Method for training a predictive model, method for predicting data, device and computer program

Abstract

The present invention provides a method, apparatus, computer device, and storage medium for training a predictive model. The method includes the steps of: acquiring a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample; determining a current training sample from the training sample set based on the training sample weight; inputting the current target energy feature corresponding to the current training sample into a pre-trained predictive model to perform basic training, and obtaining a basic predictive model upon completion of the basic training; and updating the training sample weight corresponding to each training sample based on the basic predictive model until model training is completed, and repeatedly executing the process to obtain a target predictive model, the target predictive model being used to predict interaction state information corresponding to input protein information and input compound information. Using this method, the predictive accuracy of the target predictive model obtained by training can be improved.

Description

[Technical field]

[0001] The present invention claims priority to a Chinese patent application filed on April 1, 2021, bearing application number 2021103559296 and entitled "Prediction model training method, data prediction method, device and storage medium," the entire contents of which are incorporated herein by reference.

[0002] FIELD OF THE DISCLOSURE The present invention relates to the field of computer technology, and in particular to a method for training a predictive model, a data prediction method, an apparatus, a computer device, and a storage medium. [Background technology]

[0003] With the development of artificial intelligence technology, the use of machine learning algorithms to predict the affinity between a compound and a target protein has emerged. Currently, models built using machine learning algorithms can predict the change in affinity with a compound after the mutation of a target protein, and determine whether the target protein is resistant to the compound, thereby providing a reference for doctors to use drugs. However, the prediction models built using current machine learning algorithms have problems such as low accuracy and poor generalization ability of the models. Summary of the Invention

[0004] Therefore, in order to address the above technical issues, there is a need to provide a predictive model training method, a data prediction method, an apparatus, a computer device, and a storage medium that can improve the training accuracy of a predictive model and improve the prediction accuracy.

[0005] In one aspect, a method for training a predictive model, executed by a computing device, includes the steps of: acquiring a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set including wild-type protein information, mutant protein information, and compound information, the target energy feature being acquired based on a wild-type energy feature and a mutant energy feature, the wild-type energy feature being acquired by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature being acquired by performing binding energy feature extraction based on the mutant protein information and the compound information. determining a current training sample from the training sample set based on the training sample weights; inputting current target energy features corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed; updating training sample weights corresponding to each training sample based on the basic prediction model until model training is completed, and performing the step of determining a current training sample from the training sample set based on the training sample weights again to obtain a target prediction model, wherein the target prediction model is used to predict interaction state information corresponding to input protein information and input compound information.

[0006] In another aspect, a training apparatus for a predictive model includes a sample acquisition module for acquiring a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set including wild-type protein information, mutant protein information, and compound information, the target energy feature is acquired based on a wild-type energy feature and a mutant energy feature, the wild-type energy feature is acquired by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature is acquired by performing binding energy feature extraction based on the mutant protein information and the compound information. a training module for inputting current target energy features corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed; and an iterative module for updating training sample weights corresponding to each of the training samples based on the basic prediction model and performing the step of determining a current training sample from the training sample set based on the training sample weights again until model training is completed, to obtain a target prediction model, wherein the target prediction model is used to predict interaction state information corresponding to input protein information and input compound information.

[0007] In another aspect, a computing device includes a memory having computer readable instructions stored thereon and a processor, the processor, when executing the computer readable instructions, includes a step of acquiring a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set including wild-type protein information, mutant protein information, and compound information, the target energy feature being acquired based on a wild-type energy feature and a mutant energy feature, the wild-type energy feature being acquired by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature being acquired based on the mutant protein information and the compound information. a step of determining a current training sample from the training sample set based on the training sample weights; a step of inputting current target energy features corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed; and a step of updating training sample weights corresponding to each training sample based on the basic prediction model and determining a current training sample from the training sample set based on the training sample weights again until model training is completed, to obtain a target prediction model, wherein the target prediction model is used to predict interaction state information corresponding to input protein information and input compound information.

[0008] In another aspect, a computer readable storage medium having stored thereon computer readable instructions, which when executed by a processor, includes a step of obtaining a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set including wild-type protein information, mutant protein information, and compound information, the target energy feature being obtained based on a wild-type energy feature and a mutant energy feature, the wild-type energy feature being obtained by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature being obtained based on the mutant protein information and the compound information. a step of determining a current training sample from the training sample set based on the training sample weights; a step of inputting current target energy features corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed; a step of updating training sample weights corresponding to each training sample based on the basic prediction model until model training is completed, and performing again the step of determining a current training sample from the training sample set based on the training sample weights to obtain a target prediction model, wherein the target prediction model is used to predict interaction state information corresponding to input protein information and input compound information.

[0009] The above-mentioned prediction model training method, device, computer device and storage medium include: obtaining a training sample set, the training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information and compound information; determining a current training sample from the training sample set based on the training sample weight; inputting the current target energy feature corresponding to the current training sample into the pre-trained prediction model to perform basic training; obtaining a basic prediction model when the basic training is completed; updating the training sample weight corresponding to each training sample based on the basic prediction model until the model training is completed; and performing the steps of determining a current training sample from the training sample set based on the training sample weight again to obtain a target prediction model; the target prediction model is used to predict the interaction state information corresponding to the input protein information and the input compound information. That is, by continuously updating the training sample weight in the iterative process and using the training sample weight to determine the current training sample from the training sample set, the quality of the training sample is ensured; and then using the current training sample to train the prediction model, the trained target prediction model can improve the prediction accuracy and generalization.

[0010] In another aspect, a data prediction method includes a step of acquiring data to be predicted, the data to be predicted including wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted; a step of performing binding energy feature extraction based on the wild-type protein information to be predicted and the compound information to be predicted to acquire a wild-type energy feature to be predicted; a step of performing binding energy feature extraction based on the mutant protein information to be predicted and the compound information to be predicted to acquire a mutant energy feature to be predicted; and a step of determining a target energy feature to be predicted based on the wild-type energy feature to be predicted and the mutant energy feature to be predicted. and inputting the target energy features to be predicted into a target prediction model to perform prediction and obtain interaction state information, wherein the target prediction model is obtained by again performing the steps of obtaining a training sample set, determining a current training sample from the training sample set based on training sample weights, inputting current target energy features corresponding to the current training sample into a pre-trained prediction model to perform basic training, and upon completion of basic training, obtaining a basic prediction model, updating the training sample weights based on the basic prediction model until model training is completed, and determining a current training sample from the training sample set based on the training sample weights.

[0011] In another aspect, a data prediction device includes a data acquisition module for acquiring data to be predicted, the data to be predicted including wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted; a feature extraction module for extracting binding energy features based on the wild-type protein information to be predicted and the compound information to be predicted, acquiring wild-type energy features to be predicted, and extracting binding energy features based on the mutant protein information to be predicted and the compound information to be predicted, acquiring mutant energy features to be predicted; and a target energy feature extraction module for determining a target energy feature to be predicted based on the wild-type energy feature to be predicted and the mutant energy feature to be predicted. and a prediction module that inputs the target energy features to be predicted into a target prediction model to perform prediction and obtain interaction state information, wherein the target prediction model is obtained by again executing the steps of obtaining a training sample set, determining a current training sample from the training sample set based on training sample weights, inputting current target energy features corresponding to the current training sample into a pre-training prediction model to perform basic training, and upon completion of basic training, obtaining a basic prediction model, updating the training sample weights based on the basic prediction model until model training is completed, and determining a current training sample from the training sample set based on the training sample weights.

[0012] In another aspect, a computer device includes a memory having computer readable instructions stored thereon and a processor, the processor, when executing the computer readable instructions, includes a step of acquiring data to be predicted, the data to be predicted including wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted; a step of performing binding energy feature extraction based on the wild-type protein information to be predicted and the compound information to be predicted to acquire wild-type energy features to be predicted; a step of performing binding energy feature extraction based on the mutant protein information to be predicted and the compound information to be predicted to acquire mutant energy features to be predicted; and a step of performing binding energy feature extraction based on the wild-type energy features to be predicted and the mutant energy features to be predicted. determining a target energy feature to be predicted based on a target energy feature; inputting the target energy feature to be predicted into a target prediction model to perform prediction and obtain interaction state information, wherein the target prediction model is obtained by obtaining a training sample set, determining a current training sample from the training sample set based on training sample weights, inputting a current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training, and when the basic training is completed, obtaining a basic prediction model, updating the training sample weights based on the basic prediction model until model training is completed, and performing again the steps of determining a current training sample from the training sample set based on the training sample weights.

[0013] In another aspect, a computer readable storage medium having computer readable instructions stored thereon, the computer readable instructions, when executed by a processor, includes a step of obtaining data to be predicted, the data to be predicted including wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted; performing binding energy feature extraction based on the wild-type protein information to be predicted and the compound information to be predicted to obtain wild-type energy features to be predicted; performing binding energy feature extraction based on the mutant protein information to be predicted and the compound information to be predicted to obtain mutant energy features to be predicted; and performing binding energy feature extraction based on the wild-type energy features to be predicted and the mutant energy features to be predicted. a step of determining a target energy feature to be predicted based on a pre-training feature; and a step of inputting the target energy feature to be predicted into a target prediction model to perform prediction and obtain interaction state information, wherein the target prediction model is obtained by acquiring a training sample set, determining a current training sample from the training sample set based on training sample weights, inputting a current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training, and when the basic training is completed, acquiring a basic prediction model, updating the training sample weights based on the basic prediction model until model training is completed, and again executing the steps of determining a current training sample from the training sample set based on the training sample weights.

[0014] The above data prediction method, device, computer device, and storage medium obtain data to be predicted, determine a target energy feature to be predicted, input the target energy feature to be predicted into a target prediction model to perform prediction, and obtain interaction state information. The target prediction model is obtained by obtaining a training sample set, determining a current training sample from the training sample set based on the training sample weight, inputting the current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training, obtaining a basic prediction model when the basic training is completed, updating the training sample weight based on the basic prediction model, and determining a current training sample from the training sample set based on the training sample weight again until model training is completed. That is, prediction is performed using the target prediction model to obtain interaction state information, and the target prediction model obtained by training can improve the accuracy of prediction, so that the accuracy of the obtained interaction state information can be improved. [Brief description of the drawings]

[0015] In order to more clearly describe the technology according to the embodiments of the present invention, the following briefly introduces drawings necessary for describing the embodiments. Note that the following drawings are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative work. [Figure 1] FIG. 1 is a diagram of an application environment of a method for training a predictive model in one embodiment. [Diagram 2] FIG. 1 is a schematic diagram of a method flow for training a predictive model according to one embodiment. [Diagram 3] FIG. 1 is a schematic diagram of a pre-training flow for an initial predictive model according to one embodiment. [Figure 4] FIG. 2 is a schematic diagram of a flow of obtaining initial interaction state information according to one embodiment. [Diagram 5] FIG. 2 is a schematic diagram of a flow of obtaining a target energy signature according to one embodiment. [Figure 6] FIG. 2 is a schematic diagram of a flow diagram for obtaining a wild-type energy signature according to one embodiment. [Figure 7] FIG. 1 is a schematic diagram of the flow of obtaining a mutation energy signature according to one embodiment. [Figure 8] FIG. 2 is a schematic diagram of a flow of obtaining a goal-based predictive model according to one embodiment. [Figure 9] FIG. 2 is a schematic diagram of a flow diagram for obtaining a basic predictive model according to one embodiment. [Figure 10] FIG. 4 is a schematic diagram of a flow of obtaining updated sample weights according to one embodiment. [Figure 11] 1 is a schematic diagram of a flow of a data prediction method according to one embodiment. [Figure 12] 1 is a schematic diagram of an application scenario flow of a data prediction method according to a specific embodiment; [Figure 13] FIG. 1 is a schematic diagram of a method for training a predictive model according to one specific embodiment. [Figure 14] FIG. 1 is a schematic diagram of a method for training a predictive model according to one specific embodiment. [Figure 15] FIG. 1 is a schematic diagram of a comparative test result according to one specific embodiment. [Figure 16] FIG. 16 is a schematic diagram of the index of precision and recall curves according to the specific embodiment of FIG. 15; [Figure 17] FIG. 1 is a block diagram of a configuration of a predictive model training device according to one embodiment. [Figure 18] FIG. 1 is a block diagram of a configuration of a data prediction device according to an embodiment. [Figure 19] FIG. 1 is a diagram showing the internal configuration of a computer device according to one embodiment. [Figure 20] FIG. 11 is a diagram showing the internal configuration of a computer device according to another embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0016] In order to clarify the objectives, technical means and advantages of the present invention, the following will further describe the embodiments of the present invention in detail with reference to the drawings. Note that the specific examples described in this specification are merely for the purpose of explaining the present invention and are not intended to limit the present invention.

[0017] The method for training a prediction model according to the present invention may be applied to the application environment shown in FIG. 1. Here, the terminal 102 communicates with the server 104 through a network. The server 104 receives a model training command sent by the terminal 102. The server 104 obtains a training sample set from the database 106 according to the model training command. The training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information, the target energy feature is obtained based on the wild-type energy feature and the mutant energy feature, the wild-type energy feature is obtained by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature is obtained by performing binding energy feature extraction based on the mutant protein information and the compound information. The server 104 determines a current training sample from the training sample set based on the training sample weight. The server 104 inputs the current target energy feature corresponding to the current training sample into the pre-trained prediction model to perform basic training, and obtains a basic prediction model when the basic training is completed. The server 104 updates the training sample weights corresponding to each training sample based on the basic prediction model until the model training is completed, and executes the steps of determining a current training sample from the training sample set based on the training sample weights again to obtain a target prediction model. The target prediction model is used to predict interaction state information corresponding to the input protein information and the input compound information. Here, the terminal 102 may be, but is not limited to, various personal computers, laptop computers, smartphones, tablet computers, and portable wearable devices. The server 104 may be realized using an independent server or a server cluster consisting of multiple servers.

[0018] In one embodiment, a method for training a prediction model is provided, as shown in Figure 2. The method is described as being applied to the server in Figure 1 as an example, but the method may also be applied to a terminal or a system including a terminal and a server, and is realized by the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0019] In step 202, a training sample set is obtained. The training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information, the target energy feature is obtained based on the wild-type energy feature and the mutant energy feature, the wild-type energy feature is obtained by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature is obtained by performing binding energy feature extraction based on the mutant protein information and the compound information.

[0020] Here, the protein means a target protein such as a protein kinase. The compound means a drug capable of interacting with the target protein. For example, a tyrosine kinase inhibitor. The protein information is used to represent specific information of the target protein, and may include the protein structure, the physicochemical properties of the protein, and the like. The wild-type protein information means information about an individual obtained from nature, i.e., a protein that has been non-artificially mutated. The mutant protein information means information about a protein in which a mutation, such as a mutation in a drug structure, has occurred. The compound information means information about a compound capable of interacting with a protein, and may include the compound structure, the physicochemical properties of the compound, and the like. The training sample weight means a weight corresponding to a training sample that is used to represent the quality of the corresponding training sample. A high-quality training sample can improve the quality of training when training a machine learning model. The binding energy feature means a feature during the interaction between a protein and a compound that is used to represent the interaction energy information between the target protein and the compound molecule, and may include structural features, physicochemical properties, and energy features that are features obtained by feature selection, and the like. The wild-type energy feature means a binding energy feature extracted when a wild-type protein interacts with a compound. The mutant energy feature refers to the binding energy feature extracted when the mutant protein interacts with the compound. The target energy feature is used to represent the difference between the mutant energy feature and the wild-type energy feature.

[0021] Specifically, the server may directly obtain the training sample set from the database. The training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information, the target energy feature is obtained based on the wild-type energy feature and the mutant energy feature, the wild-type energy feature is obtained by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature is obtained by performing binding energy feature extraction based on the mutant protein information and the compound information. The server may collect each training sample from the Internet, extract the target energy feature corresponding to each training sample, and initialize the training sample weight corresponding to each training sample. The server may obtain the training sample set from a third-party server that provides a data service, for example, from a third-party cloud server.

[0022] In one embodiment, the server may obtain protein information, mutant protein information, and compound information, perform binding energy feature extraction based on the wild-type protein information and the compound information to obtain wild-type energy features, perform binding energy feature extraction based on the mutant protein information and the compound information to obtain mutant energy features, and calculate the difference between the wild-type energy features and the mutant energy features to obtain target energy features. In addition, the training sample weights corresponding to the initialization may be, for example, random initialization, zero initialization, Gaussian distribution initialization, etc.

[0023] In step 204, a current training sample is determined from the training sample set based on the training sample weights.

[0024] Here, the current training sample is the training sample used in the current training.

[0025] Specifically, the server selects a training sample from the training sample set based on the training sample weight corresponding to each training sample to obtain a current training sample. For example, a training sample whose training sample weight is greater than a predetermined weight threshold may be the current training sample. In one specific embodiment, the training sample weight may be set to 0 and 1, that is, the training sample weight corresponding to each training sample may be initialized to 0 or 1. When the training sample weight is 1, the corresponding training sample is the current training sample. In one embodiment, the server may select multiple training samples from the training sample set based on the training sample weight to obtain a current training sample set including multiple training samples. The current training sample set is used to train the basic prediction model.

[0026] In step 206, the current target energy feature corresponding to the current training sample is input into the pre-trained prediction model to perform basic training, and when the basic training is completed, a basic prediction model is obtained.

[0027] Here, the current target energy feature refers to the target energy feature corresponding to the current training sample. The pre-trained prediction model refers to a pre-trained prediction model, which may be constructed using a random forest algorithm and used to predict the change in affinity between a protein and a compound before and after mutation. The base prediction model is obtained by training using the corresponding current training sample while retaining the training sample weights.

[0028] Specifically, the server may input the current target energy feature corresponding to the pre-training sample into the pre-training prediction model to make a prediction, obtain a prediction result, calculate a loss based on the prediction result, reversely update the pre-training prediction model based on the loss, and repeatedly execute the steps of inputting the current target energy feature corresponding to the pre-training sample into the pre-training prediction model to make a prediction until a basic training completion condition is reached, and the prediction model that has reached the basic training completion condition may be the basic prediction model. Here, the basic training completion condition refers to a condition for obtaining the basic prediction model, including that the training reaches an upper limit of a predetermined number of repetitions, or that the loss reaches a predetermined threshold, or that the parameters of the model do not change, etc.

[0029] In step 208, it is determined whether the training of the model is complete. If the training of the model is complete, execute step 208a. If the training of the model is not complete, execute step 208b and then return to and execute step 204.

[0030] In step 208a, a target prediction model is obtained, which is used to predict interaction state information corresponding to the input protein information and the input compound information.

[0031] Step 208b updates the training sample weights corresponding to each training sample based on the underlying prediction model, and returns to performing the step of determining a current training sample from the training sample set based on the training sample weights.

[0032] Here, completion of model training refers to a condition under which a target prediction model is obtained, and the target prediction model is a model that is finally trained to predict interaction state information corresponding to input protein information and input compound information. The interaction state information is used to represent the change in binding free energy between compounds before and after protein mutation. The binding free energy refers to the interaction that exists between a ligand and a receptor.

[0033] Specifically, when the server obtains the basic prediction model, it further judges whether model training completion is reached, and the model training completion condition may include that the number of iterations reaches a preset upper limit of the number of model training iterations. If the model training completion condition is not reached, the basic prediction model is used to update the training sample weights corresponding to each training sample without changing the parameters of the basic prediction model, the target energy features corresponding to each training sample are input into the basic prediction model, the loss corresponding to each training sample is obtained, and the training sample weights corresponding to each training sample are updated according to the loss corresponding to each training sample. When the training sample weights are updated, the step of determining a current training sample from the training sample set based on the training sample weights is repeatedly performed until the model training completion condition is reached, and the basic prediction model at the time when the model training completion condition is reached is taken as a target prediction model, and the target prediction model is used to predict the interaction state information of the input protein information corresponding to the input compound information.

[0034] The above-mentioned prediction model training method includes: obtaining a training sample set, the training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information; determining a current training sample from the training sample set based on the training sample weight; inputting the current target energy feature corresponding to the current training sample into the pre-trained prediction model to perform basic training; obtaining a basic prediction model when the basic training is completed; updating the training sample weight corresponding to each training sample based on the basic prediction model until the model training is completed; and performing the steps of determining a current training sample from the training sample set based on the training sample weight again to obtain a target prediction model; the target prediction model is used to predict the interaction state information corresponding to the input protein information and the input compound information. That is, by continuously updating the training sample weight in the iterative process and using the training sample weight to determine the current training sample from the training sample set, the quality of the training sample is ensured; and then using the current training sample to train the prediction model, the trained target prediction model can improve the prediction accuracy and generalization.

[0035] In one embodiment, as shown in FIG. 3, before step 202, ie, before obtaining the training sample set, further includes the following steps:

[0036] In step 302, each training sample and an interaction state label corresponding to each training sample are obtained. The training sample set includes wild-type protein information, mutant protein information, and compound information.

[0037] In step 304, binding initial energy feature extraction is performed based on the wild-type protein information and the compound information to obtain the wild-type initial energy feature.

[0038] Here, the binding initial energy feature means an extracted unscreened feature, and may include a non-physical model feature, a feature based on physical empirical potential, etc. Here, the non-physical model feature includes a crystalline protein compound structural feature, a physicochemical property of a ligand and a residue, an energy feature calculated based on an empirical or a descriptor scoring function, etc. The feature based on physical and empirical potential energy means an energy feature calculated by a modeling program based on a mixture of physical and empirical potential energy. The wild-type initial energy feature means a binding initial energy feature extracted when interacting with wild-type protein information and compound information.

[0039] Specifically, the server may obtain each training sample from the database, and each training sample may be a sample used for pre-training. Each training sample may be the same as or different from the training sample in the training sample set. The server may collect each training sample from the Internet, and the server may obtain each training sample from a server that provides data services. Each training sample includes wild-type protein information, mutant protein information, and compound information. In this case, the server may perform feature extraction for each training sample, that is, perform binding initial energy feature extraction using the wild-type protein information and compound information to obtain the wild-type initial energy feature corresponding to each training sample.

[0040] In step 306, binding initial energy feature extraction is performed based on the mutant protein information and compound information to obtain mutant initial energy features, and target initial energy features corresponding to each training sample are determined based on the wild-type initial energy features and the mutant initial energy features.

[0041] Here, the mutant initial energy feature refers to a binding initial energy feature extracted during interaction between mutant protein information and compound information, and the target initial energy feature is used to represent the difference between the wild-type initial energy feature and the mutant initial energy feature.

[0042] Specifically, the server performs binding initial energy feature extraction for the mutant protein information and the compound information, obtains the mutant initial energy feature, calculates the difference between the wild-type initial energy feature and the mutant initial energy feature, and sets the difference as the target initial energy feature. For example, the server may calculate the difference between structural features and set the difference as the target structural feature. Alternatively, the server may calculate the difference between physicochemical properties and set the difference between physicochemical properties as the target structural feature.

[0043] In step 308, the target initial energy features corresponding to each training sample are input into an initial prediction model to perform prediction, and the initial interaction state information corresponding to each training sample is obtained. The initial prediction model is constructed using a random forest algorithm.

[0044] Here, the initial prediction model refers to a prediction model with model parameter initialization, and the model parameter initialization may be random initialization or zero initialization. The initial prediction model is constructed using a random forest algorithm, and the random forest refers to a classifier that uses multiple trees to train and predict samples. For example, the initial prediction model may be constructed using an ExtraTree (extreme random tree) algorithm. The initial interaction state information refers to interaction state information predicted using the initial prediction model.

[0045] Specifically, the server creates an initial prediction model for initializing model parameters using a random forest algorithm in advance, then inputs target initial energy features corresponding to each training sample into the initial prediction model for prediction, and obtains initial interaction state information corresponding to each output training sample.

[0046] In step 310, a loss calculation is performed based on the initial interaction state information corresponding to each training sample and the interaction state label corresponding to each training sample to obtain initial loss information corresponding to each training sample.

[0047] Here, the interaction state label means the true interaction state information, and each training sample has a corresponding interaction state label. The initial loss information is used to represent the error between the initial interaction state information and the interaction state label.

[0048] Specifically, the server uses a preset loss function to calculate the loss between the initial interaction state information and the interaction state label corresponding to each training sample, and obtains the initial loss information corresponding to each training sample, where the loss function may be a mean square error loss function, a mean absolute error loss function, etc.

[0049] In step 312, the initial prediction model is updated based on the initial loss information until pre-training is completed, and the step of inputting the target energy features corresponding to each training sample into the initial prediction model and making predictions is executed again, and the feature importance corresponding to the pre-trained prediction model and the target initial energy features is obtained.

[0050] Here, the completion of pre-training is the condition for obtaining a pre-trained prediction model, which means that the number of pre-training iterations reaches a preset number of times, or the loss of pre-training reaches a preset threshold, or the parameters of the pre-trained prediction model do not change. The feature importance is used to represent the importance level of the target initial energy feature, and the higher the feature importance, the more important the corresponding feature is and the greater its contribution during model training.

[0051] Specifically, the server uses the initial loss information to calculate a gradient, then uses the gradient to inversely update the initial prediction model to obtain an updated prediction model, and determines whether pre-training is complete. If pre-training is not complete, the server takes the updated prediction model as the initial prediction model, and iteratively executes the step of inputting the target energy features corresponding to each training sample into the initial prediction model and making predictions again until pre-training is complete, and the updated prediction model obtained in the last iteration is the pre-trained prediction model. The pre-trained prediction model is constructed using a random forest algorithm, so that when the pre-trained prediction model is completed, the feature importance corresponding to the target initial energy feature can be directly obtained. Each feature of the target initial energy feature has a corresponding feature importance.

[0052] In step 314, a training sample weight corresponding to each training sample is determined based on the loss information corresponding to each training sample upon completion of pre-training, and a target energy feature is selected from the target initial energy features based on the feature importance.

[0053] Specifically, the server may use loss information corresponding to each training sample when pre-training is completed to determine a training sample weight corresponding to each training sample. For example, the loss information corresponding to each training sample is compared with a weight loss threshold, and if the loss information is greater than the weight loss threshold, the corresponding training sample is a good quality sample, and the corresponding training sample weight may be set to 1. If the loss information is equal to or less than the weight loss threshold, the corresponding training sample is a poor quality sample, and the corresponding training sample weight may be set to 0. Feature selection is performed from the target initial energy features according to feature importance to obtain target energy features, and the target energy features are features to be obtained when the pre-trained prediction model is further trained.

[0054] In the above embodiment, after pre-training using each training sample to obtain a pre-trained model, the training sample weights corresponding to each training sample are determined based on the loss information corresponding to each training sample upon completion of pre-training, and feature selection is performed from the target initial energy feature based on the feature importance to obtain the target energy feature, thereby improving the training efficiency during further training and ensuring the training accuracy.

[0055] In one embodiment, as shown in FIG. 4 , step 308, i.e., inputting target initial energy features corresponding to each training sample into an initial prediction model to make predictions and obtain initial interaction state information corresponding to each training sample, where the initial prediction model is constructed using a random forest algorithm, includes the following steps:

[0056] In step 402, the target initial energy features corresponding to each training sample are input into an initial prediction model.

[0057] In step 404, the initial prediction model takes the target initial energy features corresponding to each training sample as the current set to be split, calculates initial feature importance corresponding to the target initial energy features, determines initial splitting features from the target initial energy features based on the initial feature importance, splits the target initial energy features corresponding to each training sample based on the initial splitting features, obtains each splitting result including the target initial energy features corresponding to each split sample, takes each splitting result as the current set to be split, and iteratively performs the steps of calculating initial feature importance corresponding to the target initial energy features again until the splitting is completed, and obtains initial interaction state information corresponding to each training sample.

[0058] Here, the initial feature importance is the feature importance corresponding to the target initial energy feature, the initial split feature is the feature for performing decision tree splitting, the split result is the split of the target initial energy feature, and the split sample is the training sample corresponding to the target initial energy feature in the split result.

[0059] Specifically, the server inputs the target initial energy feature corresponding to each training sample into the initial prediction model, and the initial prediction model scores the input feature to obtain the initial feature importance corresponding to the target initial energy feature. Here, the initial feature importance may be calculated using information gain, information gain rate, Gini coefficient, mean square error, etc. An initial split feature is determined from the target initial energy feature based on the initial feature importance, and the target initial energy feature corresponding to each training sample is split based on the initial split feature, that is, the target initial energy feature exceeding the initial split feature is one part, and the target initial energy feature not exceeding the initial split feature is the other part, and a split result is obtained. The split result includes the target initial energy feature corresponding to each split sample. Each split result is taken as the current set to be split until the split is completed, and the steps of calculating the initial feature importance corresponding to the target initial energy feature are repeatedly performed to obtain the initial interaction state information corresponding to each training sample. Here, the split is completed means that each tree node cannot be split, that is, the leaf node only corresponds to one target initial energy feature. The initial interaction state information is the interaction state information predicted by the initial prediction model.

[0060] In the above embodiment, the target initial energy features corresponding to each training sample are input into an initial prediction model, the initial prediction model calculates initial feature importance corresponding to the target initial energy features, determines initial splitting features from the target initial energy features based on the initial feature importance, splits the target initial energy features corresponding to each training sample based on the initial splitting features, obtains each splitting result including the target initial energy features corresponding to each split sample, and takes each splitting result as the current set to be split until the splitting is completed. The steps of calculating initial feature importance corresponding to the target initial energy features are then repeatedly performed to obtain initial interaction state information corresponding to each training sample, thereby improving the accuracy of the obtained initial interaction state information.

[0061] In one embodiment, step 202, ie, obtaining a training sample set, the training sample set including a training sample weight corresponding to each said training sample, includes the steps of:

[0062] A confidence level corresponding to each training sample is obtained, and a training sample weight corresponding to each training sample is determined based on the confidence level.

[0063] Here, confidence is used to represent the level of quality of a training sample: the higher the confidence, the higher the quality of the corresponding training sample, and the better the performance of a model trained using a training sample with high confidence.

[0064] Specifically, when acquiring each training sample, the server may simultaneously acquire the reliability corresponding to each training sample. Then, the difference between the reliability may be directly used as the training sample weight for each training sample. Here, this reliability may be artificially set, or may be obtained by performing a reliability evaluation on each training sample in advance. In one embodiment, the reliability corresponding to each training sample may be compared with a preset reliability threshold, and if the reliability threshold is exceeded, the weight of the corresponding training sample may be set to 1, and this training sample may be used as the current training sample. If the reliability threshold is not exceeded, the corresponding training sample weight may be set to 0.

[0065] In the above embodiment, the reliability is obtained, and the training sample weight corresponding to each training sample is determined based on the reliability, thereby improving the efficiency of obtaining the training sample weight.

[0066] In one embodiment, as shown in FIG. 5, step 202, i.e., obtaining a training sample set, where the training sample set includes a target energy feature corresponding to each training sample, includes the following steps:

[0067] In step 502, binding energy feature extraction is performed based on the wild-type protein information and compound information to obtain the wild-type energy feature.

[0068] In step 504, binding energy feature extraction is performed based on the mutant protein information and compound information to obtain mutant energy features.

[0069] Here, the wild-type energy features include, but are not limited to, wild-type protein features, compound features, and energy features when wild-type protein information and compound information interact. The wild-type protein features are used to represent features corresponding to wild-type protein information, and include, but are not limited to, wild-type protein structure features and wild-type protein physicochemical property features. The compound features include, but are not limited to, compound structure features and compound physicochemical property features. The mutant energy features include, but are not limited to, mutant protein features, compound features, and energy features when mutant protein information and compound information interact. The mutant protein features are used to represent features corresponding to mutant protein information, and include, but are not limited to, mutant protein structure features and mutant protein physicochemical property features.

[0070] Specifically, the server performs feature extraction using wild-type protein information and compound information, extracts wild-type protein features and compound features, and extracts energy features when the wild-type protein interacts with the compound, and sets the wild-type protein features, compound features, and energy features as wild-type energy features. The server performs feature extraction using mutant protein information, extracts mutant protein features, and then extracts energy features when the mutant protein interacts with the compound, and sets the extracted mutant protein feature, compound feature, and energy features as mutant energy features.

[0071] In step 506, the difference between the wild-type and mutant energy signatures is calculated to obtain a target energy signature.

[0072] Specifically, the server calculates the difference between the wild-type energy feature and the mutant-type energy feature, for example, calculates the difference between the wild-type protein feature and the mutant-type protein feature, calculates the difference between the energy feature when the wild-type protein interacts with the compound and the energy feature when the mutant-type protein interacts with the compound, and obtains the target energy feature. In one specific embodiment, the feature difference between the wild-type energy feature and the mutant-type energy feature may be calculated to obtain the target energy feature.

[0073] In the above aspect, the accuracy of the obtained target energy characteristic can be improved by extracting the wild-type energy characteristic and the mutant energy characteristic, calculating the difference between the wild-type energy characteristic and the mutant energy characteristic, and obtaining the target energy characteristic.

[0074] In one embodiment, the wild-type energy feature includes a first wild-type energy feature and a second wild-type energy feature.

[0075] As shown in FIG. 6, step 502, ie, performing binding energy feature extraction based on wild-type protein information and compound information to obtain wild-type energy features, includes the following steps:

[0076] In step 602, binding energy feature extraction is performed using a non-physical scoring function based on wild-type protein information and compound information to obtain a first wild-type energy feature.

[0077] Here, a non-physical scoring function means an empirical or descriptor-based scoring function that is based on some a priori assumptions or fits experimental data to obtain energy features that have no obvious interpretable physical meaning. A first wild-type energy feature means a first partial energy feature that is extracted.

[0078] Specifically, the server may extract binding energy features using a preset non-physical scoring function, calculate wild-type protein information and compound information using the non-physical scoring function, obtain the calculation results, and use the calculation results as first wild-type energy features. Here, the energy features may be extracted using a scoring function (a function for evaluating the rationality of a theoretically obtained receptor-ligand binding mode).

[0079] In step 604, binding energy feature extraction is performed using a physical function based on the wild-type protein information and compound information to obtain a second wild-type energy feature.

[0080] Here, the physical function means the energy function based on the physics of mixing and the empirical potential energy, which has a clear physical meaning, and the energy function family consists of force field functions based on fitting of experimental data, quantification calculation functions based on the first principle, and solvent models based on continuous media.

[0081] Specifically, the server performs binding energy feature extraction on the wild-type protein information and the compound information using a preset physical function to obtain a second wild-type energy feature. For example, the energy feature may be calculated using an energy function in a modeling program Rosetta (a software library for polymer modeling with Monte Carlo simulation annealing as the core of the algorithm) based on mixed physical and empirical potential energy.

[0082] In step 606, fusion is performed based on the first wild-type energy signature and the second wild-type energy signature to obtain a wild-type energy signature.

[0083] Specifically, the server calculates a characteristic difference between the first wild-type energy characteristic and the second wild-type energy characteristic to obtain the wild-type energy characteristic.

[0084] In the above embodiment, the first wild-type energy feature and the second wild-type energy feature are extracted, and fusion is performed based on the first wild-type energy feature and the second wild-type energy feature to obtain a wild-type energy feature. The first wild-type energy feature and the second wild-type energy feature can better represent the interaction energy information between the wild-type target protein and the compound molecule, so that the obtained wild-type energy feature can be more accurate.

[0085] In one embodiment, the variant energy signature comprises a first variant energy signature and a second variant energy signature.

[0086] As shown in FIG. 7, step 504, ie, performing binding energy feature extraction based on mutant protein information and compound information to obtain mutant energy features, includes the following steps:

[0087] In step 702, binding energy feature extraction is performed using a non-physical function based on mutant protein information and compound information to obtain a first mutant energy feature.

[0088] In step 704, a binding energy feature extraction is performed using a physical function based on the mutant protein information and the compound information to obtain a second mutant energy feature.

[0089] Specifically, the server extracts binding energy features between the mutant protein information and compound information using a predetermined non-physical function to obtain a first mutant energy feature, and then extracts binding energy features between the mutant protein information and compound information using a predetermined physical function to obtain a second mutant energy feature.

[0090] In step 706, fusion is performed based on the first variant energy signature and the second variant energy signature to obtain a variant energy signature.

[0091] Specifically, the server calculates a feature difference between the first variant energy signature and the second variant energy signature to obtain a variant energy signature.

[0092] In the above embodiment, the first variant energy feature and the second variant energy feature are extracted, and fusion is performed based on the first variant energy feature and the second variant energy feature to obtain a variant energy feature. The first variant energy feature and the second variant energy feature can better represent the interaction energy information between the variant target protein and the compound molecule, so that the obtained variant energy feature is more accurate.

[0093] In one embodiment, as shown in FIG. 8, step 204, ie, determining a current training sample from the training sample set based on the training sample weights, includes the following steps.

[0094] In step 802, protein family information is obtained, and the training sample set is divided based on the protein family information to obtain each training sample group.

[0095] Here, proteins that have similar amino acid sequences in vivo and are sufficiently similar in structure and function constitute a "protein family," and members of the same protein family are called "homologous proteins." Protein family information is information about a protein family, and a training sample set is obtained by dividing each training sample corresponding to the same protein family.

[0096] Specifically, the server may directly obtain protein family information from a database, or the protein family information may be obtained from the Internet or from a third-party server providing data services. In one embodiment, the server may divide protein families whose protein information in the training samples has similar structures or sequences into the same training sample group, and obtain each training sample group.

[0097] In step 804, a current training sample is selected from each training sample group based on the training sample weight to obtain a current training sample set.

[0098] Specifically, the server uses the training sample weights to select a current training sample from each training sample group, i.e., sequentially selects a current training sample according to the training sample weights in the training sample groups, and selects from each training sample group to obtain a current training sample set.

[0099] Step 206, i.e., inputting the current target energy features corresponding to the current training sample into the pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed, includes the following steps:

[0100] In step 806, the current target energy features corresponding to each current training sample in the current training sample set are input into the pre-trained prediction model to perform basic training, and when the basic training is completed, a target basic prediction model is obtained.

[0101] Specifically, for basic training, the server inputs the current target energy features corresponding to each current training sample in the current training sample set into a pre-trained prediction model, and obtains a target basic prediction model when the basic training is completed.

[0102] In the above embodiment, the training sample set is divided according to protein family information to obtain each training sample group, and then a current training sample is selected from each training sample group according to the training sample weight to obtain a current training sample set, and the current target energy feature corresponding to each current training sample in the current training sample set is input into the pre-trained prediction model to perform basic training and obtain a basic prediction model. That is, by selecting a current training sample from each training sample group, the selected training sample is not concentrated in one local area, but distributed throughout the space, so that the global information contained in the training sample can be learned when training the model, and the comprehensiveness of the learning knowledge in the model training process can be ensured, and the convergence speed in the model training process can be improved, and the generalization of the model obtained by training can be improved.

[0103] In one specific embodiment, the basic form of the pre-trained predictive model is shown in Equation (1).

[0104]

number

[0105] In one embodiment, the step of selecting a current training sample from each training sample group based on the training sample weight to obtain a current training sample set includes the steps of:

[0106] Obtain current learning parameters, determine a number of selected samples and a sample distribution based on the current learning parameters, select a current training sample from each training sample group according to a training sample weight based on the number of selected samples and the sample distribution, and obtain a target current training sample set.

[0107] Here, the current learning parameters are the learning parameters used in the current training, which are used to control the selection of the current training samples; the number of selected samples is the number of training samples currently selected; the sample distribution is the distribution of the selected current training samples within each training sample group; and the target current training sample set is the set of current training samples selected using the current learning parameters.

[0108] Specifically, the server obtains current training sample parameters, and the initial values ​​of the current training sample parameters may be preset. The server uses the current learning parameters to calculate the number of samples selected during the current training and the sample distribution. Then, the server selects a current training sample from each training sample group according to the training sample weight based on the number of selected samples and the sample distribution, and obtains a target current training sample set.

[0109] In the above embodiment, the current learning parameters are used to further control the selection of training samples to obtain a target current training sample set, so that the selected training samples can be more accurate, and thus the prediction model obtained by training can be more accurate and the generalization ability of the prediction model can be improved.

[0110] In one embodiment, as shown in FIG. 9 , step 206, i.e., inputting the current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed, includes the following steps:

[0111] In step 902, the current target energy feature corresponding to the current training sample is input into the pre-trained prediction model for prediction, and the current interaction state information is obtained.

[0112] Here, the current interaction state information is used to represent the change in the protein-compound interaction before and after the mutation in the current training sample obtained by prediction.

[0113] Specifically, the server directly uses the current target energy feature corresponding to the current training sample as the input of the pre-trained prediction model, and the pre-trained prediction model makes a prediction based on the input current target energy feature and outputs the current interaction state information which is the prediction result.

[0114] In step 904, the error between the current interaction state information and the interaction state label corresponding to the current training sample is calculated to obtain current loss information.

[0115] Here, the current loss information is the error between the predicted result corresponding to the current training sample and the true result.

[0116] Specifically, the server obtains an interaction state label corresponding to the current training sample, which may be preset. The interaction state label may be a change in the interaction between the protein and the compound before and after the mutation measured by an experiment. Then, the server uses a preset loss function to calculate an error between the current interaction state information and the interaction state label corresponding to the current training sample, and obtains current loss information.

[0117] In step 906, until the completion condition of basic training is reached, the pre-trained prediction model is updated based on the current loss information, the current target energy feature corresponding to the current training sample is input into the pre-trained prediction model for prediction, and the step of obtaining the current interaction state information is performed again to obtain the basic prediction model.

[0118] Specifically, the server uses the current loss information to reverse-update the parameters of the pre-trained prediction model through a gradient descent algorithm, inputs the current target energy feature corresponding to the current training sample into the pre-trained prediction model for prediction, and iteratively performs the steps of obtaining current interaction state information again until a preset number of basic training iterations is reached or the model parameters no longer change, and the pre-trained prediction model of the last iteration is used as the basic prediction model.

[0119] In one specific embodiment, the optimization function corresponding to the pre-trained predictive model is a regression optimization function, as shown in Equation (2).

[0120]

number

[0121] In the above embodiment, after keeping the training sample weights constant, the current training sample is selected to train the pre-trained prediction model to obtain the base prediction model, so that the trained base prediction model can be made more accurate.

[0122] In one embodiment, as shown in FIG. 10, step 208b, ie, updating the training sample weights corresponding to each training sample based on the underlying prediction model, includes the following steps.

[0123] In step 1002, the target energy features corresponding to each training sample are input into a basic prediction model to obtain basic interaction state information corresponding to each training sample.

[0124] Here, each training sample is each training sample in the training sample set. The basic interaction state information is interaction state information corresponding to each training sample predicted by the basic prediction model. This interaction state information may be a relative difference between the binding free energy of a wild-type protein and a compound and the binding free energy of a mutant protein and a compound.

[0125] Specifically, when obtaining the basic prediction model by training, the server keeps the parameters in the basic prediction model unchanged and updates the training sample weights corresponding to each training sample in the training sample set, that is, the server inputs the target energy features corresponding to each training sample into the basic prediction model, and obtains the basic interaction state information corresponding to each output training sample.

[0126] In step 1004, the error between the basic interaction state information corresponding to each training sample and the interaction state label corresponding to each training sample is calculated to obtain basic loss information.

[0127] Here, the basic loss information means the error between the prediction result of the basic prediction model and the actual result.

[0128] Specifically, the server uses a preset loss function to calculate the error of each training sample, i.e., the error between the basic interaction state information and the interaction state label, and obtains basic loss information corresponding to each training sample.

[0129] In step 1006, update the training sample weights based on the basic loss information to obtain updated sample weights corresponding to each training sample.

[0130] Specifically, the server updates each training sample weight using the basic loss information corresponding to each training sample, and the server may directly use the basic loss information corresponding to each training sample as the updated sample weight corresponding to each training sample.

[0131] In one embodiment, step 1006, ie, updating the training sample weights based on the basic loss information and obtaining updated sample weights corresponding to each training sample, includes the following steps.

[0132] Obtain current learning parameters, calculate an update threshold based on the current learning parameters, compare the update threshold with basic loss information corresponding to each training sample, obtain a comparison result corresponding to each training sample, and determine an update sample weight corresponding to each training sample based on the comparison result corresponding to each training sample.

[0133] Here, the update threshold is a threshold for updating the weights of the training samples.

[0134] Specifically, the server obtains the current learning parameters and uses the current learning parameters to determine the update threshold. The update threshold is compared with the basic loss information corresponding to each training sample. If the basic loss information exceeds the update threshold, it means that the prediction error corresponding to the training sample is large, and in this case, the corresponding training sample weight is updated to the first training sample weight. If the basic loss information does not exceed the update threshold, it means that the error is small, and in this case, the corresponding training sample weight is updated to the second training sample weight. Then, when selecting the current training sample, the training sample corresponding to the weight of the second training sample is selected as the current training sample.

[0135] In one embodiment, the current learning parameters include a diversity learning parameter and a difficulty learning parameter. Calculating the update threshold based on the current learning parameters includes the following steps.

[0136] Obtain each training sample group, determine a current training sample group from each training sample group, and calculate a sample rank corresponding to the current training sample group. Calculate a weight value based on the sample rank, and weight the diversity learning parameter using the weight value to obtain a target weight value. Calculate the sum of the target weight value and the difficulty learning parameter to obtain an update threshold value.

[0137] Here, the difficulty learning parameter is a learning parameter for evaluating the difficulty, and the difficulty learning parameter is used to determine the reliability of the training sample selected during training. The diversity learning parameter is a learning parameter for evaluating the diversity. The diversity learning parameter is used to determine the distribution of the training samples selected during training in the training sample set. The sample rank is the rank of the training sample in the pre-training sample group, and the rank of a vector group is the number of vectors included in its largest unrelated group. The current training sample group refers to the training sample group whose weights currently need to be updated.

[0138] Specifically, the server obtains each training sample set, determines a current training sample set from each training sample set, and calculates a sample rank corresponding to the current training sample set. Calculates weights based on the sample ranks, and uses the weights to weight the diversity learning parameter to obtain a target weight. Calculates the sum of the target weight value and the difficulty learning parameter to obtain an update threshold for the current training sample group. In one specific embodiment, the training samples in each training sample group are sorted in ascending order according to the basic loss information. Obtain each ordered training sample group, determine a current training sample group in the ordered training sample group, and calculate an update threshold corresponding to the current training sample set.

[0139] In one specific embodiment, the training sample weights corresponding to the training samples may be updated using the following equation (3):

[0140]

number

[0141] In the above embodiment, by constantly updating the sample weights and reselecting the current training sample for training, training samples with large errors can be used in the training process to train, thereby avoiding the training sample with large errors from adversely affecting the training process, and improving the accuracy of the target prediction model obtained by training.

[0142] In one embodiment, after updating the training sample weights corresponding to each training sample based on the basic prediction model, the method includes the following steps:

[0143] A current learning parameter is obtained, the current learning parameter is updated according to a predetermined increment, an updated learning parameter is obtained, and the updated learning parameter is the current learning parameter.

[0144] Specifically, the server may preset an update condition for the current learning parameter, for example, may preset an increase amount of the current learning parameter for each weight update. Then, the current learning parameter is updated by a predetermined increase amount, an updated learning parameter is obtained, and the updated learning parameter is set as the current learning parameter. In one embodiment, the server may obtain a preset increasing number of samples, update the current learning parameter by the preset increasing number of samples, obtain the updated learning parameter, and set the updated learning parameter as the current learning parameter. Then, when the loss information obtained by training becomes large from small after the number of samples is increased, the training is completed, and the prediction model obtained by training when the number of samples is not increased is set as the target prediction model finally obtained.

[0145] In one embodiment, a data prediction method is provided, as shown in Fig. 11. The method is described as being applied to the server in Fig. 1 as an example, but the method may also be applied to a terminal or a system including a terminal and a server, and is realized by the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0146] In step 1102, data to be predicted is obtained. The data to be predicted includes wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted.

[0147] Here, the wild-type protein information to be predicted is wild-type protein information for which interaction state information needs to be predicted. The mutant protein information to be predicted is mutant protein information for which interaction state information needs to be predicted. The predicted compound information is compound information for which interaction state information needs to be predicted.

[0148] Specifically, the server may obtain the data to be predicted from the Internet, or may obtain the data to be predicted from a terminal. The server may obtain the prediction data directly from a database. In one embodiment, the server may obtain the prediction data sent by a third-party server. The third-party server may be a server that provides business services. The data to be predicted includes wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted. In one embodiment, the server may obtain the mutant protein information to be predicted and the compound information to be predicted from a terminal, and then obtain the wild-type protein information to be predicted corresponding to the mutant protein information to be predicted from a database to obtain the data to be predicted.

[0149] In step 1104, binding energy feature extraction is performed based on the wild-type protein information to be predicted and the compound information to be predicted, thereby obtaining the wild-type energy feature to be predicted, and binding energy feature extraction is performed based on the mutant protein information to be predicted and the compound information to be predicted, thereby obtaining the mutant energy feature to be predicted.

[0150] Here, the wild-type energy feature to be predicted means the energy feature when the extracted wild-type protein information to be predicted and the compound information to be predicted interact with each other, and the mutant-type energy feature to be predicted means the energy feature extracted when the mutant protein information to be predicted and the compound information to be predicted interact with each other.

[0151] Specifically, the server performs binding energy feature extraction based on the wild-type protein information to be predicted and the compound information to be predicted, and obtains the wild-type energy feature to be predicted. For example, a structural feature may be extracted based on the protein structure in the wild-type protein information to be predicted and the compound structure in the compound information to be predicted, and then a physicochemical property may be extracted based on the physicochemical property in the wild-type protein information to be predicted and the physicochemical property in the compound information to be predicted. The physicochemical property is an index that measures the properties of a chemical substance, including chemical properties such as melting and boiling point, state at room temperature, color, and acidity. The energy feature of the interaction between the wild-type protein information to be predicted and the compound information to be predicted is calculated using a scoring function, and the energy feature is calculated using an energy function based on the mixed physical and empirical potential energy, and the wild-type energy feature to be predicted is obtained. Next, binding energy feature extraction is performed based on the mutant protein information to be predicted and the compound information to be predicted to obtain the mutant energy feature to be predicted, for example, structural features are extracted based on the protein structure of the mutant protein information to be predicted and the compound structure of the compound information to be predicted, then physicochemical property features are extracted based on the physicochemical properties of the mutant protein information to be predicted and the physicochemical properties of the compound information to be predicted, energy features are extracted using a scoring function, and energy features are extracted using an energy function based on physical and empirical potential energies to obtain the mutant energy feature to be predicted.

[0152] In step 1106, a target energy signature to be predicted is determined based on the wild-type energy signature to be predicted and the mutant-type energy signature to be predicted.

[0153] Specifically, the server calculates the difference between each feature value of the wild-type energy feature to be predicted and the feature value corresponding to the mutant-type energy feature to be predicted to obtain the target energy feature to be predicted.

[0154] In step 1108, the target energy feature to be predicted is input into a target prediction model to perform prediction and obtain interaction state information. The target prediction model is obtained by again performing the steps of obtaining a training sample set, determining a current training sample from the training sample set based on the training sample weights, inputting a current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, obtaining a basic prediction model when the basic training is completed, updating the training sample weights based on the basic prediction model, and determining a current training sample from the training sample set based on the training sample weights until the model training is completed.

[0155] Here, the target prediction model may be a model trained by any embodiment of the above-mentioned prediction model training method, that is, the target prediction model is obtained by obtaining a training sample set, determining a current training sample from the training sample set based on the training sample weight, inputting the current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, obtaining a basic prediction model when the basic training is completed, updating the training sample weight according to the basic prediction model, and determining a current training sample from the training sample set based on the training sample weight again until the model training is completed.

[0156] Specifically, the server inputs the target energy feature to be predicted into the target prediction model for prediction, and obtains the output interaction state information. In one specific embodiment, the interaction state information means the relative difference in binding free energy between the mutant protein to be predicted and the wild-type protein to be predicted, respectively, with respect to the compound to be predicted. Then, the relative difference in binding free energy is compared with the resistance threshold, and if the relative difference in binding free energy exceeds the resistance threshold, it means that the mutant protein to be predicted has resistance and cannot continue to be used. If the relative difference in binding free energy does not exceed the resistance threshold, it means that the mutant protein to be predicted does not have resistance and can still be used normally.

[0157] The above data prediction method, device, computer device, and storage medium obtain data to be predicted, determine a target energy feature to be predicted, input the target energy feature to be predicted into a target prediction model to perform prediction, and obtain interaction state information. The target prediction model is obtained by obtaining a training sample set, determining a current training sample from the training sample set based on the training sample weight, inputting the current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training, obtaining a basic prediction model when the basic training is completed, updating the training sample weight based on the basic prediction model, and determining a current training sample from the training sample set based on the training sample weight again until model training is completed. That is, prediction is performed using the target prediction model to obtain interaction state information, and the target prediction model obtained by training can improve the accuracy of prediction, so that the accuracy of the obtained interaction state information can be improved.

[0158] The present invention further provides an application scenario for applying the above data prediction method. Figure 12 is a schematic diagram of the flow of the application scenario of the data prediction method according to one specific embodiment. Specifically, in the application scenario for predicting that the mutation of a target protein causes drug resistance, the server obtains prediction data sent to the terminal, including two different types of target protein information, including wild-type protein information and mutant protein information, and compound information. Then, the wild-type protein information and mutant protein information, and compound information are used to extract features having reference value for predicting affinity after protein mutation, including features of non-physical models and features based on physical and empirical potentials. The properties of the non-physical models, such as the crystalline protein-ligand structure, the physicochemical properties of the ligand and residues, and some energy features calculated based on the scoring function of empirical or descriptors, are energy features calculated using the mixed physical and empirical potential-based modeling program Rosetta, based on the properties based on physical and empirical potentials. Then, feature selection is performed, that is, corresponding features are selected from the features extracted by the target energy features obtained through feature selection during training, the target energy features to be predicted are selected, the target energy features to be predicted are input into the target prediction model for prediction, and the predicted binding free energy difference is obtained. This binding free energy difference is compared with the drug resistance threshold, and if the binding free energy difference exceeds the resistance threshold, it means that this protein mutation is a protein mutation that causes drug resistance. If the binding free energy difference does not exceed the drug resistance threshold, it means that this protein mutation is a protein mutation that does not cause drug resistance. In this case, the prediction result is sent to the terminal and displayed.

[0159] In one specific embodiment, a method for training a prediction model is provided. As shown in Figure 13, the method specifically includes the following steps:

[0160] In step 1302, a training sample set is obtained. The training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information, the target energy feature is obtained based on the wild-type energy feature and the mutant energy feature, the wild-type energy feature is obtained by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature is obtained by performing binding energy feature extraction based on the mutant protein information and the compound information.

[0161] In step 1304, obtain protein family information, divide the training sample set according to the protein family information, obtain each training sample group, obtain current learning parameters, determine the number of selected samples and the sample distribution according to the current learning parameters, select a current training sample from each training sample group according to the training sample weight based on the number of samples and the sample distribution, and obtain a target current training sample set.

[0162] In step 1306, the target energy features corresponding to each training sample in the target current training sample set are input into a basic prediction model, basic interaction state information corresponding to each training sample is obtained, the error between the basic interaction state information corresponding to each training sample and the interaction state label corresponding to each training sample is calculated, and basic loss information is obtained.

[0163] In step 1308, calculate a sample rank corresponding to the current training sample group, calculate a weight value according to the sample rank, weight the diversity learning parameter using the weight value to obtain a target weight value, calculate the sum of the target weight value and the difficulty learning parameter, and obtain an update threshold value for each training sample group.

[0164] In step 1310, compare the update threshold with the basic loss information corresponding to the training samples in each training sample group to obtain a comparison result corresponding to each training sample, and determine an update sample weight corresponding to each training sample in each training sample group based on the comparison result corresponding to each training sample.

[0165] In step 1312, update the current learning parameters according to a predetermined increment to obtain updated learning parameters, take the updated learning parameters as the current learning parameters, and re-execute the step of determining the number of selected samples and the sample distribution based on the current learning parameters until model training is completed, and obtain the target prediction model.

[0166] The present invention further provides application scenarios for applying the above prediction model training method, specifically as follows:

[0167] 14 is a schematic diagram of a method for training a prediction model according to a specific embodiment. Specifically, the method includes the following steps:

[0168] Obtain input data and training sample group information. The input data includes each training sample, and the corresponding training sample weight is 0 or 1. The training sample group information indicates the training sample group to which the training sample in the input data belongs. At this time, initialize the model parameters and learning parameters of the prediction model.

[0169] Then, the model parameters are trained while keeping the training sample weights corresponding to the training samples fixed, that is, a training sample whose training sample weight is 1 is selected based on the initialized learning parameters, a current training sample is obtained, a current target energy feature corresponding to the current training sample is extracted, and the current target energy feature is input into the initialized prediction model to perform basic training, and a basic prediction model is obtained after the basic training is completed.

[0170] Next, the sample weights are updated while keeping the parameters of the basic prediction model fixed, that is, the training sample weights are updated for each training sample using equation (3) to obtain the updated sample weights.

[0171] At this time, the initialized learning parameters are further updated, and the step of training the model parameters is repeated again while keeping the training sample weights corresponding to the training samples fixed until the model training is completed, and the model parameters and training sample weights of the predictive model at the time of completion of training are output, that is, the target predictive model is obtained.

[0172] In this embodiment, a comparison test is performed on the trained target prediction model.Specifically, training and testing are performed using the resistance reference dataset Platinum (Platinum is a comprehensive database of drug resistance information, developed to study and understand the effect of missense mutations on the interaction of ligand and proteome) and TKI.Here, the target prediction model is trained using the dataset Platinum, and then tested using the dataset TKI. The features corresponding to the affinity change after the protein mutation to be predicted were generated by employing non-physical model tools such as RDKit (RDKit is an open source toolkit for chemoinformatics, which uses machine learning methods to generate compound descriptors, generate fingerprints, calculate compound structure similarity, and display 2D and 3D molecules based on compound 2D and 3D molecular manipulation), Biopython (Biopython provides an online repository for developers who use and study bioinformatics), FoldX (to calculate protein binding free energy), PLIP (a tool for analyzing non-covalent interactions of protein ligands), and AutoDock (an open source molecular simulation software, most mainly applied to performing ligand-protein molecular docking). Then, energy features are calculated using Rosetta, a modeling program based on mixed physical and empirical potential energy. Next, features are selected to obtain the final selected features. Specifically, the table of the number of features finally selected is shown in Table 1 below.

[0173] Table 1. Feature count table

[0174] [Table 1] Here, the results of testing the target prediction model obtained by training are shown in FIG. 15. In FIG. 15, a scatter plot of the experimentally measured and predicted ΔΔG values ​​is shown, where ΔΔG is the relative difference in the binding free energy between the ligand and the receptor, that is, the corresponding binding free energy difference when the protein before and after the mutation binds to each compound. Here, the first row of FIG. 15 is a schematic diagram of the result of using only non-physical model features to predict the relative difference in binding free energy, and the second row of FIG. 15 is a schematic diagram of the result of using both non-physical model features and physical and empirical potential features to predict the relative difference in binding free energy. The first column is a scatter plot of the relative difference value of binding free energy obtained by testing using the conventional technique 1. The second column is a scatter plot of the relative difference value of binding free energy obtained by testing using the conventional technique 2. The third column is a scatter plot of the relative difference value of binding free energy obtained by testing using the technique of the present invention. Here, RMSE (Root Mean Square Error), Pearson (Pearson Correlation Coefficient is a measure of whether two data sets are on one line, and measures the linear relationship between variables), and AUPRC (area under the curve reduction precision recall curve) were used as evaluation indices. The average, minimum, and maximum values ​​of each index of RMSE, Pearson, and AUPRC are calculated as shown in Table 2.

[0175] Table 2 Evaluation index table

[0176] [Table 2] Here, the average value of the RMSE (the smaller the better) index of the present invention for all features is 0.73, the minimum value is 0.72, and the maximum value is 0.74, and it can be seen that the equal division error is obviously small for other features. The Pearson (the larger the better) index in the present invention is obviously superior to other prior arts. The AUPRC index in the present invention is also superior to other conventional techniques. Therefore, the present invention clearly improves the accuracy of prediction compared to the conventional techniques. Furthermore, FIG. 16 shows a schematic diagram of the AUPRC index in the comparison test results. Here, the first circle from left to right of each curve indicates the accuracy and recall of predicting the corresponding drug resistance result when the test sample obtains the corresponding drug resistance result when △△G>1.36kcal / mol is set as the threshold. The second circle from the left of each curve indicates the accuracy and recall corresponding to the resistance judgment result to be predicted when the top 15% △△G test sample is used as the resistance judgment result. As is clear from these, the technology of the present invention can clearly improve the performance for distinguishing whether there is resistance or not.

[0177] It should be noted that although the steps of the flowcharts of Figures 2 to 14 are shown in the order indicated by the arrows, they are not necessarily performed in the order indicated by the arrows. The execution of these steps is not limited to a strict order and may be performed in other orders unless expressly stated herein. Furthermore, at least some of the steps of Figures 2 to 14 may include multiple steps or multiple stages that are not necessarily performed at the same time but may be performed at different times, and the order of execution of these steps or stages is not necessarily consecutive but may be alternated or alternating with other steps or at least some of the steps or stages of other steps.

[0178] In one embodiment, as shown in Fig. 17, a predictive model training apparatus 1700 is provided. The predictive model training apparatus 1700 may be part of a computer device using software modules or hardware modules, or a combination of these. The apparatus specifically includes a sample acquisition module 1702, a sample determination module 1704, a training module 1706, and an iteration module 1708.

[0179] The sample acquisition module 1702 acquires a training sample set, the training sample set includes each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information, the target energy feature is acquired based on the wild-type energy feature and the mutant energy feature, the wild-type energy feature is acquired by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature is acquired by performing binding energy feature extraction based on the mutant protein information and the compound information.

[0180] The sample determination module 1704 determines a current training sample from the training sample set based on the training sample weights.

[0181] The training module 1706 performs base training by inputting the current target energy feature corresponding to the current training sample into the pre-trained prediction model, and obtains a base prediction model when the base training is completed.

[0182] The iterative module 1708 updates the training sample weights corresponding to each training sample based on the basic prediction model, and again executes the steps of determining a current training sample from the training sample set based on the training sample weights until the model training is completed, to obtain a target prediction model, which is used to predict the interaction state information corresponding to the input protein information and the input compound information.

[0183] In one embodiment, the predictive model training apparatus 1700 further includes the following modules:

[0184] The pre-training module obtains each training sample and an interaction state label corresponding to each training sample, the training sample set includes wild-type protein information, mutant protein information, and compound information, performs binding initial energy feature extraction according to the wild-type protein information and the compound information to obtain a wild-type initial energy feature, performs binding initial energy feature extraction according to the mutant protein information and the compound information to obtain a mutant initial energy feature, determines a target initial energy feature corresponding to each training sample according to the wild-type initial energy feature and the mutant initial energy feature, inputs the target initial energy feature corresponding to each training sample into an initial prediction model to make prediction, and obtains initial interaction state information corresponding to each training sample; The initial prediction model is constructed using a random forest algorithm, and a loss calculation is performed based on the initial interaction state information corresponding to each training sample and the interaction state label corresponding to each training sample, and initial loss information corresponding to each training sample is obtained. Until pre-training is completed, the initial prediction model is updated based on the initial loss information, and the step of inputting the target energy features corresponding to each training sample into the initial prediction model for prediction is executed again, and the pre-training prediction model and feature importance corresponding to the target initial energy features are obtained. Based on the loss information corresponding to each training sample at the completion of pre-training, a training sample weight corresponding to each training sample is determined, and a target energy feature is selected from the target initial energy features based on the feature importance.

[0185] In one embodiment, the pre-training module inputs the target initial energy features corresponding to each training sample into an initial prediction model, which takes the target initial energy features corresponding to each training sample as a current set to be split, calculates initial feature importance corresponding to the target initial energy features, determines initial split features from the target initial energy features according to the initial feature importance, splits the target initial energy features corresponding to each training sample according to the initial split features, obtains each split result including the target initial energy features corresponding to each split sample, takes each split result as a current set to be split, and iteratively performs the steps of calculating initial feature importance corresponding to the target initial energy features again until the split is completed, and obtains initial interaction state information corresponding to each training sample.

[0186] In one embodiment, the sample obtaining module 1702 obtains a confidence measure corresponding to each training sample, and determines a training sample weight corresponding to each training sample based on the confidence measure.

[0187] In one embodiment, the sample acquisition module 1702 performs binding energy feature extraction based on wild-type protein information and compound information to obtain a wild-type energy feature, performs binding energy feature extraction based on mutant protein information and compound information to obtain a mutant energy feature, calculates the difference between the wild-type energy feature and the mutant energy feature, and obtains a target energy feature.

[0188] In one embodiment, the wild-type energy feature includes a first wild-type energy feature and a second wild-type energy feature. The sample acquisition module 1702 performs binding energy feature extraction using a non-physical scoring function based on the wild-type protein information and the compound information to obtain a first wild-type energy feature, performs binding energy feature extraction using a physical function based on the wild-type protein information and the compound information to obtain a second wild-type energy feature, and performs fusion based on the first wild-type energy feature and the second wild-type energy feature to obtain a wild-type energy feature.

[0189] In one embodiment, the variant energy feature includes a first variant energy feature and a second variant energy feature. The sample acquisition module 1702 performs binding energy feature extraction using a non-physical function based on the variant protein information and the compound information to obtain a first variant energy feature, performs binding energy feature extraction using a physical function based on the variant protein information and the compound information to obtain a second variant energy feature, and performs fusion based on the first variant energy feature and the second variant energy feature to obtain a variant energy feature.

[0190] In one embodiment, the sample determination module 1704 obtains protein family information, divides the training sample set based on the protein family information, obtains each training sample group, selects a current training sample from each training sample group based on the training sample weight, and obtains a current training sample set.

[0191] The training module 1706 performs basic training by inputting current target energy features corresponding to each current training sample in the current training sample set into the pre-trained prediction model, and obtains a target basic prediction model when the basic training is completed.

[0192] In one embodiment, the sample determination module 1704 obtains current learning parameters, determines a selection sample number and a sample distribution based on the current learning parameters, selects a current training sample from each training sample group according to a training sample weight based on the selection sample number and the sample distribution, and obtains a target current training sample set.

[0193] In one embodiment, the training module 1706 inputs a current target energy feature corresponding to a current training sample into the pre-trained prediction model for prediction, obtains current interaction state information, calculates an error between the current interaction state information and an interaction state label corresponding to the current training sample, obtains current loss information, updates the pre-trained prediction model based on the current loss information until a completion condition of basic training is reached, inputs a current target energy feature corresponding to the current training sample into the pre-trained prediction model for prediction, and again performs the steps of obtaining current interaction state information to obtain a basic prediction model.

[0194] In one embodiment, the iterative module 1708 inputs target energy features corresponding to each training sample into a basic prediction model, obtains basic interaction state information corresponding to each training sample, calculates an error between the basic interaction state information corresponding to each training sample and the interaction state label corresponding to each training sample, obtains basic loss information, updates the training sample weights based on the basic loss information, and obtains updated sample weights corresponding to each training sample.

[0195] In one embodiment, the iterative module 1708 obtains current learning parameters, calculates an update threshold based on the current learning parameters, compares the update threshold with basic loss information corresponding to each training sample, obtains a comparison result corresponding to each training sample, and determines an update sample weight corresponding to each training sample based on the comparison result corresponding to each training sample.

[0196] In one embodiment, the current learning parameters include a diversity learning parameter and a difficulty learning parameter. The iteration module 1708 obtains each training sample group, determines a current training sample group from each training sample group, calculates a sample rank corresponding to the current training sample group, calculates a weight value based on the sample rank, weights the diversity learning parameter using the weight value, obtains a target weight value, calculates the sum of the target weight value and the difficulty learning parameter, and obtains an update threshold value.

[0197] In one embodiment, the iterative module 1708 obtains a current learning parameter, updates the current learning parameter according to a predetermined increment, obtains an updated learning parameter, and sets the updated learning parameter as the current learning parameter.

[0198] In one embodiment, as shown in Figure 18, a data prediction device 1800 is provided. The device may be part of a computer device using software modules, hardware modules, or a combination of these. The device specifically includes a data acquisition module 1802, a feature extraction module 1804, a target feature determination module 1806, and a prediction module 1808.

[0199] The data acquisition module 1802 acquires data to be predicted. The data to be predicted includes wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted.

[0200] The feature extraction module 1804 performs binding energy feature extraction based on the wild-type protein information to be predicted and the compound information to be predicted, thereby obtaining the wild-type energy feature to be predicted, and performs binding energy feature extraction based on the mutant protein information to be predicted and the compound information to be predicted, thereby obtaining the mutant energy feature to be predicted.

[0201] The target characteristic determination module 1806 determines a target energy characteristic to be predicted based on the wild-type energy characteristic to be predicted and the mutant energy characteristic to be predicted.

[0202] The prediction module 1808 inputs the target energy feature to be predicted into a target prediction model to perform prediction and obtain the interaction state information. The target prediction model is obtained by again performing the steps of obtaining a training sample set, determining a current training sample from the training sample set based on the training sample weights, inputting the current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training, obtaining a basic prediction model when the basic training is completed, updating the training sample weights based on the basic prediction model, and determining a current training sample from the training sample set based on the training sample weights until the model training is completed.

[0203] Specific limitations of the prediction model training device and the data prediction device may refer to the limitations of the prediction model training method and the data prediction method described above, and will not be described here. Each module in the data prediction device may be realized in whole or in part by software, hardware, or a combination thereof. Each module may be incorporated in a processor in a computer device in a hardware manner, or may be stored in a memory in a computer device in a software manner to facilitate the processor calling the modules to execute operations corresponding to the modules.

[0204] In one embodiment, the computer device may be a server having an internal block diagram as shown in FIG. 19. The computer device includes a processor, a memory, and a network interface connected by a system bus. Here, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium and a memory. The non-volatile storage medium stores an operating system, computer-readable instructions, and a database. The memory provides an environment for the operation of the operating system and the computer-readable instructions in the non-volatile storage medium. The database of the computer device is used to store training sample data and data to be predicted. The network interface of the computer device is for communicating with an external terminal via a network connection. The computer-readable instructions are executed by the processor to realize the method for training a prediction model or the method for predicting data.

[0205] In one embodiment, a computer device is provided. The computer device may be a terminal having an internal structure diagram as shown in FIG. 20. The computer device includes a processor, a memory, a communication interface, a display, and an input device, which are connected via a system bus. Here, the processor of the computer device is used to provide calculation and control capabilities. The memory of the computer device includes a non-volatile storage medium, a memory. The non-volatile storage medium stores an operating system and computer-readable instructions. The memory provides an environment for the operation of the operating system and the computer-readable instructions in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal by wired or wireless communication, which may be realized by WIFI, a carrier network, NFC (near field communication), or other technologies. When the computer-readable instructions are executed by the processor, the method for training a prediction model and the method for predicting data are realized. The display device of the computer device is a liquid crystal display device or an electronic ink display device, and the input device of the computer device is a touch layer covering the display device, which may be a key, a trackball, or a trackpad installed on the casing of the computer device, or an external keyboard, trackpad, mouse, etc.

[0206] Those skilled in the art will appreciate that the structures shown in Figures 19 and 20 are merely block diagrams of partial structures related to the technology of the present invention and do not constitute limitations on the computer device to which the technology of the present invention is applied, and that a particular computer device may include more or fewer components than those shown, may combine certain components, or may have a different arrangement of components.

[0207] In one embodiment, a computing device is also provided that includes a memory having computer readable instructions stored thereon and a processor that, when executing the computer readable instructions, performs the steps of the above-described method embodiments.

[0208] In one embodiment, a computer readable storage medium is provided that stores computer readable instructions that are executed by a processor to perform the steps of the above-described method embodiments.

[0209] In one embodiment, a computer program product or computer program is provided that includes computer instructions stored in a computer-readable storage medium, the computer instructions being read by a processor of the computer device from the computer-readable storage medium, and the processor executing the computer instructions such that the computer device performs the steps of the various embodiments of the method described above.

[0210] Those skilled in the art will understand that implementing all or part of the process of the method of the above-mentioned embodiment can be achieved by instructing the relevant hardware by a computer program that can be stored in a non-volatile computer-readable storage medium that can include the process such as the above-mentioned embodiment of the method when executed. Here, any reference to memory, storage, database, or other medium used in the embodiments provided in the present invention can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include Read-Only Memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, etc. Volatile memory can include Random Access Memory (RAM) or external cache. By way of example and not limitation, RAM can be in various forms such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0211] The technical features of the above embodiments may be combined in any combination. For the sake of brevity, all possible combinations of the technical features of the above embodiments have not been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope described in this specification.

[0212] The above embodiments only show some embodiments of the present application, and the description is more specific and detailed, but therefore should not be understood as limiting the patent scope of the invention. It should be noted that a person skilled in the art can make some variations and modifications within the scope of the present application without departing from the concept of the present application. Therefore, the patent protection scope of the present application is determined by the appended claims.

Claims

1. 1. A method for training a predictive model, implemented by a computing device, comprising: obtaining a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training samples including wild-type protein information, mutant protein information, and compound information, the target energy feature being obtained based on a wild-type energy feature and a mutant energy feature, the wild-type energy feature being obtained by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature being obtained by performing binding energy feature extraction based on the mutant protein information and the compound information; determining a current training sample from the training sample set based on the training sample weights; inputting the current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed; and updating a training sample weight corresponding to each of the training samples based on the basic prediction model and determining a current training sample from the training sample set based on the training sample weight until model training is completed, to obtain a target prediction model, wherein the target prediction model is used to predict interaction state information corresponding to input protein information and input compound information; The pre-trained prediction model takes the target energy features as input and the interaction state information as output; The basic prediction model takes the target energy feature as an input and the interaction state information and the training sample weights as outputs; The method, wherein the target prediction model has the training samples and the target energy features as inputs and the interaction state information as output.

2. Prior to obtaining the training sample set, obtaining each of the training samples and an interaction state label corresponding to each of the training samples, the training samples including wild-type protein information, mutant protein information, and compound information; Extracting initial binding energy features based on the wild-type protein information and the compound information to obtain initial binding energy features of the wild-type protein; Extracting initial binding energy features based on the mutant protein information and the compound information to obtain mutant initial energy features, and determining target initial energy features corresponding to each of the training samples based on the wild-type initial energy features and the mutant initial energy features; inputting the target initial energy features corresponding to each of the training samples into an initial prediction model to perform prediction, and obtaining initial interaction state information corresponding to each of the training samples, wherein the initial prediction model is constructed using a random forest algorithm; performing a loss calculation based on initial interaction state information corresponding to each of the training samples and an interaction state label corresponding to each of the training samples to obtain initial loss information corresponding to each of the training samples; updating the initial prediction model based on the initial loss information until pre-training is completed, inputting the target energy features corresponding to each training sample into the initial prediction model to perform prediction again, and obtaining feature importance corresponding to the pre-trained prediction model and the target initial energy features; 2. The method of claim 1, further comprising: determining a training sample weight corresponding to each of the training samples based on loss information corresponding to each of the training samples upon completion of pre-training; and selecting a target energy feature from the target initial energy features based on the feature importance.

3. inputting the target initial energy features corresponding to each of the training samples into an initial prediction model to perform prediction, and obtaining initial interaction state information corresponding to each of the training samples, the initial prediction model being constructed using a random forest algorithm; inputting the target initial energy features corresponding to each training sample into an initial prediction model; 3. The method of claim 2, wherein the initial prediction model comprises the steps of: taking target initial energy features corresponding to each of the training samples as a current set to be split; calculating initial feature importance corresponding to the target initial energy features; determining initial splitting features from the target initial energy features based on the initial feature importance; splitting the target initial energy features corresponding to each of the training samples based on the initial splitting features; obtaining each splitting result including the target initial energy features corresponding to each split sample; taking each of the splitting results as a current set to be split and iteratively performing the steps of calculating initial feature importance corresponding to the target initial energy features again until splitting is completed; and obtaining initial interaction state information corresponding to each of the training samples.

4. obtaining the training sample set, the training sample set including a training sample weight corresponding to each of the training samples; The method of claim 1 , comprising obtaining a confidence measure corresponding to each of the training samples, and determining a training sample weight corresponding to each of the training samples based on the confidence measure.

5. obtaining a training sample set, the training sample set including a target energy feature corresponding to each of the training samples; Extracting binding energy features based on the wild-type protein information and the compound information to obtain the wild-type energy features; Extracting binding energy features based on the mutant protein information and the compound information to obtain the mutant energy features; and calculating a difference between the wild-type energy signature and the mutant energy signature to obtain a target energy signature.

6. the wild-type energy characteristic comprises a first wild-type energy characteristic and a second wild-type energy characteristic; The step of extracting binding energy features based on the wild-type protein information and the compound information to obtain the wild-type energy features includes: performing binding energy feature extraction using a non-physical scoring function based on the wild-type protein information and the compound information to obtain a first wild-type energy feature; Extracting binding energy features using a physical function based on the wild-type protein information and the compound information to obtain a second wild-type energy feature; and performing fusion based on the first wild-type energy signature and the second wild-type energy signature to obtain the wild-type energy signature.

7. the variant energy signature comprises a first variant energy signature and a second variant energy signature; The step of extracting binding energy features based on the mutant protein information and the compound information to obtain the mutant energy features includes: Extracting binding energy features using a non-physical function based on the mutant protein information and the compound information to obtain a first mutant energy feature; Extracting binding energy features using a physical function based on the mutant protein information and the compound information to obtain a second mutant energy feature; and performing a fusion based on the first variant energy signature and the second variant energy signature to obtain the variant energy signature.

8. determining a current training sample from the training sample set based on the training sample weights, obtaining protein family information, and dividing the training sample set based on the protein family information to obtain training sample groups; selecting a current training sample from each of the training sample groups based on the training sample weights to obtain a current training sample set; The step of inputting the current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed, includes:

2. The method of claim 1, comprising: inputting current target energy features corresponding to each current training sample in the current training sample set into a pre-trained predictive model to perform basic training, and obtaining a target basic predictive model when the basic training is completed.

9. selecting a current training sample from each of the training sample groups based on the training sample weights to obtain a current training sample set, obtaining a current learning parameter and determining a selection sample number and a sample distribution based on the current learning parameter; and selecting a current training sample from each of the training sample groups according to the training sample weight based on the selected sample number and the sample distribution to obtain a target current training sample set.

10. The step of inputting the current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed, includes: inputting a current target energy feature corresponding to the current training sample into the pre-trained prediction model for prediction, and obtaining current interaction state information; calculating an error between the current interaction state information and an interaction state label corresponding to the current training sample to obtain current loss information; The method of claim 1 , further comprising: updating the pre-trained prediction model based on the current loss information, inputting a current target energy feature corresponding to a current training sample into the pre-trained prediction model for prediction, and performing the step of obtaining current interaction state information again until a completion condition of basic training is reached, and obtaining the basic prediction model.

11. updating a training sample weight corresponding to each of the training samples based on the base prediction model, inputting target energy features corresponding to each of the training samples into the basic prediction model to obtain basic interaction state information corresponding to each of the training samples; calculating an error between basic interaction state information corresponding to each of the training samples and an interaction state label corresponding to each of the training samples to obtain basic loss information; updating the training sample weights based on the underlying loss information to obtain updated sample weights corresponding to each of the training samples.

12. The step of updating the training sample weights based on the basic loss information and obtaining updated sample weights corresponding to each of the training samples includes: obtaining current learning parameters and calculating an update threshold based on the current learning parameters; comparing the updated threshold with basic loss information corresponding to each of the training samples to obtain a comparison result corresponding to each of the training samples; and determining updated sample weights corresponding to each of the training samples based on the comparison results corresponding to each of the training samples.

13. The current learning parameters include a diversity learning parameter and a difficulty learning parameter; Calculating an update threshold based on the current learning parameters includes: obtaining each training sample group, determining a current training sample group from each training sample group, and calculating a sample rank corresponding to the current training sample group; calculating a weight value based on the sample rank, and weighting the diversity learning parameters using the weight value to obtain a target weight value; and calculating a sum of the target weight value and the difficulty learning parameter to obtain the update threshold value.

14. After updating the training sample weights corresponding to each of the training samples based on the basic prediction model, The method of claim 1 , further comprising the steps of obtaining a current learning parameter, updating the current learning parameter according to a predetermined increment, obtaining an updated learning parameter, and making the updated learning parameter the current learning parameter.

15. 1. A method for predicting data, comprising: A step of obtaining data to be predicted, the data to be predicted including wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted; Extracting binding energy features based on the wild-type protein information to be predicted and the compound information to be predicted to obtain wild-type energy features to be predicted, and extracting binding energy features based on the mutant protein information to be predicted and the compound information to be predicted to obtain mutant energy features to be predicted; determining a target energy signature to be predicted based on the wild-type energy signature to be predicted and the mutant energy signature to be predicted; inputting the target energy feature to be predicted into a target prediction model to perform prediction and obtain interaction state information, wherein the target prediction model is obtained by obtaining a training sample set, determining a current training sample from the training sample set based on training sample weights, inputting a current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training, and when the basic training is completed, obtaining a basic prediction model, updating the training sample weights based on the basic prediction model until model training is completed, and performing the steps of determining a current training sample from the training sample set based on the training sample weights again; The pre-trained prediction model takes the target energy features as input and the interaction state information as output; The basic prediction model takes the target energy feature as an input and the interaction state information and the training sample weights as outputs; The method, wherein the target prediction model has the training samples and the target energy features as inputs and the interaction state information as output.

16. A device for training a predictive model, comprising: a sample acquisition module for acquiring a training sample set, the training sample set including each training sample, a training sample weight corresponding to each training sample, and a target energy feature corresponding to each training sample, the training samples including wild-type protein information, mutant protein information, and compound information, the target energy feature being acquired based on a wild-type energy feature and a mutant energy feature, the wild-type energy feature being acquired by performing binding energy feature extraction based on the wild-type protein information and the compound information, and the mutant energy feature being acquired by performing binding energy feature extraction based on the mutant protein information and the compound information; a sample determination module that determines a current training sample from the training sample set based on the training sample weights; a training module for inputting a current target energy feature corresponding to the current training sample into a pre-trained prediction model to perform basic training, and obtaining a basic prediction model when the basic training is completed; an iterative module for updating a training sample weight corresponding to each of the training samples based on the basic prediction model and determining a current training sample from the training sample set based on the training sample weight until model training is completed, and obtaining a target prediction model, the target prediction model being used to predict interaction state information corresponding to input protein information and input compound information; The pre-trained prediction model takes the target energy features as input and the interaction state information as output; The basic prediction model takes the target energy feature as an input and the interaction state information and the training sample weights as outputs; The target prediction model has the training samples and the target energy features as inputs and the interaction state information as output.

17. A data prediction device, comprising: A data acquisition module for acquiring data to be predicted, the data to be predicted including wild-type protein information to be predicted, mutant protein information to be predicted, and compound information to be predicted; a feature extraction module that performs binding energy feature extraction based on the wild-type protein information to be predicted and the compound information to be predicted, thereby obtaining wild-type energy features to be predicted, and performs binding energy feature extraction based on the mutant protein information to be predicted and the compound information to be predicted, thereby obtaining mutant energy features to be predicted; a target characteristic determination module for determining a target energy characteristic to be predicted based on the wild-type energy characteristic to be predicted and the mutant energy characteristic to be predicted; a prediction module for inputting the target energy feature to be predicted into a target prediction model to perform prediction and obtain interaction state information, wherein the target prediction model is obtained by: acquiring a training sample set; determining a current training sample from the training sample set based on a training sample weight; inputting a current target energy feature corresponding to the current training sample into a pre-training prediction model to perform basic training; when the basic training is completed, acquiring a basic prediction model; updating the training sample weight based on the basic prediction model until model training is completed; and performing the steps of determining a current training sample from the training sample set based on the training sample weight again; The pre-trained prediction model takes the target energy features as input and the interaction state information as output; The basic prediction model takes the target energy feature as an input and the interaction state information and the training sample weights as outputs; The target prediction model has the training samples and the target energy features as inputs and the interaction state information as output.

18. A computer program comprising computer readable instructions which, when executed by a processor, implement the steps of the method according to any of the claims 1 to 15.

Images