Adaptive drug repositioning method based on feature-driven isomorphic graph learning

Abstract

The application discloses an adaptive drug repositioning method based on feature-driven isomorphic graph learning, relates to the technical fields of drug repositioning and graph neural networks, and comprises the following steps: acquiring drug-disease association data and corresponding graph network data information; initializing feature embedding of disease nodes and drug nodes according to the drug-disease association data and the corresponding graph network data information, and constructing a heterogeneous graph and an isomorphic graph; performing a heterogeneous graph convolution operation through disease-to-drug direction and drug-to-disease direction propagation, and calculating the importance weight between the disease nodes and the drug nodes by using an attention mechanism to generate heterogeneous information; adaptively weighting the edges in the isomorphic graph through a learnable parameter mask, controlling the binarization degree of the weight through L1 regularization and entropy regularization, performing an isomorphic graph convolution operation, and generating isomorphic information. The application effectively reduces the noise interference in the pre-defined graph by adaptively weighting the edges in the isomorphic graph through the learnable parameter mask.

Description

Technical Field

[0001] This invention relates to the fields of drug relocation and graph neural network technology, and in particular to an adaptive drug relocation method based on feature-driven isomorphic graph learning. Background Technology

[0002] Drug repositioning, as a highly efficient drug development strategy, significantly reduces the time and economic costs of new drug development by exploring new therapeutic uses for existing drugs. Traditional drug development is time-consuming, expensive, and has a low success rate; while drug repositioning, by utilizing the safety and pharmacokinetic data of existing drugs, rapidly explores their potential in new indications and has become a hot topic in biomedical research. For example, in public health emergencies, drugs such as remdesivir were rapidly used for antiviral treatment through repositioning strategies, highlighting their significant value. Early drug repositioning methods were mainly based on computational models such as molecular docking, signaling pathway analysis, or side effect similarity; however, these methods often rely on shallow features and struggle to capture the complex nonlinear relationships between drugs and diseases, resulting in limited predictive accuracy.

[0003] With the development of deep learning technology, graph neural networks (GNNs) have been introduced into the field of drug relocation, providing a new approach to modeling multi-relationship data. GNNs can naturally represent drugs and diseases as nodes in a graph and characterize their interactions through edge connections. However, these existing GNN-based drug relocation methods still have shortcomings. On the one hand, many methods directly utilize predefined isomorphic graphs, which often contain a large number of task-irrelevant noisy relationships and lack adaptive learning mechanisms. The model cannot dynamically adjust the importance of edges during training, making it difficult to focus on task-relevant key connections. On the other hand, existing methods typically employ simple concatenation or weighted summation strategies when handling the fusion of heterogeneous and isomorphic graphs, failing to fully learn feature-driven adaptive relationships. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an adaptive drug relocation method based on feature-driven isomorphic graph learning to solve the problems of isomorphic graph noise interference and coarse heteromorphic-isomorphic fusion.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an adaptive drug relocation method based on feature-driven isomorphic graph learning, comprising: acquiring drug-disease association data and corresponding graph network data information; initializing feature embeddings of disease nodes and drug nodes based on the drug-disease association data and corresponding graph network data information, and constructing heterogeneous and isomorphic graphs; performing heterogeneous graph convolution operations through disease-to-drug and drug-to-disease propagation, and calculating importance weights between disease nodes and drug nodes using an attention mechanism to generate heterogeneous information; and adaptively weighting the edges in the isomorphic graph using a learnable parameter mask. The binarization of weights is controlled by L1 regularization and entropy regularization, and isomorphic graph convolution is performed to generate isomorphic information. A dual-path MLP is used to process heterogeneous and isomorphic information respectively to generate heterogeneous path features and isomorphic path features. Deep feature learning is performed on the heterogeneous and isomorphic path features to generate heterogeneous path optimization features and isomorphic path optimization features. The heterogeneous path optimization features and isomorphic path optimization features are fused, and the predicted probability of drug-disease association is generated through a projection layer. Multi-objective joint optimization training is performed through a multivariate loss function to output the prediction result of drug relocation.

[0007] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the drug-disease association data includes a disease node set, a drug node set, a disease-disease similarity matrix, a drug-drug similarity matrix, and a disease-drug association matrix. The disease node set is defined as ,in The number of disease nodes, Indicates disease nodes; The drug node set is defined as ,in Number of drug nodes Indicates drug nodes; The disease-disease similarity matrix is ​​defined as follows: ; The drug-drug similarity matrix is ​​defined as follows: ; The disease-drug association matrix is ​​defined as follows: ,in Indicates known associated labels, Indicates an unknown associated tag.

[0008] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the specific steps for initializing the feature embeddings of disease nodes and drug nodes are as follows: For any disease node in the disease node set It is initialized through Embedding, with a set of disease nodes as input and the initial disease embedding as output. The initial disease embeddings of all disease nodes are integrated to form a disease node embedding matrix; For any drug node in the drug node set Initialization is performed using Embedding to generate the initial drug embedding; The initial drug embeddings of all drug nodes are integrated to form a drug node embedding matrix.

[0009] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the specific steps for constructing the heteromorphic and isomorphic graphs are as follows: Heterogeneous graphs are constructed by defining disease-drug bipartite graphs and drug-disease bipartite graphs; The disease-drug bipartite graph is defined as follows: ,in Let the disease-drug edge set represent known associations; , Represents the disease node index. Represents the drug node index; The drug-disease bipartite graph is defined as follows: ,in For drug-disease border set, ; in for The inverse graph represents the connection between diseases and drugs; Construct isomorphic graphs by defining disease-disease similarity graphs and drug-drug similarity graphs; The disease-disease similarity graph is defined as follows: ,in For disease-disease boundary sets, This indicates a similarity between diseases; The drug-drug similarity map is defined as follows: ,in For drug-drug boundary sets, This indicates the similarity between drugs.

[0010] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the method involves performing heterogeneous graph convolution operations through disease-to-drug and drug-to-disease propagation, and utilizing an attention mechanism to calculate the importance weights between disease nodes and drug nodes to generate heterogeneous information. The specific steps are as follows: The attention weights from disease to drug and from drug to disease are calculated using the attention weight calculation formula. For the propagation from disease to drug, the drug node features are updated by the attention weights from disease to drug. For the spread of drugs to diseases, the disease node features are updated by the attention weights of drugs to diseases. Perform convolution operations on all drug node features in the disease-to-drug direction to generate disease-drug node features; Perform convolution operations on all disease node features from drug to disease direction to generate drug-disease node features, and treat the disease-drug node features and drug-disease node features as heterogeneous information.

[0011] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the steps are as follows: Edges in the isomorphic graph are adaptively weighted using a learnable parameter mask, and the binarization degree of the weights is controlled by L1 regularization and entropy regularization. An isomorphic graph convolution operation is then performed to generate isomorphic information. Learnable mask weights for any edge in an isomorphic graph are computed using MLP. Weighted GCN is used to perform convolution operations on disease-disease similarity maps and drug-disease similarity maps to generate disease-disease node features and drug-disease node features. Disease-disease node features and drug-drug node features are used as isomorphic information.

[0012] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the method employs a dual-path MLP to process heterogeneous and isomorphic information respectively, generating heterogeneous path features and isomorphic path features. The specific steps are as follows: Extracting disease-drug node features and drug-drug node features from heterogeneous information in the disease-drug direction; Heterogeneous path features are generated by multiplying the disease-drug node features and the drug-disease node features element-wise. Extract disease-disease node features of disease nodes and drug-drug node features of drug nodes from isomorphic information; The disease-disease node features and drug-drug node features are multiplied element-wise to generate isomorphic path features.

[0013] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the specific steps for deepening feature learning on heterogeneous path features and isomorphic path features to generate heterogeneous path optimization features and isomorphic path optimization features are as follows. Deeper feature learning is performed on heterogeneous path features and homogeneous path features using MLP. The expression for deeper feature learning of heterogeneous path features is as follows: ; in, Indicates heterogeneous path optimization characteristics; This represents the heterogeneous weights of the MLP; Indicates the heterogeneous bias of the MLP; This represents a multilayer perceptron for processing heterogeneous information. Indicates heterogeneous path characteristics; The expression for deepening feature learning of isomorphic path features is as follows: ; in, This indicates the characteristics of isomorphic path optimization. Indicates the isomorphic weights of the MLP; Indicates the isomorphic bias of the MLP; This represents a multilayer perceptron that processes isomorphic information. This represents isomorphic path characteristics.

[0014] As a preferred embodiment of the feature-driven isomorphic graph learning-based adaptive drug relocation method of the present invention, the steps of fusing heterogeneous path optimization features and isomorphic path optimization features, and generating predicted probabilities of drug-disease associations through a projection layer, are as follows: The heterogeneous path optimization features and the homogeneous path optimization features are concatenated along the node dimension to generate a fused feature; The fused features are reduced to a low-dimensional space by a linear projection layer to generate projected dimensionality-reduced features. The projected dimensionality-reduced features are passed through a prediction layer to output the predicted probability of the drug-disease association, expressed as follows: ; in, This indicates the predicted probability of a drug-disease association. This represents the weight vector of the prediction layer; Indicates the bias of the prediction layer; Represents the sigmoid function; Represents the dimensionality reduction features of projection; The drug relocation prediction model is composed of heterogeneous graph convolution, isogeneous graph convolution, dual-path MLP, projection layer, and prediction layer.

[0015] As a preferred embodiment of the adaptive drug relocation method based on feature-driven isomorphic graph learning described in this invention, the steps of performing multi-objective joint optimization training through a multivariate loss function to output the prediction result of drug relocation are as follows: The multivariate loss function includes binary cross-entropy loss, L1 regularization term, entropy regularization term and feature fusion MSE loss; The total loss function is defined as: ; in, Indicates the total loss; Indicates the BCE loss weight; This represents the binary cross-entropy loss; Indicates the weight of L1 regularization; This represents the loss due to the L1 regularization term; The weights representing entropy regularization; This represents the loss due to entropy regularization. The weights representing the fusion loss; Indicates fusion loss; The parameters of the drug repositioning prediction model are optimized using the backpropagation algorithm to minimize the total loss. Once the drug repositioning prediction model has converged during training, the parameters of the drug repositioning prediction model are saved. For the drug-disease pair to be predicted Drug-disease response The features are input into the trained drug relocation prediction model, and the prediction results of drug relocation are generated through forward propagation.

[0016] The beneficial effects of this invention are as follows: By introducing a learnable parameter mask mechanism to adaptively weight edges in a isomorphic graph, noise interference in the predefined graph is effectively reduced, solving the problems of not being able to dynamically adjust edge importance and difficulty in focusing on key connections of the task; by deepening feature learning of heterogeneous and isomorphic information through dual-path MLP, the expressive power of the network is enhanced; by combining multivariate loss functions for multi-objective joint optimization, both prediction accuracy and network sparsity are taken into account; the method of this invention can adaptively learn key relationships in graph structure, improve the accuracy and robustness of drug relocation, and is efficient and easy to implement, making it suitable for large-scale drug-disease association prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of an adaptive drug relocation method based on feature-driven isomorphism learning.

[0019] Figure 2 This is a flowchart for graph-based integration.

[0020] Figure 3 This is a flowchart of feature processing based on dual-path MLP.

[0021] Figure 4 This is a flowchart based on the output drug repositioning prediction results. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides an adaptive drug relocation method based on feature-driven isomorphism graph learning, comprising the following steps: S1. Obtain drug-disease association data and corresponding graph network data information.

[0026] Drug-disease association data includes disease node sets, drug node sets, disease-disease similarity matrices, drug-drug similarity matrices, and disease-drug association matrices.

[0027] The disease node set is defined as ,in The number of disease nodes, Indicates a disease node.

[0028] The drug node set is defined as ,in Number of drug nodes This indicates a drug node.

[0029] Disease-disease similarity matrix is ​​defined as .

[0030] The drug-drug similarity matrix is ​​defined as follows: .

[0031] The disease-drug association matrix is ​​defined as follows: ,in Indicates known associated labels, Indicates an unknown associated tag.

[0032] S2. Based on the drug-disease association data and the corresponding graph network data information, initialize the feature embedding of disease nodes and drug nodes, and construct heterogeneous graphs and isogeneous graphs.

[0033] For any disease node in the disease node set It is initialized using Embedding, with a set of disease nodes as input and an initial disease embedding as output. ,in Indicates the first Embedded disease nodes; This represents the disease node index.

[0034] The initial disease embeddings of all disease nodes are integrated to form a disease node embedding matrix, defined as follows: ,in This is a transpose operation.

[0035] For any drug node in the drug node set Initialization is performed using Embedding to generate the initial drug embedding. ,in Indicates the first Embedded drug nodes; This represents the index of the drug node.

[0036] The initial drug embeddings of all drug nodes are integrated to form a drug node embedding matrix, defined as follows: ,in This is a transpose operation.

[0037] Heterogeneous graphs are constructed by defining disease-drug bipartite graphs and drug-disease bipartite graphs, where the disease-drug bipartite graph is defined as follows: , Let the disease-drug boundary set represent known associations. This indicates the connection between drugs and diseases.

[0038] The drug-disease bipartite diagram is defined as follows: ,in For drug-disease border set, ;in for The inverse graph represents the connection between diseases and drugs.

[0039] An isomorphic graph is constructed by defining disease-disease similarity graphs and drug-disease similarity graphs, where a disease-disease similarity graph is defined as... , For disease-disease boundary sets, This indicates a similarity between diseases.

[0040] Drug-drug similarity graph is defined as ,in For drug-drug boundary sets, This indicates the similarity between drugs.

[0041] S3. Perform heterogeneous graph convolution operations through the propagation from disease to drug and from drug to disease, and use the attention mechanism to calculate the importance weights between disease nodes and drug nodes to generate heterogeneous information.

[0042] Define disease To the medicine The set of edges is ,drug To disease The set of edges is .

[0043] For the transmission of disease towards medication, the attention weight in the disease-to-medicine direction is calculated using the attention weight calculation formula, the expression of which is: ; in, This represents the attention weighting from disease to medication; This is the attention vector; yes DP The weight matrix of the orientation is used for feature transformation; Indicates a splicing operation; The activation function is set to a slope of 0.2. Represents the neighbors of the drug node; Indicates the transpose operation; This indicates the initial drug embedding.

[0044] The drug node features are updated using attention weights from disease to drug direction, expressed as follows: ; in, Indicates drug In the Features of the layer; Indicates the current layer; This represents the attention weighting from disease to medication; The feature transformation matrix representing the disease-to-drug direction; This indicates the initial disease embedding.

[0045] For the propagation of drugs towards the disease, the attention weight for drug-disease propagation is calculated, expressed as follows: ; in, This represents the attention weighting from drug administration to disease management. This is the attention vector; Feature transformation matrix from drug to disease direction; Represents the neighbors of the disease node; This indicates the transpose operation.

[0046] The disease node features are updated using attention weights from drug to disease direction, expressed as follows: ; in, Indicates disease In the Characteristics of the layer; This indicates the attention weighting of drugs to diseases; The feature transformation matrix representing the direction from drug to disease; Indicates drug Its characteristics.

[0047] Perform convolution operations on all drug node features in the disease-to-drug direction to generate disease-drug node features. , This indicates a hidden dimension.

[0048] Perform convolution operations on all disease node features from drug to disease direction to generate drug-disease node features. Disease drug node characteristics Drug-related disease node characteristics As heterogeneous information.

[0049] S4. By adaptively weighting the edges in the isomorphic graph using the learnable parameter mask, and controlling the binarization of the weights through L1 regularization and entropy regularization, the isomorphic graph convolution operation is performed to generate isomorphic information.

[0050] Calculate any edge in an isomorphic graph using MLP. Learnable mask weights It is the source node. It is the target node, and its expression is: ; in, This represents learnable mask weights; and Representing arbitrary edges Feature embedding of source and target nodes; This represents a multilayer perceptron, using independent methods for disease-disease edges. For drug-drug edge, use ; Representing temperature parameters, through temperature parameters Control the sharpness of the weights; This represents the sigmoid function.

[0051] Using a weighted GCN to perform convolution operations on the disease-disease similarity map, the convolution formula for generating disease-disease node features is as follows: ; in, Representing disease-disease node features; This represents a weighted adjacency matrix with added self-loops; express The degree matrix; Represents the GCN weight matrix; Indicates the activation function; The embedding matrix represents the disease nodes.

[0052] Weighted GCN is used to perform convolution operations on the drug-drug similarity graph to generate drug-drug node features. Disease-disease node features Drug-drug node features As isomorphic information.

[0053] S5. Use dual-path MLP to process heterogeneous and homogeneous information respectively to generate heterogeneous path features and homogeneous path features; perform in-depth feature learning on heterogeneous path features and homogeneous path features to generate heterogeneous path optimization features and homogeneous path optimization features.

[0054] Extracting drug nodes from heterogeneous information Characteristics in the disease-drug direction and disease nodes Characteristics in the drug-disease direction ;Will and Heterogeneous path features are generated by element-wise multiplication. .

[0055] Extracting disease nodes from isomorphic information Features and drug nodes Features ;Will and Generate isomorphic path features by multiplying element by element. .

[0056] Deeper feature learning is performed on heterogeneous path features and homogeneous path features using MLP. The expression for deeper feature learning of heterogeneous path features is as follows: ; in, Indicates heterogeneous path optimization characteristics; This represents the heterogeneous weights of the MLP; Indicates the heterogeneous bias of the MLP; A multilayer perceptron for processing heterogeneous information.

[0057] The expression for deepening feature learning of isomorphic path features is as follows: ; in, This indicates the characteristics of isomorphic path optimization. Indicates the isomorphic weights of the MLP; Indicates the isomorphic bias of the MLP; This represents a multilayer perceptron that processes isomorphic information.

[0058] S6. It integrates heterogeneous path optimization features and homogeneous path optimization features, and generates the predicted probability of drug-disease association through the projection layer.

[0059] Heterogeneous path optimization features and homogeneous path optimization features are concatenated along the node dimension to generate , This indicates the fusion feature.

[0060] Through the linear projection layer Dimensionality reduction to a lower-dimensional space, generating projected dimensionality-reduced features. The expression is, ; in, Represents the dimensionality reduction features of projection; This represents the weight vector of the projection layer; Indicates the bias of the projection layer; This represents the activation function.

[0061] Projection dimensionality reduction features The prediction layer outputs the predicted probability of the drug-disease association, expressed as follows: ; in, This indicates the predicted probability of a drug-disease association. This represents the weight vector of the prediction layer; Indicates the bias of the prediction layer; This represents the sigmoid function.

[0062] The drug relocation prediction model is composed of heterogeneous graph convolution, isogeneous graph convolution, dual-path MLP, projection layer, and prediction layer.

[0063] S7. Perform multi-objective joint optimization training using a multivariate loss function to output the prediction results of drug repositioning.

[0064] Multivariate loss functions include binary cross-entropy loss, L1 regularization term, entropy regularization term, and feature fusion MSE loss.

[0065] The expression for the total loss function is: ; in, Indicates the total loss; Indicates the BCE loss weight; This represents the binary cross-entropy loss; Indicates the weight of L1 regularization; This represents the loss due to the L1 regularization term; The weights representing entropy regularization; This represents the loss due to entropy regularization. The weights representing the fusion loss; This indicates the loss of fusion.

[0066] The parameters of the drug repositioning prediction model are optimized using the backpropagation algorithm to minimize the total loss. Once the drug repositioning prediction model has converged during training, its parameters are saved.

[0067] For the drug-disease pair to be predicted Drug-disease response The features are input into the trained drug relocation prediction model, and the prediction results of drug relocation are generated through forward propagation.

[0068] The expression for the binary cross-entropy loss is, ; in, For the first The true label of each sample; For the first The predicted probability of each sample; The number of samples; Indicates the sample index; the value range is 1 to... .

[0069] The expression for the L1 regularization term is, ; in, This represents the loss due to the L1 regularization term; and These are the weights of disease-disease and drug-drug edges, respectively. and Let the number of edges be denoted as '('). and Let it be the set of edges.

[0070] The expression for the entropy regularization term is, ; in, This represents the loss due to the entropy regularization term.

[0071] The expression for the feature fusion MSE loss is as follows: ; in, Indicates fusion loss; and Representing heterogeneity and isogeneity respectively The path feature vector of each sample.

[0072] In summary, this invention addresses the problems of dynamically adjusting edge importance and focusing on key connections in a predefined graph by introducing a learnable parameter mask mechanism to adaptively weight edges in the isomorphic graph. This effectively reduces noise interference in the predefined graph. Furthermore, it enhances the network's expressive power by deepening feature learning of both heterogeneous and isomorphic information through a dual-path MLP. Finally, it balances prediction accuracy with network sparsity by combining a multivariate loss function for multi-objective joint optimization. This method adaptively learns key relationships in the graph structure, improving the accuracy and robustness of drug relocation. It is also highly efficient and easy to implement, making it suitable for large-scale drug-disease association prediction.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive drug relocation method based on feature-driven isomorphic graph learning, characterized in that: include, Acquire drug-disease association data and corresponding graph network data information; Based on drug-disease association data and corresponding graph network data, the feature embeddings of disease nodes and drug nodes are initialized, and heterogeneous and homogeneous graphs are constructed. Heterogeneous graph convolution operations are performed through disease-to-drug and drug-to-disease propagation, and attention mechanisms are used to calculate the importance weights between disease nodes and drug nodes to generate heterogeneous information. By adaptively weighting the edges in the isomorphic graph using a learnable parameter mask, and controlling the binarization of the weights through L1 regularization and entropy regularization, isomorphic graph convolution operations are performed to generate isomorphic information. A dual-path MLP is used to process heterogeneous and homogeneous information respectively, generating heterogeneous path features and homogeneous path features; Deepen feature learning on heterogeneous path features and homogeneous path features to generate heterogeneous path optimization features and homogeneous path optimization features; The heterogeneous path optimization features and homogeneous path optimization features are integrated, and the predicted probability of drug-disease association is generated through the projection layer. Multi-objective joint optimization training is performed using a multivariate loss function to output the prediction results of drug relocation.

2. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 1, characterized in that: The drug-disease association data includes a set of disease nodes, a set of drug nodes, a disease-disease similarity matrix, a drug-drug similarity matrix, and a disease-drug association matrix; The disease node set is defined as ,in The number of disease nodes, Indicates disease nodes; The drug node set is defined as ,in Number of drug nodes Indicates drug nodes; The disease-disease similarity matrix is ​​defined as follows: ; The drug-drug similarity matrix is ​​defined as follows: ; The disease-drug association matrix is ​​defined as follows: ,in Indicates known associated labels, Indicates an unknown associated tag.

3. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 2, characterized in that: The specific steps for initializing the feature embedding of disease nodes and drug nodes are as follows. For any disease node in the disease node set It is initialized through Embedding, with a set of disease nodes as input and the initial disease embedding as output. The initial disease embeddings of all disease nodes are integrated to form a disease node embedding matrix; For any drug node in the drug node set Initialization is performed using Embedding to generate the initial drug embedding; The initial drug embeddings of all drug nodes are integrated to form a drug node embedding matrix.

4. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 3, characterized in that: The specific steps for constructing heterogeneous and homogeneous graphs are as follows. Heterogeneous graphs are constructed by defining disease-drug bipartite graphs and drug-disease bipartite graphs; The disease-drug bipartite graph is defined as follows: ,in Let the disease-drug edge set represent known associations; , Represents the disease node index. Represents the drug node index; The drug-disease bipartite graph is defined as follows: ,in For drug-disease border set, ; in for The inverse graph represents the connection between diseases and drugs; Construct isomorphic graphs by defining disease-disease similarity graphs and drug-drug similarity graphs; The disease-disease similarity graph is defined as follows: ,in For disease-disease boundary sets, This indicates a similarity between diseases; The drug-drug similarity map is defined as follows: ,in For drug-drug boundary sets, This indicates the similarity between drugs.

5. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 4, characterized in that: The process involves performing heterogeneous graph convolution operations along the disease-to-drug and drug-to-disease propagation paths, and utilizing an attention mechanism to calculate the importance weights between disease nodes and drug nodes to generate heterogeneous information. The specific steps are as follows: The attention weights from disease to drug and from drug to disease are calculated using the attention weight calculation formula. For the propagation from disease to drug, the drug node features are updated by the attention weights from disease to drug. For the spread of drugs to diseases, the disease node features are updated by the attention weights of drugs to diseases. Perform convolution operations on all drug node features in the disease-to-drug direction to generate disease-drug node features; Perform convolution operations on all disease node features from drug to disease direction to generate drug-disease node features, and treat the disease-drug node features and drug-disease node features as heterogeneous information.

6. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 5, characterized in that: The method involves adaptively weighting the edges in the isomorphic graph using a learnable parameter mask, controlling the binarization of the weights through L1 regularization and entropy regularization, and then performing an isomorphic graph convolution operation to generate isomorphic information. The specific steps are as follows: Learnable mask weights for any edge in an isomorphic graph are computed using MLP. Weighted GCN is used to perform convolution operations on disease-disease similarity maps and drug-disease similarity maps to generate disease-disease node features and drug-disease node features. Disease-disease node features and drug-drug node features are used as isomorphic information.

7. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 6, characterized in that: The method employs a dual-path MLP to process heterogeneous and homogeneous information separately, generating heterogeneous path features and homogeneous path features. The specific steps are as follows. Extracting disease-drug node features and drug-drug node features from heterogeneous information in the disease-drug direction; Heterogeneous path features are generated by multiplying the disease-drug node features and the drug-disease node features element-wise. Extract disease-disease node features of disease nodes and drug-drug node features of drug nodes from isomorphic information; The disease-disease node features and drug-drug node features are multiplied element-wise to generate isomorphic path features.

8. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 7, characterized in that: The specific steps for deepening feature learning on heterogeneous and homogeneous path features to generate heterogeneous path optimization features and homogeneous path optimization features are as follows. Deeper feature learning is performed on heterogeneous path features and homogeneous path features using MLP. The expression for deeper feature learning of heterogeneous path features is as follows: ； in, Indicates heterogeneous path optimization characteristics; This represents the heterogeneous weights of the MLP; Indicates the heterogeneous bias of the MLP; This represents a multilayer perceptron for processing heterogeneous information. Indicates heterogeneous path characteristics; The expression for deepening feature learning of isomorphic path features is as follows: ； in, This indicates the characteristics of isomorphic path optimization. Indicates the isomorphic weights of the MLP; Indicates the isomorphic bias of the MLP; This represents a multilayer perceptron that processes isomorphic information. This represents isomorphic path characteristics.

9. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 8, characterized in that: The process involves fusing heterogeneous path optimization features and homogeneous path optimization features, and generating predicted probabilities of drug-disease associations through a projection layer. The specific steps are as follows: The heterogeneous path optimization features and the homogeneous path optimization features are concatenated along the node dimension to generate a fused feature; The fused features are reduced to a low-dimensional space by a linear projection layer to generate projected dimensionality-reduced features. The projected dimensionality-reduced features are passed through a prediction layer to output the predicted probability of the drug-disease association, expressed as follows: ； in, This indicates the predicted probability of a drug-disease association. This represents the weight vector of the prediction layer; Indicates the bias of the prediction layer; Represents the sigmoid function; Represents the dimensionality reduction features of projection; The drug relocation prediction model is composed of heterogeneous graph convolution, isogeneous graph convolution, dual-path MLP, projection layer, and prediction layer.

10. The adaptive drug relocation method based on feature-driven isomorphic graph learning as described in claim 9, characterized in that: The process involves multi-objective joint optimization training using a multivariate loss function to output the predicted drug relocation results. The specific steps are as follows: The multivariate loss function includes binary cross-entropy loss, L1 regularization term, entropy regularization term, and feature fusion MSE loss; The total loss function is defined as: ; in, Indicates the total loss; Indicates the BCE loss weight; This represents the binary cross-entropy loss; Indicates the weight of L1 regularization; This represents the loss due to the L1 regularization term; The weights representing entropy regularization; This represents the loss due to entropy regularization. The weights representing the fusion loss; Indicates fusion loss; The parameters of the drug repositioning prediction model are optimized using the backpropagation algorithm to minimize the total loss. Once the drug repositioning prediction model has converged during training, the parameters of the drug repositioning prediction model are saved. For the drug-disease pair to be predicted Drug-disease response The features are input into the trained drug relocation prediction model, and the prediction results of drug relocation are generated through forward propagation.

Images