A method for training a molecular binding model, a molecular screening method and a device

Abstract

The present application provides a method for training a molecular binding model, a molecular screening method and a device, which can be applied to the field of intelligent medical care to solve the problem of low molecular screening accuracy. The method includes: using the molecular binding model to be trained, based on protein feature information and molecular feature information, determining the binding activity feature information, embedded feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule, wherein the embedded feature information is used to characterize the degree of binding between the sample protein molecule and the sample candidate molecule, the co-crystal feature information is used to characterize whether there is a co-crystal structure between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction; based on the binding activity feature information, embedded feature information and co-crystal feature information, determining the training loss of the molecular binding model to be trained.

Description

Technical Field

[0001] The present application relates to the field of biomedicine technology, and in particular to a method for training a molecular binding model, a molecular screening method and a device. Background Art

[0002] With the continuous development of science and technology, more and more equipment can use activity prediction models to predict the binding activity of compounds after virtual binding of molecules. Therefore, based on the predicted binding activity of compounds, some molecules can be screened out for real drug binding experiments.

[0003] When predicting the binding activity of a compound after virtual molecular binding, the activity prediction model usually makes predictions based on the chemical property characteristics of the molecule extracted by the feature extraction model, or the structural characteristics of the molecule itself.

[0004] However, the molecular binding process is not simply putting two molecules together. There may be some unpredictable situations such as interactions between molecules. Only from the perspective of the molecules themselves, the predicted binding activity of the compounds after the virtual binding of the molecules is less accurate, resulting in lower accuracy of molecular screening. Summary of the invention

[0005] The embodiments of the present application provide a method for training a molecular binding model, a molecular screening method, an apparatus, a computer device, and a storage medium, which are used to solve the problem of low accuracy in molecular screening.

[0006] In a first aspect, a method for training a molecular binding model is provided, comprising:

[0007] Obtaining protein characteristic information of each sample protein molecule, and obtaining molecular characteristic information of each sample candidate molecule;

[0008] Based on each protein feature information and each molecule feature information, the molecular binding model to be trained is subjected to multiple rounds of iterative training until the training loss of the molecular binding model to be trained meets the training target, and the molecular binding model to be trained is output as the target molecular binding model, wherein the following operations are performed for multiple rounds of iterative training respectively:

[0009] Using the molecular binding model to be trained, based on protein feature information and molecular feature information, determining the binding activity feature information, embedding feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule, wherein the binding activity feature information is used to characterize the activity of the sample protein molecule after virtual binding with the sample candidate molecule, the embedding feature information is used to characterize the degree of binding between the sample protein molecule and the sample candidate molecule, and the co-crystal feature information is used to characterize whether there is a co-crystal structure between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction;

[0010] Based on the binding activity feature information, the embedding feature information and the co-crystal feature information, a training loss of the molecular binding model to be trained is determined.

[0011] In a second aspect, a molecular screening method is provided, comprising:

[0012] Obtaining target protein molecules and target candidate molecules;

[0013] Using a feature extraction model, performing feature extraction processing on the target protein molecule and the target candidate molecule to obtain target protein feature information of the target protein molecule and target molecule feature information of the target candidate molecule;

[0014] Using the target molecule binding model, based on the target protein characteristic information and the target molecule characteristic information, determining the binding activity characteristic information between the target protein molecule and the target candidate molecule;

[0015] An activity prediction model is used to predict the activity value of the target protein molecule and the target candidate molecule after virtual binding based on the binding activity characteristic information.

[0016] In a third aspect, a device for training a molecular binding model is provided, comprising:

[0017] Acquisition module: used to obtain protein characteristic information of each sample protein molecule, and to obtain molecular characteristic information of each sample candidate molecule;

[0018] Processing module: used for performing multiple rounds of iterative training on the molecular binding model to be trained based on each protein feature information and each molecule feature information, until the training loss of the molecular binding model to be trained meets the training target, and outputting the molecular binding model to be trained as the target molecular binding model, wherein for multiple rounds of iterative training, the following operations are performed respectively:

[0019] Using the molecular binding model to be trained, based on protein feature information and molecular feature information, determining the binding activity feature information, embedding feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule, wherein the binding activity feature information is used to characterize the activity of the sample protein molecule after virtual binding with the sample candidate molecule, the embedding feature information is used to characterize the degree of binding between the sample protein molecule and the sample candidate molecule, and the co-crystal feature information is used to characterize whether there is a co-crystal structure between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction;

[0020] Based on the binding activity feature information, the embedding feature information and the co-crystal feature information, a training loss of the molecular binding model to be trained is determined.

[0021] Optionally, the acquisition module is specifically used for:

[0022] Obtaining a training sample set, wherein the training sample set includes various training samples, and each training sample includes a sample protein molecule and a sample candidate molecule;

[0023] A feature extraction model is used to perform feature extraction processing on each training sample to obtain protein feature information of each sample protein molecule and molecular feature information of each sample candidate molecule.

[0024] Optionally, the acquisition module is specifically used for:

[0025] For each training sample, the following operations are performed respectively:

[0026] Based on the sample protein molecules in the training sample, determining the adjacency matrix of the sample protein molecules, wherein the adjacency matrix of the sample protein molecules is used to characterize each sample amino acid molecule contained in the sample protein molecule, and the molecular structure distance between every two sample amino acid molecules;

[0027] Based on the sample candidate molecules in the training sample, determining an adjacency matrix of the sample candidate molecules, wherein the adjacency matrix of the sample candidate molecules is used to characterize each sample candidate atom contained in the sample candidate molecules, and the chemical bond structure between each sample candidate atom;

[0028] Feature extraction processing is performed on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample candidate molecule respectively to obtain protein feature information of the sample protein molecule and molecular feature information of the sample candidate molecule.

[0029] Optionally, the processing model is specifically used for:

[0030] Based on the protein characteristic information and the molecule characteristic information, predicting the binding distance between each sample amino acid molecule contained in the sample protein molecule and each sample candidate atom contained in the sample candidate molecule after the sample protein molecule is virtually bound to the sample candidate molecule;

[0031] Based on the obtained binding distances, embedding characteristic information and co-crystal characteristic information between the sample protein molecule and the sample candidate molecule are determined.

[0032] Optionally, the processing module is specifically used for:

[0033] Determining the embedding feature information between the sample protein molecule and the sample candidate molecule based on the minimum binding distance between the designated sample amino acid molecule and each sample candidate atom;

[0034] Based on each sample amino acid molecule, each sample candidate atom and the binding distance between each amino acid molecule and each sample candidate atom, the co-crystal characteristic information between the sample protein molecule and the sample candidate molecule is determined.

[0035] Optionally, the processing module is specifically used for:

[0036] Determining a first training loss of the molecular binding model to be trained based on a first error value between the embedded feature information and the embedded target;

[0037] Determining a second training loss of the molecular binding model to be trained based on a second error value between the co-crystal feature information and the co-crystal target;

[0038] Determining a third training loss of the molecular binding model to be trained based on a third error value between the binding activity feature information and the activity target;

[0039] Based on the obtained first training loss, second training loss and third training loss, the training loss of the molecular binding model to be trained is determined.

[0040] The processing module is specifically used for:

[0041] The trained neighbor matching model is used to match each sample amino acid molecule contained in the sample protein molecule with each sample candidate atom contained in the sample candidate molecule to obtain matching feature information between the sample protein molecule and the sample candidate molecule as the co-crystal target, wherein the matching feature information is used to characterize the matching distance between each sample amino acid molecule and each sample candidate atom.

[0042] Using a cross entropy function, determining a second error value between the eutectic characteristic information and the eutectic target;

[0043] The second error value is used as a second training loss of the molecular binding model to be trained.

[0044] Optionally, the processing module is specifically used for:

[0045] Using an activity prediction model, based on the binding activity characteristic information, predicting the sample activity value after the sample protein molecule is virtually bound to the sample candidate molecule;

[0046] Based on a third error value between the sample activity value and the corresponding reference activity value, a third training loss of the molecular binding model to be trained is determined.

[0047] Optionally, the processing module is further used for:

[0048] Determining that the training loss of the molecular binding model to be trained does not meet the training target, and adjusting the model parameters of the molecular binding model to be trained;

[0049] When it is determined that the training loss of the molecular binding model to be trained meets the training target, the molecular binding model to be trained is output as the trained target molecular binding model.

[0050] Optionally, the processing module is specifically used for:

[0051] If the obtained training loss includes a first training loss, a second training loss, and a third training loss, the first training loss is determined based on the embedding feature information, the second training loss is determined based on the eutectic feature information, and the third training loss is determined based on the binding activity feature information, then determining whether the first training loss converges, whether the second training loss converges, and whether the third training loss converges, respectively;

[0052] If at least one of the first training loss, the second training loss, and the third training loss does not converge, the model parameters of the molecular binding model to be trained are adjusted.

[0053] In a fourth aspect, a molecular screening device is provided, comprising:

[0054] Acquisition module: used to obtain target protein molecules and target candidate molecules;

[0055] Processing module: used for performing feature extraction processing on the target protein molecule and the target candidate molecule by using a feature extraction model to obtain target protein feature information of the target protein molecule and target molecule feature information of the target candidate molecule;

[0056] The processing module is also used to: use the target molecule binding model to determine the binding activity characteristic information between the target protein molecule and the target candidate molecule based on the target protein characteristic information and the target molecule characteristic information;

[0057] The processing module is also used to: use an activity prediction model to predict the activity value of the target protein molecule and the target candidate molecule after virtual binding based on the binding activity characteristic information.

[0058] A fifth aspect provides a computer device, including:

[0059] A memory for storing program instructions;

[0060] The processor is used to call the program instructions stored in the memory and execute the method as described in the first aspect according to the obtained program instructions.

[0061] In a sixth aspect, a computer-readable storage medium is provided, wherein the storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the method described in the first aspect.

[0062] In an embodiment of the present application, after obtaining the protein characteristic information of each sample protein molecule and the molecular characteristic information of each sample candidate molecule, the binding activity characteristic information, embedding characteristic information and co-crystal characteristic information between the sample protein molecule and the sample candidate molecule are further obtained based on the protein characteristic information and the molecular characteristic information. The molecular binding model is trained based on the binding activity characteristic information, the embedding characteristic information and the co-crystal characteristic information. The molecular binding model is trained from the perspective of multiple characteristic information, so that when the activity prediction model predicts the binding activity after the virtual binding of the molecules, it not only considers the chemical properties of the molecules contained in the protein characteristic information and the molecular characteristic information, or the structural characteristics of the molecules themselves, but also considers the molecular interactions such as the embedding characteristics and co-crystal characteristics between the molecules, so that the activity prediction model is more accurate in predicting the activity value based on the molecular binding model. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic diagram of the principle of the method for training a molecular binding model provided in the embodiment of the present application Figure 1 ;

[0064] Figure 2 An application scenario of the method for training a molecular binding model provided in an embodiment of the present application;

[0065] Figure 3A schematic diagram of a process for training a molecular binding model provided in an embodiment of the present application Figure 1 ;

[0066] Figure 4a A molecular structure diagram of the method for training a molecular binding model provided in an embodiment of the present application Figure 1 ;

[0067] Figure 4b A molecular structure diagram of the method for training a molecular binding model provided in an embodiment of the present application Figure 2 ;

[0068] Figure 4c A schematic diagram of the principle of the method for training a molecular binding model provided in the embodiment of the present application Figure 2 ;

[0069] Figure 4d A schematic diagram of the principle of the method for training a molecular binding model provided in the embodiment of the present application Figure 3 ;

[0070] Figure 4e A schematic diagram of a principle of the method for training a molecular binding model provided in an embodiment of the present application;

[0071] Figure 5a A schematic diagram of a process for training a molecular binding model provided in an embodiment of the present application Figure 2 ;

[0072] Figure 5b A molecular structure diagram of the method for training a molecular binding model provided in an embodiment of the present application Figure 3 ;

[0073] Figure 5c A molecular structure diagram 4 of the method for training a molecular binding model provided in an embodiment of the present application;

[0074] Figure 5d A molecular structure diagram 5 of the method for training a molecular binding model provided in an embodiment of the present application;

[0075] Figure 5e A molecular structure diagram of the method for training a molecular binding model provided in an embodiment of the present application Figure 6 ;

[0076] Figure 6 A schematic diagram of a principle of a method for training a molecular binding model provided in an embodiment of the present application is shown in FIG5;

[0077] Figure 7 A schematic diagram of a molecular screening method provided in the present application Figure 3 ;

[0078] Figure 8 A schematic diagram of a structure of a device for training a molecular binding model provided in an embodiment of the present application Figure 1 ;

[0079] Fig. 9 A schematic diagram of a molecular screening device provided in an embodiment of the present application Figure 2 ;

[0080] Fig.10 A structural diagram of a device for training a molecular binding model or a molecular screening device provided in an embodiment of the present application Figure 3 . DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0082] Some of the terms used in the embodiments of the present application are explained below to facilitate understanding by those skilled in the art.

[0083] (1) Half inhibitory concentration IC50 and pIC50:

[0084] IC50 can also be called half inhibition rate, which is the concentration of drug or inhibitor required to inhibit a specified biological process by half. There is a certain mathematical relationship between pIC50 and IC50, and pIC50 is the opposite of the logarithm of IC50.

[0085] (2) Protein molecules and molecules:

[0086] Protein molecules are the material basis of life, organic macromolecules, the basic organic matter that constitutes cells, and the main bearers of life activities. Amino acid molecules are the basic building blocks of protein molecules.

[0087] Molecules are composed of atoms, which are combined together according to a certain bonding order and spatial arrangement. The bonding order and spatial arrangement relationship are the molecular structure.

[0088] The embodiments of the present application relate to artificial intelligence (AI) technology, and are designed based on computer vision (CV) and machine learning (ML) in artificial intelligence technology.

[0089] Artificial intelligence 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 in computer science that 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.

[0090] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. Basic artificial intelligence technologies generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, mechatronics and other technologies. Artificial intelligence software technologies mainly include computer vision technology, speech processing technology, natural language processing technology, as well as machine learning / deep learning, autonomous driving, smart transportation and other major directions.

[0091] Computer vision technology Computer vision is a science that studies how to make machines "see". To put it more specifically, it refers to machine vision such as using cameras and computers to replace human eyes to identify, track and measure targets, and further perform graphic processing so that the computer processing becomes an image that is more suitable for human eye observation or transmission to instrument detection. As a scientific discipline, computer vision studies related theories and technologies, and attempts to establish an artificial intelligence system that can obtain information from images or multi-dimensional data. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous positioning and mapping, automatic driving, smart transportation and other technologies, as well as common biometric recognition technologies such as face recognition and fingerprint recognition.

[0092] Machine learning is a multi-disciplinary subject that involves 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 self-learning.

[0093] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields, such as common smart homes, smart wearable devices, virtual assistants, smart speakers, driverless cars, autonomous driving, drones, robots, smart medical care, smart customer service, Internet of Vehicles, autonomous driving, smart transportation, etc. I believe that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0094] The following is a brief introduction to the application fields of the method for training molecular binding models provided in the embodiments of the present application.

[0095] With the continuous development of science and technology, more and more equipment can use activity prediction models to predict the binding activity of compounds after virtual binding of molecules. Therefore, based on the predicted binding activity of the compounds, some molecules can be screened out for real drug binding experiments, avoiding unnecessary drug binding experiments and wasting resources.

[0096] When predicting the binding activity of compounds after virtual molecular binding, the activity prediction model is usually based on the chemical properties of the molecules extracted by the feature extraction model, or based on the structural characteristics of the molecules themselves, to predict the binding activity between the two molecules.

[0097] However, in the process of molecular binding, it is not simply putting two molecules together. There may be some unpredictable situations such as interactions between molecules. The binding activity of the compound predicted after the two molecules are virtually bound is low in accuracy based only on the perspective of the molecules themselves, resulting in low accuracy in molecular screening. As a result, some molecules that need to be bound are not screened out for drug binding experiments, while some molecules that do not need to be bound are screened out for drug binding experiments, which not only wastes resources but also affects the progress of drug research and development.

[0098] In order to solve the problem of low molecular screening accuracy, this application proposes a method for training molecular binding models. Please refer to Figure 1 The method obtains protein feature information of each sample protein molecule and molecular feature information of each sample candidate molecule. Based on each protein feature information and each molecule feature information, the molecular binding model to be trained is trained for multiple rounds of iterations until the training loss of the molecular binding model to be trained meets the training target, and then the molecular binding model to be trained is output as the target molecular binding model.

[0099] Among them, for multiple rounds of iterative training, the following operations are performed respectively: using the molecular binding model to be trained, based on the protein feature information and the molecular feature information, determining the binding activity feature information, embedded feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule, wherein the binding activity feature information is used to characterize the activity of the sample protein molecule after virtual binding with the sample candidate molecule, the embedded feature information is used to characterize the degree of binding between the sample protein molecule and the sample candidate molecule, and the co-crystal feature information is used to characterize whether there is a co-crystal structure between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction, and based on the binding activity feature information, the embedded feature information and the co-crystal feature information, the training loss of the molecular binding model to be trained is determined.

[0100] In an embodiment of the present application, after obtaining the protein characteristic information of each sample protein molecule and the molecular characteristic information of each sample candidate molecule, the binding activity characteristic information, embedding characteristic information and co-crystal characteristic information between the sample protein molecule and the sample candidate molecule are further obtained based on the protein characteristic information and the molecular characteristic information. The molecular binding model is trained based on the binding activity characteristic information, the embedding characteristic information and the co-crystal characteristic information. The molecular binding model is trained from the perspective of multiple characteristic information, so that when the activity prediction model predicts the binding activity after the virtual binding of the molecules, it not only considers the chemical properties of the molecules contained in the protein characteristic information and the molecular characteristic information, or the structural characteristics of the molecules themselves, but also considers the molecular interactions such as the embedding characteristics and co-crystal characteristics between the molecules, so that the activity prediction model is more accurate in predicting the activity value based on the molecular binding model.

[0101] The following describes the application scenarios of the method for training molecular binding models provided in this application.

[0102] Please refer to Figure 2 , which is an application scenario of the method for training a molecular binding model provided in an embodiment of the present application. The application scenario includes a client 101 and a server 102. The client 101 and the server 102 can communicate with each other, and the communication method can be to communicate using wired communication technology, such as communicating by connecting a network cable or a serial port line; or to communicate using wireless communication technology, such as communicating through Bluetooth or wireless fidelity (wireless fidelity, WIFI) and other technologies, without specific limitation.

[0103] The client 101 generally refers to a device that can instruct the server 102 to train a molecular binding model or perform molecular screening, such as a terminal device, a third-party application that the terminal device can access, or a web page that the terminal device can access. Terminal devices include but are not limited to mobile phones, computers, smart medical devices, smart appliances, etc. The server 102 generally refers to a device that can train a molecular binding model or perform molecular screening, such as a terminal device or a server. The server includes but is not limited to a cloud server, a local server, or an associated third-party server. Both the client 101 and the server 102 can use cloud computing to reduce the occupation of local computing resources; cloud storage can also be used to reduce the occupation of local storage resources.

[0104] As an embodiment, the client 101 and the server 102 may be the same device. In the embodiment of the present application, the client 101 and the server 102 are taken as examples to be different devices.

[0105] The following is based on Figure 2 , the method for training a molecular binding model provided in an embodiment of the present application is specifically introduced, taking the client 101 as the target client and the server 102 as the server as an example.

[0106] Please refer to Figure 3 , which is a flow chart of the method for training a molecular binding model provided in an embodiment of the present application.

[0107] S301, obtaining protein feature information of each sample protein molecule, and obtaining molecular feature information of each sample candidate molecule.

[0108] The server can receive protein feature information of each sample protein molecule sent by other devices, and obtain molecular feature information of each sample candidate molecule, and can also perform feature extraction processing on each sample protein molecule and each sample candidate molecule, and obtain protein feature information of each sample protein molecule, and obtain molecular feature information of each sample candidate molecule, etc., without specific restrictions. Please refer to Figure 4a , which is a possible structural diagram of a sample protein molecule. Please refer to Figure 4b , which is a possible structural diagram of the sample candidate molecule.

[0109] The embodiment of the present application is introduced by taking the example that the server uses a feature extraction model to extract protein feature information of each sample protein molecule, and obtains molecular feature information of each sample candidate molecule.

[0110] The server may first obtain a training sample set, where the training sample set includes various training samples, and each training sample includes a sample protein molecule and a sample candidate molecule.

[0111] The server can extract each sample protein molecule and each sample candidate molecule that have been subjected to drug binding experiments based on the database, thereby obtaining a training sample set; the server can also search for each sample protein molecule and each sample candidate molecule that have been subjected to drug binding experiments in network resources, thereby obtaining a training sample set, and there is no specific restriction on the method of obtaining the training sample set.

[0112] After obtaining the training sample set, the server can use the feature extraction model to perform feature extraction processing on each training sample, and obtain the protein feature information of each sample protein molecule, and the molecular feature information of each sample candidate molecule. The feature extraction model is used to perform feature extraction processing on molecules to obtain the structural information of molecules. The feature extraction model server can input the three-dimensional structure diagrams of the sample protein molecules and the sample candidate molecules into the feature extraction model, and perform feature extraction processing on each training sample; it can also input the multi-angle two-dimensional planar structure diagrams of the sample protein molecules and the sample candidate molecules into the feature extraction model, and perform feature extraction processing on each training sample, etc., without specific restrictions.

[0113] As an embodiment, when the server adopts a feature extraction model to perform feature extraction processing on each training sample separately, it can calculate through the adjacency matrix to introduce the process of feature extraction processing for one training sample. The process of feature extraction processing for other training samples is similar and will not be repeated here.

[0114] The server determines the adjacency matrix of the sample protein molecules based on the sample protein molecules in the training sample. Please refer to Figure 4c The adjacency matrix of the sample protein molecule is used to characterize each sample amino acid molecule contained in the sample protein molecule and the molecular structure distance between each two sample amino acid molecules. When determining the adjacency matrix of the sample protein molecule, the server can determine the adjacency matrix of the sample protein molecule based on each sample amino acid molecule contained in the sample protein molecule and the molecular structure distance between each two sample amino acid molecules; it can also establish the adjacency matrix of the sample protein molecule based on each sample amino acid molecule and the molecular structure distance greater than a preset distance threshold; it can also take each sample amino acid molecule as a vertex, after determining the molecular structure distance between each two sample amino acid molecules in the sample protein molecule, determine whether each molecular structure distance is greater than a preset distance threshold, and establish the corresponding edge between the two sample amino acid molecules with the molecular structure distance greater than the preset distance threshold, so as to obtain the relationship diagram between the vertex and the edge corresponding to the sample protein molecule. The server establishes the adjacency matrix of the sample protein molecule based on the obtained relationship diagram, etc., without specific limitation.

[0115] The server determines the adjacency matrix of the sample candidate molecules based on the sample candidate molecules in the training sample. Please refer to Figure 4d The adjacency matrix of the sample candidate molecule is used to characterize the various sample candidate atoms contained in the sample candidate molecule, as well as the chemical bond structure between each sample candidate atom. When determining the adjacency matrix of the sample candidate molecule, the server can determine the adjacency matrix of the sample candidate molecule based on the various sample candidate atoms contained in the sample candidate molecule, as well as the chemical bond structure between each sample candidate atom; it can also establish the adjacency matrix of the sample candidate molecule based on each sample candidate atom and the chemical bond between the sample candidate atoms in the chemical bond structure; it can also use each sample candidate atom as a vertex and the chemical bond between the sample candidate atoms as an edge to obtain a relationship graph between the vertices and the edges corresponding to the sample candidate atoms. The server establishes the adjacency matrix of the sample candidate molecule based on the obtained relationship graph, etc., without specific limitation.

[0116] The server performs feature extraction processing on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample candidate molecule respectively. Please refer to Figure 4e , obtain the protein feature information of the sample protein molecules and the molecular feature information of the sample candidate molecules.

[0117] S302, based on each protein feature information and each molecule feature information, perform multiple rounds of iterative training on the molecular binding model to be trained until the training loss of the molecular binding model to be trained meets the training target, and output the molecular binding model to be trained as the target molecular binding model.

[0118] After obtaining the protein feature information of each sample protein molecule and the molecular feature information of each sample candidate molecule, the server can perform multiple rounds of iterative training on the molecular binding model to be trained based on the protein feature information and the molecular feature information. The process of each round of iterative training is similar. The following is an introduction to the process of one round of iterative training. Please refer to Figure 5a .

[0119] S501, using the molecular binding model to be trained, based on the protein feature information and the molecular feature information, determining the binding activity feature information, embedding feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule.

[0120] After obtaining the protein feature information and the molecular feature information, the server can use the molecular binding model to be trained to predict the binding activity feature information between the sample protein molecule and the sample candidate molecule. The binding activity feature information is used to predict the binding activity feature after the sample protein molecule and the sample candidate molecule are virtually bound. At the same time, the server can also use the molecular binding model to be trained to predict the embedded feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule, and train the molecular binding model based on the embedded feature information and co-crystal feature information, so that the molecular binding model can have the performance of identifying the embedded relationship and co-crystal structure after the virtual binding between molecules, so that when predicting the binding activity feature information after the virtual binding between molecules, more accurate binding activity feature information can be predicted, and when the activity prediction model predicts the activity value based on the binding activity feature information obtained by the molecular binding model, more accurate binding activity can be obtained.

[0121] The following is an introduction to the process of using the molecular binding model to be trained to predict the embedding feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule.

[0122] The server can predict the binding distances between each sample amino acid molecule contained in the sample protein molecule and each sample candidate atom contained in the sample candidate molecule after the sample protein molecule and the sample candidate molecule are virtually bound based on the protein feature information and the molecular feature information. The server can determine the embedding feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule based on each obtained binding distance.

[0123] As an embodiment, the residues of the specified sample amino acid molecules can bind to the cavities of the sample candidate atoms to form an embedding relationship. Therefore, the server can determine the embedding feature information based on each binding distance, and then introduce supervision information to the molecular binding model to be trained through graph matching, so that the trained molecular binding model has the performance of identifying intermolecular embedding features, thereby improving the accuracy of predicting the binding activity after virtual binding of molecules based on the trained target molecule binding model.

[0124] After obtaining the binding distances between each sample amino acid molecule contained in the sample protein molecule and each sample candidate atom contained in the sample candidate molecule, the server can determine the embedded feature information between the sample protein molecule and the sample candidate molecule based on the minimum binding distance between the specified sample amino acid molecule and each sample candidate atom. The specified sample amino acid molecule can be a targeted sample amino acid molecule in the sample protein molecule, or a sample amino acid molecule specified according to the scenario, etc., without specific restrictions. By training the minimum binding distance between the specified sample amino acid molecule and each sample candidate atom to stabilize to the minimum value, the residues of the sample amino acid molecule are bound to the cavity of the corresponding sample candidate atom to form an embedding relationship. Please refer to Figure 5b , which is a structural diagram of an embedding relationship, where the sample candidate molecule is completely embedded in the sample protein molecule. Please refer to Figure 5c , which is a structural diagram of the embedding relationship, where the sample candidate molecule is partially embedded in the sample protein molecule.

[0125] As an embodiment, if the sample amino acid molecule matches the sample candidate atom, then there is a co-crystal structure between the sample amino acid molecule and the sample candidate atom. Therefore, the server can determine the co-crystal characteristic information based on each binding distance, and then introduce supervision information to the molecular binding model to be trained through neighbor matching, so that the trained molecular binding model has the performance of identifying intermolecular co-crystal characteristics, thereby improving the accuracy of predicting the binding activity after virtual binding between molecules based on the trained target molecule binding model.

[0126] After obtaining the binding distances between each sample amino acid molecule contained in the sample protein molecule and each sample candidate atom contained in the sample candidate molecule, the server can determine the co-crystal feature information between the sample protein molecule and the sample candidate molecule based on the binding distances between each sample amino acid molecule, each sample candidate atom and each sample amino acid molecule. The co-crystal feature information can be represented in the form of an adjacency matrix. By training the binding distances between the matched sample candidate atoms and the sample amino acid molecules to be stable and close to the co-crystal target, a co-crystal structure is formed between the matched sample candidate atoms and the sample amino acid molecules. Please refer to Figure 5d , which is a structural diagram of the eutectic structure, please refer to Figure 5e , which is a structural schematic diagram of the eutectic structure.

[0127] S502, determining a training loss of a molecular binding model to be trained based on the binding activity feature information, the embedding feature information and the co-crystal feature information.

[0128] After obtaining the binding activity feature information, the embedding feature information and the co-crystal feature information, the server can determine the training loss of the molecular binding model to be trained based on the binding activity feature information, the embedding feature information and the co-crystal feature information, and train the molecular binding model to be trained based on the training loss.

[0129] The server can perform weighted fusion processing on the binding activity feature information, the embedding feature information and the co-crystal feature information to obtain the fused feature information, and determine the training loss of the molecular binding model to be trained through the error value between the fused feature information and the training target.

[0130] The server may also determine the first training loss of the molecular binding model to be trained based on the first error value between the embedded feature information and the embedded target, determine the second training loss of the molecular binding model to be trained based on the second error value between the co-crystal feature information and the co-crystal target, determine the third training loss of the molecular binding model to be trained based on the third error value between the binding activity feature information and the activity target, and determine the training loss of the molecular binding model to be trained based on the obtained first training loss, second training loss, and third training loss. The server may determine the training loss of the molecular binding model to be trained based on the weighted sum of the first training loss, the second training loss, and the third training loss; it may also directly determine the first training loss, the second training loss, and the third training loss as the training loss of the molecular binding model to be trained, etc.

[0131] As an embodiment, the server can determine the binding distance between the specified sample amino acid molecule represented by the embedded feature information and the corresponding sample candidate atom, and the first error value between the reference binding distance through a graph matching model. The server determines the first training loss of the molecular binding model to be trained based on the first error value. The reference binding distance can be obtained from the PDBbind data set.

[0132] As an embodiment, the neighbor matching model is relatively easy to train because of its relatively simple function and more training data. The accuracy of the matching matrix obtained by using the trained neighbor matching model has reached a relatively high level. At this time, the trained neighbor matching model can be used to determine the co-crystal target. The server can use the trained neighbor matching model to match each sample amino acid molecule contained in the sample protein molecule with each sample candidate atom contained in the sample candidate molecule, and obtain the matching feature information between the sample protein molecule and the sample candidate molecule as the co-crystal target, wherein the matching feature information is used to characterize the matching distance between each sample amino acid molecule and each sample candidate atom. The server can use the cross entropy function to determine the second error value between the co-crystal feature information and the co-crystal target, and use the second error value as the second training loss of the molecular binding model to be trained. Among them, the trained neighbor matching model can be obtained by training based on the data in the PDBbind data set, and the cross entropy function can be replaced by other functions, without specific limitation.

[0133] As an embodiment, each training sample contained in the training sample set may include, in addition to sample protein molecules and sample candidate molecules, reference activity values ​​after virtual binding of sample protein molecules and sample candidate molecules, and the reference activity values ​​are used to characterize the real activity of sample protein molecules and sample candidate molecules after drug binding experiments. If the sample protein molecules and sample candidate molecules are obtained based on drug binding experiments, then the reference activity values ​​are obtained based on the experimental results of the drug binding experiments of the sample protein molecules and sample candidate molecules. If the sample protein molecules and sample candidate molecules are obtained from network resources, then the reference activity values ​​are also obtained from the corresponding network resources, etc., without specific limitations. Among them, the training sample set can be obtained from the PDBbind data set.

[0134] The server can use an activity prediction model to predict the sample activity value after virtual binding of the sample protein molecule and the sample candidate molecule based on the binding activity feature information, and determine the third training loss of the molecular binding model to be trained based on the third error value between the sample activity value and the corresponding reference activity value.

[0135] The server can respectively determine whether the first training loss meets the first training target, whether the second training loss meets the second training target, and whether the third training loss meets the third training target. When the first training loss, the second training loss, and the third training loss all meet their respective corresponding training targets, it is determined that the training loss of the molecular binding model to be trained meets the training target.

[0136] As an embodiment, the training target may be a pre-set target error value or a convergence state, and is not specifically limited. After obtaining the first training loss, the second training loss, and the third training loss, it may be determined whether the first training loss converges, whether the second training loss converges, and whether the third training loss converges, respectively, so as to determine whether the training loss of the molecular binding model to be trained meets the training target. When the first training loss, the second training loss, and the third training loss all converge, it is determined that the training loss of the molecular binding model to be trained meets the training target.

[0137] S503: When it is determined that the obtained training loss does not meet the training target, the model parameters of the molecular binding model to be trained are adjusted.

[0138] If the obtained training loss does not meet the training goal, it means that the prediction ability of the molecular binding model to be trained is not accurate enough and further training is needed. The current model parameters need to be further adjusted, so adjust the model parameters of the molecular binding model to be trained.

[0139] As an embodiment, if at least one of the first training loss, the second training loss and the third training loss does not converge, it is determined that the obtained training loss does not meet the training objective, and the model parameters of the molecular binding model to be trained are adjusted.

[0140] S504: When it is determined that the obtained training loss meets the preset training target, the molecular binding model to be trained is output as the trained target molecular binding model.

[0141] If the training loss obtained meets the training goal, it means that the prediction ability of the molecular binding model to be trained is relatively accurate and the current model parameters do not need to be adjusted anymore. Then the molecular binding model to be trained is output as the trained target molecular binding model.

[0142] As an example, the molecular binding model to be trained may be trained in a gradient training manner, and the binding activity information may include pIC50.

[0143] The training process in the embodiment of the present application is implemented in an end-to-end form. There is no need for artificial sub-problem division. Instead, it is completely handed over to the model to directly learn the mapping from raw data to expected output, which improves the intelligence of the training model and avoids the problem of low prediction accuracy of the trained model due to unpredictable situations when humans are involved.

[0144] After obtaining the target molecule structure model, the server can combine the feature extraction model, the target molecule structure model and the activity prediction model to predict the binding activity of the target protein molecule and the target candidate molecule after virtual binding. After obtaining the target protein molecule and the target candidate molecule, the server uses the feature extraction model to extract features of the target protein molecule and the target candidate molecule respectively to obtain target protein feature information and target molecule feature information. Based on the target protein feature information and the target molecule feature information, the server uses the target molecule structure model to obtain the binding activity feature information of the target protein molecule and the target candidate molecule after virtual binding. Based on the binding activity feature information, the server uses the activity prediction model to predict the activity value of the target protein molecule and the target candidate molecule after virtual binding.

[0145] The following is an example of the method for training a molecular binding model provided in the embodiment of the present application. Please refer to Figure 6 .

[0146] The server obtains the sample protein molecule and the sample candidate molecule, and uses a feature extraction model to respectively determine the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample candidate molecule, and uses the feature extraction model to perform feature extraction processing on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample candidate molecule, respectively, to obtain protein feature information and molecule feature information.

[0147] The server adopts the molecular binding model to be trained, and determines the binding activity feature information, embedding feature information and co-crystal feature information between the sample protein molecule and the sample candidate molecule based on the protein feature information and the molecular feature information; obtains the first error value between the embedded feature information and the embedded target based on graph matching; obtains the second error value between the co-crystal target and the co-crystal target based on neighbor matching; the server adopts the activity prediction model, obtains the sample activity value based on the binding activity feature information, obtains the reference activity value through the training sample, and determines the third error value between the sample activity value and the reference activity value.

[0148] The server determines a first training loss of the molecular binding model to be trained based on the first error value, determines a second training loss of the molecular binding model to be trained based on the second error value, and determines a third training loss of the molecular binding model to be trained based on the third error value.

[0149] When the server determines that there is a non-convergent training loss among the training losses, it is determined that the training loss of the molecular binding model to be trained is non-convergent, and the parameters of the molecular binding model to be trained are adjusted; when it is determined that all the training losses have converged, it is determined that the training loss of the molecular binding model to be trained has converged, the molecular binding model to be trained is output, and the trained molecular binding model is obtained.

[0150] Based on the same inventive concept, the present embodiment provides a molecular screening method, which uses the target molecule binding model trained by the method for training the molecular binding model introduced above to perform molecular screening, please refer to Figure 7 , which is a schematic diagram of a molecular screening method.

[0151] S701, obtaining target protein molecules and target candidate molecules.

[0152] When performing molecular screening, the target protein molecule and the target candidate molecule can be obtained first. There are multiple methods for obtaining the target protein molecule and the target candidate molecule. As an embodiment, the target client can obtain the target protein molecule and the target candidate molecule in response to the molecule selection operation triggered by the target object. The target client sends the target protein molecule and the target candidate molecule to the server, and the server can receive the target protein molecule and the target candidate molecule sent by the target client.

[0153] As an embodiment, a protein molecule set and a candidate molecule set may be pre-set in the server. When performing molecular screening, the server may select a protein molecule from the protein molecule set as a target protein molecule and select a candidate molecule from the candidate molecule set as a target candidate molecule. When the server selects from the protein molecule set or the candidate molecule set, it may be a random selection or a selection based on the property information of the molecules contained in the set and in accordance with a preset selection strategy, etc., without any specific limitation.

[0154] S702, using a feature extraction model to perform feature extraction processing on the target protein molecule and the target candidate molecule to obtain target protein feature information of the target protein molecule and target molecule feature information of the target candidate molecule.

[0155] S703, using the trained target molecule binding model, based on the target protein feature information and the target molecule feature information, determine the binding activity feature information between the target protein molecule and the target candidate molecule.

[0156] The trained target molecule binding model can be trained by other devices and then sent to the server. The server receives the trained target molecule binding model sent by other devices. It can also be the trained target molecule binding model obtained by the server for the molecular binding model to be trained. There is no specific limitation.

[0157] S704, using an activity prediction model, based on the binding activity characteristic information, predicting the activity value of the target protein molecule and the target candidate molecule after virtual binding.

[0158] After obtaining the virtual target compound, and the virtual target embedding information and virtual target co-crystal information corresponding to the virtual target compound, the server uses the target molecular structure model to obtain the predicted activity information corresponding to the virtual target compound based on the virtual target compound, the virtual target embedding information and the virtual target co-crystal information. This process is similar to the process described above of using the molecular binding model to be trained to obtain the sample activity information corresponding to the sample compound based on the sample compound, the sample embedding information and the sample co-crystal information, and will not be repeated here.

[0159] S705, based on the predicted target activity information and the preset molecular screening conditions, determine whether to select the target protein molecule and the target candidate molecule for combination.

[0160] After obtaining the predicted marker activity information, the server can determine whether to select the target protein molecule and the target candidate molecule for combination based on the predicted marker activity information and the preset molecular screening condition. The preset molecular screening condition can be a preset activity threshold. When the predicted marker activity information is greater than or equal to the preset activity threshold, the target protein molecule and the target candidate molecule are selected for combination; when the predicted marker activity information is less than the preset activity threshold, the target protein molecule and the target candidate molecule are not selected for combination.

[0161] The preset molecular screening condition may also be a preset sorting sequence number. After obtaining the predicted marker activity information of each group of target protein molecules and target candidate molecules, the activity values ​​contained in each predicted marker activity information are sorted in descending order, and the target protein molecules and target candidate molecules ranked before the preset sorting sequence number are selected for combination.

[0162] Based on the same inventive concept, the present application embodiment provides a device for training a molecular binding model, which is equivalent to the server discussed above and can implement the functions corresponding to the method for training a molecular binding model. Figure 8 , the device includes an acquisition module 801 and a processing module 802, wherein:

[0163] Acquisition module 801: used to obtain protein feature information of each sample protein molecule, and obtain molecular feature information of each sample candidate molecule;

[0164] Processing module 802: for performing multiple rounds of iterative training on the molecular binding model to be trained based on each protein feature information and each molecule feature information, until the training loss of the molecular binding model to be trained meets the training target, and outputting the molecular binding model to be trained as the target molecular binding model, wherein the following operations are performed for multiple rounds of iterative training respectively:

[0165] Using the molecular binding model to be trained, based on the protein feature information and the molecular feature information, determining the binding activity feature information, the embedding feature information and the co-crystal feature information between the sample protein molecule and the sample candidate molecule, wherein the binding activity feature information is used to characterize the activity of the sample protein molecule after virtual binding with the sample candidate molecule, the embedding feature information is used to characterize the degree of binding between the sample protein molecule and the sample candidate molecule, and the co-crystal feature information is used to characterize whether there is a co-crystal structure between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction;

[0166] The training loss of the molecular binding model to be trained is determined based on the binding activity feature information, the embedding feature information and the co-crystal feature information.

[0167] In a possible embodiment, the acquisition module 801 is specifically used for:

[0168] Obtaining a training sample set, wherein the training sample set includes various training samples, and each training sample includes a sample protein molecule and a sample candidate molecule;

[0169] The feature extraction model is used to perform feature extraction processing on each training sample to obtain the protein feature information of each sample protein molecule and the molecular feature information of each sample candidate molecule.

[0170] In a possible embodiment, the acquisition module 801 is specifically used for:

[0171] For each training sample, perform the following operations:

[0172] Based on the sample protein molecules in the training sample, determining the adjacency matrix of the sample protein molecules, wherein the adjacency matrix of the sample protein molecules is used to characterize each sample amino acid molecule contained in the sample protein molecule, and the molecular structure distance between every two sample amino acid molecules;

[0173] Based on the sample candidate molecules in the training samples, determining the adjacency matrix of the sample candidate molecules, wherein the adjacency matrix of the sample candidate molecules is used to characterize each sample candidate atom contained in the sample candidate molecules, and the chemical bond structure between each sample candidate atom;

[0174] Feature extraction processing is performed on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample candidate molecule respectively to obtain protein feature information of the sample protein molecule and molecular feature information of the sample candidate molecule.

[0175] In a possible embodiment, the processing model is specifically used to:

[0176] Based on the protein feature information and the molecular feature information, predict the binding distances between each sample amino acid molecule contained in the sample protein molecule and each sample candidate atom contained in the sample candidate molecule after the sample protein molecule is virtually bound to the sample candidate molecule;

[0177] Based on the obtained binding distances, the embedding characteristic information and co-crystal characteristic information between the sample protein molecule and the sample candidate molecule are determined.

[0178] In a possible embodiment, the processing module 802 is specifically configured to:

[0179] Determine the embedded feature information between the sample protein molecule and the sample candidate molecule based on the minimum binding distance between the designated sample amino acid molecule and each sample candidate atom;

[0180] Based on each sample amino acid molecule, each sample candidate atom and the binding distance between each amino acid molecule and each sample candidate atom, the co-crystal characteristic information between the sample protein molecule and the sample candidate molecule is determined.

[0181] In a possible embodiment, the processing module 802 is specifically configured to:

[0182] Determining a first training loss of a molecular binding model to be trained based on a first error value between the embedded feature information and the embedded target;

[0183] Determining a second training loss of the molecular binding model to be trained based on a second error value between the co-crystal feature information and the co-crystal target;

[0184] Determining a third training loss of the molecular binding model to be trained based on a third error value between the binding activity feature information and the activity target;

[0185] Based on the obtained first training loss, second training loss and third training loss, a training loss of the molecular binding model to be trained is determined.

[0186] The processing module 802 is specifically used for:

[0187] The trained neighbor matching model is used to match each sample amino acid molecule contained in the sample protein molecule with each sample candidate atom contained in the sample candidate molecule to obtain matching feature information between the sample protein molecule and the sample candidate molecule as the co-crystal target, wherein the matching feature information is used to characterize the matching distance between each sample amino acid molecule and each sample candidate atom.

[0188] Using a cross entropy function, determining a second error value between the eutectic feature information and the eutectic target;

[0189] The second error value is used as a second training loss of the molecular binding model to be trained.

[0190] In a possible embodiment, the processing module 802 is specifically configured to:

[0191] An activity prediction model is used to predict the sample activity value after the sample protein molecule is virtually combined with the sample candidate molecule based on the binding activity characteristic information;

[0192] Based on a third error value between the sample activity value and the corresponding reference activity value, a third training loss of the molecular binding model to be trained is determined.

[0193] In a possible embodiment, the processing module 802 is further configured to:

[0194] Determining that the training loss of the molecular binding model to be trained does not meet the training target, and adjusting the model parameters of the molecular binding model to be trained;

[0195] When it is determined that the training loss of the molecular binding model to be trained meets the training target, the molecular binding model to be trained is output as the trained target molecular binding model.

[0196] In a possible embodiment, the processing module 802 is specifically configured to:

[0197] If the obtained training loss includes a first training loss, a second training loss, and a third training loss, the first training loss is determined based on the embedding feature information, the second training loss is determined based on the eutectic feature information, and the third training loss is determined based on the binding activity feature information, then determining whether the first training loss converges, whether the second training loss converges, and whether the third training loss converges, respectively;

[0198] If at least one of the first training loss, the second training loss, and the third training loss does not converge, the model parameters of the molecular binding model to be trained are adjusted.

[0199] Based on the same inventive concept, the present application embodiment provides a molecular screening device, which is equivalent to the server discussed above and can realize the functions corresponding to the aforementioned molecular screening method. Fig. 9 , the device includes an acquisition module 901 and a processing module 902, wherein:

[0200] Acquisition module 901: used to obtain target protein molecules and target candidate molecules;

[0201] Processing module 902: for performing feature extraction processing on the target protein molecule and the target candidate molecule using a feature extraction model to obtain target protein feature information of the target protein molecule and target molecule feature information of the target candidate molecule;

[0202] The processing module 902 is also used to: determine the binding activity characteristic information between the target protein molecule and the target candidate molecule based on the target protein characteristic information and the target molecule characteristic information by using the target molecule binding model;

[0203] The processing module 902 is also used to: use the activity prediction model to predict the activity value of the target protein molecule and the target candidate molecule after virtual binding based on the binding activity characteristic information.

[0204] Based on the same inventive concept, an embodiment of the present application provides a computer device, and the computer device 1000 is introduced below.

[0205] Please refer to Fig.10 The above-mentioned device for training a molecular binding model or a molecular screening device can run on a computer device 1000, and the current version and historical version of the data storage program and the application software corresponding to the data storage program can be installed on the computer device 1000. The computer device 1000 includes a display unit 1040, a processor 1080 and a memory 1020, wherein the display unit 1040 includes a display panel 1041 for displaying an interactive operation interface by a user, etc.

[0206] In a possible embodiment, the display panel 1041 may be configured in the form of a liquid crystal display (LCD) or an organic light-emitting diode (OLED).

[0207] The processor 1080 is used to read the computer program and then execute the method defined by the computer program. For example, the processor 1080 reads the data storage program or file, so as to run the data storage program on the computer device 1000 and display the corresponding interface on the display unit 1040. The processor 1080 may include one or more general-purpose processors and may also include one or more DSPs (Digital Signal Processors) to perform related operations to implement the technical solutions provided in the embodiments of the present application.

[0208] The memory 1020 generally includes internal memory and external memory, and the internal memory can be a random access memory (RAM), a read-only memory (ROM), and a cache (CACHE) etc. The external memory can be a hard disk, an optical disk, a USB disk, a floppy disk or a tape drive etc. The memory 1020 is used to store computer programs and other data, and the computer program includes the application corresponding to each client etc., and other data may include the data generated after the operating system or the application is run, and the data include system data (such as configuration parameters of the operating system) and user data. In the embodiment of the present application, program instructions are stored in the memory 1020, and the processor 1080 executes the program instructions stored in 1020, realizing any method or molecular screening method of a training molecule binding model discussed in the foregoing figure.

[0209] The display unit 1040 is used to receive input digital information, character information or contact touch operation / contactless gesture, and generate signal input related to user settings and function control of the computer device 1000. Specifically, in the embodiment of the present application, the display unit 1040 may include a display panel 1041. The display panel 1041, such as a touch screen, can collect user touch operations on or near it (such as operations performed by the user using fingers, stylus, or any other suitable object or accessory on or on the display panel 1041), and drive corresponding connection devices according to a pre-set program.

[0210] In a possible embodiment, the display panel 1041 may include two parts: a touch detection device and a touch controller. The touch detection device detects the touch position of the player, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device, converts it into the touch point coordinates, and then sends it to the processor 1080, and can receive and execute the command sent by the processor 1080.

[0211] The display panel 1041 may be implemented in various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the display unit 1040, the computer device 1000 may further include an input unit 1030, which may include a graphic input device 1031 and other input devices 1032, wherein the other input devices may include, but are not limited to, one or more of a physical keyboard, a function key (such as a volume control key, a switch key, etc.), a trackball, a mouse, a joystick, etc.

[0212] In addition to the above, the computer device 1000 may also include a power supply 1090 for supplying power to other modules, an audio circuit 1060, a near field communication module 1070, and an RF circuit 109. The computer device 1000 may also include one or more sensors 1050, such as an acceleration sensor, a light sensor, a pressure sensor, etc. The audio circuit 1060 specifically includes a speaker 1061 and a microphone 1062, etc. For example, the computer device 1000 can collect the user's voice through the microphone 1062 and perform corresponding operations.

[0213] As an embodiment, the number of processors 1080 may be one or more, and the processor 1080 and the memory 1020 may be coupled or relatively independently configured.

[0214] As an example, Fig.10 The processor 1080 in the embodiment may be used to implement the following Figure 8 The functions of the acquisition module 801 and the processing module 802 in Fig. 9 The functions of the acquisition module 901 and the processing module 902.

[0215] As an example, Fig.10 The processor 1080 in can be used to implement the corresponding functions of the server 102 discussed above.

[0216] A person skilled in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, etc. Various media that can store program codes.

[0217] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected 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, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0218] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A method for training a molecular binding model, characterized in that include: Obtaining protein characteristic information of each sample protein molecule, and obtaining molecular characteristic information of each sample candidate molecule; Based on each protein feature information and each molecule feature information, the molecular binding model to be trained is subjected to multiple rounds of iterative training until the training loss of the molecular binding model to be trained meets the training target, and the target molecular binding model is output, wherein the following operations are performed for the multiple rounds of iterative training respectively: Using the molecular binding model to be trained, based on protein feature information and molecular feature information, determining binding activity feature information, embedding feature information and co-crystal feature information, wherein the binding activity feature information characterizes the activity of the sample protein molecule after virtual binding with the sample candidate molecule, the embedding feature information characterizes the degree of binding between the sample protein molecule and the sample candidate molecule, and the co-crystal feature information characterizes whether a co-crystal structure exists between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction; Using a trained neighbor matching model, each sample amino acid molecule contained in the sample protein molecule is matched with each sample candidate atom contained in the sample candidate molecule, and the obtained matching feature information is used as a co-crystal target, wherein the matching feature information characterizes the matching distance between each sample amino acid molecule and each sample candidate atom; The training loss of the molecular binding model to be trained is determined based on the binding activity feature information, the embedding feature information, and a second error value between the co-crystal feature information and the co-crystal target.

2. The method according to claim 1, characterized in that Obtain protein feature information of each sample protein molecule, and obtain molecular feature information of each sample candidate molecule, including: Obtaining a training sample set, wherein the training sample set includes various training samples, and each training sample includes a sample protein molecule and a sample candidate molecule; A feature extraction model is used to perform feature extraction processing on each training sample to obtain protein feature information of each sample protein molecule and molecular feature information of each sample candidate molecule.

3. The method according to claim 2, characterized in that Using a feature extraction model, each training sample is subjected to feature extraction processing to obtain protein feature information of each sample protein molecule and molecular feature information of each sample candidate molecule, including: For each training sample, the following operations are performed respectively: Based on the sample protein molecules in the training sample, determining the adjacency matrix of the sample protein molecules, wherein the adjacency matrix of the sample protein molecules is used to characterize each sample amino acid molecule contained in the sample protein molecule, and the molecular structure distance between every two sample amino acid molecules; Determine an adjacency matrix of the sample candidate molecules based on the sample candidate molecules in the training sample, wherein the adjacency matrix of the sample candidate molecules is used to characterize each sample candidate atom contained in the sample candidate molecules, and the chemical bond structure between each sample candidate atom; Feature extraction processing is performed on the adjacency matrix of the sample protein molecule and the adjacency matrix of the sample candidate molecule respectively to obtain protein feature information of the sample protein molecule and molecular feature information of the sample candidate molecule.

4. The method according to claim 1, characterized in that: Using the molecular binding model to be trained, based on the protein feature information and the molecular feature information, determining the embedding feature information and the co-crystal feature information, including: Based on the protein characteristic information and the molecule characteristic information, predicting the binding distance between each sample amino acid molecule contained in the sample protein molecule and each sample candidate atom contained in the sample candidate molecule after the sample protein molecule is virtually bound to the sample candidate molecule; Based on the obtained binding distances, embedding characteristic information and co-crystal characteristic information between the sample protein molecule and the sample candidate molecule are determined.

5. The method according to claim 4, characterized in that Determining the embedding characteristic information and co-crystal characteristic information between the sample protein molecule and the sample candidate molecule based on the obtained binding distances, including: Determine the embedded feature information between the sample protein molecule and the sample candidate molecule based on the minimum binding distance between the designated sample amino acid molecule and each sample candidate atom; Based on each sample amino acid molecule, each sample candidate atom and the binding distance between each amino acid molecule and each sample candidate atom, the co-crystal characteristic information between the sample protein molecule and the sample candidate molecule is determined.

6. The method according to claim 1, characterized in that Determining the training loss of the molecular binding model to be trained based on the binding activity feature information, the embedding feature information, and a second error value between the co-crystal feature information and the co-crystal target, comprising: Determining a first training loss of the molecular binding model to be trained based on a first error value between the embedded feature information and the embedded target; Determining a second training loss of the molecular binding model to be trained based on a second error value between the co-crystal feature information and the co-crystal target; Determining a third training loss of the molecular binding model to be trained based on a third error value between the binding activity feature information and the activity target; Based on the obtained first training loss, second training loss and third training loss, the training loss of the molecular binding model to be trained is determined.

7. The method according to claim 6, characterized in that Determining a second training loss of the molecular binding model to be trained based on a second error value between the co-crystal feature information and the co-crystal target includes: Using a cross entropy function, determining a second error value between the eutectic characteristic information and the eutectic target; The second error value is used as a second training loss of the molecular binding model to be trained.

8. The method according to claim 1, characterized in that Each training sample also includes a reference activity value after a sample protein molecule and a sample candidate molecule are virtually combined, and based on a third error value between the binding activity characteristic information and the activity target, a third training loss of the molecular binding model to be trained is determined, including: Using an activity prediction model, based on the binding activity characteristic information, predicting the sample activity value after the sample protein molecule is virtually bound to the sample candidate molecule; Based on a third error value between the sample activity value and the corresponding reference activity value, a third training loss of the molecular binding model to be trained is determined.

9. The method according to claim 1, characterized in that: After determining the training loss of the molecular binding model to be trained, the method further includes: Determining that the training loss of the molecular binding model to be trained does not meet the training target, and adjusting the model parameters of the molecular binding model to be trained; When it is determined that the training loss of the molecular binding model to be trained meets the training target, the molecular binding model to be trained is output as the trained target molecular binding model.

10. The method according to claim 9, characterized in that Determining that the training loss of the molecular binding model to be trained does not meet the training target, and adjusting the model parameters of the molecular binding model to be trained, including: If the obtained training loss includes a first training loss, a second training loss, and a third training loss, the first training loss is determined based on the embedding feature information, the second training loss is determined based on the eutectic feature information, and the third training loss is determined based on the binding activity feature information, then determining whether the first training loss converges, whether the second training loss converges, and whether the third training loss converges, respectively; If at least one of the first training loss, the second training loss, and the third training loss does not converge, the model parameters of the molecular binding model to be trained are adjusted.

11. A molecular screening method, characterized in that: The target molecule binding model is obtained by training based on the method according to any one of claims 1 to 10, comprising: Obtaining target protein molecules and target candidate molecules; Using a feature extraction model, performing feature extraction processing on the target protein molecule and the target candidate molecule to obtain target protein feature information of the target protein molecule and target molecule feature information of the target candidate molecule; Using the target molecule binding model, based on the target protein characteristic information and the target molecule characteristic information, determining the binding activity characteristic information between the target protein molecule and the target candidate molecule; An activity prediction model is used to predict the activity value of the target protein molecule and the target candidate molecule after virtual binding based on the binding activity characteristic information.

12. A device for training a molecular binding model, characterized in that: include: Acquisition module: used to obtain protein characteristic information of each sample protein molecule, and to obtain molecular characteristic information of each sample candidate molecule; Processing module: used to perform multiple rounds of iterative training on the molecular binding model to be trained based on each protein feature information and each molecule feature information, until the training loss of the molecular binding model to be trained meets the training target, and output the target molecular binding model, wherein for multiple rounds of iterative training, the following operations are performed respectively: Using the molecular binding model to be trained, based on protein feature information and molecular feature information, determining binding activity feature information, embedding feature information and co-crystal feature information, wherein the binding activity feature information characterizes the activity of the sample protein molecule after virtual binding with the sample candidate molecule, the embedding feature information characterizes the degree of binding between the sample protein molecule and the sample candidate molecule, and the co-crystal feature information characterizes whether a co-crystal structure exists between the sample protein molecule and the sample candidate molecule, and the molecular binding model to be trained is used for prediction; Using a trained neighbor matching model, each sample amino acid molecule contained in the sample protein molecule is matched with each sample candidate atom contained in the sample candidate molecule, and the obtained matching feature information is used as a co-crystal target, wherein the matching feature information characterizes the matching distance between each sample amino acid molecule and each sample candidate atom; The training loss of the molecular binding model to be trained is determined based on the binding activity feature information, the embedding feature information, and a second error value between the co-crystal feature information and the co-crystal target.

13. A molecular screening device, characterized in that: include: Acquisition module: used to obtain target protein molecules and target candidate molecules; Processing module: used for performing feature extraction processing on the target protein molecule and the target candidate molecule by using a feature extraction model to obtain target protein feature information of the target protein molecule and target molecule feature information of the target candidate molecule; The processing module is also used to: determine the binding activity characteristic information between the target protein molecule and the target candidate molecule based on the target protein characteristic information and the target molecule characteristic information by using a target molecule binding model; the target molecule binding model is obtained by training based on the method according to any one of claims 1 to 10; The processing module is also used to: use an activity prediction model to predict the activity value of the target protein molecule and the target candidate molecule after virtual binding based on the binding activity characteristic information.

14. A computer device, characterized in that: include: A memory for storing program instructions; The processor is used to call the program instructions stored in the memory, and execute the method according to any one of claims 1 to 11 according to the obtained program instructions.

15. A computer-readable storage medium, characterized in that: The storage medium stores computer-executable instructions, and the computer-executable instructions are used to enable a computer to execute the method according to any one of claims 1 to 11.

Images