgetting performance of the model.

[0042] Combining these three aspects, the formula for selecting data samples from old task data and adding them to the core set is as follows: ; where u is the index of the obtained data sample.

[0043] Select the most beneficial data samples and add them to the core set through the above formula. These samples not only have representativeness and diversity, but also maintain the understanding of past tasks, thus effectively alleviating the catastrophic forgetting problem.

[0044] Example 1: Performance verification of the parallel deep graph Bayesian optimization framework proposed by the present invention; Figure 5 For screening out the top three molecules with the best properties by the parallel deep graph Bayesian optimization framework proposed by the present invention, we can see that the parallel deep graph Bayesian optimization framework proposed by the present invention successfully screened out the top three molecules with the best properties in a dataset of 2 million molecules, and their true values are -112.87 kcal / mol, -89.92 kcal / mol, and -87.25 kcal / mol respectively.

[0045] Figure 6 Shows the total running time of the framework screening task, Figure 7Shows the running time of the framework proxy model training part. The five data in the figure are the BO-Ser serial framework (serial deep graph Bayesian optimization framework), BO-Par parallel framework using 4 computing nodes, BO-Par-incre parallel framework using incremental learning to train the proxy model but without an anti-forgetting strategy, BO-Par-ema-ran parallel framework using incremental learning and EMA anti-forgetting strategy but with randomly selected replay samples, and BO-Par-ema-ocs parallel framework using incremental learning and EMA and OCS sample selection methods (the parallel deep graph Bayesian optimization framework proposed in the present invention). The total running time and the time used for each part in the 2-million molecular library screening task are: the time for loading training data, the training time, the gradient calculation time when using the sample selection method, the time for loading prediction data, the prediction time, the acquisition time, and the evaluation time. For a 2-million molecular compound library, when using the BO-Ser serial framework for the screening task, setting the number of iterations to 5 and the total number of black-box evaluated molecules to 6% (i.e., 120,000), the total time required is 238.6 h, among which the proxy model training time is 13.97 h, the prediction time is 50.3 h, and the black-box evaluation time is 166.67 h; when using the BO-Par parallel framework, the total time required is 68.9 h, among which the proxy model training time is 14.09 h, the prediction time is 11.2 h, and the black-box evaluation time is 41.67 h. It can be seen that using parallel screening greatly reduces the proxy model prediction time and the black-box evaluation time, and the total time required to complete the screening task has decreased by 71.12%. In addition, compared with the BO-Ser serial framework, the proxy model training time of the BO-Par-ema-ocs parallel framework is reduced from 13.97 h to 4.49 h. At this time, the total time required for the screening task is 59.4 h, reducing the time cost by 75.1%. It can be seen that the parallel deep graph Bayesian optimization framework proposed in the present invention can greatly reduce the screening time cost and improve the screening efficiency.

[0046] For a 2-million molecular compound library, when using the deep graph Bayesian optimization framework for the screening task, setting the number of iterations to 5, only docking 6% of the molecules in the whole library, i.e., 120,000 molecules, can screen out more than 95% of the top 1000 most promising candidate molecules. Figure 8 Shows the line chart of the number of successfully matched top 1000 molecules with the best properties screened by the framework. The specific numbers are shown in Table 1: Table 1 Specific numbers of successfully matched top 1000 molecules with the best properties

[0047] The results show that the optimal molecule ratio screened by the BO-Ser serial framework is 96.1%; the optimal molecule ratio screened by the BO-Par parallel framework is 95.8%; there are significant losses in the screening performance of the BO-Par-incre parallel framework and the BO-Par-ema-ran parallel framework, and the optimal molecule ratios screened are 81.4% and 89.7% respectively; the optimal molecule ratio screened by the BO-Par-ema-ocs parallel framework is 93.7%. It can be seen that the parallel deep graph Bayesian optimization framework proposed in the present invention can maintain the screening performance while greatly reducing the screening time cost.

[0048] The above are only the preferred embodiments of the present invention. It should be noted that for those skilled in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent.

Claims

1. A molecular virtual screening method based on a parallel deep graph Bayesian optimization framework, characterized in that: The following steps are involved: Environment setup: Deploy the parallel deep graph Bayesian optimization framework on different GPUs of different servers, and treat each GPU as an independent computing node. Assign a proxy model to each computing node, and the parameters of each proxy model are different. Data distribution: Initially, the total database is randomly and evenly divided into several equal parts equal to the number of proxy models. Each part corresponds to a computing node, and each computing node branch runs independently on its corresponding computing node. Perform virtual screening tasks, including: A portion of molecules is randomly selected from the database of each computing node for black-box evaluation, and the true value of the molecules obtained by evaluation is used as the initial training data; the initial training data is used to train the proxy model of each computing node. After training, the proxy model predicts the affinity of the remaining molecules in the database, and calculates the predicted mean and variance of each data sample through the dropout technology; the acquisition function collects the molecules with the best properties for black-box evaluation based on the predicted mean and variance of the proxy model; after completing the black-box evaluation, each computing node branch uses the data evaluated by other computing nodes together with its own data as a new training data set; in each iteration, the training process of the proxy model is: based on the new training data set, combined with a part of the old task data selected from the core set, retraining is performed on the basis of the proxy model trained in the previous iteration; during the training process, the EMA strategy is used to alleviate the catastrophic forgetting problem; the replay strategy maintains the core set to store the parameters of the proxy model of the previous iteration, that is, the old task data. When the model learns a new task, the OCS sampling method is used to select the most suitable data samples from the old task data and put them into the core set; Repeat the above steps until the preset number of iterations is reached; after the preset number of iterations is reached, the top k best molecules in the total database are screened out based on the true value data obtained from all black box evaluations.

2. The molecular virtual screening method based on the parallel deep graph Bayesian optimization framework according to claim 1 is characterized in that: The specific formula of the EMA is as follows: ； in Indicates the parameters saved by the proxy model at the current t-th iteration; is a parameter with a value between 0 and 1; The parameters obtained after training the proxy model for the tth iteration; The parameters saved for the previous iteration of the proxy model.

3. The molecular virtual screening method based on the parallel deep graph Bayesian optimization framework according to claim 1 is characterized in that: The OCS sampling method selects the most appropriate data samples into the core set from three aspects, as follows: Representativeness: The representativeness of a data sample is evaluated by calculating the similarity between the gradient of the data sample and the average gradient of the old task dataset. The specific formula is as follows: ； in is the similarity score between the data sample gradient and the average gradient of the old task dataset; For data samples The gradient of , indicating that the tth iteration belongs to the current iteration data set The nth data sample of , each data point consists of its smile type x and its true affinity value y, They represent the smile formula and its true affinity value of the nth data sample in the tth iteration respectively; is the average gradient of the old task dataset; is the transpose of the average gradient of the old task dataset; Diversity: The diversity of a data sample is evaluated by calculating the gradient difference between the data sample and other data samples. The specific formula is as follows: ； in For data samples Diversity score; Indicates that in the data set, except Data samples other than It is a dataset The number of data samples in ; p Indicates the subscripts of data points other than the nth data point; n represents the nth data point; represents the gradient of the p-th data point at the t-th iteration; represents the transpose of the gradient of the p-th data point at the t-th iteration; Similarity: The similarity is evaluated by calculating the cosine similarity score of the average gradient of the data sample and the randomly selected data sample in the core set. The specific formula is as follows: ； in Represents data samples Similarity score of is a set of data samples randomly sampled from the core set C; Represents a set of data samples randomly sampled from the core set The average gradient of Represents a set of data samples randomly sampled from the core set Transpose of the mean gradient; Finally, the formula for selecting the most appropriate data samples from the old task data and putting them into the core set is as follows: ； where u is the index of the obtained data sample.

4. The molecular virtual screening method based on the parallel deep graph Bayesian optimization framework according to claim 1 is characterized in that: The acquisition function adopts the EI acquisition function, and the specific form of the EI acquisition function is as follows: ； in Indicates the current best point; Indicates that at the current point x, it is expected to exceed the current optimal point Expected improvement value; Represents the predicted mean of the target function at point x by the surrogate model; Represents the normalized distance of the predicted mean relative to the current optimal value; represents the predicted standard deviation at point x; Represents the cumulative distribution function of the standard normal distribution; Represents the probability density function of the standard normal distribution.

Images