Compound safety prediction method and device and computing equipment cluster

Abstract

The invention discloses a compound safety prediction method and device and a computing device cluster, the method is applied to a cloud service center, the cloud service center comprises a plurality of cloud application instances, and the method comprises the steps that the cloud application instances predict compound safety according to a plurality of compounds, a plurality of targets and a plurality of off-target prediction models; determining a first probability between each compound and each target spot; inputting the first probability into an adverse drug reaction (ADR) prediction model to obtain a second probability between each compound and each ADR; according to the number of the off-target targets, the second probability and the risk score of each ADR, safety scores of the multiple compounds are determined, and the safety of the compounds is predicted. In the process, the off-target prediction result is used as the feature to be input into ADR prediction, combination of off-target prediction and ADR prediction is achieved, the action mechanism of ADR is disclosed from the target point level, the safety level of the compound is quantified, and a basis is provided for safety evaluation of the compound in the drug development stage.

Description

Technical Field

[0001] The present application relates to the field of drug research and development, and in particular to a compound safety prediction method, device and computing equipment cluster. Background Art

[0002] Assessing the safety of drugs is a key step in the drug development process. One of the main causes of drug loss is off-target toxicity, which refers to the phenomenon that drugs interact with non-therapeutic targets and produce undesirable toxic effects. Therefore, it is very important to predict whether the compounds in the drug bind to specific off-target targets. In addition to considering the interaction between the compounds in the drug and the off-target targets, it is also necessary to consider whether the compounds in the drug cause adverse drug reactions (ADR) to comprehensively evaluate the safety of the drug. The current safety prediction of compounds consists of two separate parts, a separate off-target prediction and a separate ADR prediction. Separate off-target predictions can use off-target prediction models based on deep learning methods, and separate ADR predictions make extensive use of machine learning methods.

[0003] In current safety prediction methods, off-target prediction and ADR prediction are separated. However, since there is a causal relationship between drug off-target and adverse drug reactions caused by drugs in the body, and ADR is usually caused by the binding of compounds in the drug to specific targets, the existing safety prediction methods for compounds are not accurate enough and lack the explainability of off-target and ADR. Summary of the invention

[0004] The present application provides a compound safety prediction method, apparatus and computing device cluster, which realizes the combination of compound off-target prediction and ADR prediction by inputting the off-target prediction results of the compound as features into the adverse drug response (ADR) prediction, and determines the safety scoring standard for quantifying the safety of the compound based on the above prediction results. It can achieve accurate prediction of compound off-target targets and ADR, reveal the mechanism of action of ADR from the target level, and thus help determine the safety of the compound in the early stage of the drug.

[0005] In a first aspect, the present application provides a compound safety prediction system, which includes: user equipment, a network, and a cloud service center.

[0006] In a specific implementation, the user equipment may be a mobile phone, a tablet computer (pad), a desktop computer, or other equipment with data receiving and sending functions.

[0007] In a specific implementation, the cloud service center includes a cloud platform and a hardware resource pool, wherein the hardware resource pool includes multiple cloud servers and multiple cloud application instances, wherein each cloud server includes multiple cloud application instances. In the above cloud service center, each cloud application instance can be one of a bare metal server (BMS), a container, and a virtual machine (VM).

[0008] In a specific implementation, the cloud service center and the user device may be connected via a network, which may be a wide area network, a local area network, a point-to-point connection, or any combination of the above methods, which is not specifically limited here.

[0009] In a specific implementation, the cloud application instance is used to obtain compounds that need to be predicted for safety sent by a user device, or to obtain character strings that describe the atomic arrangement and bond structure corresponding to each compound. The cloud application instance is used to determine the first probability between each of the multiple compounds and each of the multiple targets based on the received multiple compounds, multiple targets, and multiple off-target prediction models; the cloud application instance is used to input the first probability into the adverse drug reaction ADR prediction model to obtain the second probability between each compound and each ADR; the cloud application instance is used to determine the safety scores of the multiple compounds based on the number of off-target targets, the second probability, and the risk score of each ADR obtained, so as to predict the safety of each of the multiple compounds, wherein the off-target target is the target corresponding to the first probability greater than the probability threshold.

[0010] In a possible implementation, multiple off-target prediction models are obtained by training a multi-task model based on a neural network in a cloud application instance, and have a one-to-one correspondence with multiple types to which multiple targets belong, wherein each of the multiple types includes one or more of the multiple targets.

[0011] In a possible embodiment, the adverse drug reaction ADR prediction model is obtained based on a multi-layer perceptron or a multi-label model training according to the drug-ADR interaction data, wherein the drug-ADR interaction data includes a plurality of first counts corresponding to the case where the compound is associated with the ADR, and a plurality of second counts corresponding to the case where the compound is not associated with the ADR.

[0012] In one possible implementation, the specific process of the cloud application instance determining the safety score of a compound is as follows: the cloud application instance is used to sort the second probabilities from large to small, and determine the ranking corresponding to each second probability; the cloud application instance is used to calculate the ADR score corresponding to each compound, wherein an ADR score is determined based on the risk score of the ADR corresponding to one or more second probabilities corresponding to a compound and the ranking corresponding to one or more second probabilities; the cloud application instance is used to determine the safety scores of multiple compounds based on the ADR score corresponding to each compound and the number of off-target targets, wherein when the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe, and when the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe.

[0013] A cloud application example in a compound safety prediction system provided in the first aspect includes a compound safety prediction application device provided in the present application, and the compound safety prediction application device includes: a receiving unit, used to receive multiple compounds; an off-target prediction unit, used to determine a first probability between each compound in the multiple compounds and each target in the multiple targets based on multiple compounds, multiple targets and multiple off-target prediction models; an adverse drug reaction ADR prediction unit, used to input the first probability into the ADR prediction model to obtain a second probability between each compound and each ADR; a safety prediction unit, used to determine the safety scores of the multiple compounds based on the number of off-target targets, the second probability and the risk score of each ADR obtained by the receiving unit, so as to predict the safety of the multiple compounds, wherein the off-target target is a target corresponding to the first probability greater than the probability threshold.

[0014] In a possible implementation, the safety prediction unit is specifically used to sort the second probabilities from large to small to determine the ranking corresponding to each second probability; it is also used to calculate the ADR score corresponding to each compound, wherein an ADR score is determined based on the risk score of the ADR corresponding to one or more second probabilities corresponding to a compound and the ranking corresponding to one or more second probabilities; it is also used to determine the safety scores of multiple compounds based on the ADR score corresponding to each compound and the number of off-target targets, wherein when the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe, and when the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe.

[0015] In the second aspect, the present application provides a compound safety prediction method, which is applied to a cloud service center, and the cloud service center includes multiple cloud application instances. The method includes: the cloud application instance determines the first probability between each compound in the multiple compounds and each target in the multiple targets according to the received multiple compounds, multiple targets and multiple off-target prediction models; the cloud application instance inputs the first probability into the adverse drug reaction ADR prediction model to obtain the second probability between each compound and each ADR; the cloud application instance determines the safety scores of the multiple compounds according to the number of off-target targets, the second probability and the risk score of each ADR obtained, so as to predict the safety of each of the multiple compounds, wherein the off-target target is the target corresponding to the first probability greater than the probability threshold.

[0016] In summary, the compound safety prediction method provided by the present application realizes the combination of off-target prediction and ADR prediction by inputting the results of the off-target prediction model as features into the ADR prediction model, and determines the compound safety scoring standard according to the above prediction results to predict the safety of the compound. It can achieve accurate prediction of off-target targets and ADR, and can reveal the mechanism of action of ADR from the target level, so as to determine the safety of the compound in the early stage of drug development and improve efficiency.

[0017] In a possible embodiment, multiple off-target prediction models are obtained by training a multi-task model based on a neural network, and there is a one-to-one correspondence with multiple types to which multiple targets belong, wherein each of the multiple types includes one or more targets among the multiple targets. In the above process, each different target type corresponds to a different multi-task model. In the multi-task framework, except for the parameters of the last layer of neural network, which are task-specific, the previous neural network parameters are shared between tasks, which can improve the generalization performance and efficiency of the model. Multiple off-target prediction models can predict the binding probability between each compound and each target, that is, the first probability.

[0018] In a possible embodiment, the adverse drug reaction ADR prediction model is obtained based on the interaction data between the drug and ADR based on a multi-layer perceptron, or a multi-label model training, wherein the interaction data between the drug and ADR includes a first count corresponding to the case where the compound is associated with the ADR, and a second count corresponding to the case where the compound is not associated with the ADR. In the above process, the adverse drug reaction ADR prediction model uses the first probability of binding between each compound and each target as an input feature to obtain the second probability between each compound and each adverse drug reaction mentioned in this application, thereby combining off-target prediction and ADR prediction, thereby improving the accuracy of compound safety prediction and enhancing the interpretability of off-target and ADR.

[0019] In a possible implementation, the specific process of a cloud application instance determining the safety scores of multiple compounds based on the number of off-target targets, the second probability, and the risk score of each ADR obtained is as follows. The cloud application instance sorts the second probabilities from large to small to determine the ranking corresponding to each second probability; the cloud application instance calculates the ADR score corresponding to each compound, wherein an ADR score is determined based on the risk score of the ADR corresponding to one or more second probabilities corresponding to a compound and the ranking corresponding to one or more second probabilities; the cloud application instance determines the safety scores of multiple compounds based on the ADR score corresponding to each compound and the number of off-target targets, wherein when the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe, and when the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe. In the above process, the risk score of each ADR can be directly obtained from the database, and the specific calculation formula of the ADR score is as follows:

[0020]

[0021] Among them, ADR i Severity score is the risk score of the ADR corresponding to the ith position after the second probability ranking. The specific calculation formula of the safety score of the compound is as follows:

[0022] Safety score=ADR score+log N (targetnumber+k)

[0023] Where N is 2 or 10, k is usually a constant of 1, and target number is the number of off-target targets corresponding to the compound. The above calculation process quantifies the safety level of the compound, provides a basis for safety assessment in the drug development process, enhances the accuracy of compound safety score calculation, and improves the efficiency of determining compound safety in the early stages of drug development.

[0024] In a possible implementation, the process of compound safety prediction further includes: the cloud application instance sends prediction data to a user device for display in a result display page, wherein the prediction data includes a first probability, a second probability, and one or more data of safety scores and safety of each of the multiple compounds. The cloud application instance includes one of a bare metal server BMS, a container, and a virtual machine.

[0025] In the third aspect, the present application provides a computing device cluster, which is applied to a compound safety prediction method provided in the first aspect, and the computing device cluster includes at least one computing device, each computing device includes a processor, a memory, a communication interface and a bus. Among them, the processor, the memory, and the communication interface can be interconnected through an internal bus or can communicate through other means such as wireless transmission. The memory can store computer instructions, and the processor is used to execute the method provided in the second aspect. The communication interface is used to receive multiple compounds sent by a user device and receive data such as the risk score of the ADR obtained from the database.

[0026] In a fourth aspect, the present application provides a computer program product comprising instructions, which, when executed by a storage device, causes the storage device to execute the method provided in the second aspect.

[0027] In a fifth aspect, the present application provides a computer-readable storage medium, which includes computer program instructions. When the computer program instructions are executed by a storage device, the storage device executes the method provided in the second aspect.

[0028] Based on the implementations provided in the above aspects, this application can also be further combined to provide more implementations. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments.

[0030] Figure 1 It is a schematic diagram of the structure of a compound safety prediction system provided in an embodiment of the present application;

[0031] Figure 2 is a schematic diagram of the structure of another compound safety prediction system provided in an embodiment of the present application;

[0032] Figure 3 This is a flow chart of a compound safety prediction method provided in an embodiment of the present application;

[0033] Figure 4 It is a schematic diagram of a process of feature extraction and generation of an off-target prediction model provided in an embodiment of the present application;

[0034] Figure 5 This is a schematic diagram of a process for predicting adverse drug reactions provided in an embodiment of the present application;

[0035] Figure 6 is a schematic diagram of the association between a target and an ADR provided in an embodiment of the present application;

[0036] Figure 7This is a schematic diagram of a process for compound safety evaluation provided in an embodiment of the present application;

[0037] Figure 8 is a schematic diagram of the structure of a computing device provided in an embodiment of the present application;

[0038] Fig. 9 is a schematic diagram of the structure of a computing device cluster provided in an embodiment of the present application;

[0039] Fig.10 It is a structural diagram of one or more computing devices connected via a network provided by an embodiment of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present invention will be described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0041] First, the application scenarios involved in this application are explained.

[0042] At present, evaluating the safety of drugs is a key step in the drug development process. One of the main reasons for drug loss is off-target toxicity, which refers to the phenomenon that drugs interact with non-therapeutic targets and produce undesirable toxic effects. Therefore, it is very important to predict whether the compounds in the drug bind to specific off-target targets. At the same time, it is also necessary to consider whether the compounds in the drug cause adverse drug reactions (ADR) to comprehensively evaluate the safety of the drug. Early compound safety prediction methods include in vitro methods such as organ chips, or in vivo methods such as animal models, but these methods are inefficient and cannot accurately predict the reaction of compounds in the human body. With the application of computational methods in the field of drug development, the current safety prediction of compounds consists of two separate parts, separate off-target prediction and separate ADR prediction. Among them, off-target prediction is to predict whether the compound binds to a specific off-target target, which is essentially a target prediction of the compound. The more off-target targets the compound binds to, the less safe the compound is. At present, the off-target prediction performance based on neural networks is better. ADR prediction predicts the relationship between compounds and ADRs by using different machine learning classifiers or building knowledge graphs. The more ADRs a compound corresponds to and the higher the risk score corresponding to the ADR, the less safe the compound is.

[0043] In the current methods for predicting compound safety, off-target prediction and ADR prediction are separated. However, since there is a causal relationship between drug off-target and adverse drug reaction (ADR) caused by the drug in the body, and ADR is usually caused by the binding of the compound in the drug to a specific target, the existing compound safety prediction methods are not accurate enough and lack the explainability of off-target and ADR.

[0044] In order to solve the above problems, the present application provides a compound safety prediction system.

[0045] like Figure 1 As shown, Figure 1 It is a structural diagram of a compound safety prediction system provided in an embodiment of the present application, wherein the compound safety prediction system includes: a user device 110, a network 120 and a cloud service center 130.

[0046] In a specific implementation, the user device 110 may be a mobile phone, a tablet computer (pad), a desktop computer, etc., a device with data receiving and sending functions. The type and quantity of the user device are not specifically limited here.

[0047] In a specific implementation, the cloud service center 130 includes a cloud platform 131 and a hardware resource pool 132. The hardware resource pool 132 includes a cloud server 1, a cloud server 2, a cloud application instance 3, and a cloud application instance 4. Among them, cloud server 1 includes cloud application instance 11 and cloud application instance 12, and cloud server 2 includes cloud application instance 21 and cloud application instance 22. It should be understood that cloud server 1 and cloud server 2 may include more cloud application instances, and the hardware resource pool 132 may also include more cloud servers and cloud application instances. In the above-mentioned cloud service center, each cloud application instance may be one of a bare metal server (BMS), a container, and a virtual machine (VM), which is not specifically limited here.

[0048] In a specific implementation, the cloud service center 130 and the user device 110 can be connected via a network 120. The network 120 can be a wide area network, a local area network, a point-to-point connection, or any combination of the above methods, which is not specifically limited here.

[0049] In a possible embodiment, before the cloud application instance performs safety prediction of the compound, it is necessary to collect data and pre-process the data. The cloud application instance is used to establish an off-target panel. In the embodiment of the present application, the off-target panel includes 91 targets, including 50 G protein-coupled receptor protein-related targets, 16 ion channel-related targets, 9 enzyme-related receptors, 6 kinase-related targets, 4 nuclear receptor-related targets, 4 transporter-related targets and 2 other categories of targets. The off-target panel may also include more or fewer numbers, more or fewer types of targets, and the present application does not specifically limit this.

[0050] Since off-target prediction is essentially the target prediction of compounds, the cloud application instance collects off-target prediction model training data from databases such as ChEMBL and PubChem according to the gene names of the targets included in the off-target panel; afterwards, the cloud application instance also collects drug-ADR interaction data from the SIDER database for the training of the ADR prediction model; and collects compounds related to toxic and side effects with toxic and non-toxic labels so that the ADR prediction model can distinguish between toxic and non-toxic compounds; and randomly collects some compounds and some compounds with Anatomical Therapeutic Chemical (ATC) codes from the ZINC database for the verification of the ADR prediction model. The cloud application instance performs data cleaning, negative sample construction, and data partitioning on the above-collected data for subsequent off-target prediction model training and ADR prediction model training.

[0051] In one possible implementation, a cloud application instance is used to determine a first probability between each of a plurality of compounds and each of a plurality of targets based on a plurality of compounds, a plurality of targets, and a plurality of off-target prediction models that require safety testing and are received from a user device; the cloud application instance is used to input the first probability into an adverse drug reaction ADR prediction model to obtain a second probability between each compound and each ADR; the cloud application instance is used to determine safety scores of a plurality of compounds based on the number of off-target targets, the second probability, and the risk score of each ADR obtained, so as to predict the safety of each of the plurality of compounds, wherein an off-target target is a target corresponding to a first probability that is greater than a probability threshold.

[0052] In a possible implementation, multiple off-target prediction models are obtained by training a multi-task model based on a neural network in a cloud application instance, and have a one-to-one correspondence with multiple types to which multiple targets belong, wherein each of the multiple types includes one or more of the multiple targets.

[0053] In a possible embodiment, the adverse drug reaction ADR prediction model is obtained based on a multi-layer perceptron or a multi-label model training according to the drug-ADR interaction data, wherein the drug-ADR interaction data includes a plurality of first counts corresponding to the case where the compound is associated with the ADR, and a plurality of second counts corresponding to the case where the compound is not associated with the ADR.

[0054] In one possible implementation, a cloud application instance is used to sort the second probabilities from large to small to determine the ranking corresponding to each second probability; the cloud application instance is used to calculate the ADR score corresponding to each compound, wherein an ADR score is determined based on the risk score of the ADR corresponding to one or more second probabilities corresponding to a compound and the ranking corresponding to one or more second probabilities; the cloud application instance is used to determine the safety scores of multiple compounds based on the ADR scores corresponding to each compound and the number of off-target targets, wherein when the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe, and when the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe.

[0055] In summary, the compound safety prediction system provided in the present application uses a cloud application example to input off-target prediction results as features into ADR prediction, thereby realizing the combination of off-target prediction and ADR prediction, and can achieve accurate prediction of off-target targets and ADRs, and can reveal the mechanism of action of ADRs from the target level. In addition, the compound safety scoring standard is determined based on the above prediction results to predict the safety of the compound, which can more accurately determine the safety of the compound in the early stage of drug development and improve efficiency.

[0056] like Figure 2 As shown, Figure 2 It is a structural diagram of another compound safety prediction system provided in an embodiment of the present application. The compound safety prediction system includes: a user device 110 and a cloud server 2, the cloud server 2 includes a cloud application instance 21, and the cloud application instance 21 includes an application device 210.

[0057] In a specific implementation, the application device 210 includes a compound safety prediction application device provided in an embodiment of the present application, and the application device includes a receiving unit 211, an off-target prediction unit 212, a drug adverse reaction prediction unit 213 and a safety prediction unit 214.

[0058] In one possible embodiment, a receiving unit is used to receive multiple compounds; an off-target prediction unit is used to determine a first probability between each compound in the multiple compounds and each target in the multiple targets based on the multiple compounds, the multiple targets and the multiple off-target prediction models; an adverse drug reaction ADR prediction unit is used to input the first probability into the ADR prediction model to obtain a second probability between each compound and each ADR; and a safety prediction unit is used to determine the safety scores of the multiple compounds based on the number of off-target targets, the second probability and the risk score of each ADR obtained by the receiving unit, so as to predict the safety of the multiple compounds, wherein the off-target target is a target corresponding to the first probability greater than a probability threshold.

[0059] In a possible implementation, the safety prediction unit is specifically used to sort the second probabilities from large to small to determine the ranking corresponding to each second probability; it is also used to calculate the ADR score corresponding to each compound, wherein an ADR score is determined based on the risk score of the ADR corresponding to one or more second probabilities corresponding to a compound and the ranking corresponding to one or more second probabilities; it is also used to determine the safety scores of multiple compounds based on the ADR score corresponding to each compound and the number of off-target targets, wherein when the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe, and when the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe.

[0060] It should be understood that the cloud server 2 may include multiple cloud application instances, each of which may include multiple application devices such as a compound safety prediction application device, which is not specifically limited here. The cloud application instance 21 may be any cloud application instance in the cloud server 2 that can implement the function of the application device 210, and the cloud server 2 may also be any cloud server in the cloud service center that can implement the function of the application device 210, which is not specifically limited here.

[0061] In summary, the compound safety prediction application device provided in the present application combines off-target prediction and ADR prediction through the mutual cooperation of a receiving unit, an off-target prediction unit, an adverse drug reaction ADR prediction unit and a safety prediction unit, thereby achieving the effect of revealing the mechanism of action of ADR from the target level, and can more accurately judge the safety of the compound and improve the efficiency of the drug development process.

[0062] like Figure 3 As shown, Figure 3 This is a flow chart of a compound safety prediction method provided in an embodiment of the present application. The compound safety prediction method is applied to Figure 1 or Figure 2 In the system shown, the method includes the following steps.

[0063] Step S310: The cloud application instance receives multiple compounds sent by the user device, and determines a first probability between each compound and each target based on the multiple compounds, multiple targets, and multiple off-target prediction models.

[0064] Before calculating the first probability, the cloud application instance collects off-target model training data from databases such as ChEMBL and PubChem according to the multiple targets included in the established off-target panel, and obtains multiple off-target prediction models based on the multi-task model training of neural networks. Among them, since each compound may have different inhibitory effects on different targets at different concentrations and affect the target activity, the off-target model training data includes interaction data between multiple compounds and multiple targets.

[0065] Specifically, the process of data preparation in the cloud application instance is as follows: the acquired off-target model training data is marked according to whether the compound inhibits the target activity. When the concentration of the compound inhibiting the target activity is less than or equal to 1 micromole per liter, the data is marked as 1, and when the concentration of the compound inhibiting the target activity is greater than 1 micromole per liter, the data is marked as 0. Afterwards, the data set is divided into K folds by cross-validation method, each fold contains an equal number of samples, and the category distribution is similar to the overall data set. The samples in each fold are divided into training set, validation set and test set for subsequent determination of the off-target prediction model.

[0066] Figure 4 Schematic diagram of a process of feature extraction and generation of off-target prediction model provided in the embodiment of the present application. Figure 4 As shown, in a possible embodiment, cloud application examples are used to represent small molecules of compounds input in the form of simplified molecular input line representation system (simplified molecular input line entry system, SMILES), which are converted into molecular fingerprints by the characterization method in DeepChem, that is, molecular data is converted into feature representations suitable for machine learning models. Afterwards, multi-task off-target model training is carried out according to the above-mentioned features, data after data preparation, and deep neural network (deep neural network, DNN), and corresponding off-target prediction models are generated for each type of off-target target. Taking the 91 targets provided in the present application as an example, the targets belong to seven protein types respectively, so seven off-target prediction models are generated, and taskN in the figure is task7.

[0067] like Figure 4As shown, in another possible implementation, the cloud application instance is used to convert the small molecules representing the compounds input in the form of SMILES into molecular graphs through the characterization method in DeepChem and then input them into the multi-task graph neural network, attentive fingerprints (Attentive FP). After several layers of atomic-level attention representation and several layers of molecular-level attention representation, an N-dimensional molecular representation is obtained, which is then passed through a layer of neural network, such as Figure 4 The classifier (softmax) shown in the figure performs task classification, that is, obtains the N tasks shown in the figure, so as to train each type of off-target target to obtain the corresponding off-target prediction model. Among them, the above-mentioned multi-task graph neural network can also include a graph convolutional network (graph convolutional network, GCN), a graph attention network (graph attention network, GAT) and a message passing neural network (Message Passing Neural Network, MPNN), etc., which are not specifically limited in this application.

[0068] The above two possible implementation methods are multi-task learning for multiple target types, and multiple off-target prediction models are trained. During the training process, except for the parameters of the last layer of neural network, which are specific to different tasks, other neural network parameters can be shared among multiple tasks. Therefore, the generalization performance and efficiency of the off-target prediction model can be improved by the above method. In addition to the above two methods, multiple off-target prediction models can also be trained by other possible implementation methods, which are not specifically limited in this application.

[0069] After obtaining multiple off-target prediction models, the cloud application instance determines the first probability between each received compound and each target provided in each embodiment of the present application. The first probability is the binding probability between the compound and the target, which is a value in the range of [0, 1].

[0070] Step S320: The cloud application instance inputs the first probability into the adverse drug reaction prediction model to obtain a second probability between each compound and each ADR.

[0071] The cloud application example first trains the ADR model based on the drug-ADR interaction data collected from the SIDER database and the compounds related to toxic side effects with toxic or non-toxic labels, based on the multilayer perceptron (MLP) or multi-label K-nearest neighbors (KNN) model. Among them, the drug-ADR interaction data includes the first count 1 of the compound and ADR when they are associated, and the second count 0 of the compound and ADR when they are not associated. Each compound can correspond to one or more associated ADR labels. The cloud application example also verifies the model by randomly collecting some compounds and some compounds with Anatomical Therapeutic Chemical (ATC) codes from the ZINC database, and finally obtains the adverse drug reaction prediction model. This application does not specifically limit other methods for obtaining adverse drug reaction prediction models.

[0072] After obtaining the adverse drug reaction prediction model, the cloud application instance inputs the first probability into the adverse drug reaction prediction model to obtain a second probability between each compound and each ADR, where the second probability is a predicted probability between the compound and the ADR.

[0073] like Figure 5 As shown, Figure 5 This is a schematic diagram of a process for predicting adverse drug reactions provided by an embodiment of the present application. The cloud application example inputs the off-target profiles determined in step S310, that is, the first probability representing the probability of binding between the compound and each target, into the adverse drug reaction prediction model, thereby predicting the possible adverse drug reactions of the compound based on the interacting targets corresponding to the compound. Specifically, Figure 5 The X matrix is ​​determined based on off-target profiles, where drug1~drugn represents multiple compounds, d1~dm represents the first probability of each compound corresponding to different targets, and m is the number of targets. S1~sx in the Y matrix indicates whether the compound is associated with the ADR provided in the embodiment of the present application. In the case of association, the first count is 1, and in the case of no association, the second count is 0, and x is the number of ADRs. Finally, the adverse drug reaction prediction model outputs the second probability.

[0074] Since the interaction between the target and the compound is one of the causes of adverse drug reactions, the above process inputs the first probability obtained by the off-target prediction model into the adverse drug reaction prediction model, thereby realizing the combination of off-target prediction and ADR prediction, which can improve the accuracy of compound safety prediction and enhance the interpretability of off-target and ADR.

[0075] In a specific embodiment, Figure 6 As shown, Figure 6 It is a schematic diagram of the association between a target and an ADR provided in an embodiment of the present application. Among them, the off-target panel includes known targets, and targets 1 to 4 are targets obtained from databases such as ChEMBL, PubChem and DrugBank that overlap with the 91 targets provided in the embodiments of the present application. The off-target targets determined by the off-target prediction are multiple off-target targets whose first probability is greater than the probability threshold obtained by performing off-target prediction according to the above steps, which also include the above-mentioned targets 1 to 4, and the probability is the corresponding first probability, and the probability is followed by one or more adverse drug reactions corresponding to each target. This embodiment is only based on Figure 6 The targets A, B, C, D, 1 to 4 and some adverse drug reactions are shown as examples for illustration, and in fact include more known targets, off-target targets and adverse drug reactions. As can be seen from the figure, the predicted adverse drug reaction corresponding to the new off-target target B contains hyperglycemia, and the predicted adverse drug reaction corresponding to the new off-target target D contains increased blood sugar, which can explain why the first compound causes the adverse drug reaction of hyperglycemia. Other situations are similar to the above process, and this application will not go into details one by one.

[0076] Step S330: The cloud application instance determines the safety scores of the multiple compounds according to the number of off-target targets, the second probability, and the acquired risk score of each ADR, so as to predict the safety of each of the multiple compounds.

[0077] The cloud application instance calculates the ADR score corresponding to each compound based on the second probability and the risk score of each ADR obtained, and then determines the safety scores of multiple compounds in combination with the number of off-target targets, so as to predict the safety of each of the multiple compounds. Among them, the off-target target is the target corresponding to the first probability greater than the probability threshold. Since the first probability is a value in the range of [0, 1], when the probability threshold is 0.5, the target corresponding to the first probability greater than 0.5 is the off-target target. The off-target target is the target that interacts with the compound. The probability threshold can also take other values, and this application does not make specific restrictions on this.

[0078] Specific as Figure 7 As shown, Figure 7 This is a schematic diagram of the process of compound safety evaluation provided in an embodiment of the present application.

[0079] In a possible implementation, the cloud application instance sorts the second probabilities from large to small, determines the ranking corresponding to each second probability, and calculates the ADR score corresponding to each compound in combination with the obtained risk score of each ADR. The specific calculation formula of the ADR score is as follows:

[0080]

[0081] Among them, ADR i Severity score is the risk score of the ADR corresponding to the ith position after the second probability ranking. Then, the safety scores of multiple compounds are determined according to the ADR scores corresponding to each compound and the number of off-target targets corresponding to each compound. The specific calculation formula for the safety score of the compound is as follows:

[0082] Safety score=ADR score+log N (targetnumber+k)

[0083] Wherein, N is 2 or 10, k is usually a constant of 1, and target number is the number of off-target targets corresponding to the compound. There is a significant difference between the safety scores of toxic compounds and non-toxic compounds, and the safety scores of toxic compounds are higher. Therefore, a safety threshold can be defined. When the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe. When the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe. This application does not make specific limitations on the above safety threshold. In addition to the above safety judgment of the compound based on the safety score, Figure 7 Compounds withdrawn from the DrugBank database and compounds warned or marked in the ChEMBL database are also considered, and these compounds are determined as unsafe compounds. In a possible embodiment, there are currently more methods for judging the safety of compounds, which are not specifically limited in this application.

[0084] The above method proposes a calculation formula for the safety score of the compound, quantifies the safety level of the compound, provides a basis for safety assessment in the drug development process, and can enhance the accuracy of the calculation of the compound safety score, thereby improving the efficiency of determining the safety of the compound in the early stages of drug development.

[0085] In addition to the above steps, the cloud application instance also sends the prediction results of the off-target prediction model determined in the compound safety prediction process and the prediction results of the adverse drug reaction model to the user device for display in the result visualization interface. In addition to the above prediction results, the safety scores and safety judgment results corresponding to multiple compounds can also be sent to the user device. The cloud application instance can also send more types of data to the user device, which is not specifically limited in this application.

[0086] In summary, the compound safety prediction method provided in the present application realizes the combination of off-target prediction and ADR prediction by inputting the results of the off-target prediction model as features into the ADR prediction model, and determines the compound safety scoring standard based on the above prediction results to predict the safety of the compound. It can achieve accurate prediction of off-target targets and ADRs, and can reveal the mechanism of action of ADR from the target level, thereby determining the safety of compounds in the early stages of drug development and improving efficiency.

[0087] In a specific embodiment, the cloud application instance is pre-trained to obtain an off-target prediction model and an adverse drug reaction prediction model, and three compounds, compound A, compound B and compound C, are received. According to the off-target prediction model, the first probabilities of compound A, compound B and compound C corresponding to 91 targets are calculated. Afterwards, the first probabilities of compound A, compound B and compound C corresponding to each target are input as features into the adverse drug reaction prediction model to obtain the second probabilities of compound A, compound B and compound C corresponding to each ADR provided in the present application. The present application provides three ADRs, namely ADR1, ADR2 and ADR3. The second probabilities are sorted in order from large to small, and the three second probabilities corresponding to compound A and ADR1, ADR2 and ADR3 are determined to be in the first, fourth and fifth positions respectively, and the risk scores of ADR1, ADR2 and ADR3 are obtained. The ADR score calculation formula provided in step S330 is used for calculation. The specific process is as follows:

[0088]

[0089] Afterwards, when the number of off-target targets corresponding to compound A is 21, its safety score is determined in combination with the ADR score of compound A. The specific calculation formula is as follows:

[0090] Safety score-A=ADR score-A+log222

[0091] Afterwards, the safety of compound A can be determined by comparing Safety score-A with the safety threshold. Similarly, the safety prediction process of compound B and compound C is similar to the above process, which will not be described in detail in this application.

[0092] like Figure 8 As shown, Figure 8 A computing device provided by an embodiment of the present invention can be applied to Figure 1 or Figure 2 In the cloud server or cloud application example in the provided compound safety prediction system, the computing device 800 includes: a processor 810, a memory 820, a communication interface 830, and a bus 840. The processor, the memory, and the communication interface may be interconnected via an internal bus or may communicate via other means such as wireless transmission.

[0093] The processor 810 may be composed of at least one general-purpose processor, such as a central processing unit (CPU), or a combination of a CPU and a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The PLD may be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof. The processor 810 is used to execute various types of digital storage instructions.

[0094] The memory 820 may be a volatile memory, such as a random access memory (RAM), a dynamic RAM (DRAM), a static RAM (SRAM), a synchronous dynamic RAM (SDRAM), a double data rate RAM (DDR), a cache, etc. The memory may also include a combination of the above types. The memory 820 includes program code, and the processor 810 can execute the program code to execute Figure 3 All the steps included in the implementation Figure 2 Function in .

[0095] The communication interface 830 can receive compounds sent by the user device, or a character string representing the compound structure, and can send the prediction results of the off-target panel, the prediction results of adverse drug reactions, and the safety prediction results to the user device, which is not specifically limited in this application.

[0096] The bus 840 may be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 The fact that only one line is used in the diagram does not mean that there is only one bus or only one type of bus.

[0097] Need to explain, Figure 8 This is only one possible implementation of the embodiment of the present application. In actual applications, the computing device 800 may also include more or fewer components, which is not limited here.

[0098] The embodiment of the present application also provides a computing device cluster. The computing device cluster includes at least one computing device. The computing device can be a server, such as a central server, an edge server, or a local server in a local data center. In some embodiments, the computing device can also be a terminal device such as a desktop computer, a laptop computer, or a smart phone.

[0099] like Fig. 9 As shown, the computing device cluster includes at least one computing device 800. The memory 820 in one or more computing devices 800 in the computing device cluster may store the same instructions for executing a compound safety prediction method provided in an embodiment of the present application.

[0100] In some possible implementations, the memory 820 of one or more computing devices 800 in the computing device cluster may also store partial instructions for executing the above-mentioned compound safety prediction method. In other words, the combination of one or more computing devices 800 can jointly execute instructions for executing the compound safety prediction method.

[0101] It should be noted that the memory 820 in different computing devices 800 in the computing device cluster can store different instructions, and the instructions stored in the memory 820 in different computing devices 800 can realize the functions of one or more units among the receiving unit, the off-target prediction unit, the adverse drug reaction prediction unit and the safety prediction unit.

[0102] In some possible implementations, one or more computing devices in the computing device cluster may be connected via a network, which may be a wide area network or a local area network. Fig.10 A possible implementation is shown. Fig.10 As shown, two computing devices 800A and 800B are connected via a network, computing device 800A includes a processor 810A, a memory 820A, a communication interface 830A and a bus 840A, and computing device 800B includes a processor 810B, a memory 820B, a communication interface 830B and a bus 840B. Specifically, the network is connected via the communication interface in each computing device. In this type of possible implementation, the memory 820A in the computing device 800A stores instructions for executing the functions of the receiving unit and the off-target prediction unit. At the same time, the memory 820B in the computing device 800B stores instructions for executing the functions of the adverse drug reaction prediction unit and the safety prediction unit. It should be understood that Fig.10 The functions of the computing device 800A shown in FIG. 8 may also be completed by multiple computing devices 800. Similarly, the functions of the computing device 800B may also be completed by multiple computing devices 800.

[0103] The embodiment of the present application further provides a computer-readable storage medium, in which instructions are stored. When the instructions are executed on a processor, Figure 3 The method flow shown is realized.

[0104] The embodiment of the present application further provides a computer program product. When the computer program product is run on a processor, Figure 3 The method flow shown is realized.

[0105] When the computer program instructions are loaded or executed on the computer, the process or function according to the embodiment of the present invention is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0106] Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. Computer-readable storage media can be any available medium that can be accessed by a computer or a data storage node such as a server, data center, etc. that includes at least one set of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., high-density digital video discs (DVDs), or semiconductor media).

[0107] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for predicting the safety of a compound, characterized in that: Applied to a cloud service center, the cloud service center includes multiple cloud application instances, and the method includes: The cloud application instance determines a first probability between each compound in the plurality of compounds and each target in the plurality of targets according to the received plurality of compounds, the plurality of targets, and the plurality of off-target prediction models; The cloud application instance inputs the first probability into a drug adverse reaction ADR prediction model to obtain a second probability between each compound and each ADR; The cloud application instance determines the safety scores of the multiple compounds based on the number of off-target targets, the second probability, and the acquired risk score of each ADR, so as to predict the safety of each of the multiple compounds, wherein the off-target target is a target corresponding to the first probability that is greater than a probability threshold.

2. The method according to claim 1, characterized in that The multiple off-target prediction models are obtained by training the cloud application instance based on a multi-task model of a neural network, and have a one-to-one correspondence with the multiple types to which the multiple targets belong, wherein each of the multiple types includes one or more of the multiple targets.

3. The method according to claim 2, characterized in that The adverse drug reaction ADR prediction model is obtained based on a multi-layer perceptron or a multi-label model training according to the interaction data between the drug and the ADR, wherein the interaction data between the drug and the ADR includes a plurality of first counts corresponding to a case where the compound is associated with the ADR, and a plurality of second counts corresponding to a case where the compound is not associated with the ADR.

4. The method according to claim 3, characterized in that The cloud application instance determines the safety scores of the multiple compounds according to the number of off-target targets, the second probability, and the acquired risk score of each ADR, including: The cloud application instance sorts the second probabilities from large to small, and determines a ranking corresponding to each second probability; The cloud application instance calculates the ADR score corresponding to each compound, wherein an ADR score is determined according to the risk score of the ADR corresponding to each of the one or more second probabilities corresponding to a compound and the ranking corresponding to each of the one or more second probabilities; The cloud application instance determines the safety scores of the multiple compounds based on the ADR score corresponding to each compound and the number of off-target targets, wherein when the safety score corresponding to the compound is higher than the safety threshold, the compound is predicted to be unsafe, and when the safety score corresponding to the compound is less than or equal to the safety threshold, the compound is predicted to be safe.

5. The method according to any one of claims 1 to 4, characterized in that: The method further comprises: The cloud application instance sends the prediction data to the user device for display, wherein the prediction data includes the first probability, the second probability, and one or more data of the safety scores and safety of each of the multiple compounds.

6. The method according to claim 5, characterized in that The cloud application instance includes one of a bare metal server BMS, a container and a virtual machine.

7. A compound safety prediction application device, the device is applied to a cloud application instance, the cloud application instance includes a bare metal server BMS, a container and a virtual machine, the application device includes: A receiving unit, for receiving a plurality of compounds; An off-target prediction unit, configured to determine a first probability between each compound in the plurality of compounds and each target point in the plurality of targets according to the plurality of compounds, the plurality of targets and the plurality of off-target prediction models; An adverse drug reaction ADR prediction unit, used for inputting the first probability into an ADR prediction model to obtain a second probability between each compound and each ADR; A safety prediction unit is used to determine the safety scores of the multiple compounds according to the number of off-target targets, the second probability and the risk score of each ADR obtained by the receiving unit, so as to predict the safety of the multiple compounds, wherein the off-target target is a target corresponding to the first probability greater than a probability threshold.

8. A computing device cluster, characterized in that: comprising at least one computing device, each computing device comprising a processor and a memory; The processor of the at least one computing device is configured to execute instructions stored in the memory of the at least one computing device, so that the computing device cluster executes the method according to claims 1 to 6.

9. A computer program product comprising instructions, characterized in that When the instruction is executed by the management device, the management device executes the method according to claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The method comprises computer program instructions, and when the computer program instructions are executed by a management device, the management device performs the method according to claims 1 to 6.

Images