Compound activity prediction method, network training method, device, medium and equipment

Abstract

The present application discloses a compound activity prediction method, a network training method, a device, a medium and equipment, which can be applied to artificial intelligence, machine learning, deep neural network, meta-learning, drug analysis, lead compound discovery and other scenarios. The method includes: obtaining the target target protein corresponding to the test compound; determining the target feature of the target target protein according to the information of the tested active compound corresponding to the target target protein; inputting the target feature into the trained functional region determination network for processing to determine the target functional region corresponding to each layer of the basic neural network, and determining the activity prediction model of the target target protein according to the target functional region corresponding to each layer of the network in the basic neural network; predicting the test compound according to the activity prediction model of the target target protein to obtain the activity prediction result of the test compound for the target target protein.

Description

Technical Field

[0001] The present application relates to the field of information processing technology, and in particular to a compound activity prediction method, a network training method, a device, a medium and equipment. Background Art

[0002] Drug screening is a step in the modern drug development process to test and obtain compounds with specific physiological activity. It is mainly a process of selecting compounds with high activity against a specific target from a large number of compounds or new compounds through standardized experimental methods. The process of drug screening is essentially the process of conducting pharmacological activity experiments on compounds. With the development of drug development technology, the physiological activity experiments of new compounds have gradually changed from early verification experiments to screening experiments, which are so-called drug screening.

[0003] In the process of virtual screening based on molecular structure, it is necessary to detect the activity of candidate compounds against the target protein. The larger the activity value, the better the inhibitory effect of the candidate compound on the target protein, and the more likely it is to be selected as a drug against the target protein. At present, the activity detection method generally tests the compound repeatedly through multiple experiments to determine the activity value of the candidate compound against a target protein. However, the method of testing the activity of compounds through experiments involves a lot of repetitive work and consumes a lot of manpower and material resources. Summary of the invention

[0004] The embodiments of the present application provide a compound activity prediction method, a network training method, an apparatus, a medium and a device, which improve the accuracy of the activity prediction of the compound to be tested in the current target protein.

[0005] In one aspect, a method for predicting compound activity is provided, the method comprising: obtaining a target protein corresponding to the compound to be tested;

[0006] Determining the target feature of the target protein according to the measured active compound information corresponding to the target protein;

[0007] Inputting the target feature into a trained functional region determination network for processing to determine the target functional region corresponding to each layer of the basic neural network, wherein the trained functional region determination network is used to predict the selection probability of the target functional region;

[0008] Determining the activity prediction model of the target protein according to the target functional regions corresponding to each layer of the network in the basic neural network;

[0009] The test compound is predicted according to the activity prediction model of the target protein to obtain the activity prediction result of the test compound for the target protein.

[0010] In another aspect, a network training method for predicting compound activity is provided, the method comprising:

[0011] Acquire training sample data, wherein the training sample data includes historical data sets of all historical target proteins, wherein each of the historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein;

[0012] Using a target feature network, according to the molecular structure features of the historically tested active compounds of the historically tested target protein and the activity data of the historically tested active compounds on the historically tested target protein, the target features of the historically tested target protein are determined;

[0013] Inputting the target features of the historical target protein into a functional region determination network to determine the historical functional region corresponding to each layer of the network in the basic neural network;

[0014] Determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network;

[0015] The target feature network, the basic neural network and the functional area determination network are trained using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data to obtain a trained target feature network, a trained basic neural network and a trained functional area determination network.

[0016] In another aspect, a compound activity prediction device is provided, the device comprising:

[0017] An acquisition unit, used for acquiring the target protein corresponding to the compound to be tested;

[0018] A determination unit, configured to determine the target feature of the target protein according to the information of the active compound detected corresponding to the target protein; and

[0019] Inputting the target feature into a trained functional region determination network for processing to determine the target functional region corresponding to each layer of the basic neural network, wherein the trained functional region determination network is used to predict the selection probability of the target functional region; and

[0020] Determining the activity prediction model of the target protein according to the target functional regions corresponding to each layer of the network in the basic neural network;

[0021] The prediction unit is used to predict the test compound according to the activity prediction model of the target protein to obtain the activity prediction result of the test compound for the target protein.

[0022] In another aspect, a network training device for predicting compound activity is provided, the device comprising:

[0023] An acquisition unit, used for acquiring training sample data, wherein the training sample data includes historical data sets of all historical target proteins, wherein each of the historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein;

[0024] a determination unit, configured to determine the target feature of the historical target protein by using a target feature network, according to the molecular structure features of the historically tested active compounds of the historical target protein, and the activity data of the historically tested active compounds on the historical target protein; and

[0025] Inputting the target features of the historical target protein into a functional region determination network to determine the historical functional region corresponding to each layer of the basic neural network; and

[0026] Determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network;

[0027] A training unit is used to train the target feature network, the basic neural network and the functional area determination network using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data to obtain a trained target feature network, a trained basic neural network and a trained functional area determination network.

[0028] On the other hand, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, wherein the computer program is suitable for being loaded by a processor to execute the steps in the method for predicting compound activity as described in any of the above embodiments.

[0029] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the steps in the method for predicting compound activity as described in any of the above embodiments by calling the computer program stored in the memory.

[0030] On the other hand, a computer program product is provided, comprising computer instructions, which, when executed by a processor, implement the steps in the method for predicting compound activity as described in any of the above embodiments.

[0031] The embodiment of the present application obtains the target target protein corresponding to the compound to be tested; according to the measured active compound information corresponding to the target target protein, the target feature of the target target protein is determined; the target feature is input into the trained functional region determination network for processing to determine the target functional region corresponding to each layer of the network in the basic neural network, and the trained functional region determination network is used to predict the selected probability of the target functional region; according to the target functional region corresponding to each layer of the network in the basic neural network, the activity prediction model of the target target protein is determined; the compound to be tested is predicted according to the activity prediction model of the target target protein to obtain the activity prediction result of the compound to be tested for the target target protein. Compared with the existing activity prediction method that directly uses the data of other historical target proteins for training and prediction, the embodiment of the present application proposes a functional regionalized meta-learning algorithm to extract the target feature of the target protein to improve the accuracy of activity prediction in the current target protein. Further, by improving the accuracy of activity prediction, the quality of virtual screening of drug molecules can be guaranteed to a certain extent, and better and more accurate lead compounds can be found, so as to carry out the discovery and development of subsequent lead compounds and candidate compounds.

[0032] The embodiment of the present application obtains training sample data, and the training sample data includes historical data sets of all historical target proteins, wherein each historical data contains at least one historical target protein and activity data of historically tested active compounds on the historical target protein; a target feature network is used to determine the target features of the historical target protein according to the molecular structure features of the historically tested active compounds of the historical target protein, and the activity data of the historically tested active compounds on the historical target protein; the target features of the historical target protein are input into the functional region determination network to determine the historical functional regions corresponding to each layer of the network in the basic neural network; the network parameters of the basic neural network corresponding to the historical target protein are determined according to the historical functional regions corresponding to each layer of the network in the basic neural network; the target feature network, the basic neural network and the functional region determination network are trained using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data to obtain a trained target feature network, a trained basic neural network and a trained functional region determination network. The data of all tested compounds with known historical target proteins are fully utilized to train the activity prediction model, avoiding the problem that the learned meta-initial model is prone to overfitting and poor generalization to new data due to the small number of target proteins. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 A schematic diagram of the drug discovery process provided in the embodiments of the present application;

[0035] Figure 2 A schematic diagram of a method for predicting compound activity provided in an embodiment of the present application;

[0036] Figure 3 Another schematic diagram of the process of predicting the activity of a compound provided in an embodiment of the present application;

[0037] Figure 4 Another schematic diagram of a method for predicting compound activity provided in an embodiment of the present application;

[0038] Figure 5a A schematic diagram of another process of the method for predicting compound activity provided in the embodiments of the present application;

[0039] Figure 5b An example diagram of an application scenario of the compound activity prediction method provided in an embodiment of the present application;

[0040] Figure 6 A schematic diagram of a network training method for predicting compound activity provided in an embodiment of the present application;

[0041] Figure 7a An example diagram of an application scenario of the network training method for predicting compound activity provided in an embodiment of the present application;

[0042] Figure 7b An example diagram of an application scenario of the network training method for predicting compound activity provided in an embodiment of the present application;

[0043] Figure 8 A schematic diagram of the structure of a compound activity prediction device provided in an embodiment of the present application;

[0044] Fig. 9 A schematic diagram of the structure of a network training device for predicting compound activity provided in an embodiment of the present application;

[0045] Fig.10 A schematic structural diagram of a compound activity prediction device provided in an embodiment of the present application;

[0046] Fig.11 A schematic structural diagram of a network training device for predicting compound activity provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0048] The embodiments of the present application provide a method for predicting compound activity, a network training method, an apparatus, a medium and a device. Specifically, the method of the embodiments of the present application can be executed by a computer device, wherein the computer device can be a terminal or a server. The embodiments of the present application can be applied to scenarios such as artificial intelligence, machine learning, deep neural networks, meta-learning, drug analysis, and lead compound discovery. At the same time, it is also suitable for predicting other properties of drug molecules, including drug absorption, distribution, metabolism, excretion, and toxicity.

[0049] First, some nouns or terms that appear in the description of the embodiments of the present application are explained as follows:

[0050] Artificial Intelligence (AI): It is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision-making.

[0051] Machine Learning (ML): It is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning and other technologies.

[0052] Artificial neural networks (ANNs): An algorithmic mathematical model that imitates the behavioral characteristics of animal neural networks and performs distributed parallel information processing. It processes input information through the network parameters of a large number of nodes (or neurons) and the interconnected relationships between nodes.

[0053] Deep Neural Network (DNN): A neural network with at least one hidden layer. Similar to shallow neural networks, deep neural networks can also provide modeling for complex nonlinear systems, but the extra layers provide a higher level of abstraction for the model, thereby improving the model's capabilities. In the process of compound activity prediction in the embodiment of the present application, an activity prediction model based on machine learning or deep learning is used to learn the correlation between the characteristic information of the molecular structure of the tested active compound and the activity data of the tested active compound on the target target protein, so that the activity prediction results of the tested compound can be generated based on the input molecular structure characteristics of the tested compound. Among them, the activity prediction model is an improvement based on the deep neural network model.

[0054] Recursive Neural Network (RNN): It is an artificial neural network with a tree-like hierarchical structure and network nodes that recursively input information according to their connection order. It is one of the deep learning algorithms. In the functional area determination network, the embodiment of the present application uses a recursive neural network RNN ​​to predict the selection probability of the target functional area, divides each hidden layer of the trained basic neural network into multiple functional areas, uses RNN to predict the probability of a functional area of ​​each layer being selected, and inputs the target feature into the RNN to obtain the selected functional area, that is, the target functional area.

[0055] Morgan Molecular Fingerprints: used to describe the characteristics of the molecular structure of a substance, especially the characteristics related to activity in the molecular structure. Morgan molecular fingerprints are a kind of circular fingerprint, which is also a topological fingerprint. Similar to Extended-Connectivity Fingerprints (ECFPs), each element in the fingerprint represents a specific substructure. The embodiment of the present application uses Morgan molecular fingerprints to describe the molecular structure characteristics related to activity in the molecular structure of a compound. The Morgan molecular fingerprint of the compound can be obtained by processing the molecular structure of the compound through the Morgan algorithm.

[0056] Meta-learning: The mapping relationship between the state characteristics and quality parameters of the neural network in each stage of the machine learning framework can be mined through supervised learning, and the performance of the neural network can be optimized according to the characteristics of the new learning task. The core idea of ​​meta-learning is to learn the initial parameters of the neural network from a large number of training tasks. The initial parameters can enable the new machine learning task to quickly converge to a better solution under small sample conditions. The present application adopts a meta-learning method to form a basic activity prediction model based on the target functional region determined by the functional region determination network, and learns the initial parameters of the deep neural network model based on the activity data of a variety of target proteins with known activities. Based on the initial parameters, a relatively small amount of activity data of the target target protein can be used to train the activity prediction network model corresponding to the target target protein.

[0057] Blockchain system: It can be a distributed system formed by connecting clients and multiple nodes (any form of computing devices connected to the network, such as servers and user terminals) through network communication. The nodes form a peer-to-peer (P2P) network. The P2P protocol is an application layer protocol running on the Transmission Control Protocol (TCP). In a distributed system, any machine such as a server or terminal can join and become a node. The node includes the hardware layer, the middle layer, the operating system layer and the application layer.

[0058] See also Figure 1 , the figure shows the process of drug discovery, which may include: target identification and confirmation, lead compound discovery, lead compound discovery and optimization, candidate compound confirmation and development, and clinical trial stage. The lead compound discovery and lead compound discovery stages include virtual screening based on molecular structure and virtual screening based on target structure. The compound activity prediction method and network training method of the present application can be applied to virtual screening based on molecular structure, which can provide services for the pharmaceutical process and accelerate the discovery process of lead compounds using artificial intelligence algorithms.

[0059] The process of discovering lead compounds includes virtual screening based on molecular structure. Compared with traditional experimental screening, virtual screening by computational methods does not require the consumption of compound samples, greatly saving manpower and material resources and accelerating the screening process. Ligand-based drug design is one of the common methods of virtual screening. It refers to learning and establishing a model of the relationship between molecular structure and activity based on the known active ligand small molecule structure, so as to predict the model of new compounds.

[0060] At present, there are two main methods for predicting the activity of drug molecules for target proteins. One is to use the prediction model obtained by training each historical target protein to predict the training drug molecules with measured activity of the current target protein. In this method, the machine learning model of each historical target protein is a random forest, and the model of the target target protein is a partial least squares regression. The model expression ability of the two models is weak.

[0061] The other is multi-task learning technology, that is, all target proteins use a deep neural network model and share the underlying deep neural network parameters to achieve knowledge transfer, while the high-level parameters are unique to each target protein to adapt to the differences between proteins. This type of method requires each target protein to have unique high-level parameters due to multi-task learning, and these parameters are trained with the measured activity data of the target protein. Considering the limited measured activity data, it is easy to cause overfitting problems. In addition, learning tasks with low correlation together may lead to low model stability and even damage the prediction effect of the model.

[0062] Another type is meta-learning technology, which is to learn a common initial model, and quickly update the initial model to the prediction model of the target through the training data of each target of the historical non-target protein, and then optimize the initial model with the test data to make the updated prediction model of each target have the best performance. The initial model is finally applied to the target protein, and the final prediction model of the target target is obtained by quickly updating the training data. However, due to the small number of target proteins, the learned meta-initial model is prone to overfitting and the generalization effect to new data is poor. At the same time, the knowledge of target proteins with low similarity to the target target may be transferred, resulting in negative effects.

[0063] This application is aimed at virtual screening based on molecular structure to predict the activity data of the test compound, which can provide services to pharmaceutical companies and thus accelerate the discovery process of their lead compounds.

[0064] The key problem to be solved by the embodiments of the present application is that, in the process of using meta-learning to predict the activity value of the characteristics of the drug molecular structure, due to the small number of target proteins, the learned meta-initial model is prone to overfitting, and may transfer the knowledge of target proteins with low similarity to the target target, causing negative effects. The embodiments of the present application propose a functional regionalization meta-learning algorithm in ligand-based drug design, which uses highly correlated data in other historical target protein activity data to summarize the historical target protein data with measured activity into different functional regions according to similarity, thereby making full use of highly correlated data in other historical target protein activity data to improve the accuracy of activity prediction in the target protein.

[0065] In order to better understand the technical solution provided in the embodiment of the present application, the following briefly introduces the application scenarios to which the technical solution provided in the embodiment of the present application is applicable. It should be noted that the application scenarios introduced below are only used to illustrate the embodiment of the present application and are not limited. Take the compound activity prediction method executed by a computer device as an example, wherein the computer device can be a terminal or a server or other device.

[0066] During the compound activity prediction stage, the user can upload the information of the compound to be tested through the client, browser client or instant messaging client installed in the computer device. After the computer device obtains the uploaded compound information, the target characteristics of the target target protein are determined according to the information of the tested active compounds corresponding to the target target protein; the target characteristics are input into the trained functional region determination network for processing to determine the target functional region corresponding to each layer of the network in the basic neural network, and the trained functional region determination network is used to predict the selection probability of the target functional region; according to the target functional regions corresponding to each layer of the network in the basic neural network, the activity prediction model of the target target protein is determined; the compound to be tested is predicted according to the activity prediction model of the target target protein to obtain the activity prediction result of the compound to be tested for the target target protein.

[0067] During the training phase, the computer device acquires training sample data, which includes historical data sets of all historical target proteins, wherein each historical data contains at least one historical target protein and activity data of historically tested active compounds on the historical target protein; a target feature network is used to determine the target features of the historical target protein based on the molecular structure features of the historically tested active compounds of the historical target protein and the activity data of the historically tested active compounds on the historical target protein; the target features of the historical target protein are input into the trained functional region determination network to determine the historical functional region; based on the historical functional region corresponding to each layer of the network in the basic neural network, an activity prediction model corresponding to the historical target protein is determined; and the target feature network and the functional region determination network are trained using the training sample data.

[0068] It should be noted that the compound activity prediction process, the training process of the activity prediction model, and the actual prediction process can be completed in the server or in the terminal. When the training process and the actual prediction process of the model are completed in the server, when the trained activity prediction model needs to be used, the compound to be tested can be input into the server, and after the server actually predicts, the activity value of the compound to be tested is sent to the terminal for display. When the training process and the actual prediction process of the model are completed in the terminal, when the trained activity prediction model needs to be used, the compound to be tested can be input into the terminal, and after the terminal actually predicts, the terminal displays the activity value of the compound to be tested. When the training process of the model is completed in the server, and the actual prediction process of the model is completed in the terminal, when the trained drug analysis model needs to be used, the compound to be tested can be input into the terminal, and after the terminal actually predicts, the terminal displays the activity value of the compound to be tested. Optionally, the trained model file (model file) in the server can be transplanted to the terminal. If the activity value of the input compound to be tested needs to be predicted, the compound to be tested is input into the trained model file (model file), and the activity value of the compound to be tested can be obtained by calculation.

[0069] Among them, the embodiments of the present application can be implemented in combination with cloud technology or blockchain network technology. For example, the compound activity prediction method disclosed in the embodiments of the present application, wherein these data can be stored on the blockchain. For example, activity prediction models, basic neural networks, target feature networks, functional region determination networks, feature extraction networks, fully connected networks, recursive neural networks, target target protein tested compound data sets, historical target protein historical data sets can all be stored on the blockchain.

[0070] In order to facilitate the storage and query of the target target protein corresponding to the test compound, the target feature of the target target protein, the trained functional region determination network, the target functional region corresponding to each layer of the network in the trained basic neural network, the activity prediction model of the target target protein, and the activity prediction result of the test compound for the target target protein, optionally, the compound activity prediction method also includes: sending the target target protein corresponding to the test compound, the target feature of the target target protein, the trained functional region determination network, the target functional region corresponding to each layer of the network in the trained basic neural network, the activity prediction model of the target target protein, and the activity prediction result of the test compound for the target target protein to the blockchain network, so that the nodes of the blockchain network fill the target target protein corresponding to the test compound, the target feature of the target target protein, the trained functional region determination network, the target functional region corresponding to each layer of the network in the trained basic neural network, the activity prediction model of the target target protein, and the activity prediction result of the test compound for the target target protein into the new block, and when consensus is reached on the new block, the new block is appended to the end of the blockchain. The embodiments of the present application can store the target target protein corresponding to the test compound, the target feature of the target target protein, the trained functional region determination network, the target functional region corresponding to each layer of the trained basic neural network, the activity prediction model of the target target protein, and the activity prediction result of the test compound for the target target protein on the chain to achieve record backup. When it is necessary to obtain the activity value of the predicted compound equivalent to the target target protein, the corresponding activity prediction result of the test compound for the target target protein can be directly and quickly obtained from the blockchain, thereby improving the efficiency of compound activity prediction.

[0071] It should be noted that the order of description of the following embodiments is not intended to limit the priority order of the embodiments.

[0072] Each embodiment of the present application provides a method for predicting the activity of a compound. The method can be executed by a terminal or a server, or by both a terminal and a server. The embodiments of the present application are described by taking the compound activity prediction executed by a server as an example.

[0073] See also Figure 2 , Figure 2 A schematic diagram of a method for predicting compound activity provided in an embodiment of the present application, the method comprising:

[0074] Step 210, obtaining the target protein corresponding to the compound to be tested.

[0075] As mentioned above, the user can input the compound to be tested and the target protein through the terminal device. Among them, the terminal device can be a mobile terminal, a PDA (Personal Digital Assistant), a computer, a notebook, etc. A compound activity prediction client is installed on the terminal device, and the compound activity prediction client has a functional module for predicting the activity of the compound. When the user needs to screen the lead compound for the target protein during the new drug development process, the user opens the compound activity prediction client on the terminal device, and the user can enter the information of the compound to be tested and the target protein in the corresponding page.

[0076] Step 220, determining the target feature of the target protein according to the information of the active compounds that have been tested corresponding to the target protein.

[0077] Specifically, for the target protein input by the user, a small amount of information on active compounds tested corresponding to the target protein can be obtained, and the information on active compounds tested can be the molecular structure of the tested active compound, characteristic information on the molecular structure of the tested active compound, and activity data of the tested active compound on the target protein, etc. Among them, the data on a small amount of active compounds tested corresponding to the target protein can be stored in a database, and extracted from the database according to the target protein input by the user.

[0078] After obtaining the information of the tested active compounds corresponding to the target protein, feature extraction can be performed on the tested active compound information to determine the target features of the target protein.

[0079] Optionally, the step of determining the target feature of the target protein according to the information of the tested active compound corresponding to the target protein includes:

[0080] The molecular structure characteristics of the tested active compounds and the activity data of the tested active compounds on the target protein are input into the trained target feature network for processing to determine the target features of the target protein. The trained target feature network is trained based on the similarity between the activity data of all historical target proteins.

[0081] Among them, the molecular structure characteristics of the active compound refer to the characteristic information of the molecular structure of the active compound, and the characteristic information of the molecular structure can be represented by the Morgan molecular fingerprint of the compound to be tested. The generation process of the Morgan molecular fingerprint can include the following steps: atom initialization, iterative update, and feature generation. The molecular structure information of the compound to be tested contains the arrangement structure information of atoms. Atom initialization refers to assigning an integer identifier to each atom. For example, a fixed hash function can be applied to the connection features of an atom and the adjacent area of ​​the previous layer of the network to generate a feature representing the atom, and the output result of the hash function is used as the integer identifier of the atom. The iterative update is centered on each atom and merges the atoms in a circle around it until the specified radius is reached to form a substructure. Feature generation operates on the substructure and generates a feature list. According to the generated feature list, the Morgan molecular fingerprint of the compound to be tested is obtained.

[0082] The trained target feature network can be trained based on the similarity between the activity data of all historical target proteins. The historical target proteins include a large number of target protein data with measured activity in existing public data sets, and the historical target proteins may include data sets of target target proteins and non-target target proteins.

[0083] Optional, such as Figure 3 As shown, the step of inputting the molecular structure characteristics of the tested active compound and the activity data of the tested active compound on the target protein into the trained target feature network for processing to determine the target feature of the target protein can be achieved through steps 211 to 213, which specifically include:

[0084] Step 211, using a feature extraction network to extract features of the molecular structure features of the tested active compound to obtain intermediate features of the tested active compound.

[0085] Specifically, the feature extraction network can be composed of all the training data x of the target protein s , after the feature extraction network, the feature extraction network It can be composed of a three-layer multi-layer perceptron (MLP), and the two middle hidden layers each contain a preset number of neurons, such as 500 neurons. Feature extraction is performed on the molecular structure features of the tested active compounds to obtain the intermediate features of the tested active compounds. The intermediate features characterize the target features of the target protein to a certain extent.

[0086] Step 212, concatenate the intermediate features of the tested active compounds and the activity data of the tested active compounds on the target protein.

[0087] Step 213, input the serially connected data into a fully connected network including a preset number of neurons, and average the output results of the fully connected layer corresponding to the molecular structures of all tested active compounds to obtain the target features of the target protein.

[0088] Specifically, after obtaining the intermediate features, the intermediate features of the tested active compounds and the activity data y of the tested active compounds on the target protein are combined. s Concatenate. Concatenation means connecting the intermediate features and activity values ​​together to form a new data vector. For example, if the intermediate features are 1024-dimensional and the activity value is 1-dimensional, the concatenation results in a 1025-dimensional data vector.

[0089] The concatenated data is input into a fully connected network containing a preset number of neurons. For example, the preset number can be 64. The concatenated data is passed through a fully connected layer MF containing 64 neurons to obtain the target feature. It can be expressed as the following formula (1):

[0090]

[0091] in, It is the intermediate feature output by the feature extraction network and the activity data y of the target protein s MF() represents the output of the fully connected network.

[0092] Furthermore, the output results of the fully connected layer corresponding to the molecular structures of all tested active compounds are averaged to obtain the target features of the target protein. The target feature network can be expressed as the following formula (2):

[0093]

[0094] Where k represents the k drug molecules of the tested compound. If the drug molecules on different target proteins are different, then averaging can reduce the impact of different drug molecules. On the contrary, if they are all the same, then averaging is not necessary. a Represents the parameters of the target feature network, including the feature extraction network , and the parameters of the fully connected network MF.

[0095] In this way, the target characteristics of the target protein are characterized by combining the intermediate features extracted from the molecular structure characteristics of the tested compound and the activity values ​​corresponding to the molecular structure characteristics of the tested compound. This can avoid the possibility that the same molecular structure characteristics may appear in different targets but have different activity values, thereby improving the accuracy of the target protein characteristics characterized by the target characteristics.

[0096] Step 230, input the target feature into the trained functional area determination network for processing to determine the target functional area corresponding to each layer of the trained basic neural network. The trained functional area determination network is used to predict the selection probability of the target functional area.

[0097] Specifically, in order to determine the target functional area of ​​each layer of the activity prediction model, the present application proposes a functional area determination network to predict the probability of each functional area being selected. The functional area determination network can use a recursive neural network RNN ​​for prediction, and divide the network of each hidden layer in the trained basic neural network into multiple functional areas. Optionally, each layer of the deep neural network is decomposed into three undetermined functional areas. When the user gives a new target protein, the target feature is input into the functional area determination network to output the most relevant target functional area of ​​each layer. In this way, the data of other target proteins with measured activity can be summarized into different functional areas according to similarity, thereby making full use of highly correlated data in the activity data of other target proteins to improve the accuracy of activity prediction in the current target protein.

[0098] Among them, the trained functional region determination network can be trained by existing historical target protein data, or it can be a functional region determination network with given initial parameter values, and the functional region determination network is established in the predictive meta-learning process.

[0099] Among them, the trained basic neural network can be a multi-layer deep neural network, including an input layer, a hidden layer and an output layer. The basic neural network can be obtained by training the data of existing historical target proteins, or it can be a basic neural network with given initial parameter values, and the basic neural network is established in the predicted meta-learning process. Exemplarily, the basic neural network can be constructed as a preset number of deep neural networks, such as a four-layer deep neural network, and the three middle hidden layers each contain a preset number of neurons, such as 500 neurons. For the number of network layers of the basic neural network, the number can be selected according to actual conditions. It should be noted that too few layers may cause underfitting, and too many layers may cause overfitting, resulting in poor prediction results. Optionally, the present application can construct a four-layer deep neural network.

[0100] Optional, such as Figure 4 As shown, each layer of the trained basic neural network includes multiple undetermined functional areas, and step 230 can be implemented by steps 231 to 233, which specifically include:

[0101] Step 231, determining the series feature according to the target feature and the selected probability corresponding to all the pending functional areas of the previous layer network in the basic neural network.

[0102] Step 232, determining the latent features of the current layer network according to the series features and the latent features of the previous layer network in the basic neural network.

[0103] Step 233, using the trained functional region determination network to process the latent features of the current layer network, so as to determine the target functional region of the current layer network from multiple pending functional regions.

[0104] Specifically, in order to determine the target functional area of ​​each layer of the activity prediction model, a trained functional area determination network can be used to predict the probability of each pending functional area being selected. A recursive neural network RNN ​​can be used for prediction. RNN can be used to process ordered data and output predictions. The functional area determination network is ordered, and RNN is more suitable for use as a functional area determination network.

[0105] The basic neural network can be constructed as a preset number of deep neural networks, and each layer of the basic neural network is divided into a preset number of pending functional areas.

[0106] Optionally, each network layer in the basic neural network is divided into three to-be-determined functional areas, and the basic neural network has a four-layer network structure.

[0107] For example, the basic neural network has l layers of networks, each layer of the network has c pending functional areas, and the probability of the c pending functional areas in the lth layer being selected is The input of the network is the characteristics of the target And the probability p of all pending functional areas of the previous layer network being selected l-1 The concatenated features of the previous layer of network, and the hidden features h l-1 , and then output the hidden feature h of this layer l ,The latent feature can be expressed as formula (3):

[0108]

[0109] Among them, w b Represents the network parameters of the RNN network.

[0110] Optionally, the step of using a trained functional region determination network to process latent features of the current layer network to determine a target functional region of the current layer network from a plurality of pending functional regions includes:

[0111] The hidden features of all the pending functional areas in each layer of the basic neural network are normalized through the trained functional area determination network to obtain the corresponding selection probabilities of all the pending functional areas;

[0112] The target functional area of ​​the current layer network is determined according to the undetermined functional area corresponding to the maximum value among the selected probabilities corresponding to all the undetermined functional areas.

[0113] For example, the probability of selecting the c undetermined functional areas in the lth layer is According to the latent feature h l After normalization by softmax function, for example, It can be expressed as formula (4):

[0114]

[0115] Further, take The value with the largest probability value is expressed as And select the maximum probability value The corresponding pending functional area is expressed as As the target functional area of ​​the lth layer.

[0116] Step 240, determining the activity prediction model of the target protein according to the target functional region corresponding to each layer of the trained basic neural network.

[0117] When the target functional area of ​​each layer of the network is determined, the target functional area of ​​each layer of the network is connected according to the weight, or connected in sequence to determine the activity prediction model of the target protein. Among them, the activity prediction model can be used to predict the test compound input by the user to obtain the activity prediction result of the test compound for the target protein. In this way, by generating an activity prediction model specific to the target protein based on the correlation between a large number of historical target proteins and using the information of a small number of tested compounds of the target protein, the accuracy of activity prediction in the current target protein can be effectively improved.

[0118] Optional, see Figure 5a , according to the target functional region corresponding to each layer of the trained basic neural network, the steps of determining the activity prediction model of the target protein include:

[0119] Step 241, sequentially connect the target functional regions corresponding to each layer of the basic neural network to obtain a basic activity prediction model for the target protein.

[0120] Determine the target functional area of ​​each layer of the basic neural network Finally, this application connects each layer in sequence to obtain the basic activity prediction model of the target protein:

[0121]

[0122] For example, see Figure 5b , the target functional area predicted according to the target characteristics is Connect the above target functional areas in sequence to obtain Figure 5b The basic activity prediction model shown.

[0123] Step 242, using the gradient descent method to adjust the basic activity prediction model of the target protein at least once.

[0124] Specifically, the gradient descent method can be obtained by automatic gradient derivation using a deep learning framework such as PyTorch. For example, the basic activity prediction model of the target protein can be adjusted using the gradient descent method as shown in formula (5):

[0125]

[0126] Wherein, φ is the network parameter of the basic activity prediction model of the target protein obtained in the above step 241, and α is the learning rate of gradient optimization. Optionally, α can be set to 0.01. (x s ,y s ) is the training data set of the tested compounds of the target protein The molecular structure characteristics x of the tested active compounds output by the basic activity prediction model of the target protein before adjustment s Predicted activity of target protein; i-1 is the network parameter of the basic activity prediction model of the target protein before adjustment; i The network parameters of the basic activity prediction model for the adjusted target protein.

[0127] Assuming that the initial value of the network in the gradient descent method is θ, during the model adjustment process, the network parameters of the activity prediction model corresponding to each target protein are obtained by gradient optimization through several steps starting from the value θ.

[0128] That is, the objective function is differentiated with respect to φ, and gradient descent is performed through the differentiation to make the objective function converge, so as to quickly update the network parameter φ of the basic activity prediction model.

[0129] For example, in the fast update process, the optimization starts from the network initial value θ, and after one step of gradient calculation, φ is obtained. 1 , φ 1 Replace the network initial value θ for the second step of gradient descent, and so on.

[0130] in, It can be expressed as formula (6):

[0131]

[0132] Here, k represents the k drug molecules in the tested compounds. represents the kth molecular structure feature of the tested active compound, express The measured activity values. represents the adjusted basic activity prediction model φ i The predicted activity value.

[0133] Step 243, determining the activity prediction model of the target protein according to the adjusted basic activity prediction model.

[0134] Specifically, the basic activity prediction model φ adjusted by the gradient descent method i , determine the activity prediction model of the target protein.

[0135] Step 250 , predicting the test compound according to the activity prediction model of the target protein to obtain the activity prediction result of the test compound on the target protein.

[0136] Specifically, the test compound x q Input the data of a small number of compounds tested on the target protein into the activity prediction model to obtain the activity prediction result of the target protein. φ (x q ).

[0137] All of the above technical solutions can be arbitrarily combined to form optional embodiments of the present application, which will not be described in detail here.

[0138] The embodiment of the present application obtains the target target protein corresponding to the compound to be tested; determines the target feature of the target target protein according to the information of the measured active compound corresponding to the target target protein; inputs the target feature into the trained functional region determination network for processing to determine the target functional region corresponding to each layer of the network in the basic neural network, and the trained functional region determination network is used to predict the selected probability of the target functional region; determines the activity prediction model of the target target protein according to the target functional region corresponding to each layer of the network in the basic neural network; predicts the compound to be tested according to the activity prediction model of the target target protein to obtain the activity prediction result of the compound to be tested for the target target protein. Compared with the existing activity prediction method that directly uses the data of other historical target proteins for training and prediction, the embodiment of the present application proposes a functional regionalized meta-learning algorithm, which extracts the target feature of the target protein, and summarizes the measured activity data into different functional regions according to the similarity according to the target feature, thereby making full use of the highly correlated data in the activity data of other target proteins to improve the accuracy of activity prediction in the current target protein. Furthermore, by improving the accuracy of activity prediction, the quality of virtual screening of drug molecules can be guaranteed to a certain extent, and better and more accurate lead compounds can be discovered, thereby enabling the discovery and development of subsequent lead compounds and candidate compounds.

[0139] Each embodiment of the present application also provides a network training method for predicting compound activity. The method can be executed by a terminal or a server, or by both a terminal and a server. The embodiments of the present application are described by taking the training method executed by a server as an example.

[0140] See also Figure 6 , Figure 6 A first flow chart of a network training method for compound activity prediction provided in an embodiment of the present application, the method comprising:

[0141] Step 610, obtaining training sample data, the training sample data including historical data sets of all historical target proteins, wherein each historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein.

[0142] Specifically, the training sample data includes historical data sets of all historical target proteins. The historical target proteins include known target proteins that are currently disclosed or included in the database. Each historical data contains at least one historical target protein and the activity data of historically tested active compounds on the historical target protein.

[0143] Step 620, using a target feature network, according to the molecular structure features of the historically tested active compounds of the historical target protein, and the activity data of the historically tested active compounds on the historical target protein, determine the target features of the historical target protein.

[0144] The molecular structure characteristics of the active compounds tested in the past are the molecular structure characteristics of the active compounds tested corresponding to a certain historical target protein. The molecular structure characteristics are the same as described above, and refer to the characteristic information of the molecular structure of the active compound. The characteristic information of the molecular structure can be represented by the Morgan molecular fingerprint of the compound to be tested.

[0145] Optionally, the target feature network includes a feature extraction network and a fully connected network. The target feature network is used to determine the target feature of the historical target protein according to the molecular structure features of the historically tested active compounds of the historical target protein and the activity data of the historically tested active compounds on the historical target protein, including:

[0146] A feature extraction network is used to extract the molecular structure features of the active compounds that have been tested in the past, so as to obtain the intermediate features of the active compounds that have been tested in the past;

[0147] Concatenate the intermediate features of historically tested active compounds and the activity data of historically tested active compounds on historical target proteins;

[0148] The concatenated data are input into a fully connected network containing a preset number of neurons, and the output results of the fully connected layer corresponding to the molecular structures of all historically tested active compounds are averaged to obtain the target features of the historical target proteins.

[0149] All historical target proteins N in the training sample data are processed to obtain the target features of their corresponding historical target proteins The target feature network can refer to formula (2) in the above prediction method:

[0150]

[0151] Where k represents the k drug molecules of the tested compound. If the drug molecules on different target proteins are different, then averaging can reduce the impact of different drug molecules. On the contrary, if they are all the same, then averaging is not necessary. a Represents the parameters of the target feature network, including the feature extraction network , and the parameters of the fully connected network MF.

[0152] Step 630, input the target features of the historical target protein into the trained functional region determination network to determine the historical functional region.

[0153] Optionally, each network layer in the basic neural network includes a plurality of undetermined functional regions, and the target features of the historical target protein are input into the trained functional region determination network to determine the historical functional regions, including:

[0154] The series features are determined according to the selected probabilities corresponding to the target features and all pending functional areas of the previous layer of the basic neural network.

[0155] The latent features of the current layer network are determined according to the concatenated features and the latent features of the previous layer network in the basic neural network.

[0156] The trained functional region determination network is used to process the latent features of the current layer network to determine the target functional region of the current layer network from multiple pending functional regions.

[0157] Optionally, each network layer in the basic neural network can be divided into three to-be-determined functional areas, and the basic neural network has a four-layer network structure.

[0158] Optionally, the step of using a trained functional region determination network to process latent features of the current layer network to determine a target functional region of the current layer network from a plurality of pending functional regions includes:

[0159] The hidden features of all the pending functional areas in each layer of the basic neural network are normalized through the trained functional area determination network to obtain the corresponding selection probabilities of all the pending functional areas;

[0160] The target functional area of ​​the current layer network is determined according to the undetermined functional area corresponding to the maximum value among the selected probabilities corresponding to all the undetermined functional areas.

[0161] The specific implementation method can refer to the embodiments corresponding to the above-mentioned compound activity prediction method, which will not be described in detail here.

[0162] Step 640, determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network.

[0163] Among them, the method of determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional area corresponding to each layer of the network in the basic neural network is the same as the determination method in the above prediction method, and the determined network parameters can be expressed as:

[0164]

[0165] Optionally, the target functional regions corresponding to each layer of the basic neural network are sequentially connected, and the gradient descent method is used to perform at least one adjustment to determine the network parameters of the basic neural network corresponding to the historical target protein. The gradient descent method can refer to formula (5) in the above prediction method:

[0166]

[0167] Where φ is the network parameter of the basic neural network corresponding to the historical target protein, α is the learning rate of gradient optimization, and α can be set to 0.01. (x s ,y s ) is the training dataset of the tested compounds of the historical target protein The molecular structure features x of the tested active compounds output by the basic neural network corresponding to the historical target protein s Predicted activity against historical target proteins. i-1 is the network parameter of the basic neural network corresponding to the historical target protein after the i-1th round of parameter adjustment; φ i are the network parameters of the basic neural network corresponding to the historical target protein after the i-th round of parameter adjustment.

[0168] Assuming that the initial value of the network in the gradient descent method is θ, during the model training process, the network parameters of the basic neural network corresponding to each target protein are obtained by gradient optimization through several steps starting from the value θ.

[0169] That is, the objective function is differentiated with respect to φ, and gradient descent is performed through the differentiation to make the objective function converge, so as to quickly update φ.

[0170] For example, in the fast update process, the optimization starts from the network initial value θ, and after one step of gradient calculation, φ is obtained. 1 , φ 1 Replace the network initial value θ for the second step of gradient descent, and so on.

[0171] Step 650, using the network parameters and training sample data of the basic neural network corresponding to the historical target protein, the target feature network, the basic neural network and the functional area determination network are trained to obtain a trained target feature network, a trained basic neural network and a trained functional area determination network.

[0172] Optionally, the steps of training the target feature network, the basic neural network and the functional region determination network using the network parameters and training sample data of the basic neural network corresponding to the historical target protein to obtain the trained target feature network, the trained basic neural network and the trained functional region determination network include:

[0173] Inputting the molecular structure characteristics of the historically tested active compounds of the historical target protein into the basic neural network corresponding to the historical target protein with network parameters, and obtaining the historical predicted activity of the historically tested active compounds on the historical target protein;

[0174] The first objective function is determined based on the historical predicted activity of historically tested active compounds against historical target proteins and the activity data of historically tested active compounds against historical target proteins.

[0175] Specifically, the molecular structure characteristics x of the active compounds tested in history are k Input the basic neural network corresponding to the historical target protein to obtain the historical predicted activity of the tested active compounds against the historical target protein output by the activity prediction network model to be trained For example, the first objective function can be expressed as formula (6):

[0176]

[0177] Here, k represents the k drug molecules in the tested compounds. represents the kth molecular structure feature of the tested active compound, express The measured activity values. Indicates the molecular structure characteristics of historically tested active compounds Input the historical target protein corresponding to the basic neural network to obtain the historical predicted activity. i Represents the network parameters of the basic neural network corresponding to the historical target protein after adjustment by the gradient descent method.

[0178] The network parameters of the basic neural network are adjusted according to the determined first objective function until the training end condition is met.

[0179] Specifically, the training end condition may be that the loss value converges to a preset target value. In some other embodiments, the training end condition may also be that a preset number of training times is reached.

[0180] Optionally, the step of training the target feature network, the basic neural network and the functional region determination network using the network parameters and training sample data of the basic neural network corresponding to the historical target protein to obtain the trained target feature network, the trained basic neural network and the trained functional region determination network also includes:

[0181] (1) The historical data set is randomly divided into multiple non-overlapping historical data subsets, and a target feature network is used to extract features from the multiple historical data subsets to obtain target features of the multiple historical data subsets.

[0182] For example, all historical data sets The data are randomly divided into n non-overlapping subsets, and then the target feature network is used to extract features from the n non-overlapping subsets to obtain target features of multiple historical data subsets.

[0183] (2) According to the similarity of the target features of the same historical target protein and the similarity of the target features of different historical target proteins in the target features of the multiple historical data subsets, a second objective function is determined, and the second objective function is optimized to train the target feature network. For example, the second objective function can use the objective function of contrastive learning, which is expressed as formula (7):

[0184]

[0185] in, and Represented as two subsets of the same historical target protein i. and Representation Subset and The characteristics of their respective targets. and Represents two subsets of historical target protein i and historical target protein j respectively and The corresponding target features. N means there are N historical target proteins.

[0186] By the second objective function The optimization of minimizing the second objective function can make the formula The larger the similarity of the features from different subsets of the same historical target protein, the higher the similarity of the target features from different subsets of the same historical target protein. At the same time, the smaller the similarity of the features from all subsets of different historical target proteins.

[0187] Optionally, the step of training the target feature network, the basic neural network and the functional region determination network using the network parameters and training sample data of the basic neural network corresponding to the historical target protein to obtain the trained target feature network, the trained basic neural network and the trained functional region determination network also includes:

[0188] Determine a third objective function based on the first objective function and the second objective function;

[0189] The third objective function is optimized to train the basic neural network, the target feature network, and the functional area determination network.

[0190] For example, the third objective function can be expressed as formula (8):

[0191]

[0192] Among them, θ is the network initial value in the gradient descent method, w a is the network parameter of the target feature network, w b Determine the network parameters of the network for the functional area. s and s Represents the molecular structure characteristics and activity values ​​of the tested compounds corresponding to the historical target proteins in the historical data set. i It is represented by the network parameters of the basic neural network corresponding to the historical target protein after the i-th round of parameter adjustment. As mentioned above, it can be seen that φ i Contains the network initial value θ, and the network parameters w in the functional area to determine the network b . Represents x s The predicted activity value outputted by the basic neural network corresponding to the adjusted historical target protein. N represents N historical target proteins.

[0193] The first objective function can be expressed by the above formula (2). Represents the second objective function of the target feature network, including the network parameters w of the target feature network a , which can be expressed by the above formula (3).

[0194] When the target functional regions corresponding to each layer of the basic neural network are connected sequentially, the historical target protein network parameter φ is obtained. i Then, calculate the first objective function And the second objective function Finally, the third objective function is obtained And by optimizing the third objective function, such as minimizing the calculation, the network parameters of the basic neural network, the network parameters of the target feature network, and the network parameters of the functional area determination network are trained to obtain a trained target feature network, a trained basic neural network, and a trained functional area determination network.

[0195] Furthermore, the third objective function is adjusted at least once using a gradient descent method to update network parameters of the basic neural network, network parameters of the target feature network, and network parameters of the functional area determination network.

[0196] The network parameters of the basic neural network include the network initial value θ shared globally by all historical target proteins, and the parameters of the target feature network are w a , the network parameters of the functional area are determined as w b .

[0197] For example, gradient descent training can be expressed as (9)-(11):

[0198]

[0199]

[0200]

[0201] Among them, θ on the left side of the equation of formula (9) is the initial value of the network after optimization by gradient descent, and θ on the right side of the equations of formulas (9)-(11) is the initial value of the network before optimization in gradient descent; w a is the network parameter of the target feature network, w b Determine the network parameters of the network for the functional area, β is the learning rate for gradient optimization, and α can be set to 0.001, optionally. is the objective function The gradient with respect to θ. is the objective function About w a gradient. is the objective function About w b gradient. is the third objective function, please refer to the above formula (8).

[0202] Specifically, using this gradient descent method to update network parameters can perform finite-step optimization, for example, performing 1-step update, or 5-step update.

[0203] Optionally, during the training process, formulas (1)-(11) may be cycled multiple times to update the above network parameters to obtain a trained target feature network, a trained basic neural network, and a trained functional region determination network. For example, the cycle may be repeated 1,000 times.

[0204] See also Figure 7a , Figure 7a Through training sample data Target feature network Get ti, and input the functional area to determine the network, and select it by the maximum probability Determine the target functional area and select the maximum probability value The corresponding pending functional area is expressed as As the target functional area of ​​the first layer. Finally, the target functional areas of each layer of the network are connected in sequence to obtain the network parameters corresponding to the basic neural network of the target protein. Then the network parameters are updated quickly by gradient descent. express It can be calculated by the above formula (1).

[0205] Through the above training process, we can get the optimized θ,w a ,w b , when actually predicting the activity value of the test compound, by optimizing θ,w a ,w b An activity prediction model is established for the target protein, and the test compound is predicted based on the activity prediction model of the target protein to obtain the activity prediction result of the test compound for the target protein.

[0206] In addition, considering the different complexities of different functional areas, the required network width and complexity are different, and different functional areas may share some parameters. In the above embodiment, the functional area is determined, and each area is a small network module. In other embodiments, each dimension of all pending functional areas in the lth layer can be determined as the target functional area, and the output of the functional area determination network is determined from the probability of each pending functional area being selected to the predicted probability of each dimension being selected. In this way, multiple dimensions can be selected, and at the same time, the criterion for each dimension to be selected is that the probability corresponding to the dimension exceeds a preset threshold, that is, it is selected.

[0207] In order to more clearly show the difference between the present application and the activity prediction method of the prior art, please refer to Figure 7b , Figure 7b There are three methods for predicting activity values, the two on the left are existing prediction methods, and the right is an example of an embodiment of the present application. Among them, all target proteins on the far left use the prediction model x obtained by training each historical target protein to predict the training drug molecules with measured activity of the current target protein. In the middle, by learning a common initial model, the historical target proteins 1-n do not include the target protein of current concern. For each target of the historical target proteins 1-n, the initial model is quickly updated to the prediction model of the target through its training data, and then the initial model is optimized with the test data, so that the updated prediction model performance of each target is optimal, and the prediction model structure of each target is the same, and each target corresponds to different model network parameters.

[0208] The rightmost is an example of an embodiment of the present application, in which the target features of each of the historical target proteins 1-n are extracted, and the functional region determination network determines the target functional region corresponding to each historical protein target, and connects the target functional region of each layer of the network. Further, the target feature network, the basic neural network and the functional region determination network are trained using the network parameters and training sample data of the basic neural network corresponding to the historical target protein to obtain a trained target feature network, a trained basic neural network and a trained functional region determination network. When the user inputs the test compound and the target target protein, an activity prediction model can be generated based on the trained target feature network, the trained basic neural network and the trained functional region determination network, and a small amount of tested compound information of the target target protein to predict the activity value of the test compound to the target target protein.

[0209] The embodiment of the present application obtains training sample data, and the training sample data includes historical data sets of all historical target proteins, wherein each historical data contains at least one historical target protein and activity data of historical active compounds tested on historical target proteins; a target feature network is used to determine the target features of historical target proteins according to the molecular structure features of historical active compounds tested on historical target proteins and the activity data of historical active compounds tested on historical target proteins; the target features of historical target proteins are input into the functional region determination network to determine the historical functional regions corresponding to each layer of the network in the basic neural network; the network parameters of the basic neural network corresponding to the historical target protein are determined according to the historical functional regions corresponding to each layer of the network in the basic neural network; the network parameters of the basic neural network corresponding to the historical target protein and the training sample data are used to train the target feature network, the basic neural network and the functional region determination network to obtain a trained target feature network, a trained basic neural network and a trained functional region determination network. The activity prediction model is trained by making full use of the data of the tested compounds of all known historical target proteins, thereby avoiding the problem that the learned meta-initial model is prone to overfitting and the generalization effect to new data is poor due to the small number of target proteins. Secondly, compared with the prior art, different types of target proteins have large differences in interaction with small molecules, so that different target protein pharmacophores are different, and the knowledge of target proteins with low similarity to the target target is transferred to cause negative effects. At the same time, tasks with low correlation are learned together, which may lead to low stability of the model and even damage the prediction effect of the model. This application proposes a functional regionalization meta-learning algorithm in ligand-based drug design, and summarizes other target protein data with measured activity into different functional areas according to similarity, so as to make full use of highly correlated data in other target protein activity data to improve the accuracy of activity prediction in the current target protein. Thirdly, this application uses a fully connected network and a recursive neural network to determine the respective target feature network and the functional region determination network, and optimizes the three network parameters during training to improve the prediction accuracy of the activity prediction model formed by the target target protein, and the model has a strong expression ability.

[0210] All of the above technical solutions can be arbitrarily combined to form optional embodiments of the present application, which will not be described in detail here.

[0211] In order to better implement the compound activity prediction method of the present application embodiment, the present application embodiment also provides a compound activity prediction device. Figure 8 , Figure 8 A schematic diagram of the structure of a compound activity prediction device provided in an embodiment of the present application. The compound activity prediction device 800 may include:

[0212] An acquisition unit 810 is used to acquire the target protein corresponding to the compound to be tested;

[0213] A determination unit 820 is used to determine the target feature of the target protein according to the measured active compound information corresponding to the target protein; and

[0214] Inputting the target feature into the trained functional region determination network for processing to determine the target functional region corresponding to each layer of the basic neural network, and the trained functional region determination network is used to predict the selection probability of the target functional region; and

[0215] According to the target functional regions corresponding to each layer of the basic neural network, the activity prediction model of the target protein is determined;

[0216] The prediction unit 830 is used to predict the test compound according to the activity prediction model of the target protein to obtain the activity prediction result of the test compound on the target protein.

[0217] Optionally, the determination unit 820 can be used to input the molecular structure characteristics of the tested active compound and the activity data of the tested active compound on the target target protein into a trained target feature network for processing to determine the target characteristics of the target target protein. The trained target feature network is trained based on the similarity between the activity data of all historical target proteins.

[0218] Optionally, the determination unit 820 can also be used to use a feature extraction network to extract features from the molecular structure features of the tested active compound to obtain intermediate features of the tested active compound; concatenate the intermediate features of the tested active compound and the activity data of the tested active compound on the target target protein; input the concatenated data into a fully connected network containing a preset number of neurons, and average the output results of the fully connected layers corresponding to the molecular structures of all tested active compounds to obtain the target features of the target target protein.

[0219] Optionally, the determination unit 820 can also be used to determine the series features based on the target features and the selected probabilities corresponding to all the pending functional areas of the previous layer of the network in the basic neural network; determine the hidden features of the current layer of the network based on the series features and the hidden features of the previous layer of the network in the basic neural network; use the trained functional area determination network to process the hidden features of the current layer of the network to determine the target functional area of ​​the current layer of the network from multiple pending functional areas.

[0220] Optionally, the determination unit 820 can also be used to normalize the latent features of all the pending functional areas of each layer in the basic neural network through the trained functional area determination network to obtain the selection probabilities corresponding to all the pending functional areas; determine the target functional area of ​​the current layer network according to the pending functional area corresponding to the maximum value of the selection probabilities corresponding to all the pending functional areas.

[0221] Optionally, the determination unit 820 can also be used to sequentially connect the target functional regions of each layer in the basic neural network to obtain a basic activity prediction model of the target protein; use the gradient descent method to adjust the basic activity prediction model of the target protein at least once; and determine the activity prediction model of the target protein based on the adjusted basic activity prediction model.

[0222] The present application also provides a network training device for predicting compound activity. Fig. 9 , Fig. 9 A schematic diagram of the structure of a network training device for predicting compound activity provided in an embodiment of the present application. The network training device 900 for predicting compound activity may include:

[0223] An acquisition unit 910 is used to acquire training sample data, where the training sample data includes historical data sets of all historical target proteins, wherein each historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein;

[0224] A determination unit 920 is used to determine the target feature of the historical target protein using a target feature network according to the molecular structure features of the historically tested active compounds of the historical target protein and the activity data of the historically tested active compounds on the historical target protein; and

[0225] Inputting the target features of the historical target protein into a functional region determination network to determine the historical functional region corresponding to each layer of the basic neural network; and

[0226] Determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network;

[0227] The training unit 930 is used to train the target feature network, the basic neural network and the functional area determination network using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data to obtain a trained target feature network, a trained basic neural network and a trained functional area determination network.

[0228] Optionally, the training unit 930 can be used to input the molecular structure characteristics of the historically measured active compounds of the historical target protein into the basic neural network corresponding to the historical target protein with the network parameters to obtain the historical predicted activity of the historically measured active compounds to the historical target protein; determine a first objective function based on the historical predicted activity of the historically measured active compounds to the historical target protein and the activity data of the historically measured active compounds to the historical target protein; and adjust the network parameters of the basic neural network based on the determined first objective function until the training end conditions are met.

[0229] Optionally, the training unit 930 can also be used to randomly divide the historical data set into multiple non-overlapping historical data subsets; use the target feature network to extract features from the multiple historical data subsets to obtain target features of the multiple historical data subsets; determine a second objective function based on the similarity of target features of the same historical target protein in the target features of the multiple historical data subsets, and the similarity of target features of different historical target proteins; optimize the second objective function to train the target feature network.

[0230] Optionally, the training unit 930 can also be used to determine a third objective function based on the first objective function and the second objective function; optimize the third objective function to train the basic neural network, the target feature network, and the functional area determination network.

[0231] It should be noted that the functions of each module in the compound activity prediction device 800 and the network training device 900 in the embodiment of the present application can correspond to the specific implementation methods of any embodiment in the above-mentioned method embodiments, and will not be repeated here.

[0232] Each unit in the compound activity prediction device 800 and the network training device 900 can be implemented in whole or in part by software, hardware, or a combination thereof. Each unit can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the corresponding operations of each unit.

[0233] The compound activity prediction device 800 and the network training device 900 can be integrated in a terminal or server with a storage device and a processor installed to have computing power, or the drug analysis device 600 is the terminal or server. The terminal can be a smart phone, a tablet computer, a laptop computer, a smart TV, a smart speaker, a wearable smart device, a personal computer (PC) and other devices, and the terminal can also include a client, which can be a video client, a browser client or an instant messaging client. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0234] Fig.10 The schematic structural diagram of the compound activity prediction device 800 provided in the embodiment of the present application, the compound activity prediction device 800 may include: a communication interface 801, a memory 802, a processor 803 and a communication bus 804. The communication interface 801, the memory 802, and the processor 803 communicate with each other through the communication bus 804. The communication interface 801 is used for the device 800 to communicate data with an external device. The memory 802 can be used to store software programs and modules, and the processor 803 runs the software programs and modules stored in the memory 802, such as the software programs of the corresponding operations in the aforementioned method embodiments.

[0235] Optionally, the processor 803 may call the software program and module stored in the memory 802 to perform the following operations:

[0236] Obtain the target protein corresponding to the compound to be tested;

[0237] Determine the target characteristics of the target protein based on the information of the tested active compounds corresponding to the target protein;

[0238] Input the target feature into the trained functional region determination network for processing to determine the target functional region corresponding to each layer of the trained basic neural network, and the trained functional region determination network is used to predict the selection probability of the target functional region;

[0239] Determine the activity prediction model of the target protein based on the target functional region corresponding to each layer of the trained basic neural network;

[0240] The activity prediction model of the target protein is used to predict the activity of the test compound to obtain the activity prediction result of the test compound on the target protein.

[0241] Optionally, the compound activity prediction device 800 can be integrated in a terminal or server with a storage device and a processor installed and having computing power, or the compound activity prediction device 800 is the terminal or server. The terminal can be a smart phone, a tablet computer, a laptop computer, a smart TV, a smart speaker, a wearable smart device, a personal computer and other devices. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0242] Fig.11 A schematic structural diagram of a network training device 900 provided in an embodiment of the present application is shown in FIG. Fig.11 As shown, the network training device 900 may include: a communication interface 901, a memory 902, a processor 903 and a communication bus 904. The communication interface 901, the memory 902, and the processor 903 communicate with each other through the communication bus 904. The communication interface 901 is used for the device 800 to communicate data with an external device. The memory 902 can be used to store software programs and modules, and the processor 903 runs the software programs and modules stored in the memory 902, such as the software programs of the corresponding operations in the aforementioned method embodiments.

[0243] Optionally, the processor 930 may call software programs and modules stored in the memory 902 to perform the following operations:

[0244] Acquire training sample data, wherein the training sample data includes historical data sets of all historical target proteins, wherein each of the historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein;

[0245] Using a target feature network, according to the molecular structure features of the historically tested active compounds of the historically tested target protein and the activity data of the historically tested active compounds on the historically tested target protein, the target features of the historically tested target protein are determined;

[0246] Inputting the target features of the historical target protein into a functional region determination network to determine the historical functional region corresponding to each layer of the network in the basic neural network;

[0247] Determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network;

[0248] The target feature network, the basic neural network and the functional area determination network are trained using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data to obtain a trained target feature network, a trained basic neural network and a trained functional area determination network.

[0249] Optionally, the network training device 900 can be integrated in a terminal or server with a storage device and a processor installed and having computing power, or the compound activity prediction device 800 is the terminal or server. The terminal can be a smart phone, a tablet computer, a laptop computer, a smart TV, a smart speaker, a wearable smart device, a personal computer and other devices. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0250] Optionally, the present application further provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0251] The present application also provides a computer-readable storage medium for storing a computer program. The computer-readable storage medium can be applied to a computer device, and the computer program enables the computer device to execute the corresponding process in the compound activity prediction method in the embodiment of the present application, which will not be described in detail for the sake of brevity.

[0252] The present application also provides a computer program product, which includes computer instructions, which are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the corresponding process in the compound activity prediction method in the embodiment of the present application, which will not be repeated here for the sake of brevity.

[0253] The present application also provides a computer program, which includes computer instructions, which are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the corresponding process in the compound activity prediction method in the embodiment of the present application, which will not be repeated here for the sake of brevity.

[0254] It should be understood that the processor of the embodiment of the present application may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method embodiment can be completed by the hardware integrated logic circuit or software instructions in the processor. The above processor can be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in the embodiment of the present application can be directly embodied as a hardware decoding processor to perform, or the hardware and software modules in the decoding processor are combined and performed. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, and other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0255] It can be understood that the memory in the embodiments of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0256] It should be understood that the above-mentioned memory is exemplary but not restrictive. For example, the memory in the embodiments of the present application may also be static random access memory (static RAM, SRAM), dynamic random access memory (dynamic RAM, DRAM), synchronous dynamic random access memory (synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (synch link DRAM, SLDRAM) and direct memory bus random access memory (Direct Rambus RAM, DR RAM), etc. That is to say, the memory in the embodiments of the present application is intended to include but not limited to these and any other suitable types of memory.

[0257] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0258] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0259] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0260] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0261] In addition, each functional unit in the embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0262] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer or a server) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include various media that can store program codes, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks, or optical disks.

[0263] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting compound activity, It is characterized in that The method comprises: Obtain the target protein corresponding to the compound to be tested; Determining the target feature of the target protein according to the measured active compound information corresponding to the target protein; Inputting the target feature into a trained functional region determination network for processing to determine the target functional region corresponding to each layer of the trained basic neural network, wherein the trained functional region determination network is used to predict the selection probability of the target functional region; Determining the activity prediction model of the target protein according to the target functional region corresponding to each layer of the trained basic neural network; Predicting the test compound according to the activity prediction model of the target protein to obtain an activity prediction result of the test compound for the target protein; Each layer of the trained basic neural network includes a plurality of undetermined functional areas, and the inputting of the target feature into the trained functional area determination network for processing to determine the target functional area corresponding to each layer of the basic neural network includes: Determine the series feature according to the target feature and the selected probabilities corresponding to all pending functional areas of the previous layer of the trained basic neural network; Determine the latent features of the current layer network according to the series features and the latent features of the previous layer network in the trained basic neural network; The trained functional region determination network is used to process the latent features of the current layer network to determine the target functional region of the current layer network from the multiple pending functional regions.

2. The method according to claim 1, It is characterized in that Determining the target feature of the target protein according to the measured active compound information corresponding to the target protein includes: The molecular structure characteristics of the tested active compound and the activity data of the tested active compound on the target protein are input into a trained target feature network for processing to determine the target characteristics of the target protein. The trained target feature network is trained based on the similarity between the activity data of all historical target proteins.

3. The method according to claim 2, It is characterized in that The target feature network includes a feature extraction network and a fully connected network. The molecular structure features of the tested active compound and the activity data of the tested active compound on the target protein are input into the target feature network for processing to determine the target features of the target protein, including: Using the feature extraction network to extract features of the molecular structure features of the tested active compound to obtain intermediate features of the tested active compound; Concatenating the intermediate features of the tested active compound and the activity data of the tested active compound on the target protein; The serially connected data are input into the fully connected network including a preset number of neurons, and the output results of the fully connected layer corresponding to the molecular structures of all tested active compounds are averaged to obtain the target features of the target protein.

4. The method according to claim 1, It is characterized in that The step of using the trained functional region determination network to process the latent features of the current layer network to determine the target functional region of the current layer network from the multiple pending functional regions includes: Normalizing the latent features of all pending functional areas of each layer of the trained basic neural network through the trained functional area determination network to obtain the corresponding selection probabilities of all pending functional areas; The target functional area of ​​the current layer network is determined according to the undetermined functional area corresponding to the maximum value among the selected probabilities corresponding to all the undetermined functional areas.

5. The method according to any one of claims 1 to 4, It is characterized in that The step of determining the activity prediction model of the target protein according to the target functional regions corresponding to each layer of the network in the basic neural network includes: Sequentially connecting the target functional regions of each layer of the trained basic neural network to obtain a basic activity prediction model for the target protein; Using a gradient descent method to adjust the basic activity prediction model of the target protein at least once; According to the adjusted basic activity prediction model, an activity prediction model of the target protein is determined.

6. A network training method for compound activity prediction, It is characterized in that The method comprises: Acquire training sample data, wherein the training sample data includes historical data sets of all historical target proteins, wherein each of the historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein; Using a target feature network, according to the molecular structure features of the historically tested active compounds of the historically tested target protein and the activity data of the historically tested active compounds on the historically tested target protein, the target features of the historically tested target protein are determined; Inputting the target features of the historical target protein into a functional region determination network to determine the historical functional region corresponding to each layer of the network in the basic neural network; Determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network; Using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data, the target feature network, the basic neural network and the functional region determination network are trained to obtain a trained target feature network, a trained basic neural network and a trained functional region determination network; Each layer of the basic neural network includes a plurality of undetermined functional regions, and the target features of the historical target protein are input into the functional region determination network to determine the historical functional regions corresponding to each layer of the basic neural network, including: Determine the series feature according to the target feature and the selected probabilities corresponding to all pending functional areas of the previous layer of the network in the basic neural network; Determine the latent features of the current layer network according to the series features and the latent features of the previous layer network in the basic neural network; The functional region determination network is used to process the latent features of the current layer network to determine the historical functional region of the current layer network from a plurality of pending functional regions.

7. The method according to claim 6, It is characterized in that The method of training the target feature network, the basic neural network and the functional region determination network by using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data comprises: Inputting the molecular structure characteristics of the historically tested active compounds of the historical target protein into the basic neural network corresponding to the historical target protein with the network parameters to obtain the historical predicted activity of the historically tested active compounds on the historical target protein; Determining a first objective function according to the historical predicted activity of the historically tested active compound on the historical target protein and the activity data of the historically tested active compound on the historical target protein; The network parameters of the basic neural network are adjusted according to the determined first objective function until a training end condition is met.

8. The method according to claim 7, It is characterized in that The training of the target feature network, the basic neural network and the functional region determination network using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data also includes: Randomly dividing the historical data set into a plurality of disjoint historical data subsets; Using the target feature network to extract features from the multiple disjoint historical data subsets to obtain target features of the multiple historical data subsets; Determining a second objective function according to the similarity of the target features of the same historical target protein and the similarity of the target features of different historical target proteins in the target features of the multiple historical data subsets; The second objective function is optimized to train the target feature network.

9. The method according to claim 8, It is characterized in that The training of the target feature network, the basic neural network and the functional region determination network using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data also includes: Determine a third objective function according to the first objective function and the second objective function; The third objective function is optimized to train the basic neural network, the target feature network, and the functional area determination network.

10. The method according to claim 9, It is characterized in that The training process also includes: The third objective function is adjusted at least once using a gradient descent method to update the network parameters of the basic neural network, the network parameters of the target feature network, and the network parameters of the functional area determination network.

11. A compound activity prediction device, It is characterized in that The device comprises: An acquisition unit, used for acquiring the target protein corresponding to the compound to be tested; A determination unit, configured to determine the target feature of the target protein according to the information of the active compound detected corresponding to the target protein; and Inputting the target feature into a trained functional region determination network for processing to determine the target functional region corresponding to each layer of the trained basic neural network, wherein the trained functional region determination network is used to predict the selection probability of the target functional region; and Determining the activity prediction model of the target protein according to the target functional regions corresponding to each layer of the network in the trained basic neural network; A prediction unit, used to predict the test compound according to the activity prediction model of the target protein to obtain an activity prediction result of the test compound for the target protein; Each layer of the network in the trained basic neural network includes multiple pending functional areas. The determination unit is also used to determine the series features based on the target features and the selection probabilities corresponding to all the pending functional areas of the previous layer of the network in the trained basic neural network; determine the hidden features of the current layer of the network based on the series features and the hidden features of the previous layer of the network in the trained basic neural network; use the trained functional area determination network to process the hidden features of the current layer of the network to determine the target functional area of ​​the current layer of the network from the multiple pending functional areas.

12. A network training device for predicting compound activity, It is characterized in that The device comprises: An acquisition unit, used for acquiring training sample data, wherein the training sample data includes historical data sets of all historical target proteins, wherein each of the historical data includes at least one historical target protein and activity data of historically tested active compounds on the historical target protein; a determination unit, configured to determine the target feature of the historical target protein by using a target feature network, according to the molecular structure features of the historically tested active compounds of the historical target protein, and the activity data of the historically tested active compounds on the historical target protein; and Inputting the target features of the historical target protein into a functional region determination network to determine the historical functional region corresponding to each layer of the basic neural network; and Determining the network parameters of the basic neural network corresponding to the historical target protein according to the historical functional regions corresponding to each layer of the network in the basic neural network; A training unit, used to train the target feature network, the basic neural network and the functional region determination network using the network parameters of the basic neural network corresponding to the historical target protein and the training sample data, so as to obtain a trained target feature network, a trained basic neural network and a trained functional region determination network; Each layer of the network in the basic neural network includes multiple pending functional areas. The determination unit is further used to determine the series features based on the target features and the selection probabilities corresponding to all the pending functional areas of the previous layer of the network in the basic neural network; determine the hidden features of the current layer of the network based on the series features and the hidden features of the previous layer of the network in the basic neural network; and use the functional area determination network to process the hidden features of the current layer of the network to determine the historical functional area of ​​the current layer of the network from the multiple pending functional areas.

13. A computer-readable storage medium, It is characterized in that The computer-readable storage medium stores a computer program, and the computer program is suitable for being loaded by a processor to execute the steps in the method according to any one of claims 1 to 10.

14. A computer device, It is characterized in that The computer device comprises a processor and a memory, wherein a computer program is stored in the memory, and the processor is used to execute the steps in the method according to any one of claims 1 to 10 by calling the computer program stored in the memory.

15. A computer program product comprising computer instructions, It is characterized in that When the computer instructions are executed by a processor, the steps in the method according to any one of claims 1 to 10 are implemented.

Images