A Link Prediction Method and Apparatus Based on Symbolic Graph Neural Network

Abstract

A link prediction method and apparatus based on symbolic graph neural networks includes acquiring a symbolic graph comprising multiple drug nodes and multiple links. The multidimensional attribute features of the multiple drug nodes in the symbolic graph constitute an attribute feature matrix. This attribute feature matrix is ​​input into a preset link prediction model. After processing by a symbolic graph neural network based on a low-pass filter, a fully connected layer, and a regression classifier within the link prediction model, the predicted link between any two drug nodes in the symbolic graph is iteratively output. The error between the predicted link and the actual link in the symbolic graph is analyzed, and the model parameters of the link prediction model are adjusted based on the error until the training termination condition is met. The trained link prediction model is then used to predict the link between the drug node pairs to be predicted. This invention designs a symbolic graph neural network based on a low-pass filter. The end-to-end prediction model can simultaneously train a feature extractor and a link type discriminator, resulting in high accuracy in link prediction.

Description

Technical Field

[0001] This invention belongs to the field of graph machine learning and graph signal processing technology, specifically relating to a link prediction method and apparatus based on symbolic graph neural networks. Background Technology

[0002] In recent years, graph signal processing and deep learning have opened up new paths for graph data analysis. Graph signal processing is an extension of traditional signal processing for processing signals in irregular domains, defining many basic concepts similar to those in traditional signal processing, such as graph Fourier transform and graph filters. Among these, graph filters are one of the most widely used tools, and graph convolutional networks inspired by graph filters are currently the most popular graph depth methods. With the rapid development of graph signal processing and deep learning, symbolic network mining based on these technologies has even greater development and application potential. Graphs, as a general data representation, can be used to simulate many complex systems in the real world, such as social networks, protein-protein interaction networks, and drug-drug interaction networks. Unsigned graphs represent whether relationships exist between entities and are currently the main graph modeling method, but they cannot distinguish the type of relationship. Symbolic graphs are a type of graph data with edge symbol attributes, and are used in fields such as social networks and drug data to characterize homogeneous and heterogeneous relationships between entities. Symbolic networks can distinguish between two levels of relationships using positive and negative signs, and have stronger expressive power.

[0003] In the link prediction problem, existing technologies mainly suffer from the inability to simultaneously integrate node attributes and graph structure characteristics, and the need to train non-end-to-end patterns in two stages: a feature extractor and a link type discriminator. Neural network models inspired by graph signal processing are a major method for analyzing graph data, avoiding the shortcomings of existing technologies. However, due to the aforementioned limitations, their extension on symbolic graphs remains very limited. Therefore, it is necessary to design a new symbolic graph filter and an end-to-end framework for graph signal processing to better apply these techniques to symbolic graph link prediction. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a link prediction method with higher accuracy based on symbolic graph neural networks.

[0005] To achieve the above objectives, based on one aspect of the present invention, a link prediction method based on symbolic graph neural networks is provided, the method comprising:

[0006] Obtain a symbol graph, which includes multiple drug nodes and multiple actual links connecting two drug nodes. Each drug node is associated with a corresponding multidimensional attribute feature. The multidimensional attribute features of multiple drug nodes in the symbol graph constitute an attribute feature matrix.

[0007] The attribute feature matrix is ​​input into a preset link prediction model. After processing by the symbol graph neural network based on a low-pass filter, the fully connected layer and the regression classifier in the link prediction model, the predicted link between any two drug nodes in the symbol graph is iteratively output. The error between the predicted link and the actual link in the symbol graph is analyzed, and the model parameters of the link prediction model are adjusted according to the error until the training termination condition is met.

[0008] The trained link prediction model is used to predict the link of drug nodes to be predicted, wherein the drug node pair to be predicted is one or more pairs of drug nodes in the symbol graph.

[0009] In one embodiment, the step of iteratively outputting the predicted link between any two drug nodes in the symbol graph after processing by the low-pass filter-based symbol graph neural network, fully connected layer, and regression classifier in the link prediction model includes:

[0010] The symbolic graph neural network based on a low-pass filter in the link prediction model is used to process the attribute feature matrix to obtain a low-dimensional feature representation corresponding to each drug node.

[0011] The fully connected layer and the regression classifier are applied to process the low-dimensional feature representation corresponding to each drug node to output the predicted link between any two drug nodes.

[0012] In one embodiment, the application of the low-pass filter-based symbolic graph neural network in the link prediction model to process the attribute feature matrix and obtain a low-dimensional feature representation corresponding to each drug node includes:

[0013] Based on the symbol diagram, obtain the eigenvector matrix and eigenvalue matrix corresponding to the symbol diagram;

[0014] The symbolic graph neural network based on a low-pass filter is applied to process the feature vector matrix and the feature value matrix to obtain the low-dimensional feature representation corresponding to each drug node.

[0015] In one embodiment, obtaining the eigenvector matrix and eigenvalue matrix corresponding to the symbolic map based on the symbolic map includes:

[0016] Based on the symbolic graph, obtain the symbolic graph adjacency matrix and symbolic graph degree matrix corresponding to the symbolic graph;

[0017] Calculate the Laplace matrix of the symbolic graph based on the adjacency matrix and the degree matrix of the symbolic graph;

[0018] Calculate the symmetric normalized Laplacian matrix of the symbolic graph based on the Laplacian matrix and the symbolic graph degree matrix;

[0019] The symmetric normalized Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvector matrix and the eigenvalue matrix.

[0020] In one embodiment, the symbolic graph neural network based on the low-pass filter is:

[0021] ,

[0022] ,

[0023] in, This represents an attribute feature matrix, where each row of the attribute feature matrix consists of multi-dimensional feature attributes of the drug node, and the number of rows equals the total number of drug nodes. Indicates a low-pass filter. This represents the activation function. This represents the trainable weight matrix used for feature transformation in the first layer. This represents the trainable weight matrix used for feature transformation in the second layer. This indicates the total number of drug nodes. The eigenvector matrix, , For the first The feature vector of each drug node For the eigenvalue matrix, , For the first The characteristic values ​​of each drug node For an ideal low-pass filter, k represents the number of low-frequency components retained by the eigenvalue of each drug node.

[0024] In one embodiment, the step of applying the fully connected layer and the regression classifier to process the low-dimensional feature representation corresponding to each drug node and outputting a predicted link between any two drug nodes includes:

[0025] Based on the low-dimensional feature representation corresponding to each drug node, the fully connected layer is applied to concatenate any two drug nodes.

[0026] The regression classifier is applied to any two drug nodes output by the fully connected layer to perform classification prediction, and the predicted link between any two drug nodes is output.

[0027] In one embodiment, analyzing the error between the predicted link and the actual link in the symbol graph includes:

[0028] The following relationship is used to analyze the error between the predicted link and the actual link in the symbol diagram:

[0029]

[0030] in, Indicates the drug node being spliced. Low-dimensional feature representation and drug nodes Low-dimensional feature representation, Indicates drug node With drug nodes The types of edges between them Represents the set of edge types. or , Indicates the type of connection. Associated weights; if a given prediction is true, then Returns 1 otherwise, returns 0. , This is the weight matrix of the symbolic graph neural network. This is the weight matrix of the fully connected layer. These are the regression coefficients of the regression classifier.

[0031] According to another aspect of the present invention, a link prediction device based on a symbolic graph neural network is provided, comprising:

[0032] The symbol graph acquisition module is used to acquire a symbol graph, which includes multiple drug nodes and multiple actual links connecting two drug nodes. Each drug node is associated with a corresponding multi-dimensional attribute feature. The multi-dimensional attribute features of multiple drug nodes in the symbol graph constitute an attribute feature matrix.

[0033] The link prediction model training module is used to input the attribute feature matrix into a preset link prediction model. After processing by the symbol graph neural network based on a low-pass filter, the fully connected layer and the regression classifier in the link prediction model, the module iteratively outputs the predicted link between any two drug nodes in the symbol graph, analyzes the error between the predicted link and the actual link in the symbol graph, and adjusts the model parameters of the link prediction model according to the error until the training termination condition is met.

[0034] The link prediction module is used to perform link prediction on the drug node pair to be predicted using the trained link prediction model. The drug node pair to be predicted is one or more pairs of drug nodes in the symbol graph.

[0035] According to another aspect of the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the link prediction method as described in any of the above.

[0036] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the link prediction method as described in any of the preceding claims.

[0037] The present invention relates to a link prediction method and apparatus based on symbolic graph neural networks, which is used to solve the problem of drug interaction prediction. It integrates drug node attributes and graph structure characteristics. For the problems of graph signal filtering and link prediction in symbolic graphs, a symbolic graph neural network based on an ideal low-pass filter is designed. The designed link prediction model includes three parts: a symbolic graph neural network based on a low-pass filter, a fully connected layer, and a regression classifier. The link prediction model is designed as an end-to-end structure. The weight matrix of the symbolic graph neural network, the weight matrix of the fully connected layer, and the regression coefficients of the regression classifier are trained simultaneously, which can train the feature extractor and the link type discriminator at the same time. The model has few parameters and high link prediction accuracy. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in one or more embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only one or more embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating an embodiment of the link prediction method based on symbolic graph neural networks of the present invention.

[0040] Figure 2 This is a schematic diagram illustrating the training process of the link prediction model in one embodiment of the present invention;

[0041] Figure 3 This is a schematic diagram of an embodiment of the link prediction device based on symbolic graph neural network of the present invention.

[0042] Figure 4 This is an internal structural diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0044] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the element or object listed following the word and its equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0045] like Figure 1 As shown, this invention provides a link prediction method based on symbolic graph neural networks, the method comprising:

[0046] S1: Obtain the symbol graph, which includes multiple drug nodes and multiple actual links connecting two drug nodes. Each drug node is associated with corresponding multidimensional attribute features. The multidimensional attribute features of multiple drug nodes in the symbol graph constitute an attribute feature matrix.

[0047] Specifically, obtaining the symbolic image requires first obtaining the drug dataset, which contains... Each drug sample in the dataset contains corresponding multidimensional attribute features. Each drug sample in the dataset is treated as a drug node, and the multidimensional attribute features of each drug sample are associated with the multidimensional attribute features of the corresponding drug node. A symbolic graph is used... express, ,in, For a set of nodes, This represents the total number of drug distribution nodes. This is a set of links (i.e., a set of edges).

[0048] In this embodiment, the multidimensional attribute features include drug-binding protein and chemical structure attributes, which are represented as 0-1 strings. Then, the link between any two drug nodes is determined based on the known relationship between any two drug samples. In other embodiments, the multidimensional attribute features of drug nodes may also include other features.

[0049] S2: Input the attribute feature matrix into the preset link prediction model. After processing by the symbol graph neural network based on low-pass filter, fully connected layer and regression classifier in the link prediction model, iteratively output the predicted link between any two drug nodes in the symbol graph, analyze the error between the predicted link and the actual link in the symbol graph, and adjust the model parameters of the link prediction model according to the error until the training termination condition is met.

[0050] In this embodiment, the link prediction model includes three parts: a symbol graph neural network based on a low-pass filter, a fully connected layer, and a regression classifier. When training the link prediction model, the existing edges in the symbol graph are divided into a training set and a test set. First, a model that can accurately classify whether the symbol of an unknown link between nodes is positive or negative is trained on the training set. Then, the test set is used to verify the classification effect of the model.

[0051] Specifically, the attribute feature matrix of the symbol graph is used as the input to the link prediction model. Each row of the attribute feature matrix consists of the multidimensional feature attributes of a drug node, and the number of rows in the attribute feature matrix is ​​equal to the total number of drug nodes.

[0052] First, the attribute feature matrix is ​​processed by a symbolic graph neural network based on a low-pass filter in the link prediction model to obtain the low-dimensional feature representation corresponding to each drug node. Then, a fully connected layer and a regression classifier are applied to process the low-dimensional feature representation corresponding to each drug node to output the predicted link between any two drug nodes.

[0053] In this embodiment, a symbolic graph neural network based on a low-pass filter in the link prediction model is used to process the attribute feature matrix to obtain a low-dimensional feature representation corresponding to each drug node, including:

[0054] Obtain the symbol map adjacency matrix corresponding to the symbol map based on the symbol map. and symbolic graph degree matrix Then, based on the symbolic graph adjacency matrix and symbolic graph degree matrix Calculate the Laplace matrix of the symbolic graph , ,in, For the symbolic graph adjacency matrix, , The positive adjacency matrix of the symbolic graph. , The negative adjacency matrix of the symbolic graph. ,in, Represents a node With nodes The weight of the positive edge between them. Represents a node With nodes The weight of the negative edge between them. , Symbolic graph degree matrix , ,in, , positive adjacency matrix The degree matrix of the symbolic graph is called the positive degree matrix of the symbolic graph. , , negative adjacency matrix The degree matrix of the symbolic graph is called the negative degree matrix of the symbolic graph. , , Used to construct diagonal matrices.

[0055] Then, according to the Laplace matrix Symbolic graph degree matrix Compute the symmetric normalized Laplacian matrix of the symbolic graph , .

[0056] Then, for the symmetric normalized Laplace matrix Perform eigenvalue decomposition to obtain the eigenvector matrix. and eigenvalue matrix , ;

[0057] in, , eigenvalues For image frequency, It is a diagonal matrix composed of eigenvalues, with the following distributions on the diagonal: .

[0058] Then, based on the eigenvector matrix and eigenvalue matrix A symbolic graph neural network based on a low-pass filter is applied to process the data and obtain the low-dimensional feature representation corresponding to each drug node.

[0059] Specifically, the symbolic graph neural network based on the low-pass filter is as follows:

[0060] ,

[0061] ,

[0062] in, Represents the attribute feature matrix, Indicates a low-pass filter. This represents the activation function. This represents the trainable weight matrix used for feature transformation in the first layer. This represents the trainable weight matrix used for feature transformation in the second layer. In this embodiment, the graph neural network based on the low-pass filter has two layers. This indicates the total number of drug nodes. The eigenvector matrix, , Indicates the first The feature vector of each drug node For the eigenvalue matrix, , Indicates the first The characteristic values ​​of each drug node For an ideal low-pass filter, k represents the number of low-frequency components retained by the eigenvalue of each drug node.

[0063] For symmetric normalized Laplace matrix Eigenvalues ​​are obtained by performing eigenvalue decomposition. In the symbol graph, the eigenvectors corresponding to large eigenvalues ​​show a strong inconsistency with the topology, meaning the graph frequencies of large eigenvalues ​​vary significantly across the graph, resulting in poor smoothness. Conversely, the eigenvectors corresponding to small eigenvalues ​​are smoother signals. Using the attribute feature matrix as the graph signal of the symbol graph, a low-pass filter is applied. It can remove eigenvalues The high-frequency signal is removed while the low-frequency signal is retained, thereby removing noise and maintaining the smoothness of the graph signal.

[0064] In this embodiment, the process of applying a fully connected layer and a regression classifier to the low-dimensional feature representation corresponding to each drug node and outputting the predicted link between any two drug nodes includes: applying a fully connected layer to concatenate any two drug nodes according to the low-dimensional feature representation corresponding to each drug node, applying a regression classifier to classify and predict any two drug nodes output by the fully connected layer, and outputting the predicted link between any two drug nodes.

[0065] Specifically, in this embodiment, the link prediction model applies the following relational formula to analyze the error between the predicted link and the actual link in the symbol diagram:

[0066]

[0067] in, Indicates the drug node being spliced. Low-dimensional feature representation and drug nodes Low-dimensional feature representation, Indicates drug node With drug nodes The types of edges between them Represents the set of edge types. , Indicates the type of connection. Associated weights; if a given prediction is true, then Returns 1 otherwise, returns 0. , This is the weight matrix of the symbolic graph neural network. This is the weight matrix of the fully connected layer. These are the regression coefficients of the regression classifier.

[0068] like Figure 2 As shown, when training the link prediction model, the low-dimensional feature representations of each drug node in the symbolic graph are first obtained through a symbolic graph neural network. Then, the output of the symbolic graph neural network is input into a fully connected layer, where any drug node pair is processed. Low-dimensional feature representation , The data is then concatenated and fed into a regression classifier, which predicts drug node pairs. The output labels, either "+", "-", or "0", indicate the pair of drug nodes. The link symbols between them, based on the output labels and the drug node pairs in the symbol diagram. The model parameters of the link prediction model are adjusted based on the actual link error. Iterative optimization calculations are performed as described above, adjusting the model parameters until the training termination condition is met. The training termination condition is generally reaching a preset number of iterations or the error being within an allowable range. The model parameters of the link prediction model include the weight matrix of the symbolic graph neural network, the weight matrix of the fully connected layer, and the regression coefficients of the regression classifier. In this embodiment, the weight matrix of the symbolic graph neural network, the weight matrix of the fully connected layer, and the regression coefficients of the regression classifier are trained simultaneously.

[0069] S3: Use the trained link prediction model to predict the link of the drug node pairs to be predicted. The drug node pairs to be predicted are one or more pairs of drug nodes in the symbol graph.

[0070] By using a low-pass filter-based symbolic graph neural network in the trained link prediction model, the node embedding result (i.e., the final low-dimensional feature representation) of each drug node in the symbolic graph can be obtained. The link prediction model of this invention, as a direct learning type neural network model, can predict unknown link types between existing nodes in the symbolic graph. During link prediction, the symbolic graph is input into the trained link prediction model. In the trained link prediction model, the nodes to be predicted... Node embedding results and Node embedding results The data is concatenated, then fed into a two-layer fully connected layer, and finally the output of the fully connected layer is processed. Inputting the data into the regression classifier will yield the link prediction results for the corresponding drug node pairs.

[0071] Based on this embodiment, a simulation example for predicting drug interactions is provided.

[0072] Specifically, the method of this embodiment is applied to four datasets obtained from the DrugBank website: D-1562, D-597, D-603, and D-548. Among them, D-1562, D-597, and D-603 are symbolic drug interaction prediction datasets, and their corresponding edge type sets... The symbols for the edges are determined based on the semantic descriptions in the DrugBank database. For example, when quinine (DrugBank, Id: DB00468) is used in combination with aprepitant (DrugBank, Id: DB00673), the serum concentration increases, while when used in combination with mitotane (DrugBank, Id: DB00648), the serum concentration decreases. Thus, the edge between quinine and aprepitant is considered to be "+", which is a positive edge, and the edge between quinine and mitotane is considered to be "-", which is a negative edge.

[0073] D-548 is a dataset for predicting drug side effects, and its corresponding set of edge types. In the D-548 dataset, 48,584 drug side effects (i.e., negative links) were obtained from TWOSIDES, indicating only whether a side effect exists between two drugs, without specifying the specific type of side effect. The 8,299 drug positive links were obtained through the similarity of attribute features between drugs. Specifically, in this simulation example, chemical structure (C) and drug-binding proteins (DBPs), as attributes of the drugs themselves, are expressed as 0-1 strings of length 881 and 1642, respectively. For chemical structures, "1" indicates the presence of the corresponding chemical substructure, and "0" indicates its absence. For drug-binding proteins, "1" indicates that the drug binds to a specific target protein, and "0" indicates that the drug does not bind to the specific target protein.

[0074] D-1562, D-597, and D-603, as symbolic drug interaction prediction datasets, have two attribute features: "chemical structure" and "drug-binding protein." D-548, as a drug side effect prediction dataset, only has the attribute feature "chemical structure." Furthermore, spectral features obtained from the Laplacian matrix eigenvalue decomposition can also be used as attribute features for drugs. In this simulation example, the link prediction method of this invention is compared with the Graph Embedding (SSE) method and Graph Convolutional Network (GCN) using the AUC and F1 evaluation metrics. Detailed descriptions of the datasets are shown in Table 1.

[0075] Table 1. Dataset Description

[0076]

[0077] In this simulation experiment, the number of iterations for each method is 1000, the embedding dimension is 32, and the feature vector dimension in SSE is 32. Similarly, in this simulation experiment, the first 32 small feature vectors are used to design a low-pass filter, and thus a symbol graph link prediction model based on the low-pass filter is obtained.

[0078] In this simulation experiment, the prediction performance of SSE and GCN on the four datasets D-1562, D-597, D-603, and D-548 is shown in Tables 2 to 5. AUC (area under the curve) is a model evaluation metric, especially one of the main offline evaluation metrics used in binary classification models. The AUC value is the area enclosed by the ROC curve and the x-axis. The AUC value represents the distance between y=1 and y=0. That is, the larger the AUC, the wider the distance between the two data points, and the higher the classification accuracy. F1 is a metric for classification problems, representing the average of precision and recall. Ours refers to the link prediction method of this invention.

[0079] Table 2: Prediction performance results on the D-1562 dataset

[0080]

[0081] Table 3: Prediction performance results on the D-597 dataset

[0082]

[0083] Table 4: Prediction performance results on the D-603 dataset

[0084]

[0085] Table 5: Prediction performance results on the D-548 dataset

[0086]

[0087] By observing the prediction performance results in Tables 2, 3, 4, and 5, it can be found that, except for the F1 score on the D-548 dataset, where the prediction performance of the method of the present invention is slightly lower than that of GCN, the prediction performance of the method of the present invention is significantly better than that of SSE and GCN on the four datasets D-1562, D-597, D-603, and D-548. The method of the present invention achieves a high standard of prediction effect.

[0088] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, an embodiment of the present invention also provides a link prediction device based on symbolic graph neural networks, used to implement the link prediction method based on symbolic graph neural networks of any of the above embodiments, such as... Figure 3 As shown, it includes:

[0089] Symbol graph acquisition module 10 is used to acquire symbol graphs. Symbol graphs include multiple drug nodes and multiple actual links connecting two drug nodes. Each drug node is associated with corresponding multi-dimensional attribute features. The multi-dimensional attribute features of multiple drug nodes in the symbol graph constitute an attribute feature matrix.

[0090] The link prediction model training module 20 is used to input the attribute feature matrix into the preset link prediction model. After processing by the symbol graph neural network based on low-pass filter, fully connected layer and regression classifier in the link prediction model, iteratively outputs the predicted link between any two drug nodes in the symbol graph, analyzes the error between the predicted link and the actual link in the symbol graph, and adjusts the model parameters of the link prediction model according to the error until the training termination condition is met.

[0091] Link prediction module 20 is used to perform link prediction on the drug node pair to be predicted using the trained link prediction model. The drug node pair to be predicted is one or more pairs of drug nodes in the symbol graph.

[0092] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, an embodiment of the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the link prediction method based on symbolic graph neural network described in any of the above embodiments.

[0093] Figure 4The illustration shows a more specific hardware diagram of an electronic device provided in this embodiment. The device may include: a processor 100, a memory 200, an input / output interface 300, a communication interface 400, and a bus 500. The processor 100, memory 200, input / output interface 300, communication interface 400, and bus 500 are interconnected within the device.

[0094] The processor 100 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention.

[0095] The memory 200 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 200 can store the operating system and other applications. When the technical solutions provided in the embodiments of the present invention are implemented by software or firmware, the relevant program code is stored in the memory 200 and is called and executed by the processor 100.

[0096] The input / output interface 300 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components within the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.

[0097] The communication interface 400 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0098] Bus 500 includes a pathway for transmitting information between various components of the device, such as processor 100, memory 200, input / output interface 300, and communication interface 400.

[0099] It should be noted that although the above-described device only shows the processor 100, memory 200, input / output interface 300, communication interface 400, and bus 500, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0100] Based on the same inventive concept, corresponding to the methods of the above embodiments, an embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the link prediction method based on symbolic graph neural networks as described in any of the above embodiments.

[0101] The computer-readable storage medium of this embodiment includes permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology; the information may be computer-readable instructions, data structures, program modules or other data. Examples of computer-readable storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transfer medium that can be used to store information that can be accessed by computer devices.

[0102] The computer program stored in the computer storage medium of the above embodiments is used to enable the computer to execute the link prediction method based on symbolic graph neural network as described in any of the above embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0103] The present invention relates to a link prediction method and apparatus based on symbolic graph neural networks, which is used to solve the problem of drug interaction prediction. It integrates drug node attributes and graph structure characteristics. For the problems of graph signal filtering and link prediction in symbolic graphs, a symbolic graph neural network based on an ideal low-pass filter is designed. The designed link prediction model includes three parts: a symbolic graph neural network based on a low-pass filter, a fully connected layer, and a regression classifier. The link prediction model is designed as an end-to-end structure. The weight matrix of the symbolic graph neural network, the weight matrix of the fully connected layer, and the regression coefficients of the regression classifier are trained simultaneously, which can train the feature extractor and the link type discriminator at the same time. The model has few parameters and high link prediction accuracy.

[0104] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of the present invention is limited to these examples; within the framework of the present invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present invention as described above, which are not provided in detail for the sake of brevity.

[0105] One or more embodiments of this invention are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this invention. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this invention should be included within the protection scope of this invention.

Claims

1. A link prediction method based on symbolic graph neural networks, characterized in that, The method includes: Obtain a symbol graph, which includes multiple drug nodes and multiple actual links connecting two drug nodes. Each drug node is associated with a corresponding multidimensional attribute feature. The multidimensional attribute features of multiple drug nodes in the symbol graph constitute an attribute feature matrix. The multidimensional attribute features include drug-binding protein and chemical structure attributes. Determine the actual link between any two drug nodes based on the known relationship between any two drug samples. The attribute feature matrix is ​​input into a preset link prediction model. After processing by the symbol graph neural network based on a low-pass filter, the fully connected layer and the regression classifier in the link prediction model, the predicted link between any two drug nodes in the symbol graph is iteratively output. The error between the predicted link and the actual link in the symbol graph is analyzed, and the model parameters of the link prediction model are adjusted according to the error until the training termination condition is met. The process of processing the symbolic graph through the low-pass filter-based symbolic graph neural network, fully connected layer, and regression classifier in the link prediction model, and iteratively outputting the predicted link between any two drug nodes in the symbolic graph, includes: The low-pass filter-based symbolic graph neural network in the link prediction model is used to process the attribute feature matrix to obtain a low-dimensional feature representation corresponding to each drug node. This includes: obtaining a feature vector matrix and a feature value matrix corresponding to the symbolic graph based on the symbolic graph; and applying the low-pass filter-based symbolic graph neural network to process the feature vector matrix and the feature value matrix to obtain the low-dimensional feature representation corresponding to each drug node. Based on the low-dimensional feature representation corresponding to each drug node, the fully connected layer and the regression classifier are applied for processing to output the predicted link between any two drug nodes. The symbolic graph neural network based on the low-pass filter is: ， ， in, This represents an attribute feature matrix, where each row of the attribute feature matrix consists of multi-dimensional feature attributes of the drug node, and the number of rows equals the total number of drug nodes. Indicates a low-pass filter. This represents the activation function. This represents the trainable weight matrix used for feature transformation in the first layer. This represents the trainable weight matrix used for feature transformation in the second layer. This indicates the total number of drug nodes. The eigenvector matrix, , For the first The feature vector of each drug node For the eigenvalue matrix, , For the first The characteristic values ​​of each drug node For an ideal low-pass filter, k represents the number of low-frequency components retained by the eigenvalue of each drug node; The analysis of the error between the predicted link and the actual link in the symbol diagram includes: The following relational analysis is applied: in, Indicates the drug node being spliced. Low-dimensional feature representation and drug nodes Low-dimensional feature representation, Indicates drug node With drug nodes The types of edges between them Represents the set of edge types. or , Indicates the type of connection. Associated weights; if a given prediction is true, then Returns 1 otherwise, returns 0. , This is the weight matrix of the symbolic graph neural network. This is the weight matrix of the fully connected layer. The regression coefficients are those of the regression classifier. The trained link prediction model is used to predict the link of drug nodes to be predicted, wherein the drug node pair to be predicted is one or more pairs of drug nodes in the symbol graph.

2. The link prediction method based on symbolic graph neural networks as described in claim 1, characterized in that, The step of obtaining the eigenvector matrix and eigenvalue matrix corresponding to the symbolic map based on the symbolic map includes: Based on the symbolic graph, obtain the symbolic graph adjacency matrix and symbolic graph degree matrix corresponding to the symbolic graph; Calculate the Laplace matrix of the symbolic graph based on the adjacency matrix and the degree matrix of the symbolic graph; Calculate the symmetric normalized Laplacian matrix of the symbolic graph based on the Laplacian matrix and the symbolic graph degree matrix; The symmetric normalized Laplacian matrix is ​​subjected to eigenvalue decomposition to obtain the eigenvector matrix and the eigenvalue matrix.

3. The link prediction method based on symbolic graph neural networks as described in claim 1, characterized in that, The step of processing the low-dimensional feature representation corresponding to each drug node using the fully connected layer and the regression classifier to output the predicted link between any two drug nodes includes: Based on the low-dimensional feature representation corresponding to each drug node, the fully connected layer is applied to concatenate any two drug nodes. The regression classifier is applied to any two drug nodes output by the fully connected layer to perform classification prediction, and the predicted link between any two drug nodes is output.

4. A link prediction device based on symbolic graph neural networks, characterized in that, include: The symbol graph acquisition module is used to acquire a symbol graph, which includes multiple drug nodes and multiple actual links connecting two drug nodes. Each drug node is associated with a corresponding multidimensional attribute feature. The multidimensional attribute features of multiple drug nodes in the symbol graph constitute an attribute feature matrix. The multidimensional attribute features include drug-binding protein and chemical structure attributes. The actual link between any two drug nodes is determined based on the known relationship between any two drug samples. The link prediction model training module is used to input the attribute feature matrix into a preset link prediction model. After processing by the symbol graph neural network based on a low-pass filter, the fully connected layer and the regression classifier in the link prediction model, the module iteratively outputs the predicted link between any two drug nodes in the symbol graph, analyzes the error between the predicted link and the actual link in the symbol graph, and adjusts the model parameters of the link prediction model according to the error until the training termination condition is met. The process of processing the symbolic graph through the low-pass filter-based symbolic graph neural network, fully connected layer, and regression classifier in the link prediction model, and iteratively outputting the predicted link between any two drug nodes in the symbolic graph, includes: The low-pass filter-based symbolic graph neural network in the link prediction model is used to process the attribute feature matrix to obtain a low-dimensional feature representation corresponding to each drug node. This includes: obtaining a feature vector matrix and a feature value matrix corresponding to the symbolic graph based on the symbolic graph; and applying the low-pass filter-based symbolic graph neural network to process the feature vector matrix and the feature value matrix to obtain the low-dimensional feature representation corresponding to each drug node. Based on the low-dimensional feature representation corresponding to each drug node, the fully connected layer and the regression classifier are applied for processing to output the predicted link between any two drug nodes. The symbolic graph neural network based on the low-pass filter is: ， ， in, This represents an attribute feature matrix, where each row of the attribute feature matrix consists of multi-dimensional feature attributes of the drug node, and the number of rows equals the total number of drug nodes. Indicates a low-pass filter. This represents the activation function. This represents the trainable weight matrix used for feature transformation in the first layer. This represents the trainable weight matrix used for feature transformation in the second layer. This indicates the total number of drug nodes. The eigenvector matrix, , For the first The feature vector of each drug node For the eigenvalue matrix, , For the first The characteristic values ​​of each drug node For an ideal low-pass filter, k represents the number of low-frequency components retained by the eigenvalue of each drug node; The analysis of the error between the predicted link and the actual link in the symbol diagram includes: The following relational analysis is applied: in, Indicates the drug node being spliced. Low-dimensional feature representation and drug nodes Low-dimensional feature representation, Indicates drug node With drug nodes The types of edges between them Represents the set of edge types. or , Indicates the type of connection. Associated weights; if a given prediction is true, then Returns 1 otherwise, returns 0. , This is the weight matrix of the symbolic graph neural network. This is the weight matrix of the fully connected layer. The regression coefficients are those of the regression classifier. The link prediction module is used to perform link prediction on the drug node pair to be predicted using the trained link prediction model. The drug node pair to be predicted is one or more pairs of drug nodes in the symbol graph.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 3.

Images