Method, device, equipment and storage medium for determining interaction information

Abstract

The present application discloses a method, apparatus, device and storage medium for determining interaction information. The method includes: obtaining basic information of a first target object, basic information of a second target object and a target interaction information prediction model, wherein the target interaction information prediction model is trained using a loss function at a global level and a loss function at a key local level; calling the target interaction information prediction model to process the basic information of the first target object and the basic information of the second target object to obtain target interaction information between the first target object and the second target object. Based on the above process, the training process of the interaction information prediction model focuses on both global information and key local information, the training effect of the interaction information prediction model is good, and the accuracy of determining the interaction information between the first target object and the second target object using the trained interaction information prediction model is high.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a method, apparatus, device, and storage medium for determining interaction information. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, the use of interaction information prediction models to determine the interaction between two targets is becoming increasingly common. For example, interaction information prediction models can be used to determine the pIC50 value between a target (i.e., a protein) and a drug (i.e., a small molecule) as interaction information for drug screening. The pIC50 value represents the negative logarithm of the concentration of the small molecule that achieves a 50% inhibitory effect on the protein.

[0003] In related technologies, only a global loss function is used to train an interaction information prediction model, and the interaction information between two targets is then determined using the trained interaction information prediction model. In this process, the global loss function only allows the training of the interaction information prediction model to focus on global information, which is relatively limited. This results in poor training results and low accuracy in the interaction information between the two targets determined using the trained interaction information prediction model. Summary of the Invention

[0004] The present invention provides a method, apparatus, device, and storage medium for determining interaction information, which can be used to improve the training effect of the interaction information prediction model. The technical solution is as follows:

[0005] In one aspect, an embodiment of the present application provides a method for determining interaction information, the method comprising:

[0006] Obtaining basic information of the first target object, basic information of the second target object, and a target interaction information prediction model, wherein the target interaction information prediction model is trained using a global-level loss function and a key local-level loss function, wherein the key local-level loss function is determined based on attention information corresponding to key sub-sample objects in the sample objects that meet a reference condition, where the key sub-sample objects in the sample objects that meet the reference condition are some sub-sample objects in all sub-sample objects;

[0007] The target interaction information prediction model is called to process the basic information of the first target object and the basic information of the second target object to obtain the target interaction information between the first target object and the second target object.

[0008] In another aspect, a device for determining interaction information is provided, the device comprising:

[0009] an acquisition unit, configured to acquire basic information of the first target object, basic information of the second target object, and a target interaction information prediction model, wherein the target interaction information prediction model is trained using a global-level loss function and a key local-level loss function, wherein the key local-level loss function is determined based on attention information corresponding to key sub-sample objects in the sample objects that meet a reference condition, wherein the key sub-sample objects in the sample objects that meet the reference condition are some sub-sample objects in all sub-sample objects;

[0010] A processing unit is configured to call the target interaction information prediction model to process the basic information of the first target and the basic information of the second target to obtain target interaction information between the first target and the second target.

[0011] In one possible implementation, the processing unit includes:

[0012] a first acquisition subunit, configured to call the target interaction information prediction model, and acquire, based on the basic information of the first target, first basic feature information corresponding to the first target and first attention information corresponding to the first target; and acquire, based on the basic information of the second target, second basic feature information corresponding to the second target and second attention information corresponding to the second target;

[0013] a second acquisition subunit, configured to acquire first global feature information corresponding to the first target object based on the first basic feature information and the first attention information; and acquire second global feature information corresponding to the second target object based on the second basic feature information and the second attention information;

[0014] The third acquisition subunit is configured to acquire target interaction information between the first target object and the second target object based on the first global feature information and the second global feature information.

[0015] In one possible implementation, the first global feature information includes the first global feature information at the node level and the first global feature information at the edge level, the second global feature information includes the second global feature information at the node level and the second global feature information at the edge level, the target interaction information includes the target interaction information at the node level and the target interaction information at the edge level, and the target interaction information prediction model includes a first prediction processing model and a second prediction processing model; the third acquisition subunit is used to call the first prediction processing model to perform prediction processing on the first global feature information at the node level and the second global feature information at the node level to obtain the node-level target interaction information between the first target object and the second target object; and call the second prediction processing model to perform prediction processing on the first global feature information at the edge level and the second global feature information at the edge level to obtain the edge-level target interaction information between the first target object and the second target object.

[0016] In one possible implementation, the basic information of the first target object is the graph information of the first target object, and the basic information of the second target object is the graph information of the second target object; the target interaction information prediction model includes a first node information transmission model, a first edge information transmission model, a second node information transmission model, and a second edge information transmission model; the first basic feature information includes the first basic feature information at the node level and the first basic feature information at the edge level, and the second basic feature information includes the second basic feature information at the node level and the second basic feature information at the edge level; the first acquisition subunit is further used to call the first node information transmission model to perform node-level feature extraction on the graph information of the first target object to obtain the first basic feature information at the node level; call the first edge information transmission model to perform edge-level feature extraction on the graph information of the first target object to obtain the first basic feature information at the edge level; call the second node information transmission model to perform node-level feature extraction on the graph information of the second target object to obtain the second basic feature information at the node level; call the second edge information transmission model to perform edge-level feature extraction on the graph information of the second target object to obtain the second basic feature information at the edge level.

[0017] In one possible implementation, the first node information transmission model includes a first node feature update layer and a first node-level feature output layer; the first edge information transmission model includes a first edge feature update layer and a first edge-level feature output layer; the first acquisition subunit is further used to call the first node feature update layer to update the features of the nodes in the graph information of the first target object to obtain target node features; call the first node-level feature output layer to output the target node features to obtain the first basic feature information at the node level; call the first edge feature update layer to update the features of the edges in the graph information of the first target object to obtain target edge features; call the first edge-level feature output layer to output the target edge features to obtain the first basic feature information at the edge level.

[0018] In one possible implementation, the apparatus further includes:

[0019] A determination unit is used to determine a first key sub-target among all sub-targets of the first target based on first attention information corresponding to the first target, wherein the first key sub-target is used to indicate a sub-target in the first target used to interact with the second target; and to determine a second key sub-target among all sub-targets of the second target based on second attention information corresponding to the second target, wherein the second key sub-target is used to indicate a sub-target in the second target used to interact with the first target.

[0020] In one possible implementation, the processing unit is further configured to call an initial interaction information prediction model to obtain attention information corresponding to a first sample object, attention information corresponding to a second sample object, and predicted interaction information between the first sample object and the second sample object, wherein the predicted interaction information is obtained based on global feature information corresponding to the first sample object and the global feature information corresponding to the second sample object;

[0021] The acquiring unit is further configured to acquire a global loss function based on the predicted interaction information and the standard interaction information between the first sample object and the second sample object;

[0022] The determining unit is further configured to determine, from the attention information corresponding to the sample objects satisfying the reference condition, attention information corresponding to a key sub-sample object in the sample object satisfying the reference condition, wherein the sample object satisfying the reference condition is at least one of the first sample object and the second sample object;

[0023] The acquisition unit is further configured to acquire a loss function at a key local level based on attention information corresponding to the key subsample objects in the sample objects that meet the reference condition;

[0024] The device further comprises:

[0025] an updating unit, configured to reversely update the parameters of the initial interaction information prediction model based on the loss function at the global level and the loss function at the key local level;

[0026] The acquisition unit is further configured to obtain a target interaction information prediction model in response to the parameter updating process satisfying a termination condition.

[0027] In one possible implementation, the predicted interaction information includes predicted interaction information at the node level and predicted interaction information at the edge level; the acquisition unit is further used to determine a first sub-loss function based on the predicted interaction information at the node level and the standard interaction information between the first sample object and the second sample object; determine a second sub-loss function based on the predicted interaction information at the edge level and the standard interaction information between the first sample object and the second sample object; determine a third sub-loss function based on the predicted interaction information at the node level and the predicted interaction information at the edge level; and determine the loss function at the global level based on the first sub-loss function, the second sub-loss function, and the third sub-loss function.

[0028] In one possible implementation, the sample objects that meet the reference condition are the first sample object and the second sample object; the determining unit is further configured to determine, from the attention information corresponding to the first sample object, attention information corresponding to a first key sub-sample object in the first sample object; and to determine, from the attention information corresponding to the second sample object, attention information corresponding to a second key sub-sample object in the second sample object.

[0029] The acquisition unit is further used to determine a fourth sub-loss function based on the attention information corresponding to the first key sub-sample object; determine a fifth sub-loss function based on the attention information corresponding to the second key sub-sample object; and determine the loss function of the key local level based on the fourth sub-loss function and the fifth sub-loss function.

[0030] In one possible implementation, the type of the first target is protein, and the type of the second target is a small molecule; the acquisition unit is further used to determine the spatial distance between the amino acids in the first target based on the structural information of the first target; determine the adjacency matrix corresponding to the first target based on the spatial distance between the amino acids, and the adjacency matrix is ​​used to indicate the association relationship between the amino acids in the first target; according to the adjacency matrix corresponding to the first target and the amino acids in the first target, obtain the graph information of the first target, and use the graph information of the first target as the basic information of the first target; based on the atoms in the second target and the chemical bond information between the atoms, obtain the graph information of the second target, and use the graph information of the second target as the basic information of the second target.

[0031] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one program code, and the at least one program code is loaded and executed by the processor to implement any of the above-mentioned methods for determining interaction information.

[0032] On the other hand, a computer-readable storage medium is provided, in which at least one program code is stored. The at least one program code is loaded and executed by a processor to implement any of the above-mentioned methods for determining interaction information.

[0033] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the aforementioned methods for determining interaction information.

[0034] The technical solutions provided by the embodiments of the present application bring at least the following beneficial effects:

[0035] In an embodiment of the present application, a loss function at the global level and a loss function at the key local level are used to train an interaction information prediction model, and then the interaction information between the first target object and the second target object is determined using the trained target interaction information prediction model. The loss function at the global level enables the model training process to focus on global information; the loss function at the key local level enables the model training process to focus on key local information. In other words, in an embodiment of the present application, the training process of the interaction information prediction model focuses on both global information and key local information, the training effect of the interaction information prediction model is better, and the accuracy of determining the interaction information between the first target object and the second target object using the trained interaction information prediction model is higher. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0037] Figure 1 is a schematic diagram of an implementation environment of a method for determining interaction information provided in an embodiment of the present application;

[0038] Figure 2 is a flow chart of a method for determining interaction information provided by an embodiment of the present application;

[0039] Figure 3 is a schematic diagram of partial image information of a first target object provided in an embodiment of the present application;

[0040] Figure 4 is a flow chart of a method for obtaining target interaction information between a first target object and a second target object provided in an embodiment of the present application;

[0041] Figure 5 is a schematic diagram of a process for acquiring target interaction information between a first target object and a second target object provided in an embodiment of the present application;

[0042] Figure 6 This is a flowchart of a method for training a target interaction information prediction model provided by an embodiment of the present application;

[0043] Figure 7 is a schematic diagram of a device for determining interaction information provided by an embodiment of the present application;

[0044] Figure 8 This is a schematic diagram of the structure of a processing unit provided in an embodiment of the present application;

[0045] Figure 9 is a schematic diagram of a device for determining interaction information provided by an embodiment of the present application;

[0046] Figure 10 It is a structural diagram of a device for determining interaction information provided in an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0048] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0049] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0050] The solution provided in the embodiments of the present application relates to machine learning technology for artificial intelligence. Machine Learning (ML) is a multi-disciplinary interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. 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 own 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 generally include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0051] 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, smart marketing, unmanned driving, autonomous driving, drones, robots, smart medical care, smart customer service, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0052] This application embodiment provides a method for determining interaction information, please refer to Figure 1 , which shows a schematic diagram of an implementation environment of the method for determining interaction information provided by an embodiment of the present application. The implementation environment includes: a terminal 11 and a server 12.

[0053] The method for determining interaction information provided in the embodiments of the present application can be executed by terminal 11 or by server 12, and this is not limited in the embodiments of the present application. In an exemplary embodiment, when the method for determining interaction information provided in the embodiments of the present application is executed by terminal 11, the interaction information obtained by terminal 11 is sent to server 12 for storage. Of course, terminal 11 can also store the obtained interaction information.

[0054] In an exemplary embodiment, when the method for determining the interaction information provided in the embodiment of the present application is executed by the server 12, the server 12 can send the obtained interaction information to the terminal 11 for storage. Of course, the server 12 can also store the obtained interaction information.

[0055] In one possible implementation, terminal 11 may be a smart device such as a mobile phone, tablet computer, or personal computer. Server 12 may be a single server, a server cluster consisting of multiple servers, or a cloud computing service center. Terminal 11 and server 12 establish a communication connection via a wired or wireless network.

[0056] Those skilled in the art should understand that the above-mentioned terminal 11 and server 12 are only examples. Other existing or future terminals or servers that are applicable to this application should also be included in the scope of protection of this application and are included here by reference.

[0057] Based on the above Figure 1 In the implementation environment shown, the present application embodiment provides a method for determining interaction information, taking the method applied to the server 12 as an example. Figure 2 As shown, the method provided in the embodiment of the present application may include the following steps:

[0058] In step 201, basic information of the first target object, basic information of the second target object and a target interaction information prediction model are obtained. The target interaction information prediction model is trained using a loss function at the global level and a loss function at the key local level.

[0059] Among them, the loss function at the key local level is determined based on the attention information corresponding to the key sub-sample objects in the sample objects that meet the reference conditions, and the key sub-sample objects in the sample objects that meet the reference conditions are some sub-sample objects in all sub-sample objects.

[0060] The first target and the second target refer to two targets whose interaction information needs to be determined using the target interaction information prediction model. The embodiment of the present application does not limit the types of the first target and the second target. For example, the first target and the second target are both proteins; or, the first target and the second target are both small molecules; or, one of the first target and the second target is a protein and the other is a small molecule. It should be noted that the first target and the second target each include at least one sub-target. When the type of a certain target is protein, the type of the sub-target included in the target is amino acid; when the type of a certain target is a small molecule, the type of the sub-target included in the target is atom.

[0061] Proteins have spatial structures, and proteins are formed by chains of amino acids folding in space; small molecules refer to molecules with smaller molecular weights, and illustratively, small molecules are molecules with a molecular weight of less than 500. In an exemplary embodiment, small molecules refer to drugs. For the case where one of the first target and the second target is a protein and the other is a small molecule, the process of determining the interaction information between the two targets realizes the virtual screening of drugs. Virtual screening plays a very important role in drug development and can greatly reduce the time and cost required for related experiments. The embodiment of the present application can use the target interaction information prediction model obtained by training to perform virtual screening of drugs, effectively utilize the characteristic information of proteins and small molecules, and improve the reliability of virtual screening of drugs.

[0062] The basic information of the first target is information used to characterize the first target and is used to input the target interaction information prediction model, and the basic information of the second target is information used to characterize the second target and is used to input the target interaction information prediction model. The basic information of the first target and the basic information of the second target provide data support for the prediction process of the target interaction information prediction model. In an exemplary embodiment, the basic information of the first target is the graph information of the first target, and the basic information of the second target is the graph information of the second target. The graph information of the target is used to represent the target in the form of a graph including nodes and edges.

[0063] In one possible implementation, the implementation method of obtaining the basic information of the first target and the basic information of the second target is related to the type of the first target and the type of the second target. In an exemplary embodiment, the type of the first target and the type of the second target include the following four cases:

[0064] Case 1: The first target is a protein, and the second target is a small molecule.

[0065] In one possible implementation, in this case 1, the implementation method of obtaining the basic information of the first target object and the basic information of the second target object includes the following steps 1 to 4:

[0066] Step 1: Based on the structural information of the first target, determine the spatial distance between amino acids in the first target.

[0067] When the type of the first target is protein, the structural information of the first target refers to the structural information of the protein corresponding to the first target. The structural information of the protein is recorded in the structural file of the protein, and the structural information of the first target can be obtained in the structural file of the protein corresponding to the first target.

[0068] When the first target is a protein, the sub-targets included in the first target are amino acids. The structural information of the first target includes basic information about each amino acid in the first target. Based on this basic information, the spatial distance between amino acids in the first target can be determined. It should be noted that the spatial distance between amino acids in the first target includes the spatial distance between any two amino acids in the first target.

[0069] In one possible implementation, after determining the spatial distances between amino acids in the first target, the spatial distances between the amino acids can be normalized to obtain a standardized spatial distance between the amino acids. The standardized spatial distance between any two amino acids can provide a reference for obtaining basic features of an edge connecting the nodes corresponding to the two amino acids.

[0070] In an exemplary embodiment, for amino acids i and j in the protein corresponding to the first target, the standard spatial distance between amino acids i and j is obtained based on Formula 1:

[0071]

[0072] in, represents the standard spatial distance between amino acids i and j; d ij represents the spatial distance between amino acid i and amino acid j; d′ represents the scaling scale, and exemplarily, the value of d′ is

[0073] Step 2: Based on the spatial distance between the amino acids, determine the adjacency matrix corresponding to the first target, where the adjacency matrix is ​​used to indicate the association relationship between the amino acids in the first target.

[0074] After determining the spatial distances between the amino acids in the first target, an adjacency matrix is ​​determined based on the spatial distances between the amino acids to indicate the association relationship between the amino acids in the first target. In one possible implementation, the adjacency matrix is ​​composed of values ​​indicating the association relationship between any two amino acids. The value indicating the association relationship between amino acid i and amino acid j is determined based on Formula 2:

[0075]

[0076] Among them, A ij represents a value indicating the association relationship between amino acid i and amino acid j; d ij represents the spatial distance between amino acid i and amino acid j; d0 represents the distance threshold, and exemplarily, the value of d0 is Based on Formula 2, a value indicating the association relationship between any two amino acids can be determined, thereby obtaining an adjacency matrix corresponding to the first target.

[0077] Step 3: According to the adjacency matrix corresponding to the first target and the amino acids in the first target, the graph information of the first target is obtained, and the graph information of the first target is used as the basic information of the first target.

[0078] The graph information of the first target is used to represent the first target in the form of a graph. In one possible implementation, based on the adjacency matrix corresponding to the first target and the amino acids in the first target, the graph information of the first target is obtained by treating the amino acids in the first target as nodes and constructing edges between the nodes based on the adjacency matrix to obtain the graph information of the first target.

[0079] In one possible implementation, the process of constructing edges between nodes according to the adjacency matrix is ​​as follows: for two nodes corresponding to any two amino acids, if the value determined based on the adjacency matrix to indicate the association relationship between the any two amino acids is 1, then an edge is constructed between the two nodes corresponding to the any two amino acids; if the value determined based on the adjacency matrix to indicate the association relationship between the any two amino acids is 0, then no edge is constructed between the two nodes corresponding to the any two amino acids.

[0080] For example, the partial image information of the first target object is as follows: Figure 3 As shown. Figure 3 In the partial graph information shown, amino acid A, amino acid B, amino acid C, amino acid D and amino acid E are all nodes. Since the spatial distance between amino acid A and amino acid B is less than the distance threshold (i.e., d AB <d0), so there is an edge between the two nodes corresponding to amino acid A and amino acid B. Similarly, there is an edge between the two nodes corresponding to amino acid B and amino acid C, between the two nodes corresponding to amino acid C and amino acid D, and between the two nodes corresponding to amino acid C and amino acid E.

[0081] After acquiring the image information of the first object, the image information of the first object is used as the basic information of the first object, thereby obtaining the basic information of the first object.

[0082] Step 4: Based on the atoms in the second target and the chemical bond information between the atoms, obtain the graph information of the second target, and use the graph information of the second target as the basic information of the second target.

[0083] The graph information of the second target object is used to represent the second target object in the form of a graph. When the second target object is a small molecule, the sub-target objects included in the second target object are atoms. In one possible implementation, based on the atoms in the second target object and the chemical bond information between the atoms, the graph information of the second target object is obtained by treating the atoms in the second target object as nodes, constructing edges between the nodes based on the chemical bond information between the atoms, and obtaining the graph information of the second target object.

[0084] In an exemplary embodiment, the process of constructing edges between nodes based on the chemical bond information between atoms is as follows: for any two atoms, if the chemical bond information between the atoms indicates that there is a chemical bond connection between the any two atoms, then an edge is constructed between the nodes corresponding to the any two atoms; if the chemical bond information between the atoms indicates that there is no chemical bond connection between the any two atoms, then no edge is constructed between the nodes corresponding to the any two atoms.

[0085] After acquiring the image information of the second object, the image information of the second object is used as the basic information of the second object, thereby obtaining the basic information of the second object.

[0086] Case 2: The first target is a small molecule, and the second target is a protein.

[0087] In one possible implementation, in this scenario 2, the basic information of the first target is obtained by obtaining graph information of the first target based on information about atoms and chemical bonds between atoms in the first target, and using the graph information of the first target as the basic information of the first target. The method for obtaining the basic information of the first target is similar to step 4 in scenario 1 and is not further described here.

[0088] The basic information of the second target is obtained by determining the spatial distances between amino acids in the second target based on the structural information of the second target; determining an adjacency matrix corresponding to the second target based on the spatial distances between the amino acids, where the adjacency matrix indicates the relationships between the amino acids in the second target; and obtaining graph information of the second target based on the adjacency matrix corresponding to the second target and the amino acids in the second target, and using the graph information of the second target as the basic information of the second target. The method for obtaining the basic information of the second target is similar to steps 1 to 3 under scenario 1 and is not further described here.

[0089] Case 3: The type of the first target is protein, and the type of the second target is also protein.

[0090] In one possible implementation, in this case 3, the implementation method for obtaining the basic information of the first target is: based on the structural information of the first target, determining the spatial distance between the amino acids in the first target; based on the spatial distance between the amino acids, determining the adjacency matrix corresponding to the first target, the adjacency matrix is ​​used to indicate the association relationship between the amino acids in the first target; according to the adjacency matrix corresponding to the first target and the amino acids in the first target, obtaining the graph information of the first target, and using the graph information of the first target as the basic information of the first target.

[0091] The implementation method for obtaining the basic information of the second target is: based on the structural information of the second target, determining the spatial distance between the amino acids in the second target; based on the spatial distance between the amino acids, determining the adjacency matrix corresponding to the second target, the adjacency matrix is ​​used to indicate the association relationship between the amino acids in the second target; according to the adjacency matrix corresponding to the second target and the amino acids in the second target, obtaining the graph information of the second target, and using the graph information of the second target as the basic information of the second target.

[0092] The above-mentioned implementation method of obtaining the basic information of the first target object and the implementation method of obtaining the basic information of the second target object are both referred to steps 1 to 3 under situation 1, and will not be repeated here.

[0093] Case 4: The first target is a small molecule, and the second target is also a small molecule.

[0094] In one possible implementation, in this case 4, the method for obtaining the basic information of the first target is: based on the information of the atoms in the first target and the chemical bonds between the atoms, the graph information of the first target is obtained, and the graph information of the first target is used as the basic information of the first target. The method for obtaining the basic information of the second target is: based on the information of the atoms in the second target and the chemical bonds between the atoms, the graph information of the second target is obtained, and the graph information of the second target is used as the basic information of the first target. The above-mentioned methods for obtaining the basic information of the first target and the methods for obtaining the basic information of the second target are both referred to step 4 under case 1 and are not repeated here.

[0095] It should be noted that the above is merely an exemplary description of obtaining basic information about the first target and the second target, and the embodiments of the present application are not limited thereto. In exemplary embodiments, the basic information about the first target and the second target may also be other information besides image information, as long as it can provide data support for the prediction process of the target interaction information prediction model, and the embodiments of the present application are not limited thereto.

[0096] In one possible implementation, after obtaining the basic information of the first target object and the basic information of the second target object, the basic information of the first target object and the basic information of the second target object can be stored so as to facilitate subsequent acquisition of the basic information of the first target object and the basic information of the second target object by direct extraction.

[0097] In order to determine the interaction information between the first target and the second target, in addition to obtaining the basic information of the first target and the basic information of the second target, it is also necessary to obtain a target interaction information prediction model. The target interaction information prediction model refers to a trained interaction information prediction model. The target interaction information prediction model is trained using a loss function at the global level and a loss function at the key local level. Among them, the loss function at the global level is used to make the model training process focus on global information, and the loss function at the key local level is used to make the model training process focus on key local information. In other words, the process of training to obtain the target interaction information prediction model pays attention to both global information and key local information, and the training effect is better.

[0098] The process of training the target interaction information prediction model will be Figure 6The details of the global loss function and the key local loss function will be described in detail in the embodiment shown in FIG. Figure 6 The embodiment shown is described in detail and will not be described here in detail.

[0099] It should be noted that in step 201, the target interaction information prediction model may be obtained by directly extracting a trained target interaction information prediction model or by training the target interaction information prediction model, and this is not limited in this embodiment of the present application. In the case of directly extracting a trained target interaction information prediction model, the process of training the target interaction information prediction model has already been completed before executing step 201, and the trained target interaction information prediction model has been stored.

[0100] In step 202, a target interaction information prediction model is called to process basic information of the first target and basic information of the second target to obtain target interaction information between the first target and the second target.

[0101] The target interaction information between the first target and the second target is information that is predicted using a target interaction information prediction model and is used to reflect the interaction between the first target and the second target. The meaning of the target interaction information between the first target and the second target is related to the types of the first target and the second target. In an exemplary embodiment, for example, when the first target and the second target are both small molecules, the interaction information between the first target and the second target indicates the chemical reaction information between the two small molecules; when the first target and the second target are both proteins, the interaction information between the first target and the second target indicates the binding information between the two proteins; when one of the first target and the second target is a protein and the other is a small molecule, the interaction information between the first target and the second target indicates protein-small molecule activity information, and the protein-small molecule activity information can be used to screen small molecules.

[0102] The present embodiment does not limit the representation format of protein-small molecule activity information. For example, protein-small molecule activity information is represented using pIC50 values. Here, pIC50 = -1g(IC50), where IC50 represents the concentration of the small molecule that achieves a 50% inhibitory effect on the protein.

[0103] After obtaining the basic information of the first target object, the basic information of the second target object and the target interaction information prediction model, the target interaction information prediction model can be called to process the basic information of the first target object and the basic information of the second target object, thereby obtaining the target interaction information between the first target object and the second target object.

[0104] In one possible implementation, see Figure 4 The implementation process of step 202 includes the following steps 2021 to 2023:

[0105] Step 2021: Call the target interaction information prediction model, and based on the basic information of the first target object, obtain the first basic feature information corresponding to the first target object and the first attention information corresponding to the first target object; based on the basic information of the second target object, obtain the second basic feature information corresponding to the second target object and the second attention information corresponding to the second target object.

[0106] In one possible implementation, based on the basic information of the first target object, a process of obtaining first basic feature information corresponding to the first target object and first attention information corresponding to the first target object includes the following steps 21a and 221b:

[0107] Step 2021a: Based on the basic information of the first target object, obtain first basic feature information corresponding to the first target object.

[0108] In one possible implementation, the target interaction information prediction model includes a first feature extraction model, which is used to extract the first basic feature information corresponding to the first target object based on the basic information of the first target object. Based on this, the process of obtaining the first basic feature information corresponding to the first target object based on the basic information of the first target object is: calling the first feature extraction model to perform feature extraction on the basic information of the first target object, and obtaining the first basic feature information corresponding to the first target object. The embodiment of the present application does not limit the model structure of the first feature extraction model, as long as the first basic feature information corresponding to the first target object can be determined based on the basic information of the first target object. In an exemplary embodiment, when the basic information of the first target object is the graph information of the first target object, the model structure of the first feature extraction model is a dual information transmission model composed of a node information transmission model and an edge information transmission model.

[0109] In one possible implementation, the basic information of the first target object is the graph information of the first target object, the first feature extraction model includes a first node information transmission model and a first edge information transmission model, and the first basic feature information includes the first basic feature information at the node level and the first basic feature information at the edge level. In this case, the process of calling the first feature extraction model to perform feature extraction on the basic information of the first target object to obtain the first basic feature information corresponding to the first target object includes: calling the first node information transmission model to perform node-level feature extraction on the graph information of the first target object to obtain the first basic feature information at the node level; calling the first edge information transmission model to perform edge-level feature extraction on the graph information of the first target object to obtain the first basic feature information at the edge level. In an exemplary embodiment, the model structures of the first node information transmission model and the first edge information transmission model are both MPNN (Message Passing Neural Network).

[0110] In one possible implementation, the first node information transfer model includes a first node feature update layer and a first node-level feature output layer. The first node information transfer model is called to perform node-level feature extraction on the graph information of the first target object. The process of obtaining the first basic feature information at the node level includes the following steps A and B:

[0111] Step A: Call the first node feature update layer to update the features of the nodes in the graph information of the first target object to obtain the target node features.

[0112] The target node features include the updated node features of each node in the graph information of the first target object obtained after processing by the first node feature update layer. The first node feature update layer in the first node information transmission model obtains the updated node features of the node by aggregating the node features around a certain node and all edge features to the node, and updates the reference number step. The reference number is set based on experience or flexibly adjusted according to the application scenario, and the embodiment of the present application does not limit this. It should be noted that a node in the graph information of the first target object corresponds to a sub-target object in the first target object. If the type of the first target object is protein, then a node in the graph information of the first target object corresponds to an amino acid; if the type of the first target object is a small molecule, then a node in the graph information of the first target object corresponds to an atom.

[0113] In an exemplary embodiment, the process of calling the first node feature update layer to update the features of the nodes in the graph information of the first target object is implemented based on Formula 3:

[0114]

[0115]

[0116]

[0117] in, represents the initial node feature of node v; σ(·) represents the activation function; x v Represents the basic features of node v, which are determined based on the basic information of the sub-target corresponding to node v. For example, when the first target is a protein, the basic features of node v are determined based on the basic information of the amino acids corresponding to node v, including but not limited to the type, sequence number, and structure of the amino acids. When the first target is a small molecule, the basic features of node v are determined based on the basic information of the atoms corresponding to node v, including but not limited to the type, sequence number, and charge of the atoms.

[0118] represents the value of the information transfer function at the d+1th step (d is an integer not less than 0) corresponding to node v; cat(·,·) represents the concatenation function; N(v) represents the set of adjacent nodes of node v; represents the node feature of node k after d steps of updating; e vk Represents the basic features of the edge from node k to node v. For example, the basic features of the edge from node k to node v are obtained based on the distance between the sub-target object corresponding to node k and the sub-target object corresponding to node v. It should be noted that when the type of the first target is protein, the distance between the amino acid corresponding to node k and the amino acid corresponding to node v refers to the standard space distance. In formula 3, e vk is the correlation feature in the information transmission process (μ attached ).

[0119] It represents the node feature of node v after d+1 steps of updating, and also represents the value of the node update function corresponding to node v in the d+1th step. The node update function uses a linear transformation plus a bias; W in and W a is a parameter of the first node feature update layer. In an exemplary embodiment, when the first node feature update layer is called to perform a multi-step update, W in and W a These two parameters are shared.

[0120] Assuming that the reference number is D (D is an integer not less than 1), after performing D-step update processing according to Formula 3, the target node features including the updated node features of each node in the graph information of the first target object are obtained, and step B is executed.

[0121] Step B: Call the first node-level feature output layer to perform output processing on the target node features to obtain the first basic feature information at the node level.

[0122] The first basic feature information at the node level includes the node-level output features corresponding to each node in the graph information of the first target object.

[0123] The processing of the first node-level feature output layer can be considered an additional information transfer step. In this first node-level feature output layer, different parameters are used to obtain the first basic feature information at the node level corresponding to the first target object. In an exemplary embodiment, the process of calling the first node-level feature output layer to output the target node feature is implemented based on Formula 4:

[0124]

[0125] in, represents the fusion feature of node υ at the node level; represents the node feature of node k after D steps of updating; x k represents the basic features of node k; in Formula 4, k represents the union of the set of adjacent nodes of node v and node v; W represents the node-level output feature corresponding to node v; o Represents the parameters of the feature output layer at the first node level; the meanings of other parameters are the same as those in Formula 3.

[0126] Based on Formula 4, the node-level output features corresponding to each node in the graph information of the first target object can be obtained, and then the first basic feature information at the node level corresponding to the first target object can be obtained.

[0127] For example, the first basic feature information at the node level corresponding to the first target object is represented as Here, n (n is an integer not less than 1) is the number of sub-reference objects included in the first target object, and is also the number of nodes in the graph information of the first target object.

[0128] In one possible implementation, the first edge information transfer model includes a first edge feature update layer and a first edge-level feature output layer. Calling the first edge information transfer model to perform edge-level feature extraction on the graph information of the first target object to obtain first basic feature information at the edge level includes the following steps a and b:

[0129] Step a: Call the first edge feature update layer to update the edge features in the graph information of the first target object to obtain the target edge features.

[0130] The target edge features include updated edge features of each edge in the graph information of the first target object obtained after processing by the first edge feature update layer. The first edge feature update layer in the first node information transfer model is used to update the edge features in the graph information of the first target object by aggregating information on the edges. In an exemplary embodiment, the process of calling the first edge feature update layer to update the edge features in the graph information of the first target object is implemented based on Formula 5:

[0131]

[0132]

[0133]

[0134] in, represents the initial edge feature of the edge from node w to node v; e vw Represents the basic features of the edge from node w to node v; The value of the information transfer function corresponding to the d+1th step (d is an integer not less than 0) of the edge from node w to node v; represents the edge feature from node k to node v after d steps of updating; x k Represents the basic features of node k. It is calculated by aggregating the features of the neighboring edge set of the edge from node w to node v and the features of the nodes corresponding to the neighboring edges in the neighboring edge set. Among them, the neighboring edge set of the edge from node w to node v refers to the set of all edges starting from node v except the edge from node w to node v. In formula 5, x k is the correlation feature in the information transmission process (μ attached ).

[0135] W′ represents the edge feature of the edge from node w to node v after d+1 steps of update. in and W b is a parameter of the first edge feature update layer. In an exemplary embodiment, when the first edge feature update layer is called to perform a multi-step update, W′ in and W b These two parameters are shared.

[0136] In an exemplary embodiment, the number of update steps during the process of calling the first edge feature update layer to update the edge features in the graph information of the first target object is the same as the number of update steps during the process of calling the first node feature update layer to update the node features in the graph information of the first target object. Assuming that the same number of update steps is D (D is an integer not less than 1), after performing the update process for D steps according to Formula 5, a target edge feature including the updated edge features of each edge in the graph information of the first target object is obtained, and step b is executed.

[0137] Step b: Call the first edge-level feature output layer to perform output processing on the target edge feature to obtain the first basic feature information at the edge level.

[0138] The first basic feature information at the edge level includes output features at the edge level corresponding to each node in the graph information of the first target object.

[0139] The processing of the first edge-level feature output layer can be considered a node information aggregation step. In this first edge-level feature output layer, the edge features are transferred to the nodes, thereby obtaining the first basic feature information of the edge level. In an exemplary embodiment, the process of calling the first edge-level feature output layer to output the target edge features is implemented based on Formula 6:

[0140]

[0141] in, represents the fusion features of node v at the edge level; represents the edge feature from node k to node v after D-step update; x k Represents the basic features of node k; Represents the output feature of the edge layer corresponding to node v; W′0 represents the parameters of the first edge layer feature output layer.

[0142] Based on Formula 6, the output features of the edge layer corresponding to each node in the graph information of the first target object can be obtained, and then the first basic feature information of the edge layer corresponding to the first target object can be obtained.

[0143] For example, the first basic feature information of the edge layer corresponding to the first target object is expressed as Here, n (n is an integer not less than 1) is the number of sub-reference objects included in the first target object, and is also the number of nodes in the graph information corresponding to the first target object.

[0144] After obtaining the first basic feature information at the node level and the first basic feature information at the edge level corresponding to the first target object, the first basic feature information corresponding to the first target object is obtained.

[0145] It should be noted that the above is only an exemplary description of obtaining the first basic feature information corresponding to the first target object based on the first feature extraction model, and this application is not limited to this. In an exemplary embodiment, the first feature extraction model is a single-branch information transmission model, and the first basic feature information corresponding to the first target object is the basic feature information of an entire object.

[0146] Step 2021b: Based on the first basic feature information corresponding to the first target object, obtain the first attention information corresponding to the first target object.

[0147] The first attention information includes the attention information of each node in the graph information of the first target object, and the attention information of any node includes the attention weight corresponding to the node at one or more angles. The embodiment of the present application does not limit the number of angles of the attention weight corresponding to the node included in the attention information of any node. The number can be set based on experience or flexibly adjusted according to the application scenario. Exemplarily, the number of angles of the attention weight corresponding to the node included in the attention information of any node is r (r is an integer not less than 1). In an exemplary embodiment, the sum of the attention weights corresponding to the nodes in the graph information of the first target object at the same angle is 1.

[0148] The first attention information corresponding to the first target object is obtained based on the first basic feature information corresponding to the first target object. In one possible implementation, for the case where the first basic feature information includes the first basic feature information at the node level and the first basic feature information at the edge level, the first attention information corresponding to the first target object includes the first attention information at the node level and the first attention information at the edge level. In this case, based on the first basic information corresponding to the first target object, the process of obtaining the first attention information corresponding to the first target object is: based on the first basic information at the node level, obtaining the first attention information at the node level corresponding to the first target object; based on the first basic information at the edge level, obtaining the first attention information at the edge level corresponding to the first target object.

[0149] The first attention information at the node level includes the attention information at the node level of each node in the graph information of the first target object, and the first attention information at the edge level includes the attention information at the edge level of each node in the graph information of the first target object. For any node in the graph information of the first target object, the attention information at the node level of any node may be the same as the attention information at the edge level of any node, or may be different from the attention information at the edge level of any node, and this embodiment of the application is not limited to this.

[0150] In one possible implementation, based on the first basic feature information at the node level, the process of obtaining the first attention information at the node level corresponding to the first target object is implemented based on Formula 7:

[0151]

[0152] in, Represents the first attention information at the node level corresponding to the first target object; Represents the first basic feature information at the node level The transpose of ; W1 and W2 are both learnable parameters; tanh(·) represents the hyperbolic tangent function. In an exemplary embodiment, W1∈R h×a , W2∈R r×h In Formula 7 above, W1 is used to perform a linear transformation, transforming the first basic feature information at the node level in the a-dimensional space into the h-dimensional space. Then, the hyperbolic tangent function tanh(·) is used for nonlinear mapping. W2 then linearly transforms the features in the h-dimensional space into the r-dimensional space (r is an integer not less than 1), resulting in the distribution of node attention weights at r different angles. For any angle, the greater the attention weight corresponding to a node at that angle, the more important the node is at that angle. Finally, the softmax(·) function is used to sum the attention weights for each angle to 1.

[0153] In one possible implementation, based on the first basic feature information at the edge level, the process of obtaining the first attention information at the edge level corresponding to the first target object is implemented based on Formula 8:

[0154]

[0155] in, Represents the first attention information at the edge level corresponding to the first target object; Represents the first basic feature information at the node level The meaning of other references is the same as that of Formula 7. In an exemplary embodiment, the two parameters W1 and W2 in Formula 8 are shared with the two parameters W1 and W2 in Formula 7, so that the first basic feature information at the node level and the first basic feature information at the edge level can interact with each other during the training process.

[0156] After obtaining the first attention information at the node level and the first attention information at the edge level, the first attention information corresponding to the first target object is obtained. In an exemplary embodiment, the above formulas 7 and 8 can be regarded as self-attention readout functions, and the first attention information corresponding to the first target object can be obtained through the self-attention readout function.

[0157] It should be noted that the above is only an exemplary description of obtaining the first attention information corresponding to the first target object based on the first basic feature information, and the embodiments of the present application are not limited to this. When the first basic information is other than the above situation, the process of obtaining the first attention information will also change.

[0158] In one possible implementation, based on the basic information of the second target object, the process of obtaining the second basic feature information corresponding to the second target object and the second attention information corresponding to the second target object includes the following steps 2021c and 2021d:

[0159] Step 2021c: Based on the basic information of the second target object, obtain second basic feature information corresponding to the second target object.

[0160] In one possible implementation, the target interaction information prediction model also includes a second feature extraction model, which is used to extract the second basic feature information corresponding to the second target object based on the basic information of the second target object. Based on this, the process of obtaining the second basic feature information corresponding to the second target object based on the basic information of the second target object is: calling the second feature extraction model to perform feature extraction on the basic information of the second target object, and obtaining the second basic feature information corresponding to the second target object. The embodiment of the present application does not limit the model structure of the second feature extraction model, as long as the second basic feature information corresponding to the second target object can be determined based on the basic information of the second target object. In an exemplary embodiment, the model structure of the second feature extraction model is the same as the model structure of the first feature extraction model. When the model structure of the first feature extraction model is a dual information transmission model composed of a node information transmission model and an edge information transmission model, the model structure of the second feature extraction model is also a dual information transmission model composed of a node information transmission model and an edge information transmission model.

[0161] In one possible implementation, the basic information of the second target object is the graph information of the second target object, the second feature extraction model includes a second node information transmission model and a second edge information transmission model, and the second basic feature information includes the second basic feature information at the node level and the second basic feature information at the edge level. In this case, the process of calling the second feature extraction model to perform feature extraction on the basic information of the second target object to obtain the second basic feature information corresponding to the second target object includes: calling the second node information transmission model to perform node-level feature extraction on the graph information of the second target object to obtain the second basic feature information at the node level; calling the second edge information transmission model to perform edge-level feature extraction on the graph information of the second target object to obtain the second basic feature information at the edge level. In an exemplary embodiment, the model structures of the second node information transmission model and the second edge information transmission model are both MPNN (Message Passing Neural Network).

[0162] In one possible implementation, the second node information transmission model includes a second node feature update layer and a second node-level feature output layer. The process of calling the second node information transmission model to extract node-level features from the graph information of the second target object to obtain the second basic feature information at the node level includes: calling the second node feature update layer to update the features of the nodes in the graph information of the second target object to obtain target node features; and calling the second node-level feature output layer to output the target node features to obtain the second basic feature information at the node level. The implementation of this process refers to steps A and B in step 2021a and is not further described here.

[0163] In one possible implementation, the second edge information transfer model includes a second edge feature update layer and a second edge-level feature output layer. The process of invoking the second edge information transfer model to extract edge-level features from the graph information of the second target object to obtain the second basic feature information at the edge level is as follows: invoking the second edge feature update layer to update the edge features in the graph information of the second target object to obtain target edge features; and invoking the second edge-level feature output layer to output the target edge features to obtain the second basic feature information at the edge level. The implementation of this process refers to steps a and b in step 2021a and is not further described here.

[0164] After obtaining the second basic feature information at the node level and the second basic feature information at the edge level corresponding to the second target object, the second basic feature information corresponding to the second target object is obtained.

[0165] It should be noted that the above is only an exemplary description of obtaining the second basic feature information corresponding to the second target object based on the second feature extraction model, and this application is not limited to this. In an exemplary embodiment, the second feature extraction model is a single-branch information transmission model, and the second basic feature information corresponding to the second target object is a whole basic feature information.

[0166] Step 2021d: Based on the second basic feature information corresponding to the second target object, obtain the second attention information corresponding to the second target object.

[0167] The second attention information includes the attention information of each node in the graph information of the second target object, and the second attention information corresponding to the second target object is obtained based on the second basic feature information corresponding to the second target object.

[0168] In one possible implementation, for the case where the second basic feature information includes the second basic feature information at the node level and the second basic feature information at the edge level, the second attention information corresponding to the second target object includes the second attention information at the node level and the second attention information at the edge level. In this case, based on the second basic information, the process of obtaining the second attention information corresponding to the second target object is: based on the second basic information at the node level, obtaining the second attention information at the node level corresponding to the second target object; based on the second basic information at the edge level, obtaining the second attention information at the edge level corresponding to the second target object. The implementation method of this process is shown in step 2021b and will not be repeated here. After obtaining the second attention information at the node level and the second attention information at the edge level, the second attention information corresponding to the second target object is obtained.

[0169] It should be noted that the parameter values ​​in the self-attention readout function used in the process of obtaining the second attention information at the node level and the second attention information at the edge level corresponding to the second target object may be different from or the same as the parameter values ​​in the self-attention readout function used in the process of obtaining the first attention information at the node level and the first attention information at the edge level corresponding to the first target object. This is not limited in the embodiments of the present application.

[0170] The second attention information at the node level includes the attention information at the node level of each node in the graph information of the second target object, and the second attention information at the edge level includes the attention information at the edge level of each node in the graph information of the second target object. For any node in the graph information of the second target object, the attention information at the node level of any node may be the same as the attention information at the edge level of any node, or may be different from the attention information at the edge level of any node, and this embodiment of the application is not limited to this.

[0171] It should be noted that the above is only an exemplary description of obtaining the second attention information corresponding to the second target object based on the second basic feature information, and the embodiments of the present application are not limited to this. When the second basic information is other than the above situation, the process of obtaining the second attention information will also change.

[0172] Step 2022: Based on the first basic feature information and the first attention information, obtain the first global feature information corresponding to the first target object; based on the second basic feature information and the second attention information, obtain the second global feature information corresponding to the second target object.

[0173] The first global feature information corresponding to the first target object refers to the global feature information corresponding to the first target object that includes attention information. In one possible implementation, the first basic feature information includes the first basic feature information at the node level and the first basic feature information at the edge level, and the first attention information includes the first attention information at the node level and the first attention information at the edge level. In this case, the first global feature information corresponding to the first target object includes the first global feature information at the node level and the first global feature information at the edge level.

[0174] In one possible implementation, the process of obtaining the first global feature information corresponding to the first target object based on the first basic feature information and the first attention information is: based on the first basic feature information at the node level and the first attention information at the node level, obtaining the first global feature information at the node level corresponding to the first target object; based on the first basic feature information at the edge level and the first attention information at the edge level, obtaining the first global feature information at the edge level corresponding to the first target object.

[0175] The first global feature information at the node level corresponding to the first target object refers to the global feature information corresponding to the first target object that includes the attention information at the node level. In one possible implementation, based on the first basic feature information at the node level and the first attention information at the node level, the process of obtaining the first global feature information at the node level corresponding to the first target object is implemented based on Formula 9:

[0176]

[0177] in, Representing first global feature information at the node level corresponding to the first target object; Representing first basic feature information at the node level corresponding to the first target object; represents the first attention information at the node level corresponding to the first target object; the flatten(·) function represents the expansion into a one-dimensional vector. In an exemplary embodiment, but Wherein, n represents the number of nodes in the graph information of the first target object, r represents the number of angles of attention weight corresponding to any node, and a represents the custom alignment value. Even if the number of nodes in the graph information of different first target objects is different, based on Formula 9, a fixed-size global feature information containing node attention weights can be obtained.

[0178] The first global feature information at the edge level corresponding to the first target object refers to the global feature information corresponding to the first target object that includes the attention information at the edge level. In one possible implementation, based on the first basic feature information at the edge level and the first attention information at the edge level, the process of obtaining the first global feature information at the edge level corresponding to the first target object is implemented based on Formula 10:

[0179]

[0180] in, Representing first global feature information of an edge layer corresponding to the first target object; Representing first basic feature information of an edge layer corresponding to the first target object; Represents the first attention information at the edge level corresponding to the first target object.

[0181] After obtaining the first global feature information at the node level and the first global feature information at the edge level, the first global feature information corresponding to the first target object is obtained.

[0182] The second global feature information corresponding to the second target object refers to the global feature information corresponding to the second target object that includes attention information. In one possible implementation, the second basic feature information includes the second basic feature information at the node level and the second basic feature information at the edge level, and the second attention information includes the second attention information at the node level and the second attention information at the edge level. In this case, the second global feature information corresponding to the second target object includes the second global feature information at the node level and the second global feature information at the edge level.

[0183] In one possible implementation, the process of obtaining the second global feature information corresponding to the second target object based on the second basic feature information and the second attention information is: based on the second basic feature information at the node level and the second attention information at the node level, obtaining the second global feature information at the node level corresponding to the second target object; based on the second basic feature information at the edge level and the second attention information at the edge level, obtaining the second global feature information at the edge level corresponding to the second target object.

[0184] The second global feature information at the node level corresponding to the second target object refers to the global feature information corresponding to the second target object, including the attention information at the node level. The second global feature information at the edge level corresponding to the second target object refers to the global feature information corresponding to the second target object, including the attention information at the edge level. The method for obtaining the second global feature information at the node level and the second global feature information at the edge level corresponding to the second target object refers to the method for obtaining the first global feature information at the node level and the first global feature information at the edge level corresponding to the first target object, which will not be repeated here. After obtaining the second global feature information at the node level and the second global feature information at the edge level, the second global feature information corresponding to the second target object is obtained.

[0185] It should be noted that the above is only an exemplary description of obtaining the first global feature information corresponding to the first target object and the second global feature information corresponding to the second target object, and the embodiments of the present application are not limited to this. In an exemplary embodiment, for the case where the first basic feature information and the second basic feature information are both basic feature information of a whole, the first attention information and the second attention information are also attention information of a whole. In this case, the first global feature information of a whole is obtained directly based on the first basic feature information and the first attention information, and the second global feature information of a whole is obtained directly based on the second basic feature information and the second attention information.

[0186] It should be noted that since the characteristics of all nodes in the graph information of the first target object are comprehensively considered in the process of obtaining the first global feature information, the first global feature information can represent the first target object from a global perspective. Similarly, the second global feature information can represent the second target object from a global perspective.

[0187] Step 2023: Based on the first global feature information and the second global feature information, obtain target interaction information between the first target object and the second target object.

[0188] In one possible implementation, the target interaction information prediction model includes a prediction processing model. Based on the first global feature information and the second global feature information, the implementation process of obtaining the target interaction information between the first target object and the second target object is: calling the prediction processing model to perform prediction processing on the first global feature information and the second global feature information to obtain the target interaction information between the first target object and the second target object.

[0189] After obtaining first global feature information corresponding to the first target and second global feature information corresponding to the second target, the first and second global feature information are input into a prediction processing model. After being processed by the prediction processing model, target interaction information between the first and second targets is obtained. In an exemplary embodiment, after the first and second global feature information are input into the prediction processing model, the prediction processing model integrates the first and second global feature information to obtain target interaction information between the first and second targets.

[0190] In one possible implementation, for a case where the first global feature information includes first global feature information at the node level and first global feature information at the edge level, and the second global feature information includes second global feature information at the node level and second global feature information at the edge level, the target interaction information between the first target object and the second target object includes target interaction information at the node level and target interaction information at the edge level, and the prediction processing model includes a first prediction processing model and a second prediction processing model. The first prediction processing model is used to obtain target interaction information at the node level based on the global feature information at the node level, and the second prediction processing model is used to obtain target interaction information at the edge level based on the global feature information at the edge level.

[0191] In the above case, the process of calling the prediction processing model to perform prediction processing on the first global feature information and the second global feature information to obtain the target interaction information between the first target object and the second target object is: calling the first prediction processing model to perform prediction processing on the first global feature information at the node level and the second global feature information at the node level to obtain the node-level target interaction information between the first target object and the second target object; calling the second prediction processing model to perform prediction processing on the first global feature information at the edge level and the second global feature information at the edge level to obtain the edge-level target interaction information between the first target object and the second target object.

[0192] In one possible implementation, the first prediction processing model is a fully connected neural network. The first prediction processing model is called to perform prediction processing on the first global feature information at the node level and the second global feature information at the node level. The process of obtaining the node-level target interaction information between the first target object and the second target object is implemented based on Formula 11:

[0193]

[0194] Among them, Pred a representing target interaction information at a node level between a first target object and a second target object; Representing first global feature information at the node level corresponding to the first target object, used to characterize the global features at the node level corresponding to the first target object obtained after being processed by the first node information transfer model and the self-attention readout function; represents the second global feature information at the node level corresponding to the second target object, which is used to characterize the global features at the node level corresponding to the second target object obtained after processing by the second node information transfer model and the self-attention readout function; cat(·,·) represents the concatenation function, which concatenates the first global feature information at the node level corresponding to the first target object and the second global feature information at the node level corresponding to the second target object, thereby combining the information of the first target object and the second target object at the node level; FCN a Represents the parameters of the first prediction processing model.

[0195] In one possible implementation, the second prediction processing model is a fully connected neural network. The second prediction processing model is called to perform prediction processing on the first global feature information at the edge level and the second global feature information at the edge level. The process of obtaining the target interaction information at the edge level between the first target object and the second target object is implemented based on Formula 12:

[0196]

[0197] Among them, Pred b representing target interaction information at an edge level between a first target and a second target; Representing first global feature information at the edge level corresponding to the first target object, used to characterize the global features of the edge level corresponding to the first target object obtained after being processed by the first edge information transfer model and the self-attention readout function; represents the second global feature information at the edge level corresponding to the second target object, which is used to characterize the global features of the edge level corresponding to the second target object obtained after processing by the second edge information transfer model and the self-attention readout function; cat(·,·) represents the concatenation function, which concatenates the first global feature information at the edge level corresponding to the first target object and the second global feature information at the edge level corresponding to the second target object, thereby combining the information of the first target object and the second target object at the edge level; FCN b Represents the parameters of the second prediction processing model.

[0198] In an exemplary embodiment, the model structure of the second prediction processing model is the same as the model structure of the first prediction processing model, both of which are fully connected neural networks, but because the second prediction processing model and the first prediction processing model process different information, the parameters of the second prediction processing model may be different from the parameters of the first prediction processing model.

[0199] After obtaining the target interaction information at the node level and the target interaction information at the edge level, the target interaction information between the first target object and the second target object is obtained.

[0200] It should be noted that the above is only an exemplary description of obtaining target interaction information between a first target object and a second target object, and the embodiments of the present application are not limited thereto. In an exemplary embodiment, when the first global feature information and the second global feature information are both global feature information as a whole, the prediction processing model is also a whole processing model. In this case, the prediction processing model is directly called to perform prediction processing on the first global feature information and the second global feature information to obtain the whole target interaction information.

[0201] In one possible implementation, after obtaining the first attention information corresponding to the first target and the second attention information corresponding to the second target, it also includes: based on the first attention information corresponding to the first target, determining a first key sub-target among all sub-targets of the first target, the first key sub-target being used to indicate the sub-target in the first target used to interact with the second target; based on the second attention information corresponding to the second target, determining a second key sub-target among all sub-targets of the second target, the second key sub-target being used to indicate the sub-target in the second target used to interact with the first target.

[0202] In one possible implementation, based on the first attention information corresponding to the first target, a first key sub-target is determined from among all sub-targets of the first target by: determining the attention information corresponding to each sub-target within the first target based on the first attention information corresponding to the first target; and selecting, from among all sub-targets of the first target, sub-targets whose corresponding attention information meets a selection condition as first key sub-targets. The first key sub-target is used to indicate a sub-target within the first target that is used to interact with the second target.

[0203] In one possible implementation, taking the first attention information corresponding to the first target object as an overall attention information as an example, the attention information corresponding to any sub-target object in the first target object is also an overall attention information. The attention information corresponding to any sub-target object in the first target object includes the attention weight of the any sub-target object at each angle. Based on this, the attention information corresponding to any sub-target object in the first target object meets the selection condition, which means that the attention weight of the any sub-target object at each angle has an attention weight that is not less than the target proportion, which meets the threshold condition, and meeting the threshold condition means that it is not less than the weight threshold. The target proportion and weight threshold are set according to experience, or flexibly adjusted according to the application scenario, and the embodiments of the present application do not limit this.

[0204] Exemplarily, the target ratio is set to 80%, and the weight threshold is set to 0.3. Assuming that the attention information corresponding to any sub-target in the first target includes the attention weight of any sub-target at 10 angles, then when there are 8 or more attention weights of any sub-target in the 10-angle attention weights that are not less than 0.3, it means that the attention information corresponding to any sub-target meets the selection conditions.

[0205] In one possible implementation, for the case where the first attention information corresponding to the first target object includes the first attention information at the node level and the first attention information at the edge level, the attention information corresponding to any sub-target object in the first target object also includes the attention information at the node level and the attention information at the edge level. Both the attention information at the node level and the attention information at the edge level corresponding to any sub-target object in the first target object include the attention weight of the any sub-target object at each angle. Based on this, the attention information corresponding to any sub-target object in the first target object meeting the selection condition may mean that the attention information at the node level corresponding to the any sub-target object meets the selection condition, or that the attention information at the edge level corresponding to the any sub-target object meets the selection condition, or that both the attention information at the node level and the attention information at the edge level corresponding to the any sub-target object meet the selection condition. The embodiments of the present application do not limit this.

[0206] The implementation method of determining the second key sub-target among all sub-targets of the second target based on the second attention information corresponding to the second target object refers to the above-mentioned implementation method of determining the first key sub-target among all sub-targets of the first target object based on the first attention information corresponding to the first target object, which will not be repeated here.

[0207] It should be noted that in the actual prediction process, it is unknown which sub-objects in the target object interact with another target object. The key sub-objects determined among all the sub-objects included in the target object are used to indicate which sub-objects in the target object interact with another target object.

[0208] Since the loss function of the key local level determined based on the attention information corresponding to the key sub-sample objects in the sample objects that meet the reference conditions is used in the process of training the interaction information prediction model, the first attention information of the first target object and the second attention information of the second target object obtained by using the target interaction information prediction model are more reliable. The key sub-target objects determined based on the attention information corresponding to a certain target object can more accurately indicate which sub-target objects in the target object interact with another target object.

[0209] It should be noted that, in an exemplary embodiment, the attention information corresponding to a certain target object may indicate that there is no key sub-target object among all the sub-target objects of the target object. In this case, it means that the target object does not meet the reference condition. At this time, it is considered that the target object as a whole interacts with another target object.

[0210] In an exemplary embodiment, one of the first and second targets is a protein, and the other is a small molecule. Using a target interaction information prediction model to predict target interaction information between the first and second targets can be applied to small molecule screening scenarios. In this case, the target interaction information prediction model can be considered a drug screening model, where drugs are small molecules that interact with proteins, and the interaction information can be considered protein-small molecule activity information.

[0211] For example, the process of acquiring target interaction information between the first target and the second target is as follows: Figure 5 As shown. The image information of the first target object 501 ( Figure 4 Only part of the graph information is shown in the figure) is input into the first feature extraction model 502 in the target interaction information prediction model. After feature extraction and processing by the self-attention readout function, the first global feature information 503 corresponding to the first target object is obtained; the graph information ( Figure 4 Only part of the figure information is shown in the figure) is input into the second feature extraction model 505 in the target interaction information prediction model, and after feature extraction and processing by the self-attention readout function, the second global feature information 506 corresponding to the second target object is obtained; the prediction processing model in the target interaction information prediction model is called to integrate the first global feature information 503 corresponding to the first target object and the second global feature information 506 corresponding to the second target object, and based on the integrated feature information 507, the target interaction information between the first target object and the second target object is obtained.

[0212] In an embodiment of the present application, a loss function at the global level and a loss function at the key local level are used to train an interaction information prediction model, and then the interaction information between the first target object and the second target object is determined using the trained target interaction information prediction model. The loss function at the global level enables the model training process to focus on global information; the loss function at the key local level enables the model training process to focus on key local information. In other words, in an embodiment of the present application, the training process of the interaction information prediction model focuses on both global information and key local information, the training effect of the interaction information prediction model is better, and the accuracy of determining the interaction information between the first target object and the second target object using the trained interaction information prediction model is higher.

[0213] Based on the above Figure 1 In the implementation environment shown, the embodiment of the present application provides a method for training a target interaction information prediction model, taking the method applied to the server 12 as an example. Figure 6 As shown, the method provided in the embodiment of the present application may include the following steps:

[0214] In step 601, an initial interaction information prediction model is called to obtain attention information corresponding to a first sample object, attention information corresponding to a second sample object, and predicted interaction information between the first sample object and the second sample object.

[0215] The predicted interaction information is obtained based on the global feature information corresponding to the first sample and the global feature information corresponding to the second sample.

[0216] The initial interaction information prediction model refers to a model that requires training for predicting interaction information. This embodiment of the present application does not limit the method for obtaining the interaction information prediction model. For example, the interaction information prediction model is designed by a developer and uploaded to a server, whereupon the server obtains the interaction information prediction model. The parameters of the target interaction information prediction model are the parameters to be updated. Training of the interaction prediction model is achieved by updating the parameters of the target interaction information prediction model.

[0217] The first sample and the second sample are a group of sample objects used to train the initial interaction information prediction model. In an exemplary embodiment, multiple groups of sample objects may be used in the process of training the initial interaction information prediction model. The embodiment of the present application takes the training of the initial interaction information prediction model using a group of sample objects as an example. In one possible implementation, the type of the first sample object is the same as Figure 2 In the embodiment shown, the first target object is of the same type as the training object, and the second target object is of the same type as the training object. Figure 2 In the illustrated embodiment, the second targets are of the same type to ensure the accuracy of the interaction information predicted by the target interaction information prediction model.

[0218] In one possible implementation, the process of calling the initial interaction information prediction model to obtain the attention information corresponding to the first sample object, the attention information corresponding to the second sample object, and the predicted interaction information between the first sample object and the second sample object is as follows: calling the initial interaction information prediction model, based on the basic information of the first sample object, obtaining the basic feature information corresponding to the first sample object and the attention information corresponding to the first sample object; based on the basic information of the second sample object, obtaining the basic feature information corresponding to the second sample object and the attention information corresponding to the second sample object; based on the basic feature information corresponding to the first sample object and the attention information corresponding to the first sample object, obtaining the global feature information corresponding to the first sample object; based on the basic feature information corresponding to the second sample object and the attention information corresponding to the second sample object, obtaining the global feature information corresponding to the second sample object; based on the global feature information corresponding to the first sample object and the global feature information corresponding to the second sample object, obtaining the predicted interaction information between the first sample object and the second sample object. For the implementation of this process, see Figure 2 Steps 2021 to 2023 in the illustrated embodiment are not repeated here.

[0219] In step 602 , a global loss function is obtained based on the predicted interaction information and the standard interaction information between the first sample and the second sample.

[0220] A global loss function is a loss function obtained by focusing on global information. The standard interaction information between the first and second specimens is used to indicate the actual interaction information between the first and second specimens. For example, the standard interaction information between the first and second specimens is in the form of a label.

[0221] In one possible implementation, when the predicted interaction information includes node-level predicted interaction information and edge-level predicted interaction information, a process of determining a global-level loss function based on the predicted interaction information and standard interaction information between the first sample and the second sample includes the following steps 6021 to 6024:

[0222] Step 6021: Determine a first sub-loss function based on the predicted interaction information at the node level and the standard interaction information between the first sample and the second sample.

[0223] The first sub-loss function is used to measure the difference between the predicted interaction information and the standard interaction information at the node level. The embodiment of the present application does not limit the form of the first sub-loss function. Exemplarily, the form of the first sub-loss function is the mean square error.

[0224] In an exemplary embodiment, based on the predicted interaction information at the node level and the standard interaction information between the first sample and the second sample, the process of determining the first sub-loss function is implemented based on Formula 13:

[0225]

[0226] in, Represents the first sub-loss function; Pred a represents the predicted interaction information at the node level; Target represents the standard interaction information; MSE(·,·) represents the mean square error function.

[0227] Step 6022: Determine a second sub-loss function based on the predicted interaction information at the edge level and the standard interaction information between the first sample and the second sample.

[0228] The second sub-loss function is used to measure the difference between the predicted interaction information and the standard interaction information at the edge level. The embodiment of the present application does not limit the form of the second sub-loss function. Exemplarily, the form of the second sub-loss function is the mean square error.

[0229] In an exemplary embodiment, based on the predicted interaction information at the edge level and the standard interaction information between the first sample and the second sample, the process of determining the second sub-loss function is implemented based on Formula 14:

[0230]

[0231] in, Represents the second sub-loss function; Pred b represents the predicted interaction information at the edge level; Target represents the standard interaction information; MSE(·,·) represents the mean square error function.

[0232] Step 6023: Determine the third sub-loss function based on the predicted interaction information at the node level and the predicted interaction information at the edge level.

[0233] The third sub-loss function is used to measure the difference between the predicted interaction information at the node level and the predicted interaction information at the edge level. The embodiment of the present application does not limit the form of the third sub-loss function. Exemplarily, the form of the third sub-loss function is the mean square error.

[0234] In an exemplary embodiment, based on the predicted interaction information at the node level and the predicted interaction information at the edge level, the process of determining the third sub-loss function is implemented based on Formula 15:

[0235] L dis =MSE(Pred a, Pred b ) (Formula 15)

[0236] Among them, L dis Represents the third sub-loss function; Pred a Represents the predicted interaction information Pred at the node level b represents the predicted interaction information at the edge level; MSE(·,·) represents the mean squared error function. The purpose of designing this third sub-loss function is to ensure that the predicted interaction information at the node level is consistent with the predicted interaction information at the edge level.

[0237] Step 6024: Determine a global loss function based on the first sub-loss function, the second sub-loss function, and the third sub-loss function.

[0238] After obtaining the first sub-loss function, the second sub-loss function and the third sub-loss function, a global loss function is determined based on the first sub-loss function, the second sub-loss function and the third sub-loss function to obtain a loss function for focusing on global information.

[0239] In an exemplary embodiment, based on the first sub-loss function, the second sub-loss function and the third sub-loss function, the way to determine the loss function at the global level is: perform weighted summation on the first sub-loss function, the second sub-loss function and the third sub-loss function to obtain the loss function at the global level. In the process of weighted summation, the weights corresponding to each sub-loss function are set according to experience or flexibly adjusted according to the application scenario, and the embodiments of the present application do not limit this. For example, the weights corresponding to each sub-loss function can be set to 1, then the loss function at the global level is the sum of the first sub-loss function, the second sub-loss function and the third sub-loss function.

[0240] According to the above content, the loss function at the global level is obtained by integrating multiple sub-loss functions, which can play a multi-supervision effect on the training process of the interaction information prediction model.

[0241] It should be noted that the above steps 6021 to 6024 are merely an exemplary description of obtaining a global-level loss function, and the present application is not limited thereto. In an exemplary embodiment, when the predicted interaction information is a whole, the loss function determined based on the predicted interaction information and the standard interaction information is directly used as the global-level loss function.

[0242] In step 603, in the attention information corresponding to the sample object meeting the reference condition, attention information corresponding to the key sub-sample object in the sample object meeting the reference condition is determined, and the sample object meeting the reference condition is at least one of the first sample object and the second sample object.

[0243] The key sub-sample objects in the sample objects that meet the reference conditions are some sub-sample objects in all the sub-sample objects.

[0244] The sample that satisfies the reference condition is at least one of the first sample and the second sample. In other words, at least one of the first sample and the second sample satisfies the reference condition, that is, the first sample satisfies the reference condition, or the second sample satisfies the reference condition, or both the first sample and the second sample satisfy the reference condition. Key subsamples within the sample that satisfies the reference condition are some of the subsamples within all the subsamples. In an exemplary embodiment, a key subsample within a sample refers to a subsample that interacts with another sample within all the subsamples included in the sample.

[0245] For a certain sample, all of the sub-samples included in the sample may interact with another sample, or only some of the sub-samples may interact with another sample. In the embodiment of the present application, the sample in which only some of the sub-samples interact with another sample among all the sub-samples included is regarded as the sample that meets the reference condition. For example, assuming that one of the two samples is a protein and the other is a small molecule, the sub-samples included in the protein-type sample are amino acid 1, amino acid 2, and amino acid 3. If only one or two of amino acids 1, 2, and 3 interact with the small molecule-type sample, then the protein-type sample is regarded as the sample that meets the reference condition; if amino acid 1, 2, and 3 all interact with the small molecule-type sample, then the protein-type sample is regarded as the sample that does not meet the reference condition.

[0246] In an embodiment of the present application, at least one of the two sample objects includes a sample object that satisfies a reference condition. In exemplary embodiments, for both the first sample object and the second sample object, information indicating which key subsample objects are included in the sample object that satisfies the reference condition can also be obtained, thereby directly determining the key subsample objects within the sample object that satisfies the reference condition. The information indicating which key subsample objects are included in the sample object that satisfies the reference condition can be considered supervisory information corresponding to the sample object that satisfies the reference condition, and this supervisory information provides a reference for obtaining the loss function at the key local level.

[0247] After the first sample and the second sample, it is possible to determine which sample or samples satisfy the reference condition. After determining the sample satisfying the reference condition, the attention information corresponding to the key sub-sample in the sample satisfying the reference condition is determined from the attention information corresponding to the sample satisfying the reference condition.

[0248] The attention information corresponding to a certain sample object includes the attention information of each node in the graph information of the sample object, and the attention information of any node includes the attention weight corresponding to the node at one or more angles. Since each node in the graph information of a certain sample object corresponds to a sub-sample object in the sample object, the attention information of a certain node can be directly used as the attention information corresponding to the sub-sample object corresponding to any node. In other words, after the key sub-sample objects in the sample object that meets the reference condition are known, the attention information corresponding to the key sub-sample objects can be directly determined in the attention information corresponding to the sample object that meets the reference condition. It should be noted that the number of key sub-sample objects in a sample object that meets the reference condition may be one or more, and determining the attention information corresponding to the key sub-sample objects refers to determining the attention information corresponding to each key sub-sample object. The attention information corresponding to any key sub-sample object includes the attention weight corresponding to the key sub-sample object at one or more angles.

[0249] In step 604, based on the attention information corresponding to the key sub-sample objects in the sample objects that meet the reference conditions, a loss function at the key local level is obtained.

[0250] After determining the attention information corresponding to the key sub-samples within the sample that meets the reference condition, a loss function at the key local level is obtained based on the attention information corresponding to the key sub-samples within the sample that meets the reference condition. Since the key sub-references are some of the sub-samples within all the sub-samples included in the sample that meets the reference condition, the loss function at the key local level obtained based on the attention information corresponding to the key sub-samples can be used to focus the training process of the interaction information prediction model on key local information.

[0251] In a possible implementation, the sample objects meeting the reference condition include the following three situations. In different situations, the implementation of step 604 is different.

[0252] Case 1: The sample that meets the reference condition is the first sample.

[0253] In this case 1, the implementation method of step 604 is: in the attention information corresponding to the first sample object, determine the attention information corresponding to the key sub-sample object in the first sample object; based on the attention information corresponding to the key sub-sample object in the first sample object, determine the loss function of the key local level.

[0254] The attention information corresponding to the first sample object includes the attention information of each node in the graph information of the first sample object. Exemplarily, in the attention information corresponding to the first sample object, the method for determining the attention information corresponding to the key sub-sample object in the first sample object is: in the attention information corresponding to the first sample object, determine the attention information of the node corresponding to the key sub-sample object in the first sample object, and use the attention information of the node corresponding to the key sub-sample object in the first sample object as the attention information corresponding to the key sub-sample object in the first sample object.

[0255] The number of key sub-sample objects in the first sample object is one or more, and the attention information corresponding to any key sub-sample object includes the attention weight corresponding to the key sub-sample object at one or more angles. It should be noted that the angles corresponding to the attention weights included in the attention information corresponding to different key sub-sample objects in the first sample object are the same. In an exemplary embodiment, since the sum of the attention weights corresponding to the same angle for each node in the graph information of the first sample object is 1, the sum of the attention weights corresponding to the same angle for each sub-sample object in the first sample object is also 1.

[0256] In an exemplary embodiment, based on the attention information corresponding to the key sub-sample object in the first sample object, the process of determining the loss function at the key local level is implemented based on Formula 16:

[0257]

[0258] Among them, L pocket represents the loss function of the key local level determined in this case 1; r represents the number of angles of attention weight corresponding to any key sub-sample object in the first sample object included in the attention information corresponding to the key sub-sample object, and r is an integer not less than 1; It represents the sum of the attention weights corresponding to the i-th angle of each key sub-sample object in the first sample object.

[0259] Case 2: The sample satisfying the reference condition is the second sample.

[0260] In this case 2, step 604 is implemented as follows: in the attention information corresponding to the second sample object, determine the attention information corresponding to the key sub-sample object in the second sample object; based on the attention information corresponding to the key sub-sample object in the second sample object, determine the loss function at the key local level.

[0261] The attention information corresponding to the second sample object includes the attention information of each node in the graph information of the second sample object. Exemplarily, in the attention information corresponding to the second sample object, the method for determining the attention information corresponding to the key sub-sample object in the second sample object is: in the attention information corresponding to the second sample object, determine the attention information of the node corresponding to the key sub-sample object in the second sample object, and use the attention information of the node corresponding to the key sub-sample object in the second sample object as the attention information corresponding to the key sub-sample object in the second sample object.

[0262] The number of key sub-sample objects in the second sample object is one or more, and the attention information corresponding to any key sub-sample object includes the attention weight corresponding to the key sub-sample object at one or more angles. It should be noted that the angles corresponding to the attention weights included in the attention information corresponding to different key sub-sample objects in the second sample object are the same. In an exemplary embodiment, since the sum of the attention weights corresponding to the same angle for each node in the graph information of the second sample object is 1, the sum of the attention weights corresponding to the same angle for each sub-sample object in the second sample object is also 1.

[0263] In an exemplary embodiment, the process of determining the loss function at the key local level based on the attention information corresponding to the key sub-sample object in the second sample object is also implemented based on the above formula 16. In the process of determining the loss function at the key local level based on the attention information corresponding to the key sub-sample object in the second sample object based on formula 16, L pocket represents the loss function of the key local level determined in this case 2; r represents the number of angles of attention weight corresponding to any key sub-sample object in the second sample object included in the attention information corresponding to the key sub-sample object, and r is an integer not less than 1; It represents the sum of the attention weights corresponding to the i-th angle of each key sub-sample object in the second sample object.

[0264] Case 3: The samples that meet the reference condition are the first sample and the second sample.

[0265] In this case 3, the implementation method of step 604 is as follows: in the attention information corresponding to the first sample object, determine the attention information corresponding to the first key sub-sample object in the first sample object; in the attention information corresponding to the second sample object, determine the attention information corresponding to the second key sub-sample object in the second sample object; based on the attention information corresponding to the first key sub-sample object, determine the fourth sub-loss function; based on the attention information corresponding to the second key sub-sample object, determine the fifth sub-loss function; based on the fourth sub-loss function and the fifth sub-loss function, determine the loss function of the key local level.

[0266] Exemplarily, the process of determining the fourth sub-loss function based on the attention information corresponding to the first key sub-sample object is implemented based on Formula 16. In the process of determining the fourth sub-loss function based on the attention information corresponding to the first key sub-sample object based on Formula 16, L pocket represents the fourth sub-loss function; r represents the number of angles of attention weight corresponding to any first key sub-sample object included in the attention information corresponding to the first key sub-sample object, and r is an integer not less than 1; It represents the sum of the attention weights corresponding to each first key sub-sample object at the i-th angle.

[0267] Exemplarily, the process of determining the fifth sub-loss function based on the attention information corresponding to the second key sub-sample object is implemented based on Formula 16. In the process of determining the fifth sub-loss function based on the attention information corresponding to the second key sub-sample object based on Formula 16, L pocket represents the fifth sub-loss function; r represents the number of angles of attention weight corresponding to any second key sub-sample object included in the attention information corresponding to the second key sub-sample object, and r is an integer not less than 1; It represents the sum of the attention weights corresponding to each second key sub-sample object at the i-th angle.

[0268] After determining the fourth sub-loss function and the fifth sub-loss function, the loss function of the key local level is determined based on the fourth sub-loss function and the fifth sub-loss function. In an exemplary embodiment, the loss function of the key local level is determined based on the fourth sub-loss function and the fifth sub-loss function as follows: the fourth sub-loss function and the fifth sub-loss function are weightedly summed to obtain the loss function of the key local level. In the process of weighted summation, the weights corresponding to the fourth sub-loss function and the fifth sub-loss function are set according to experience or flexibly adjusted according to the application scenario, and this embodiment of the present application does not limit this. Exemplarily, the weights corresponding to each sub-loss function can be set to 1, then the loss function of the key local level is the sum of the fourth sub-loss function and the fifth sub-loss function.

[0269] It should be noted that the above content takes the attention information corresponding to the sample object as a whole as an example to introduce the implementation method of step 604. The embodiments of the present application are not limited to this. In an exemplary embodiment, for the case where the attention information corresponding to the sample object includes attention information at the node level and attention information at the edge level, the attention information corresponding to the key sub-sample object includes attention information at the node level and attention information at the edge level. In this case, the process of obtaining the loss function at the key local level based on the attention information corresponding to the key sub-sample object may refer to: obtaining the loss function at the key local level based on the attention information at the node level corresponding to the key sub-sample object; it may also refer to: obtaining the loss function at the key local level based on the attention information at the edge level corresponding to the key sub-sample object; it may also refer to: obtaining the loss function at the key local level based on the attention information at the edge level corresponding to the key sub-sample object; it may also refer to: obtaining the loss function at the key local level based on the attention information at the node level and the attention information at the edge level corresponding to the key sub-sample object. The embodiments of the present application are not limited to this.

[0270] In an exemplary embodiment, the process of obtaining the loss function at the key local level based on the attention information at the node level corresponding to the key sub-sample object and the process of obtaining the loss function at the key local level based on the attention information at the edge level corresponding to the key sub-sample object can both refer to the process of obtaining the loss function at the key local level when the attention information is a whole attention information.

[0271] In an exemplary embodiment, the method for obtaining the loss function at the key local level based on the node-level attention information and the edge-level attention information corresponding to the key sub-sample object is: based on the node-level attention information corresponding to the key sub-sample object, determine the loss function at the node level; based on the edge-level attention information corresponding to the key sub-sample object, determine the loss function at the edge level; based on the node-level loss function and the edge-level loss function, determine the loss function at the key local level.

[0272] In either case, a loss function at the key local level can be obtained. The loss function at the key local level is designed to make the sum of the attention weights corresponding to each angle of the key sub-sample objects in the sample objects that meet the reference conditions as close to 1 as possible, so that the interaction information prediction model can learn more about the key sub-sample objects in the sample objects that meet the reference conditions during the training process. Since key sub-sample objects are sub-sample objects that interact with each other, the loss function at the key local level can enable the interaction information prediction model to learn more about the interacting sub-sample objects during the training process.

[0273] Furthermore, since key sub-samples are part of all sub-samples included in the sample that meets the reference condition, the loss function at the key local level focuses on local information. The design of the loss function at the key local level can utilize the attention weights of key sub-samples read out by the self-attention readout function to add supervisory signals for focusing on local information to the training process of the interaction information prediction model.

[0274] In an exemplary embodiment, when the first sample is a protein and the second sample is a small molecule, the sample that meets the reference condition is the first sample, as the entire small molecule typically interacts with a subset of amino acids in the protein. The key subsample within the first sample refers to the amino acids in the protein that bind to the small molecule. These amino acids are referred to as pocket amino acids, and information about these pocket amino acids can reflect information about the protein pocket.

[0275] In step 605 , the parameters of the initial interaction information prediction model are updated in reverse based on the loss function at the global level and the loss function at the key local level.

[0276] After obtaining the global loss function in step 602 and the key local loss function in step 604, the parameters of the initial interaction information prediction model are reversely updated based on the global loss function and the key local loss function. Each time the parameters of the interaction information prediction model are updated, the interaction information prediction model is trained once.

[0277] In one possible implementation, the process of reversely updating the parameters of the initial interaction information prediction model based on the loss function at the global level and the loss function at the key local level is as follows: based on the loss function at the global level and the loss function at the key local level, a comprehensive loss function is determined, and the comprehensive loss function is used to reversely update the parameters of the initial interaction information prediction model. In one possible implementation, the method of determining the comprehensive loss function based on the loss function at the global level and the loss function at the key local level is as follows: performing weighted summation of the loss function at the global level and the loss function at the key local level to obtain a comprehensive loss function. In the process of weighted summation, the weights corresponding to the loss function at the global level and the loss function at the key local level are set according to experience or flexibly adjusted according to the application scenario, and this is not limited in the embodiments of the present application. For example, the weights corresponding to the loss function at the global level and the loss function at the key local level can both be set to 1, and the comprehensive loss function is the sum of the loss function at the global level and the loss function at the key local level.

[0278] The embodiment of the present application does not limit the method of using the comprehensive loss function to reversely update the parameters of the initial interaction information prediction model. Exemplarily, the method of using the comprehensive loss function to reversely update the parameters of the initial interaction information prediction model is a gradient descent method.

[0279] In one possible implementation, based on the global-level loss function and the key local-level loss function, the process of reversely updating the parameters of the initial interaction information prediction model can be performed immediately after a global-level loss function and a key local-level loss function are obtained using a group of samples, or can be performed after a small batch of samples are used to obtain a global-level loss function and a key local-level loss function for the small batch. This embodiment of the application does not limit this. The number of small batches is set based on experience or flexibly adjusted according to the application scenario. This embodiment of the application does not limit this.

[0280] In one possible implementation, after initially updating the parameters of the interaction information prediction model based on the global-level loss function and the key local-level loss function, a determination is made as to whether the parameter update process satisfies a termination condition. If the parameter update process satisfies the termination condition, step 606 is executed. If the parameter update process does not satisfy the termination condition, steps 601 to 605 are continued until the parameter update process satisfies the termination condition, at which point step 606 is executed.

[0281] In an exemplary embodiment, the parameter update process satisfies the termination conditions including but not limited to the following three situations:

[0282] Case A: The number of parameter updates reaches the threshold.

[0283] The number threshold can be set based on experience or flexibly adjusted according to the application scenario, and the embodiments of the present application do not limit this.

[0284] Case B: The comprehensive loss function is less than the loss threshold.

[0285] The loss threshold can be set based on experience or flexibly adjusted according to the application scenario, and this is not limited in the embodiments of the present application.

[0286] Case C, the comprehensive loss function converges.

[0287] The convergence of the comprehensive loss function means that as the number of parameter updates increases, the fluctuation range of the comprehensive loss function in the update results of the reference number of times is within the reference range. For example, assuming the reference range is -10 -3 ~10 -3 , assuming the reference number is 10. If the fluctuation range of the comprehensive loss function in the 10 parameter update results is -10 -3 ~10 -3If , the comprehensive loss function is considered to converge.

[0288] When any of the above three situations is met, it means that the parameter update process meets the termination condition, and step 606 is executed.

[0289] In step 606, in response to the parameter updating process satisfying a termination condition, a target interaction information prediction model is obtained.

[0290] When the parameter updating process satisfies the termination condition, the interaction information prediction model obtained when the reference updating process satisfies the termination condition is used as the trained target interaction information prediction model, thereby obtaining the target interaction information prediction model.

[0291] In one possible implementation, after obtaining a trained target interaction information prediction model, the target interaction information prediction model is used to predict the interaction information between two targets. Because the target interaction information prediction model is trained using a global loss function and a key local loss function, the target interaction information prediction model can predict the interaction information between two targets with high accuracy.

[0292] In the embodiment of the present application, in addition to designing a loss function at the global level, a loss function at the key local level is also designed. In the field of drug screening, the loss function at the key local level can provide supervision for the learning of information about pocket amino acids in proteins. The embodiment of the present application can focus on the information fusion at the pocket amino acids while paying attention to the global information of the protein, and can strengthen the fusion learning of the global and local information of the protein, thereby improving the accuracy of the protein-small molecule activity information predicted by the trained drug screening model.

[0293] In the embodiment of the present application, during the training of the initial interaction information prediction model, the parameters of the initial interaction information prediction model are reversely updated based on a global-level loss function and a key local-level loss function. The global-level loss function enables the update of model parameters to focus on global information, while the key local-level loss function enables the update of model parameters to focus on local information. In other words, in the embodiment of the present application, the model training process focuses on both global and local information, which is beneficial for improving the training effect of the interaction information prediction model and, in turn, improving the accuracy of the interaction information predicted using the trained target interaction information prediction model.

[0294] See also Figure 7 , an embodiment of the present application provides a device for determining interaction information, the device comprising:

[0295] An acquisition unit 701 is configured to acquire basic information of the first target object, basic information of the second target object, and a target interaction information prediction model. The target interaction information prediction model is trained using a global-level loss function and a key local-level loss function. The key local-level loss function is determined based on attention information corresponding to key subsample objects in the sample objects that meet a reference condition. The key subsample objects in the sample objects that meet the reference condition are some subsample objects in all subsample objects.

[0296] The processing unit 702 is configured to call a target interaction information prediction model to process basic information of the first target and basic information of the second target to obtain target interaction information between the first target and the second target.

[0297] In one possible implementation, see Figure 8 , the processing unit 702 includes:

[0298] The first acquisition subunit 7021 is configured to call the target interaction information prediction model to acquire, based on the basic information of the first target, first basic feature information corresponding to the first target and first attention information corresponding to the first target; and acquire, based on the basic information of the second target, second basic feature information corresponding to the second target and second attention information corresponding to the second target;

[0299] The second acquisition subunit 7022 is configured to acquire first global feature information corresponding to the first target object based on the first basic feature information and the first attention information; and acquire second global feature information corresponding to the second target object based on the second basic feature information and the second attention information.

[0300] The third acquisition subunit 7023 is configured to acquire target interaction information between the first target and the second target based on the first global feature information and the second global feature information.

[0301] In one possible implementation, the first global feature information includes the first global feature information at the node level and the first global feature information at the edge level, the second global feature information includes the second global feature information at the node level and the second global feature information at the edge level, the target interaction information includes the target interaction information at the node level and the target interaction information at the edge level, and the target interaction information prediction model includes the first prediction processing model and the second prediction processing model; the third acquisition subunit 7023 is used to call the first prediction processing model to perform prediction processing on the first global feature information at the node level and the second global feature information at the node level to obtain the node-level target interaction information between the first target object and the second target object; and call the second prediction processing model to perform prediction processing on the first global feature information at the edge level and the second global feature information at the edge level to obtain the edge-level target interaction information between the first target object and the second target object.

[0302] In one possible implementation, the basic information of the first target object is the graph information of the first target object, and the basic information of the second target object is the graph information of the second target object; the target interaction information prediction model includes a first node information transmission model, a first edge information transmission model, a second node information transmission model and a second edge information transmission model; the first basic feature information includes the first basic feature information at the node level and the first basic feature information at the edge level, and the second basic feature information includes the second basic feature information at the node level and the second basic feature information at the edge level; the first acquisition subunit 7021 is also used to call the first node information transmission model to perform node-level feature extraction on the graph information of the first target object to obtain the first basic feature information at the node level; call the first edge information transmission model to perform edge-level feature extraction on the graph information of the first target object to obtain the first basic feature information at the edge level; call the second node information transmission model to perform node-level feature extraction on the graph information of the second target object to obtain the second basic feature information at the node level; call the second edge information transmission model to perform edge-level feature extraction on the graph information of the second target object to obtain the second basic feature information at the edge level.

[0303] In one possible implementation, the first node information transmission model includes a first node feature update layer and a first node-level feature output layer; the first edge information transmission model includes a first edge feature update layer and a first edge-level feature output layer; the first acquisition subunit 7021 is also used to call the first node feature update layer to update the features of the nodes in the graph information of the first target object to obtain the target node features; call the first node-level feature output layer to output the target node features to obtain the first basic feature information at the node level; call the first edge feature update layer to update the features of the edges in the graph information of the first target object to obtain the target edge features; call the first edge-level feature output layer to output the target edge features to obtain the first basic feature information at the edge level.

[0304] In one possible implementation, see Figure 9 , the device further comprises:

[0305] Determination unit 703 is used to determine a first key sub-target among all sub-targets of the first target based on first attention information corresponding to the first target, where the first key sub-target is used to indicate a sub-target in the first target that is used to interact with the second target; and to determine a second key sub-target among all sub-targets of the second target based on second attention information corresponding to the second target, where the second key sub-target is used to indicate a sub-target in the second target that is used to interact with the first target.

[0306] In one possible implementation, the processing unit 702 is further configured to call an initial interaction information prediction model to obtain attention information corresponding to the first sample object, attention information corresponding to the second sample object, and predicted interaction information between the first sample object and the second sample object, where the predicted interaction information is obtained based on global feature information corresponding to the first sample object and global feature information corresponding to the second sample object.

[0307] The acquiring unit 701 is further configured to acquire a global loss function based on the predicted interaction information and the standard interaction information between the first sample object and the second sample object;

[0308] The determining unit 703 is further configured to determine, from the attention information corresponding to the sample objects satisfying the reference condition, attention information corresponding to the key sub-sample object in the sample object satisfying the reference condition, where the sample object satisfying the reference condition is at least one of the first sample object and the second sample object;

[0309] The acquisition unit 701 is further configured to acquire a loss function at a key local level based on attention information corresponding to key subsample objects in the sample objects that meet the reference condition;

[0310] See also Figure 9, the device further comprises:

[0311] An updating unit 704 is configured to reversely update the parameters of the initial interaction information prediction model based on the global-level loss function and the key local-level loss function;

[0312] The acquisition unit 701 is further configured to obtain a target interaction information prediction model in response to the parameter updating process satisfying a termination condition.

[0313] In one possible implementation, the predicted interaction information includes predicted interaction information at the node level and predicted interaction information at the edge level; the acquisition unit 701 is further used to determine a first sub-loss function based on the predicted interaction information at the node level and the standard interaction information between the first sample object and the second sample object; determine a second sub-loss function based on the predicted interaction information at the edge level and the standard interaction information between the first sample object and the second sample object; determine a third sub-loss function based on the predicted interaction information at the node level and the predicted interaction information at the edge level; and determine a loss function at the global level based on the first sub-loss function, the second sub-loss function, and the third sub-loss function.

[0314] In one possible implementation, the sample objects satisfying the reference condition are the first sample object and the second sample object; the determining unit 703 is further configured to determine, from the attention information corresponding to the first sample object, the attention information corresponding to the first key sub-sample object in the first sample object; and to determine, from the attention information corresponding to the second sample object, the attention information corresponding to the second key sub-sample object in the second sample object;

[0315] The acquisition unit 701 is also used to determine the fourth sub-loss function based on the attention information corresponding to the first key sub-sample object; determine the fifth sub-loss function based on the attention information corresponding to the second key sub-sample object; and determine the loss function at the key local level based on the fourth sub-loss function and the fifth sub-loss function.

[0316] In one possible implementation, the type of the first target is protein, and the type of the second target is a small molecule; the acquisition unit 701 is further used to determine the spatial distance between amino acids in the first target based on the structural information of the first target; determine the adjacency matrix corresponding to the first target based on the spatial distance between the amino acids, and the adjacency matrix is ​​used to indicate the association relationship between the amino acids in the first target; obtain the graph information of the first target based on the adjacency matrix corresponding to the first target and the amino acids in the first target, and use the graph information of the first target as the basic information of the first target; obtain the graph information of the second target based on the atoms in the second target and the chemical bond information between the atoms, and use the graph information of the second target as the basic information of the second target.

[0317] In an embodiment of the present application, a loss function at the global level and a loss function at the key local level are used to train an interaction information prediction model, and then the interaction information between the first target object and the second target object is determined using the trained target interaction information prediction model. The loss function at the global level enables the model training process to focus on global information; the loss function at the key local level enables the model training process to focus on key local information. In other words, in an embodiment of the present application, the training process of the interaction information prediction model focuses on both global information and key local information, the training effect of the interaction information prediction model is better, and the accuracy of determining the interaction information between the first target object and the second target object using the trained interaction information prediction model is higher.

[0318] It should be noted that the apparatus provided in the above embodiments is merely illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0319] Figure 10 This is a schematic diagram of the structure of a device for determining interaction information provided in an embodiment of the present application. The device may refer to a server, which may have relatively large differences due to different configurations or performances. It may include one or more processors (Central Processing Units, CPU) 1001 and one or more memories 1002, wherein the one or more memories 1002 store at least one program code, and the at least one program code is loaded and executed by the one or more processors 1001 to implement the interaction information determination method provided in each of the above method embodiments. Of course, the server may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server may also include other components for implementing device functions, which will not be described here.

[0320] In an exemplary embodiment, a computer device is further provided, comprising a processor and a memory, wherein the memory stores at least one program code, which is loaded and executed by one or more processors to implement any of the above methods for determining interaction information.

[0321] In an exemplary embodiment, a computer-readable storage medium is further provided, in which at least one program code is stored. The at least one program code is loaded and executed by a processor of a computer device to implement any of the above-mentioned methods for determining interaction information.

[0322] In one possible implementation, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, and the like.

[0323] In an exemplary embodiment, a computer program product or computer program is also provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform any of the above-described methods for determining interaction information.

[0324] It should be noted that the terms "first," "second," and the like in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments of the application described herein can be implemented in an order other than those illustrated or described herein. The implementations described in the above exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of devices and methods consistent with certain aspects of the present application as detailed in the appended claims.

[0325] It should be understood that the term "plurality" used herein refers to two or more. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can mean: A exists alone, A and B exist simultaneously, or B exists alone. The character " / " generally indicates an "or" relationship between the associated objects.

[0326] The above are merely exemplary embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for determining interaction information, characterized in that: The method comprises: Obtaining basic information of the first target object, basic information of the second target object, and a target interaction information prediction model, wherein the target interaction information prediction model is trained using a global-level loss function and a key local-level loss function, wherein the key local-level loss function is determined based on attention information corresponding to key sub-sample objects in the sample objects that meet a reference condition, where the key sub-sample objects in the sample objects that meet the reference condition are some sub-sample objects in all sub-sample objects; Calling the target interaction information prediction model to obtain, based on the basic information of the first target, first basic feature information corresponding to the first target and first attention information corresponding to the first target; and obtaining, based on the basic information of the second target, second basic feature information corresponding to the second target and second attention information corresponding to the second target; Based on the first basic feature information and the first attention information, obtaining first global feature information corresponding to the first target object; based on the second basic feature information and the second attention information, obtaining second global feature information corresponding to the second target object; Target interaction information between the first target object and the second target object is acquired based on the first global feature information and the second global feature information.

2. The method according to claim 1, characterized in that The first global feature information includes first global feature information at the node level and first global feature information at the edge level, the second global feature information includes second global feature information at the node level and second global feature information at the edge level, the target interaction information includes target interaction information at the node level and target interaction information at the edge level, and the target interaction information prediction model includes a first prediction processing model and a second prediction processing model; The acquiring target interaction information between the first target and the second target based on the first global feature information and the second global feature information includes: Calling the first prediction processing model to perform prediction processing on the first global feature information at the node level and the second global feature information at the node level to obtain node-level target interaction information between the first target object and the second target object; The second prediction processing model is called to perform prediction processing on the first global feature information at the edge level and the second global feature information at the edge level to obtain target interaction information at the edge level between the first target object and the second target object.

3. The method according to claim 1, characterized in that The basic information of the first target object is the graph information of the first target object, and the basic information of the second target object is the graph information of the second target object; the target interaction information prediction model includes a first node information transmission model, a first edge information transmission model, a second node information transmission model, and a second edge information transmission model; the first basic feature information includes first basic feature information at the node level and first basic feature information at the edge level, and the second basic feature information includes second basic feature information at the node level and second basic feature information at the edge level; The calling the target interaction information prediction model to obtain first basic feature information corresponding to the first target based on the basic information of the first target includes: Calling the first node information transfer model to perform node-level feature extraction on the graph information of the first target object to obtain first basic feature information at the node level; Calling the first edge information transfer model to perform edge-level feature extraction on the graph information of the first target object to obtain first basic feature information at the edge level; The acquiring, based on the basic information of the second target object, second basic feature information corresponding to the second target object includes: Calling the second node information transfer model to perform node-level feature extraction on the graph information of the second target object to obtain second basic feature information at the node level; The second edge information transfer model is called to perform edge-level feature extraction on the graph information of the second target object to obtain second basic feature information at the edge level.

4. The method according to claim 3, characterized in that The first node information transfer model includes a first node feature update layer and a first node-level feature output layer; the first edge information transfer model includes a first edge feature update layer and a first edge-level feature output layer; The calling of the first node information transfer model to perform node-level feature extraction on the graph information of the first target object to obtain first basic feature information at the node level includes: Calling the first node feature update layer to update the features of the nodes in the graph information of the first target object to obtain target node features; Calling the first node-level feature output layer to perform output processing on the target node feature to obtain first basic feature information at the node level; The calling of the first edge information transfer model to perform edge-level feature extraction on the graph information of the first target object to obtain first basic feature information at the edge level includes: Calling the first edge feature update layer to update the edge features in the graph information of the first target object to obtain target edge features; The first edge-level feature output layer is called to perform output processing on the target edge feature to obtain first basic feature information of the edge level.

5. The method according to claim 1, wherein acquiring, based on the basic information of the first target object, first basic feature information corresponding to the first target object and first attention information corresponding to the first target object; After obtaining second basic feature information corresponding to the second target and second attention information corresponding to the second target based on the basic information of the second target, the method further includes: Determining, based on first attention information corresponding to the first target, a first key sub-target among all sub-targets of the first target, where the first key sub-target is used to indicate a sub-target of the first target that is used to interact with the second target; Based on the second attention information corresponding to the second target, a second key sub-target is determined among all sub-targets of the second target, where the second key sub-target is used to indicate a sub-target in the second target that is used to interact with the first target.

6. The method according to any one of claims 1 to 5, characterized in that: Before obtaining the basic information of the first target, the basic information of the second target, and the target interaction information prediction model, the method further includes: Invoking an initial interaction information prediction model to obtain attention information corresponding to a first sample object, attention information corresponding to a second sample object, and predicted interaction information between the first sample object and the second sample object, wherein the predicted interaction information is obtained based on global feature information corresponding to the first sample object and global feature information corresponding to the second sample object; Obtaining a global loss function based on the predicted interaction information and standard interaction information between the first sample substance and the second sample substance; Determining, from the attention information corresponding to the sample object meeting the reference condition, attention information corresponding to the key sub-sample object in the sample object meeting the reference condition, wherein the sample object meeting the reference condition is at least one of the first sample object and the second sample object; Obtaining a loss function at a key local level based on attention information corresponding to key subsample objects in the sample objects that meet the reference conditions; Based on the global-level loss function and the key local-level loss function, reversely updating the parameters of the initial interaction information prediction model; In response to the parameter updating process satisfying a termination condition, a target interaction information prediction model is obtained.

7. The method according to claim 6, characterized in that The predicted interaction information includes predicted interaction information at a node level and predicted interaction information at an edge level; and obtaining a global loss function based on the predicted interaction information and the standard interaction information between the first sample object and the second sample object includes: determining a first sub-loss function based on the predicted interaction information at the node level and the standard interaction information between the first sample object and the second sample object; determining a second sub-loss function based on the predicted interaction information at the edge level and the standard interaction information between the first sample object and the second sample object; determining a third sub-loss function based on the predicted interaction information at the node level and the predicted interaction information at the edge level; Determine the global loss function based on the first sub-loss function, the second sub-loss function, and the third sub-loss function.

8. The method according to claim 6, characterized in that The samples meeting the reference condition are the first sample and the second sample; The determining, from the attention information corresponding to the sample objects meeting the reference condition, the attention information corresponding to the key sub-sample objects in the sample objects meeting the reference condition includes: Determining, from the attention information corresponding to the first sample object, attention information corresponding to a first key sub-sample object in the first sample object; Determining, from the attention information corresponding to the second sample object, attention information corresponding to a second key sub-sample object in the second sample object; The acquiring of a loss function at a key local level based on attention information corresponding to key sub-sample objects in the sample objects that meet the reference condition includes: Determining a fourth sub-loss function based on attention information corresponding to the first key sub-sample object; Determining a fifth sub-loss function based on the attention information corresponding to the second key sub-sample object; Based on the fourth sub-loss function and the fifth sub-loss function, a loss function of the key local level is determined.

9. The method according to any one of claims 1 to 5, characterized in that: The type of the first target is protein, and the type of the second target is small molecule; The obtaining of basic information of the first target object and basic information of the second target object includes: determining the spatial distance between amino acids in the first target based on the structural information of the first target; Determining an adjacency matrix corresponding to the first target based on the spatial distances between the amino acids, wherein the adjacency matrix is ​​used to indicate the association relationship between the amino acids in the first target; Obtaining graph information of the first target according to the adjacency matrix corresponding to the first target and the amino acids in the first target, and using the graph information of the first target as basic information of the first target; Based on the atoms in the second object and the chemical bond information between the atoms, the graph information of the second object is obtained, and the graph information of the second object is used as the basic information of the second object.

10. A device for determining interaction information, characterized in that: The device comprises: an acquisition unit, configured to acquire basic information of the first target object, basic information of the second target object, and a target interaction information prediction model, wherein the target interaction information prediction model is trained using a global-level loss function and a key local-level loss function, wherein the key local-level loss function is determined based on attention information corresponding to key sub-sample objects in the sample objects that meet a reference condition, wherein the key sub-sample objects in the sample objects that meet the reference condition are some sub-sample objects in all sub-sample objects; a processing unit, comprising a first acquiring subunit, a second acquiring subunit, and a third acquiring subunit; The first acquisition subunit is configured to call the target interaction information prediction model to acquire, based on the basic information of the first target, first basic feature information corresponding to the first target and first attention information corresponding to the first target; and acquire, based on the basic information of the second target, second basic feature information corresponding to the second target and second attention information corresponding to the second target; The second acquisition subunit is configured to acquire first global feature information corresponding to the first target object based on the first basic feature information and the first attention information; and acquire second global feature information corresponding to the second target object based on the second basic feature information and the second attention information; The third acquisition subunit is configured to acquire target interaction information between the first target object and the second target object based on the first global feature information and the second global feature information.

11. The device according to claim 10, characterized in that The first global feature information includes first global feature information at the node level and first global feature information at the edge level, the second global feature information includes second global feature information at the node level and second global feature information at the edge level, the target interaction information includes target interaction information at the node level and target interaction information at the edge level, and the target interaction information prediction model includes a first prediction processing model and a second prediction processing model; The third acquisition sub-unit is used to call the first prediction processing model to perform prediction processing on the first global feature information at the node level and the second global feature information at the node level, and obtain the node-level target interaction information between the first target object and the second target object; and call the second prediction processing model to perform prediction processing on the first global feature information at the edge level and the second global feature information at the edge level, and obtain the edge-level target interaction information between the first target object and the second target object.

12. The device according to claim 10, characterized in that The basic information of the first target object is the graph information of the first target object, and the basic information of the second target object is the graph information of the second target object; the target interaction information prediction model includes a first node information transmission model, a first edge information transmission model, a second node information transmission model, and a second edge information transmission model; the first basic feature information includes first basic feature information at the node level and first basic feature information at the edge level, and the second basic feature information includes second basic feature information at the node level and second basic feature information at the edge level; The first acquisition subunit is configured to call the first node information transfer model to perform node-level feature extraction on the graph information of the first target object to obtain first basic feature information at the node level; Calling the first edge information transfer model to perform edge-level feature extraction on the graph information of the first target object to obtain first basic feature information at the edge level; Calling the second node information transfer model to perform node-level feature extraction on the graph information of the second target object to obtain second basic feature information at the node level; The second edge information transfer model is called to perform edge-level feature extraction on the graph information of the second target object to obtain second basic feature information at the edge level.

13. The device according to claim 12, characterized in that The first node information transfer model includes a first node feature update layer and a first node-level feature output layer; the first edge information transfer model includes a first edge feature update layer and a first edge-level feature output layer; The first acquisition subunit is configured to call the first node feature update layer to update the features of the nodes in the graph information of the first target object to obtain target node features; The first node-level feature output layer is called to output the target node feature to obtain the first basic feature information at the node level; the first edge feature update layer is called to update the edge feature in the graph information of the first target object to obtain the target edge feature; the first edge-level feature output layer is called to output the target edge feature to obtain the first basic feature information at the edge level.

14. The device according to claim 10, characterized in that The device further comprises: A determination unit is used to determine a first key sub-target among all sub-targets of the first target based on first attention information corresponding to the first target, wherein the first key sub-target is used to indicate a sub-target in the first target used to interact with the second target; and to determine a second key sub-target among all sub-targets of the second target based on second attention information corresponding to the second target, wherein the second key sub-target is used to indicate a sub-target in the second target used to interact with the first target.

15. The device according to any one of claims 10 to 14, characterized in that: The processing unit is further configured to call an initial interaction information prediction model to obtain attention information corresponding to a first sample object, attention information corresponding to a second sample object, and predicted interaction information between the first sample object and the second sample object, wherein the predicted interaction information is obtained based on global feature information corresponding to the first sample object and global feature information corresponding to the second sample object; The acquiring unit is further configured to acquire a global loss function based on the predicted interaction information and the standard interaction information between the first sample object and the second sample object; The device further comprises: a determining unit, configured to determine, from the attention information corresponding to the sample objects satisfying the reference condition, attention information corresponding to a key sub-sample object in the sample object satisfying the reference condition, wherein the sample object satisfying the reference condition is at least one of the first sample object and the second sample object; The acquisition unit is further configured to acquire a loss function at a key local level based on attention information corresponding to the key subsample objects in the sample objects that meet the reference condition; an updating unit, configured to reversely update the parameters of the initial interaction information prediction model based on the loss function at the global level and the loss function at the key local level; The acquisition unit is further configured to obtain a target interaction information prediction model in response to the parameter updating process satisfying a termination condition.

16. The device according to claim 15, characterized in that The predicted interaction information includes predicted interaction information at the node level and predicted interaction information at the edge level; the acquisition unit is used to determine a first sub-loss function based on the predicted interaction information at the node level and the standard interaction information between the first sample object and the second sample object; determine a second sub-loss function based on the predicted interaction information at the edge level and the standard interaction information between the first sample object and the second sample object; determine a third sub-loss function based on the predicted interaction information at the node level and the predicted interaction information at the edge level; and determine the loss function at the global level based on the first sub-loss function, the second sub-loss function and the third sub-loss function.

17. The device according to claim 15, characterized in that The samples meeting the reference condition are the first sample and the second sample; The determining unit is configured to determine, from the attention information corresponding to the first sample object, attention information corresponding to a first key sub-sample object in the first sample object; Determining, from the attention information corresponding to the second sample object, attention information corresponding to a second key sub-sample object in the second sample object; The acquisition unit is used to determine a fourth sub-loss function based on the attention information corresponding to the first key sub-sample object; determine a fifth sub-loss function based on the attention information corresponding to the second key sub-sample object; and determine the loss function of the key local level based on the fourth sub-loss function and the fifth sub-loss function.

18. The device according to any one of claims 10 to 14, characterized in that: The first target is a protein, and the second target is a small molecule; the acquisition unit is used to determine the spatial distance between amino acids in the first target based on the structural information of the first target; Determining an adjacency matrix corresponding to the first target based on the spatial distances between the amino acids, the adjacency matrix being used to indicate associations between the amino acids in the first target; obtaining graph information of the first target based on the adjacency matrix corresponding to the first target and the amino acids in the first target, and using the graph information of the first target as basic information of the first target; Based on the atoms in the second object and the chemical bond information between the atoms, the graph information of the second object is obtained, and the graph information of the second object is used as the basic information of the second object.

19. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one program code, and the at least one program code is loaded and executed by the processor to implement the method for determining interaction information according to any one of claims 1 to 9.

20. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one program code, and the at least one program code is loaded and executed by a processor to implement the method for determining interaction information according to any one of claims 1 to 9.

21. A computer program product, characterized in that The computer program product includes computer instructions, which are stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the method for determining interaction information as described in any one of claims 1 to 9.

Images