Drug repositioning method and system based on subgraph perception and hybrid graph neural network

Abstract

The application discloses a drug repositioning method and system based on subgraph perception and a hybrid graph neural network, and the method comprises the following steps: in a drug-disease association network, a corresponding dynamic drug-disease association subgraph is generated for a target drug-disease pair; a drug subgraph and a disease subgraph are respectively constructed for a drug node and a disease node in the drug-disease association subgraph; deep feature extraction is performed on the drug-disease association subgraph by using a multi-head dynamic graph attention neural network model; feature extraction is performed on the drug-disease association subgraph by using a GCN model and a GAT model respectively, so that the final feature representations of the drug subgraph, the disease subgraph and the drug-disease association subgraph are obtained; and the three are fused and input into a drug-disease association prediction model based on a multi-layer perception machine (MLP) network, so that the association relationship of the target drug-disease pair is predicted. The application can realize efficient and accurate drug-disease association prediction.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of drug prediction and analysis, and particularly relates to a drug repositioning method and system based on subgraph perception and a hybrid graph neural network. BACKGROUND

[0002] Drug repositioning refers to a process of accelerating drug development by exploring new indications of approved drugs. Compared with traditional new drug development, drug repositioning finds potential therapeutic uses of existing drugs beyond their original indications through re-research and analysis. Since the safety and pharmacokinetic properties of drugs have been verified through previous clinical trials and market, drug repositioning not only significantly shortens the development cycle from laboratory research to clinical application, but also effectively reduces the research and development cost and failure risk, thus becoming an efficient and low-risk alternative in the field of drug development. This strategy has shown great potential in the treatment of cancer, neurodegenerative diseases, infectious diseases and other fields, and its high success rate and low development cost have brought important commercial value and social benefits to the pharmaceutical industry and medical researchers.

[0003] In recent years, with the rapid development of deep learning technology, especially graph neural networks, drug repositioning methods based on graph learning have gradually become a research hotspot. Such methods can utilize the powerful graph structure data processing capability of graph neural networks to efficiently model the complex relationships between drugs and diseases. For example, graph convolution networks gradually aggregate node neighborhood information through stacked convolution layers to realize multi-level feature expression of drug-disease networks; graph attention networks effectively capture the differences in local structures in the network by adaptively assigning importance weights to different nodes.

[0004] However, existing drug repositioning methods fail to fully exploit the potential information in drug and disease networks and effectively combine the global and local characteristics of the network, resulting in limited accuracy and reliability of drug-disease relationship prediction. Therefore, how to develop a drug repositioning method that can dynamically capture the local structural features of the network while fully utilizing the global semantic information to achieve efficient and accurate drug-disease association prediction has become an important problem to be solved in the current technical field. SUMMARY

[0005] The purpose of the present application is to provide a drug repositioning method based on subgraph perception and a hybrid graph neural network, which can achieve efficient and accurate drug-disease association prediction.

[0006] In order to achieve the above-mentioned purpose of the present application, the technical solution provided by the present application is as follows:

[0007] In a first aspect, the present application provides a drug repositioning method based on subgraph perception and a hybrid graph neural network, comprising:

[0008] S1, in drug-disease association networks In this context, the focus is on the target drug-disease pair, starting from the target drug-disease pair... Extract all drug nodes and disease nodes related to the target node from the neighborhood, as well as the edges between these nodes, to generate a corresponding dynamic drug-disease association subgraph. ;

[0009] S2. Utilizing similarity data between drug nodes and similarity data between disease nodes, targeting... Drug subgraphs are constructed from drug nodes and disease nodes respectively. Disease subgraph ;

[0010] S3, to The application uses a multi-head dynamic graph attention neural network model for deep feature extraction to obtain... Final feature representation ;

[0011] S4. Based on the fusion of the Graph Convolutional Network (GCN) model and the Graph Attention Network (GAT) model, respectively... and The network performs feature extraction to obtain and Final feature representation and ;

[0012] S5, will , and Final feature representation , , By merging features, a unified feature representation can be constructed. ;Will Input a drug-disease association prediction model based on a multilayer perceptron (MLP) network, and output the prediction results of the association between the target drug and the disease pair.

[0013] The multi-head dynamic graph attention neural network model in step S3, the graph convolutional network (GCN) model and the graph attention network (GAT) model in step S4, and the drug-disease association prediction model based on the multilayer perceptron (MLP) network in step S5 are obtained by training and optimization based on known target drug-disease pair associations.

[0014] In a possible implementation, the method further includes: based on a known target drug-disease pair association, training and optimizing the multi-head dynamic graph attention neural network model in step S3, the graph convolution network (GCN) model and the graph attention network (GAT) model in S4, and the drug-disease association prediction model based on the multi-layer perceptron (MLP) network in step S5, including:

[0015] S1-S5 are performed on the target drug-disease pair with known association, to obtain an association prediction result of the target drug-disease pair; according to the association prediction result of the target drug-disease pair and the true label of the association of the target drug-disease pair, a cross-entropy loss function is used as a loss function, a regularization term is introduced to improve the generalization ability of the model, the learnable parameters are updated to accelerate the convergence of the model by combining the Adam optimizer, and the Dropout technology is used to reduce overfitting, so as to realize the optimization of the model.

[0016] Therefore, when processing sparse data or unbalanced data sets, problems such as overfitting or insufficient generalization ability can be avoided, the prediction effect on rare diseases is improved, and the robustness to noise information in the data is improved.

[0017] In a possible implementation, in step S1, the ;

[0018] A drug-disease association network is constructed using known drug-disease associations , wherein is a set of drug nodes, contains different drugs, is a set of disease nodes, contains different diseases; is an edge set, , wherein represents the association between the drug node and the disease node , if the drug and the disease are known to be associated, the weight of the edge is set to 1, otherwise it is 0.

[0019] In a possible implementation, in step S2, the drug subgraph , wherein is a set of nodes in , containing all drug nodes in , is an edge set in , representing Similarity data between drug nodes; disease subgraph ),in for The set of nodes in the middle contains All disease nodes in for The set of edges in the , representing Similarity data between disease nodes;

[0020] In step S2, based on the chemical structure information (such as SMILES representation) obtained from the DrugBank database, the pairwise similarity between all drug nodes is calculated using the Tanimoto similarity formula to generate a drug similarity matrix. , Middle Line 1 Column elements Indicates the first The drug node and the first Similarity between drug sections; using a drug similarity matrix. Data representing the similarity between drug nodes;

[0021] In step S2, phenotypic descriptions of diseases (such as MeSH terms) are obtained from the OMIM disease phenotype database, and the pairwise similarity between all disease nodes is obtained using a semantic calculation method based on Jaccard similarity. This generates a disease similarity matrix. , Middle Line 1 Column elements Indicates the first The disease node and the first The similarity between disease nodes; using a disease similarity matrix. Data representing the similarity between disease nodes;

[0022] Based on the similarity data between drug nodes and the similarity data between disease nodes, K-nearest neighbor similarity matrices for drugs and diseases are constructed respectively. and , Middle Line 1 Column elements and Middle Line 1 Column elements The definition is as follows:

[0023] ;

[0024] wherein, represents the K-neighbors set of the i-th drug node, if the i-th drug node belongs to the K-neighbors set, the similarity value keeps the original similarity value , otherwise, it is set to 0;

[0025]

[0026] wherein, represents the K-neighbors set of the i-th disease node, if the i-th disease node belongs to the K-neighbors set, the similarity value keeps the original similarity value , otherwise, it is set to 0;

[0027] The K-neighbors similarity matrix of drugs and diseases is obtained and After obtaining the K-neighbors similarity matrix of drugs and diseases and , the two are respectively taken as the weight values of the corresponding edges in and

[0028] .In a possible implementation, in the step S3, the multi-head dynamic graph attention neural network model comprises a multi-layer neural network.

[0029] The step S3 comprises:

[0030] S301, in the i-th attention head of the j-th layer neural network, the attention weight of the node and the node is calculated as follows:

[0031]

[0032] wherein, , represents the neighbor node set in the neighborhood of the node , and and respectively represent the feature vectors of the nodes and ; is the weight matrix of the i-th attention head of the j-th layer neural network, used for mapping the input feature to a high-dimensional latent space, is the weight matrix of the i-th attention head of the j-th layer neural network, used for mapping the input feature to a high-dimensional latent space, is the weight matrix of the i-th attention head of the j-th layer neural network, used for mapping the input feature to a high-dimensional latent space, is the weight matrix of the i-th attention head of the j-th layer neural network, used for mapping the input feature to a high-dimensional latent space, is the weight matrix of the i-th attention head of the j-th layer neural network, used for mapping the input feature to a high-dimensional latent space, is the weight matrix of the i-th attention head of the j-th layer neural network, used for mapping the input feature to a high-dimensional latent space, ​​​​​​​​a significance weight vector of the attention head connection feature; and The values of the attention weight vector can be randomly initialized and changed as the model learns.

[0033] S302, based on the attention weight, using a multi-head attention mechanism to weight and sum (dynamic aggregation) the features of the adjacent nodes of the target node, the formula is defined as:

[0034] ;

[0035] wherein, represents the node The updated feature representation in the first layer neural network, represent the number of attention heads;

[0036] In the above steps, each attention head independently calculates the attention weight and averages the features of the adjacent nodes to ensure consistency in dimensions.

[0037] S303, for using a ReLU activation function for nonlinear activation and using a normalization function for normalization processing to enhance the nonlinear ability of feature expression, obtaining the output feature of the first layer neural network; ;

[0038] S304, the output features of the multi-layer neural network are fused to obtain the feature matrix of ; the feature matrix of the first behavior , ;

[0039] That is, each row in the feature matrix corresponds to a node trained feature vector;

[0040] S305, using a differentiable pooling method to learn trainable normalization weights, grouping and aggregating node features to reduce the size of the drug-disease subgraph and gradually expand the receptive field range of the node; The feature matrix and the adjacency matrix are mapped to a smaller subgraph by the following transformation:

[0041] ;

[0042] ;

[0043] wherein, is a learnable normalized weight matrix to assign original nodes into fewer super nodes, is a , is the number of original nodes, is the number of super nodes after pooling, where element represents the weight of original node assigned to super node , is randomly initialized and changes with the learning of the model; is the adjacency matrix of , describing the connection between nodes, where element takes the weight of the corresponding edge in ; is the feature matrix of super nodes, is the adjacency matrix after pooling, describing the relationship between super nodes;

[0044] and constitute the final feature representation .

[0045] In one possible implementation, the step S4 comprises:

[0046] S401, using a GCN model to aggregate the features of the target node and its adjacent nodes; wherein the GCN model comprises a layer neural network, wherein 2; drug nodes and disease nodes in the first layer neural network are updated according to the following formula:

[0047] ;

[0048] ;

[0049] wherein, , respectively represent the local structure characteristics of the drug nodes and disease nodes in the first layer neural network, and the normalization coefficients and are defined as:

[0050] ;

[0051] ;

[0052] in, and Representing drug nodes and drug nodes In the drug subgraph The degree of the interval (the sum of in-degree and out-degree); and Representing disease nodes and disease nodes In the disease subplot The degree in;

[0053] , In the GCN model, respectively Drug subgraph in layered neural networks Disease subgraph The learnable parameter matrix has values ​​that can be randomly initialized and optimized during training using the backpropagation algorithm;

[0054] Calculated from the above formula and ;

[0055] S402. Use the GAT model to capture the interactions between neighboring nodes; where the GAT model includes... Layered neural networks, in which 2; Each layer of the neural network employs a multi-head attention mechanism for the drug nodes. and disease nodes Update the features; drug node and disease nodes In the GAT model The feature update formula for a layered neural network is defined as follows:

[0056] ;

[0057] ;

[0058] and For drug nodes and disease nodes In the GAT model The updated node features of the layered neural network This refers to the number of attention heads included in the GAT model. For the first GAT model The first layer of a neural network Drug nodes in individual attention points and drug nodes Attention weights between them For the first GAT model The first layer of a neural network Disease nodes in individual attention points and disease nodes Attention weights between them and In the GAT model, respectively Drug subgraph in layered neural networks Disease subgraph The learnable parameter matrix has values ​​that can be randomly initialized and optimized during training using the backpropagation algorithm;

[0059] Calculated from the above formula and ;

[0060] S403. Weighted fusion of the outputs of the GCN model and the GAT model to generate the final feature representation of the target node; the fusion process is defined as follows:

[0061] ;

[0062] ;

[0063] in, and Representing drug nodes and disease nodes The final characteristic representation, The fusion ratio hyperparameter is used to adjust the relative contributions of the outputs of the GCN and GAT models.

[0064] Drug subgraph The final feature representation of all drug nodes constitutes the drug subgraph. Final feature representation ;

[0065] Disease subgraph The final feature representation of all disease nodes constitutes the disease subgraph. Final feature representation .

[0066] In one possible implementation, in step S5, , and Final feature representation , , By merging features, a unified feature representation can be constructed. The feature fusion process is defined as follows:

[0067] ;

[0068] wherein, represents a concatenation operation.

[0069] In a possible implementation, the method further includes: The input is a drug-disease association prediction model based on a multi-layer perception (MLP) network, and the drug-disease association prediction model outputs a prediction result of an association between a target drug-disease pair.

[0070] The output layer of the drug-disease association prediction model predicts an association probability between a drug node and a disease node , and the greater the probability, the more likely the association:

[0071] ;

[0072] wherein, , , and are weight and coefficient parameters, and values of the weight and coefficient parameters can be randomly initialized and optimized through a back propagation algorithm in a training process.

[0073] In a possible implementation, the cross-entropy loss function is defined as follows:

[0074] ;

[0075] wherein, is a true label of an association between a drug node and a disease node , is a predicted association probability between a drug node and a disease node , is a total number of samples. The sample is a target drug-disease pair data with a known association.

[0076] A regularization term is introduced to improve the generalization ability of the model, and the regularization term is added to an objective function, and is defined as:

[0077] ;

[0078] wherein, represents a regularization coefficient, represents a learnable parameter of a kth layer of the model.

[0079] In a second aspect, the application provides a drug repositioning system based on subgraph perception and hybrid graph neural network, comprising: a data input layer, a subgraph extraction layer, a feature extraction layer, a multi-head dynamic graph attention neural network model layer, a GCN model and GAT model fusion neural network layer, a feature fusion and prediction layer.

[0080] The data input layer is used to collect and integrate data of drugs, diseases and their associated relationships to form a drug-disease association network .

[0081] The subgraph extraction layer is used to extract all drug nodes and disease nodes related to the target node and the edges between these nodes from the neighborhood of the target drug-disease pair in the drug-disease association network , with the target drug-disease pair as the center, to generate a dynamic drug-disease association subgraph .

[0082] The feature extraction layer is used to construct drug subgraphs and disease subgraphs for drug nodes and disease nodes in the dynamic drug-disease association subgraph respectively using similarity data between drug nodes and similarity data between disease nodes.

[0083] The multi-head dynamic graph attention neural network model layer is used to perform deep feature extraction on the dynamic drug-disease association subgraph to obtain the final feature representation .

[0084] The GCN model and GAT model fusion neural network layer is used to perform feature extraction on the drug subgraph and the disease subgraph respectively based on the graph convolution network GCN model and the graph attention network GAT model fusion to obtain and the final feature representation and .

[0085] The feature fusion and prediction layer is used to fuse , and the final feature representation , , to construct a unified feature representation . The drug-disease association prediction model based on the multi-layer perception machine MLP network is input, and the drug-disease association prediction model outputs the association relationship prediction result of the target drug-disease pair.​

[0086] The system implements drug repositioning based on the drug repositioning method based on subgraph perception and hybrid graph neural network described above.

[0087] In a third aspect, the present application provides an electronic device, comprising a memory and a processor.

[0088] The memory is configured to store a computer program.

[0089] The processor is configured to invoke the computer program to execute the method described above.

[0090] In a fourth aspect, the present application provides a computer readable storage medium, wherein the computer readable storage medium stores a computer program, and the computer program, when executed on an electronic device, causes the electronic device to implement the method described above.

[0091] In a fifth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program, when executed on an electronic device, causes the electronic device to implement the method described above.

[0092] The specific implementation manners of the second to fifth aspects of the present application described above can refer to the implementation manners of the first aspect described above, which will not be described here.

[0093] The beneficial effects of the present application are as follows:

[0094] The present application discloses a drug repositioning method based on subgraph perception and hybrid graph neural network, which extracts a subgraph centered on a target drug-disease pair dynamically; constructs drug subgraphs and disease subgraphs using the similarity of drugs and diseases, and comprehensively uses a multi-head dynamic graph attention mechanism and a multi-modal feature extraction technology fused with GCN model and GAT model to effectively capture the local structure and global similarity relationship of nodes; and realizes accurate prediction of drug-disease association through feature fusion and MLP network. Thus, the present application solves the problems of insufficient local feature capture, poor multi-dimensional feature fusion capability, and poor performance in sparse and unbalanced data caused by static modeling of drug-disease network in existing methods, and has the significant advantages of high accuracy, strong robustness and good generalization capability. BRIEF DESCRIPTION OF DRAWINGS

[0095] The present application will be further described below in combination with the drawings and embodiments:

[0096] Figure 1 is a flowchart of a drug repositioning method based on subgraph perception and hybrid graph neural network in an embodiment of the present application;

[0097] Figure 2 is a subgraph extraction schematic diagram in an embodiment of the present application;

[0098] Figure 3 This is a comparative diagram of predictive new drug experiments in one embodiment of this application;

[0099] Figure 4 This is a schematic diagram of the system flow of a drug relocation system based on subgraph perception and hybrid graph neural network in one embodiment of this application. Detailed Implementation

[0100] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0101] Example 1

[0102] like Figure 1 As shown in the embodiments, this application discloses a drug relocation method based on subgraph perception and hybrid graph neural networks, including:

[0103] S1, in drug-disease association networks In this context, the focus is on the target drug-disease pair, starting from the target drug-disease pair... Extract all drug nodes and disease nodes related to the target node from the neighborhood, as well as the edges between these nodes, to generate a corresponding dynamic drug-disease association subgraph. ;

[0104] In some embodiments, the Constructing drug-disease association networks using known drug-disease associations. ,in For drug node set, Include Different drugs, For disease node set, Include Different diseases; Let be the set of edges. ,in Representative drug node and disease nodes The relationship between them, if the drugs are known With disease If there is a relationship, then the edge weights Set to 1 otherwise set to 0.

[0105] In this embodiment, the dynamic drug-disease association subgraph This is achieved by centralizing the target drug-disease pair within a drug-disease association network, and then... Extracting a local subgraph consisting of all associated nodes and edges within the neighborhood, such as... Figure 2 (a) Target drug -disease As the center node, The neighbor nodes in the 2-neighborhood have , The neighbor nodes have , and these nodes are integrated to construct the target drug -disease subgraph, such as Figure 2 (b), which retains the topology and interaction around the target node and is used to capture local features.

[0106] In order to accurately represent the hierarchical structure of the dynamic drug-disease association subgraph , the target drug node and the disease node are layered and labeled according to the shortest path distance from the node to the center node (the shortest path distance from the drug node and the disease node of the target drug-disease pair), to distinguish the hierarchical information of different nodes, while retaining the relative position relationship between nodes, retaining the topology and interaction around the target node, ensuring the clarity of the subgraph structure feature expression, which can be used to capture local features.

[0107] In order to avoid data leakage in the prediction task, in the construction process of the dynamic drug-disease association subgraph , the direct edge between the target drug and the target disease is deleted (such as Figure 3 The and ), to ensure that the model will not be disturbed by known associations during the learning process.

[0108] By extracting subgraphs for feature extraction, it avoids the network being too large and noisy, and compared with the global feature aggregation method used by the full graph learning model, it improves the efficiency and avoids ignoring the key detail features of the subgraph level between specific drug-disease pairs, which usually carry important semantic information and local interaction features. Feature extraction through subgraphs can make the model focus on meaningful interactions.

[0109] S2, using the similarity data between drug nodes and the similarity data between disease nodes, constructs a drug subgraph (drug similarity network) and a disease subgraph (disease similarity network) for the drug nodes and disease nodes in ;

[0110] In some embodiments, the drug subgraph , where is the node set in , containing all drug nodes in , and is The set of edges in the , representing Similarity data between drug nodes; disease subgraph ),in for The set of nodes in the middle contains All disease nodes in for The set of edges in the , representing Similarity data between disease nodes.

[0111] In some embodiments, based on chemical structure information (such as SMILES representation) obtained from the DrugBank database, the pairwise similarity between all drug nodes is calculated using the Tanimoto similarity formula to generate a drug similarity matrix. , Middle Line number Column elements Indicates the first The drug node and the first Similarity between drug sections; using a drug similarity matrix. Data representing the similarity between drug nodes;

[0112] In some embodiments, by obtaining phenotypic descriptions of diseases (such as MeSH terms) from the OMIM disease phenotype database, and using a semantic calculation method based on Jaccard similarity, the pairwise similarity between all disease nodes is obtained. This generates a disease similarity matrix. , Middle Line number Column elements Indicates the first The disease node and the first The similarity between disease nodes; using a disease similarity matrix. Data representing the similarity between disease nodes;

[0113] Based on the similarity data between drug nodes and the similarity data between disease nodes, K-nearest neighbor similarity matrices for drugs and diseases are constructed respectively. and , Middle Line number Column elements and Middle Line number Column elements The definition is as follows:

[0114] ;

[0115] wherein, represents the K-nearest neighbor set of the i-th drug node, if the i-th drug node belongs to the K-nearest neighbor set, then the similarity value remains the original similarity value , otherwise it is set to 0. wherein,

[0116] ;

[0117] wherein, represents the K-nearest neighbor set of the i-th disease node, if the i-th disease node belongs to the K-nearest neighbor set, then the similarity value remains the original similarity value , otherwise it is set to 0. After obtaining the K-nearest neighbor similarity matrices of drugs and diseases

[0118] and , the two are respectively taken as the weight values of the corresponding edges in and . S3, for

[0119] applying a multi-head dynamic graph attention neural network model to perform deep feature extraction to obtain the final feature representation ;

[0120] In some embodiments, in the step S3, the multi-head dynamic graph attention neural network model comprises a multi-layer neural network.

[0121] In one specific embodiment, the feature extraction network is composed of three layers of neural networks, and the number of attention heads of the three layers is 6. The input dimension of the first layer network is the dimension of the original features of the nodes, and the hidden layer dimension is 128; the input dimension of the second layer network is the product of the output dimension of the first layer and the number of attention heads, and the hidden layer dimension is 64; the input dimension of the third layer is the product of the output dimension of the second layer and the number of attention heads, and the hidden layer dimension is 1.

[0122] ​​​​​Each layer of the neural network dynamically calculates the attention weight (attention coefficient) between each node and its neighbor node through the attention mechanism, thereby extracting the interaction characteristics between the nodes; the features of the adjacent nodes are weighted and summed to obtain the updated feature representation of the target node, and the ReLU activation function is used for nonlinear activation and the normalization function is used for normalization processing, so as to enhance the nonlinear ability of the feature expression; after the output features of the multi-layer neural network are fused, the differentiable pooling method is used to learn the trainable normalization weight, the node features are grouped and aggregated to obtain the final feature representation . In this way, the local structure and global topology information of the subgraph can be retained, and the local interaction relationship and global topology structure information can be comprehensively captured.

[0123] Through this step, different importance can be assigned to the neighborhood nodes, and the feature aggregation is more accurate.

[0124] Specifically, the step S3 comprises:

[0125] S301, in the first layer neural network, the attention weight of the node and the node is calculated as follows:

[0126] ;

[0127] wherein, ; represents the neighborhood node set in the neighborhood of the node ; and represent the feature vectors of the node and respectively; in the step S1, the nodes can be first marked according to the shortest path distance of the nodes to the target drug-disease pair, and then the hierarchical marking information of the nodes is one-hot encoded as the feature vector of the node is the weight matrix of the first layer neural network, which is used to map the input features to a high-dimensional latent space, is the significance weight vector of the first layer neural network in the first attention head; the values of and can be randomly initialized and changed with the learning of the model;

[0128] ​S302. Based on attention weights, a multi-head attention mechanism is used to perform weighted summation (dynamic aggregation) of the features of neighboring nodes of the target node. The formula is defined as follows:

[0129] ;

[0130] in, Represents a node In the The updated feature representation in a layered neural network. Represents the number of attention heads;

[0131] In the above steps, each attention head independently calculates attention weights and averages the features of adjacent nodes to ensure dimensionality consistency;

[0132] S303, to Nonlinear activation is performed using the ReLU activation function, and normalization is performed using a normalization function to enhance the nonlinearity of feature representation, resulting in nodes. In the Output features of layered neural networks ;

[0133] S304. Fuse the output features of the multi-layer neural network to obtain... Feature matrix The feature matrix The Behavior , ;

[0134] In other words, Each row in the vector corresponds to the feature vector of a node.

[0135] S305. Utilize differentiable pooling to learn trainable normalized weights, group and aggregate node features, and then... Feature matrix and adjacency matrix Map to a smaller subgraph using the following transformation:

[0136] ;

[0137] ;

[0138] in, It is a learnable normalized weight matrix used to distribute raw nodes among fewer supernodes. The dimension is , The original number of nodes. The number of supernodes after pooling, where elements representative node assigned to super node weight, randomly initialized and changed with the learning of the model; for adjacency matrix describing the connection between nodes, where the element takes the weight of the corresponding edge in ; represent the feature matrix of the super node, is the pooled adjacency matrix describing the relationship between super nodes;

[0139] and constitute the final feature representation .

[0140] By learning trainable normalization weights through differentiable pooling method, grouping and aggregating node features can reduce the scale of drug-disease subgraph, gradually adjust the receptive field range of nodes, avoid dilution of redundant information, and improve the efficiency and generalization ability of the model.

[0141] S4, based on the graph convolution network GCN model and the graph attention network GAT model, respectively, the feature extraction of and network, get and the final feature representation and ;

[0142] This step uses the GCN model to aggregate neighborhood information to capture the local structure characteristics of the node, and calculates the attention weight between the node and the neighbor node through the GAT model, so as to extract the global similarity relationship of the node; Therefore, the structural characteristics and similarity relationship of the drug node and the disease node can be extracted;

[0143] In one specific embodiment, the input dimension of the GCN model layer is the dimension of the drug subgraph / disease subgraph, that is, the number of columns of the similarity matrix, and the output dimension is 128. The importance weight between the node and the neighbor node is calculated through the graph attention network (GAT model), so as to extract the global similarity relationship of the node. The dimension of the GAT model layer is set to be the same as that of the GCN model layer, and the hidden layer is 64.

[0144] In some embodiments, this step includes:

[0145] S401, using the GCN model to aggregate the features of the target node and its adjacent nodes; wherein the GCN model includes layer neural network, wherein​ 2; drug node and disease node In the first layer neural network, the feature update formula is:

[0146] ;

[0147] ;

[0148] wherein, , respectively represent the local structure characteristics of the drug node and the disease node in the first layer neural network, and the normalization coefficients and are defined as:

[0149] ;

[0150] ;

[0151] wherein, and respectively represent the degrees (sum of in-degree and out-degree) of the drug node and the drug node in the drug subgraph ; and respectively represent the degrees of the disease node and the disease node in the disease subgraph ;

[0152] , are respectively the learnable parameter matrices of the first layer neural network in the GCN model for the drug subgraph and the disease subgraph , which can be randomly initialized and optimized through the back propagation algorithm in the training process;

[0153] and are calculated by the above formula;

[0154] S402, using a GAT model to capture the interaction between adjacent nodes; wherein the GAT model comprises layer neural networks, wherein 2; each layer of neural network respectively uses a multi-head attention mechanism to update the features of the drug node and the disease node ; the drug node​ and disease nodes In the GAT model, the feature update formula of the layer neural network is defined as follows:

[0155] ;

[0156] ;

[0157] and is the drug node and disease nodes In the GAT model, the updated node feature of the layer neural network, is the number of attention heads contained in the GAT model, is the attention weight between the drug node and the drug node in the th attention head of the layer neural network in the GAT model, is the attention weight between the disease node and the disease node in the th attention head of the layer neural network in the GAT model, and are respectively the learnable parameter matrices for the drug subgraph and the disease subgraph in the layer neural network in the GAT model, which can be randomly initialized and optimized through the backpropagation algorithm during the training process;

[0158] and are calculated by the above formula ;

[0159] S403, the outputs of the GCN model and the GAT model are weighted and fused to generate the final feature representation of the target node; the fusion process is defined as:

[0160] ;

[0161] ;

[0162] wherein, and represent the final feature representations of the drug node and the disease node , respectively The fusion ratio hyperparameter is used to adjust the relative contributions of the outputs of the GCN and GAT models.

[0163] Drug subgraph The final feature representation of all drug nodes constitutes the drug subgraph. Final feature representation ;

[0164] Disease subgraph The final feature representation of all disease nodes constitutes the disease subgraph. Final feature representation .

[0165] This step can improve the ability to identify hidden patterns in drug-disease associations, thereby enhancing the overall expressive power of the model.

[0166] S5, will , and Final feature representation , , By merging features, a unified feature representation can be constructed. ;Will Input a drug-disease association prediction model based on a multilayer perceptron (MLP) network, and output the prediction results of the association between the target drug and the disease pair.

[0167] Will , and Final feature representation , , A unified feature representation is obtained by fusing the data, which combines the characteristics of both drugs and disease nodes. This unified feature representation is then subjected to a nonlinear transformation using an MLP (Multi-Level Processing) to learn the complex interactions between drugs and disease nodes.

[0168] In some embodiments, , and Final feature representation , By merging features, a unified feature representation can be constructed. The feature fusion process is defined as follows:

[0169] ;

[0170] in, This represents a splicing operation.

[0171] In some embodiments, the output layer of the drug-disease association prediction model predicts the association probability between the drug node and the disease node , the greater the probability, the more likely the association:

[0172] ;

[0173] wherein, , , and are weight and coefficient parameters, the values of which can be randomly initialized and optimized in the training process by the back propagation algorithm.

[0174] In some embodiments, the multi-head dynamic graph attention neural network model in step S3, the graph convolution network GCN model and the graph attention network GAT model in step S4, and the drug-disease association prediction model based on the multi-layer perceptron MLP network in step S5 are trained and optimized based on known target drug-disease pair association relationships.

[0175] The cross-entropy loss function is defined as follows:

[0176] ;

[0177] wherein, is the true label of the association relationship between the drug node and the disease node , is the predicted association probability between the drug node and the disease node , is the total number of samples. The sample is the target drug-disease pair data with known association relationship.

[0178] In some embodiments, a regularization term is introduced to improve the generalization ability of the model, including adding a regularization term in the objective function, which is defined as:

[0179] ;

[0180] wherein, represents the regularization coefficient, denotes the learnable parameters of the kth layer of the model.

[0181] It should be understood that the above-mentioned numbers S1-S5, S301-S305, S401-S403 are only used to distinguish and facilitate the expression of each different step, and do not necessarily constitute a limitation on the execution order between each step.

[0182] In the embodiments of the present application, data is collected: experiments are performed on three data sets, respectively Fdataset, Cdataset and Lrssl. Drug information comes from the DrugBank database, and disease information comes from the OMIM database. Each data set includes drug and disease similarity matrix, drug-disease association matrix, and drug and disease id. Drug id is taken from DrugBank, and disease id is taken from OMIM. Fdataset contains 593 drugs and 313 diseases and 1933 drug-disease associations, with a sparsity of 0.0104. The details of the three data sets are shown in Table 1 as follows:

[0183]

[0184] In the embodiments of the present application, 10-fold cross-validation is used to evaluate each data set. In addition, to prevent overfitting and improve evaluation accuracy, an equal amount of negative samples is systematically included in the training and testing process to match the number of positive samples, forming a 1:1 ratio. AUROC and AUPRC are used as the main indicators for comprehensive evaluation of the overall performance of the model.

[0185] Table 2 shows the detailed comparison of the method provided in the embodiments of the present application and six benchmark models NIMCGCN, DRWBNCF, iDrug, PSGCN, DRAGNN and WIGRL on three benchmark data sets. The benchmark models include, for example, the DRWBNCF model, which uses a weighted bilinear operation to integrate drug-disease association and similarity networks; and the PSGCN model, which extracts subgraphs of drug-disease pairs and uses a hierarchical attention mechanism to capture multi-scale features. It can be seen that the method provided in the embodiments of the present application outperforms other models in various indicators. Notably, on the sparse data set Lrssl, the method provided in the embodiments of the present application achieves an AUROC of 0.9655, which is significantly higher than the second-best PSGCN of 0.942, highlighting its ability to effectively handle sparse data.

[0186]

[0187] In the embodiments of the present application, to solve the cold start problem and evaluate the ability to predict new drug indications, a leave-one-out cross-validation experiment is performed. This experiment excludes all known drug-disease associations as the test set. The associations of the five most similar drugs determined by the similarity score are used to update the drug-disease graph, while the remaining confirmed associations are used as the training set. The results are shown in Table 3 as follows: Figure 3

[0188] Embodiment Two

[0189] As shown in Table 4, the AUROC of the method provided in the embodiments of the present application on the Fdataset is 0.9237, which is significantly better than other baseline models. Figure 4 ​As shown in the figure, an embodiment of this application discloses a drug relocation system based on subgraph perception and hybrid graph neural network, including a data input layer, a subgraph extraction layer, a feature extraction layer, a multi-head dynamic graph attention neural network model layer, a fusion neural network layer of GCN model and GAT model, and a feature fusion and prediction layer.

[0190] The data input layer is used to collect and integrate data on drugs, diseases, and their relationships.

[0191] In some embodiments, the data on drugs, diseases, and their associations include chemical structure information of drugs (such as SMILES representation), phenotypic descriptions of diseases (such as MeSH terms), and association data between drugs and diseases; these data are sourced from public databases (such as DrugBank and OMIM) and are standardized into a unified structured form through a data input layer; the data input layer constructs a set of drug nodes and a set of disease nodes, and connects drug nodes and disease nodes through associations to form a drug-disease association network;

[0192] The subgraph extraction layer is used to extract data from drug-disease association networks. Extracting dynamic drug-disease association subgraphs This is to capture the local structural features and interactions of target drug-disease pairs; specifically, the subgraph extraction layer in the drug-disease association network... In this context, the focus is on the target drug-disease pair, starting from the target drug-disease pair... In the neighborhood, extract all drug nodes and disease nodes related to the target node, as well as the edges between these nodes, to generate a dynamic drug-disease association subgraph. ;

[0193] In some embodiments, the subgraph extraction layer employs a distance coding method to hierarchically label nodes based on the shortest path distance from each node to the target drug or disease node, thereby clarifying the hierarchical relationship of nodes in the subgraph.

[0194] The feature extraction layer is used to utilize similarity data between drug nodes and similarity data between disease nodes to target... Drug nodes and disease nodes are used to construct drug subgraphs (drug similarity networks). Disease subgraph (disease similarity network) ;

[0195] Specifically, the feature extraction layer can calculate the similarity between drug nodes by analyzing the chemical structure information of drugs, generating a drug similarity matrix; simultaneously, it can calculate the similarity between disease nodes using the phenotypic information of diseases, generating a disease similarity matrix; furthermore, the feature extraction layer uses... Nearest neighbor drug similarity matrix and The proximal disease similarity matrix is used to construct a drug subgraph and a disease subgraph respectively , wherein the nodes of the network are drug or disease nodes in a dynamic drug-disease association subgraph , and the edges are similarity relationships between the nodes;

[0196] The multi-head dynamic graph attention neural network model layer is used to perform deep feature extraction on the dynamic drug-disease association subgraph to obtain the final feature representation ;

[0197] The layer calculates the attention weight between the node and the neighbor node in each layer independently through the multi-head attention mechanism, aggregates the features of the neighbor nodes by weighting, generates the node representation in different semantic spaces, and uses the differentiable pooling technology to train the normalization weight, aggregates the node features by grouping, and obtains the final feature representation , so that the key local interaction characteristics and global structure information can be retained while the subgraph size is compressed;

[0198] The GCN model and GAT model fusion neural network layer is used to perform feature extraction on the network based on the graph convolution network (GCN model) and the graph attention network (GAT model) respectively, to obtain and the final feature representation and ; and ;

[0199] The structural characteristics and similarity relationships of the nodes can be extracted based on the drug subgraph and the disease subgraph ; the layer aggregates the features of the neighbor nodes through the graph convolution network (GCN model) to capture the local structure information of the drug nodes and the disease nodes; at the same time, the attention weight between the node and its neighbor nodes is calculated by using the graph attention network (GAT model) to extract the global similarity relationship; finally, the output results of the GCN model and the GAT model are fused by weighting to generate a comprehensive feature representation

[0200] The feature fusion and prediction layer is used to fuse , and the final feature representation , , to construct a unified feature representation ; and The input is a drug-disease association prediction model based on a multi-layer perception (MLP) network, and the drug-disease association prediction model outputs a prediction result of an association relationship of a target drug-disease pair.

[0201] The system realizes drug repositioning based on the method of embodiment one.

[0202] Embodiment three

[0203] The embodiment provides an electronic device, comprising a memory and a processor.

[0204] The memory is configured to store a computer program.

[0205] The processor is configured to invoke the computer program to execute the method of embodiment one.

[0206] Embodiment four

[0207] The embodiment provides a computer readable storage medium, and the computer readable storage medium stores a computer program.

[0208] Embodiment five

[0209] The embodiment provides a computer program product, comprising a computer program.

[0210] The specific implementation manners of the system, the electronic device, the computer readable storage medium and the computer program product provided by the embodiment of the application can refer to the specific embodiments of the above method, and will not be repeated here.

[0211] Obviously, those skilled in the art should understand that each unit or each step of the above application can be realized by a general computing device, which can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and can be realized by program code executable by the computing device, so that it can be stored in the storage device and executed by the computing device, or it can be made into each integrated circuit module, or multiple modules or steps can be made into a single integrated circuit module. Thus, the application is not limited to any specific combination of hardware and software.

[0212] Unless otherwise defined, all terms and expressions used in the present specification are to be interpreted under their ordinary meaning as commonly used in the art. The specific examples of the present disclosure are intended to be illustrative only and not limiting of the true scope of the present disclosure. In order to ensure that the present disclosure is clear and concise, detailed descriptions of some of the possible combinations of technical features are not provided, but it is understood that any combination of technical features is considered to be within the scope of the present disclosure, as long as such combination is logically possible.

[0213] Those skilled in the art can make various modifications, adjustments or equivalent changes to the embodiments of the present disclosure without departing from the essential spirit and technical core of the present disclosure after reading the present specification. These modifications and changes, including but not limited to the changes in the structural forms and the sequence of steps shown in the drawings, should be considered as the extension of the present disclosure, as long as the results are still within the protection scope of the claims. In addition, the appended claims not only define the protection scope of the present disclosure, but also cover the equivalent forms or reasonable extensions of these claims.

Claims

1. A drug relocation method based on subgraph perception and hybrid graph neural networks, characterized in that, include: S1, in drug-disease association networks In this context, the focus is on the target drug-disease pair, starting from the target drug-disease pair... Extract all drug nodes and disease nodes related to the target node from the neighborhood, as well as the edges between these nodes, to generate a corresponding dynamic drug-disease association subgraph. ; S2. Utilizing similarity data between drug nodes and similarity data between disease nodes, targeting... Drug subgraphs are constructed from drug nodes and disease nodes respectively. Disease subgraph ; S3, to The application uses a multi-head dynamic graph attention neural network model for deep feature extraction to obtain... Final feature representation ; S4. Based on the fusion of the Graph Convolutional Network (GCN) model and the Graph Attention Network (GAT) model, respectively... and The network performs feature extraction to obtain and Final feature representation and ; Specifically, it includes: S401. Use the GCN model to aggregate the features of the target node and its neighboring nodes; wherein the GCN model includes Layered neural networks, in which 2; Drug nodes and disease nodes In the The feature update formula for a layered neural network is: ； ； in, , Representing the first Layered neural network drug nodes and disease nodes Local structural characteristics, normalization coefficients and Defined as: ； ； in, and Representing drug nodes and drug nodes In the drug subgraph The degree in; and Representing disease nodes and disease nodes In the disease subplot The degree in; , In the GCN model, respectively Drug subgraph in layered neural networks Disease subgraph The learnable parameter matrix; Calculated from the above formula and ; S402. Use the GAT model to capture the interactions between neighboring nodes; where the GAT model includes... Layered neural networks, in which 2; Each layer of the neural network employs a multi-head attention mechanism for the drug nodes. and disease nodes Update the features; drug node and disease nodes In the GAT model The feature update formula for a layered neural network is defined as follows: ； ； and For drug nodes and disease nodes In the GAT model The updated node features of the layered neural network This refers to the number of attention heads included in the GAT model. For the first time in the GAT model The first layer of a neural network Drug nodes in individual attention points and drug nodes Attention weights between them For the first time in the GAT model The first layer of a neural network Disease nodes in individual attention points and disease nodes Attention weights between them and In the GAT model, respectively Drug subgraph in layered neural networks Disease subgraph The learnable parameter matrix; Calculated from the above formula and ; S403. Weighted fusion of the outputs of the GCN model and the GAT model to generate the final feature representation of the target node; the fusion process is defined as follows: ； ； in, and Representing drug nodes and disease nodes The final characteristic representation, The fusion ratio hyperparameter is used to adjust the relative contributions of the outputs of the GCN and GAT models. Drug subgraph The final feature representation of all drug nodes constitutes the drug subgraph. Final feature representation ; Disease subgraph The final feature representation of all disease nodes constitutes the disease subgraph. Final feature representation ; S5, will , and Final feature representation , , By merging features, a unified feature representation can be constructed. ;Will Input a drug-disease association prediction model based on a multilayer perceptron (MLP) network, and output the prediction results of the association between the target drug and the disease pair. The multi-head dynamic graph attention neural network model in step S3, the graph convolutional network (GCN) model and the graph attention network (GAT) model in step S4, and the drug-disease association prediction model based on the multilayer perceptron (MLP) network in step S5 are obtained by training and optimization based on known target drug-disease pair associations.

2. The method according to claim 1, characterized in that, In step S1, a drug-disease association network is constructed using known drug-disease associations. ,in For drug node set, Include Different drugs, For disease node set, Include Different diseases; Let be the set of edges. ,in Representative drug node and disease nodes The relationship between them, if the drugs are known With disease If there is a relationship, then the edge weights Set to 1 otherwise set to 0.

3. The method according to claim 1, characterized in that, In step S2, the drug subgraph ,in for The set of nodes in the middle contains All drug nodes in the middle, for The set of edges in the , representing Similarity data between drug nodes; disease subgraph ),in for The set of nodes in the middle contains All disease nodes in for The set of edges in the , representing Similarity data between disease nodes; In step S2, based on the chemical structure information obtained from the DrugBank database, the pairwise similarity between all drug nodes is calculated to generate a drug similarity matrix. , Middle Line 1 Column elements Indicates the first The drug node and the first Similarity between drug nodes; Using drug similarity matrix Data representing the similarity between drug nodes; In step S2, phenotypic descriptions of diseases are obtained from the OMIM disease phenotype database, and the pairwise similarity between all disease nodes is calculated. This generates a disease similarity matrix. , Middle Line 1 Column elements Indicates the first The disease node and the first Similarity between disease segments; using a disease similarity matrix. Data representing the similarity between disease nodes; Based on the similarity data between drug nodes and the similarity data between disease nodes, K-nearest neighbor similarity matrices for drugs and diseases are constructed respectively. and , Middle Line 1 Column elements and Middle Line 1 Column elements The definition is as follows: ； in, Representing the Drug nodes Nearest neighbor set, if the first If a drug node belongs to the K-nearest neighbor set, then the similarity value remains the original similarity value. Otherwise, set it to 0; ； in, Representing the Disease nodes Nearest neighbor set, if the first If a disease node belongs to the K nearest neighbor set, then the similarity value remains the original similarity value. Otherwise, set it to 0; Obtain the K-nearest neighbor similarity matrix between drugs and diseases and After the matrix is ​​formed, the two are respectively used as and The weight of the corresponding edge in the middle.

4. The method according to claim 1, characterized in that, In step S3, the multi-head dynamic graph attention neural network model includes a multi-layer neural network; Step S3 includes: S301, in the In a layered neural network, the first In each attention head, the node With nodes Attention weights are calculated as follows: ； in, , Representative node of The set of neighboring nodes in the neighborhood. and Representing nodes respectively and eigenvectors; For the first Layer 1 neural network The weight matrix of each attention head is used to map the input features to a high-dimensional latent space. For the first In a layered neural network, the first The saliency weight vector of each attention head connection feature; S302. Based on attention weights, a multi-head attention mechanism is used to perform a weighted summation of the features of the neighboring nodes of the target node. The formula is defined as follows: ； in, Represents a node In the The updated feature representation in a layered neural network. Represents the number of attention heads; S303, to Nonlinear activation is performed using the ReLU activation function, and normalization is performed using a normalization function to enhance the nonlinearity of feature representation, resulting in nodes. In the Output features of layered neural networks ; S304. Fuse the output features of the multi-layer neural network to obtain... Feature matrix The feature matrix The Behavior , ; S305. Use differentiable pooling to learn trainable normalized weights, group and aggregate node features to reduce the size of the drug-disease subgraph, and gradually expand the receptive field of the nodes. Feature matrix and adjacency matrix Map to a smaller subgraph using the following transformation: ； ； in, It is a learnable normalized weight matrix used to distribute raw nodes among fewer supernodes. The dimension is , The original number of nodes. The number of supernodes after pooling, where elements Represents the original node Assigned to supernode The weights; for The adjacency matrix, where the elements are taken as... Weight of the corresponding edge ; The feature matrix representing the supernode, It is the adjacency matrix after pooling; and constitute Final feature representation .

5. The method according to claim 1, characterized in that, The Input a drug-disease association prediction model based on a multilayer perceptron (MLP) network, and output the prediction results of the association between the target drug and disease pair, including: The output layer of the drug-disease association prediction model predicts drug nodes using the following formula. and disease nodes The probability of association between them: ； in, , , and These are the weight and coefficient parameters.

6. A drug relocation system based on subgraph perception and hybrid graph neural networks, characterized in that, It includes a data input layer, a subgraph extraction layer, a feature extraction layer, a multi-head dynamic graph attention neural network model layer, a fusion neural network layer of GCN and GAT models, and a feature fusion and prediction layer; The data input layer is used to collect and integrate data on drugs, diseases, and their relationships to form a drug-disease association network. ; The subgraph extraction layer is used in drug-disease association networks. In this context, the focus is on the target drug-disease pair, starting from the target drug-disease pair... In the neighborhood, extract all drug nodes and disease nodes related to the target node, as well as the edges between these nodes, to generate a dynamic drug-disease association subgraph. ; The feature extraction layer is used to utilize similarity data between drug nodes and similarity data between disease nodes to target... Drug subgraphs are constructed from drug nodes and disease nodes respectively. Disease subgraph ; The multi-head dynamic graph attention neural network model layer is used for dynamic drug-disease association subgraphs. Perform deep feature extraction to obtain Final feature representation ; The aforementioned GCN and GAT model fusion neural network layer is used to fuse the graph convolutional network GCN model and the graph attention network GAT model respectively. and The network performs feature extraction to obtain and Final feature representation and ; The feature fusion and prediction layer is used to... , and Final feature representation , , By merging features, a unified feature representation can be constructed. ;Will Input a drug-disease association prediction model based on a multilayer perceptron (MLP) network, and output the prediction results of the association between the target drug and the disease pair. The system achieves drug repositioning based on the method described in any one of claims 1 to 5.

7. An electronic device, characterized in that, include: Memory and processor; The memory is used to store computer programs; The processor is configured to invoke the computer program to perform the method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed on an electronic device, causes the electronic device to perform the method as described in any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is run on an electronic device, the electronic device causes the electronic device to perform the method as described in any one of claims 1 to 5.

Images