Drug relocation method based on multi-modal diagram learning

Abstract

The invention discloses a drug relocation method based on multi-modal diagram learning, and belongs to the technical field of drug research and development. The implementation method of the invention comprises the following steps: 1, constructing a multi-modal drug relocation network consisting of sub-networks with the same number of input modals, and enabling the sub-networks to independently process single-modal data; 2, performing characterization learning and characterization fusion on the sub-networks in the multi-modal drug relocation network; 3, carrying out adaptive graph structure learning on the multi-modal drug relocation network representing fusion; 4, loss optimization is carried out on the multi-mode drug relocation network of adaptive learning, and then a drug relocation prediction result is obtained; compared with the prior art, the method has the advantages that representation learning and self-adaptive learning are carried out through similarity and incidence matrix calculation and by using an attention mechanism and graph neural network learning high-level features on the basis of the sub-network constructed on the basis of multi-modal data, and the generalization ability of drug relocation downstream tasks of multi-modal graph learning is improved.

Description

Technical Field

[0001] The present invention relates to a drug repositioning method based on multimodal graph learning, and belongs to the technical field of drug research and development. Background Art

[0002] Drug repositioning, also known as "repurposing old drugs," typically refers to the repositioning of approved drugs for new indications. Drug repositioning can accelerate the drug development process, reduce costs, and provide treatment options for unmet medical needs. Computational drug repositioning reduces drug-disease interactions by providing candidate molecules or drugs for wet lab testing, resulting in lower overall R&D costs and shorter development cycles, boosting industry R&D enthusiasm. Computational drug repositioning methods, based on known drug-disease association data, design appropriate algorithms based on features to predict potential drug-disease interactions, accelerating drug space search and reducing drug R&D costs. They enable large-scale exploration of potential drug-disease associations and improve prediction accuracy through multi-source data integration, enabling drug repositioning at a lower cost and in a shorter timeframe. The inverse molecular docking method INVDOCK performs target prediction and can be integrated with other technologies for back-end evaluation of candidate compounds identified earlier using other methods. A similarity integration method, based on ligand similarity encoded by extended connectivity fingerprints (ECFPs), quantitatively groups related protein targets. Leveraging target-specific features, a method based on protein-ligand interaction fingerprints has been developed. For ligand-based methods, the prediction accuracy is often poor because few ligands bind to known target proteins.

[0003] Although research in computational drug retargeting has achieved initial success both domestically and internationally, it still lacks depth overall. The collection and mining of information about drugs, diseases, and so on is inadequate, and models lack sufficient inference capabilities based on a single graph or a simple collection of graphs. The application of simple vector concatenation and intra-modal attention mechanisms cannot fully learn the relationships between multimodal, multi-graph data. Previous studies have manually set the similarity matrix metrics for single and multi-graphs used in multi-stage frameworks, requiring careful adjustment before implementation. Consequently, generalization performance in downstream tasks is questionable. Furthermore, in multi-stage models, the training of submodules such as representation learning, graph construction, and prediction is independent of each other, weakening the integrity and expressiveness of the entire model framework.

[0004] Therefore, how to improve the generalization ability of multimodal graph learning for drug relocalization downstream tasks has become an urgent problem to be solved. Summary of the Invention

[0005] The purpose of this paper is to solve the technical problem of how to improve the generalization ability of multimodal graph learning for downstream drug relocalization tasks. We propose a drug relocalization method based on multimodal graph learning. This method constructs an adaptive graph learning matrix and jointly optimizes it with the downstream graph neural network (GNN), thereby improving the generalization ability of downstream tasks.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A drug repositioning method based on multimodal graph learning of the present invention comprises the following steps:

[0008] Step 1: Construct a multimodal drug relocalization network consisting of subnetworks with equal number of input modalities, so that the subnetworks can independently process single modality data;

[0009] Step 1.1: Use drug data and disease data to construct a drug-disease association subnetwork;

[0010] Step 1.2: Construct drug interaction subnetwork using drug interaction data;

[0011] Step 1.3: Construct a drug-target interaction subnetwork using drug-target interaction data;

[0012] Step 1.4: Using the side effect data between drugs, the similarity of side effects between drugs is obtained using the method shown in formula (1), and a drug side effect subnetwork screened by the side effect similarity is constructed;

[0013]

[0014] Where A and B represent drug A and drug B respectively, N A is the number of side effects associated with drug A; N B is the number of side effects associated with drug B; N AB is the number of side effects shared by drug A and drug B, and the range of Tanimoto coefficient is [0,1];

[0015] Step 1.5: Using medicinal chemistry data, obtain the drug molecule fingerprint similarity using the method shown in formula (2), and construct a medicinal chemistry subnetwork screened by the drug molecule fingerprint similarity;

[0016]

[0017] Where A and B represent drug A and drug B respectively, N′ A and N′ B Represents the binary bit strings of the MACCS fragments of the two drug molecules, N′ AB is a binary bit string that both have on the fragment;

[0018] Step 1.6: Using drug data and clinical data, construct a drug clinical subnetwork that has been screened for drug clinical similarity;

[0019] Step 1.6.1: Obtain the single-layer clinical similarity between drugs using the method shown in formula (3);

[0020]

[0021] Where k represents the kth layer of ATC coding, S k represents the k-level clinical similarity between drug A and drug B;

[0022] Step 1.6.2: Obtain the single-layer clinical similarity between drugs using the method shown in formula (4);

[0023]

[0024] Among them, ATC k Represents all ATC codes on layer k, S atc (A, B) represents the clinical similarity between drug A and drug B, and n represents the number of coding levels in the ATC classification system;

[0025] Step 1.7: Using the drug target sequence data, obtain the drug target sequence similarity using the method shown in formula (5), and construct a drug target sequence subnetwork that has been screened by target sequence similarity;

[0026]

[0027] Among them, targets a and b correspond to the targets of drugs A and B respectively. When a≠b, the overall sequence similarity S is obtained by averaging the protein sequence similarities of targets a (a∈A) and b (b∈B). p ;

[0028] Step 1.8: Using the drug target gene data, obtain the drug target gene similarity using the method shown in formula (6), and construct a drug target gene sub-network screened by target gene similarity;

[0029]

[0030] Among them, genes a and b correspond to the target genes of drugs A and B respectively. The gene ontology similarity S is obtained by averaging the semantic similarity of the gene pairs a (a∈A) and b (b∈B) using the graph. GO ;

[0031] Step 2: Perform representation learning and representation fusion on the sub-networks in the multimodal drug relocalization network;

[0032] Step 2.1: Representation learning of the sub-networks in the multimodal drug relocalization network;

[0033] Step 2.1.1: Use the scaled dot product attention method for capturing and learning inter-modal shared information as shown in Equation (7) to learn inter-modal representations for the sub-networks in the multimodal drug relocalization network;

[0034]

[0035] Where, P ij Indicates the degree of attention of the i-th modality to the j-th modality, τ is the scaling factor, is the transposition operator, exp is the exponential function, q i is the query matrix, k j is the bond matrix, M is the number of modes;

[0036] Step 2.1.2: Use the method shown in formula (8) to construct a representation for capturing and learning cross-modal shared information, and perform cross-modal representation learning on the sub-network in the multimodal drug relocation network;

[0037]

[0038] F sh =W h ·Vec(V^ T ) (8)

[0039] Among them, formula (8) adds shared cross-modal information to the value vector of each modality to represent the shared information of each modality value vector, P is the inter-modal attention matrix, I represents the identity matrix, and α is a hyperparameter that controls the degree of self-retention of the original modality information; is the projection matrix used to add V^; F sh Represents shared representations between modalities;

[0040] Step 2.1.3: Construct a representation for capturing and learning intra-modal shared information using the method shown in Equation (9), and perform intra-modal representation learning on the sub-networks in the multimodal drug relocalization network;

[0041]

[0042] Among them, Vec is the vectorized symbol, P is the inter-modal attention matrix, It is a modality-specific information representation;

[0043] Step 2.2: The sub-networks in the multimodal drug relocalization network are characterized by fusion using the method shown in formula (10);

[0044] h=Concat(h sh,h sp ) (10)

[0045] Among them, h represents the overall representation, and Concat represents the feature vector concatenation operation;

[0046] Step 3: Adaptive graph structure learning of the multimodal drug relocalization network with fused representations;

[0047] Step 3.1: Construct a drug repositioning adjacency matrix for adaptive graph structure learning of a multimodal drug repositioning network using the method shown in formula (11);

[0048] A ij =Sim(h i ,h j )=cos(W A h i ,W A h j ) (11)

[0049] Among them, A ij is the adjacency matrix, Sim is the similarity calculation function, h i and h j are the information of node i and node j respectively, cos is the cosine similarity calculation function, W A is a learnable weight matrix;

[0050] Step 3.2: Optimize the drug repositioning adjacency matrix learned by the adaptive graph structure of the multimodal drug repositioning network;

[0051] Step 3.2.1: Optimize the sparsity of the drug relocation adjacency matrix using the method shown in formula (12);

[0052]

[0053] Among them, XX is the element in the original matrix, X.min is the minimum value of all elements in the matrix, X.max is the maximum value of all elements in the matrix, and X norm It is a scaled value used to process the element value interval of the adjacency matrix into a non-negative form and scale it;

[0054] Step 3.2.2: Optimize the connectivity of the drug relocation adjacency matrix using the method shown in formula (13);

[0055]

[0056] Among them, A is the adjacency matrix representing the graph structure, H is the modal representation, and N is the number of nodes.

[0057] Establish smoothness constraints on graph signals for connections between similar nodes;

[0058] Step 3.2.3: Optimize the connectivity of the drug relocation adjacency matrix using the method shown in formula (14);

[0059]

[0060] Where A is the adjacency matrix representing the graph structure, Represents a vector of all 1s, N is the number of nodes, and log is a logarithmic function is the connectivity constraint of the graph structure;

[0061] Step 3.2.4: Optimize the constraints of the drug relocation adjacency matrix using the method shown in formula (15);

[0062]

[0063] Among them, H is the modal representation, F is the Frobenius norm, γ is the hyperparameter, Used to regularize the graph, β and γ are hyperparameters that balance the regularization. Used to adjust the sparsity of graph structure;

[0064] Step 4: The adaptively learned multimodal drug relocation network is subjected to loss optimization using the method shown in formula (16) to obtain the drug relocation prediction result;

[0065]

[0066] Where, Y^ sp =f(H sp ) is the prediction result, is the regularization term, is the loss function of the GNN prediction module, γ and η are hyperparameters;

[0067] A computer-readable storage device stores a computer program, wherein the computer program, when executed, implements the drug repositioning method based on multimodal graph learning.

[0068] A drug repositioning system based on multimodal graph learning includes a storage device, a processor, and a computer program stored in the storage device and executable on the processor, wherein the processor executes the computer program to implement the drug repositioning method based on multimodal graph learning.

[0069] Compared with the prior art, the present invention has the following beneficial effects:

[0070] 1. The present invention has rich representation learning and adaptive learning. As multimodal learning cannot take into account the association and difference between data of different modalities, a modal representation learning mechanism is proposed as an important part of the model framework to synchronously learn shared information and modality-specific information between modalities. At the same time, a multimodal attention mechanism is introduced in the learning process to assign different weights to nodes in the heterogeneous network, so that important nodes can retain more effective information. At the same time, a meaningful graph structure is constructed for the downstream graph convolutional learning network prediction module. By calculating the weighted cosine similarity between nodes, a learnable matrix function is obtained. This matrix is jointly optimized with the downstream graph neural network GNN. In the GNN prediction stage, the final prediction result is generated using the sparse graph A based on modal representation and learning.

[0071] 2. Based on multimodal data, the present invention comprehensively considers problems such as different entity types, complex entity relationships, and non-uniform dimensions. First, different similarities and association matrices are calculated according to the data type, and then they are integrated into a heterogeneous network. The attention mechanism and graph neural network are used to learn high-level features and make predictions. Compared with the baseline models of technical routes such as matrix decomposition and network embedding, there are improvements. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is a schematic diagram of the process of the present invention; DETAILED DESCRIPTION

[0073] In order to better illustrate the purpose and advantages of the present invention, the following is a further description of the invention in conjunction with the accompanying drawings and examples. It should be noted that the implementation of the present invention is not limited to the following embodiments, and any form of modification or change made to the present invention will fall within the scope of protection of the present invention.

[0074] Example

[0075] like Figure 1 As shown, the drug relocation method based on multimodal graph learning in this embodiment is specifically implemented as follows:

[0076] Step 1: Construct a multimodal drug relocalization network consisting of subnetworks with equal number of input modalities, so that the subnetworks can independently process single modality data;

[0077] Step 1.1: Use drug data and disease data to construct a drug-disease association subnetwork;

[0078] In the embodiment, data collected on drug and disease information includes 10 types of relationship types, including drug-drug interactions, drug-target interactions, drug-side effect associations, drug chemical similarity relationships, drug treatment similarity, drug target sequence similarity, gene ontology biological processes, gene ontology cellular components, gene ontology molecular functions, and drug-disease associations, which are related to three entity types including drugs, diseases, and proteins, as the raw data for constructing the drug repositioning network.

[0079] Step 1.2: Construct drug interaction subnetwork using drug interaction data;

[0080] In the embodiment, specifically, drug-drug interaction data reported in clinical practice are obtained, chemical names, common names, and commercial names are standardized through a vocabulary to form a unified identifier, and a drug-drug interaction network is formed.

[0081] Step 1.3: Construct a drug-target interaction subnetwork using drug-target interaction data;

[0082] In the embodiment, obtaining high-quality drug-target interaction information, obtaining detailed biological activity data between drugs and targets, and screening drug-target interaction data that meet the standards include: obtaining data through public drug and disease databases; obtaining detailed biological activity data between drugs and targets, extracting data from public and widely used databases, and retaining data that meet the inhibition constant K i , dissociation constant K d , half inhibitory concentration IC 50 and half maximal effect concentration (EC) 50 The drug-target interactions were all less than 10 μM; proteins with unique accession numbers were retained. Specifically, proteins that met the inhibition constant K i , dissociation constant K d , half inhibitory concentration IC 50 and half maximal effect concentration (EC) 50 The drug-target interactions were all less than 10 μM; proteins with unique accession numbers were retained; proteins marked as "reviewed" in the database were retained.

[0083] Step 1.4: Using the side effect data between drugs, the similarity of side effects between drugs is obtained using the method shown in formula (1), and a drug side effect subnetwork screened by the side effect similarity is constructed;

[0084]

[0085] Where A and B represent drug A and drug B respectively, N A is the number of side effects associated with drug A; N Bis the number of side effects associated with drug B; N AB is the number of side effects shared by drug A and drug B, and the range of Tanimoto coefficient is [0,1];

[0086] Step 1.5: Using medicinal chemistry data, obtain the drug molecule fingerprint similarity using the method shown in formula (2), and construct a medicinal chemistry subnetwork screened by the drug molecule fingerprint similarity;

[0087]

[0088] Where A and B represent drug A and drug B respectively, N′ A and N′ B Represents the binary bit strings of the MACCS fragments of the two drug molecules, N′ AB is a binary bit string that both have on the fragment;

[0089] In the embodiment, a network for constructing drug chemical similarity is constructed by obtaining SMILES format information and applying Tanimoto coefficient calculation to the molecular access system MACCS to describe molecules based on sequence;

[0090] Step 1.6: Using drug data and clinical data, construct a drug clinical subnetwork that has been screened for drug clinical similarity;

[0091] Step 1.6.1: Obtain the single-layer clinical similarity between drugs using the method shown in formula (3);

[0092]

[0093] Where k represents the kth layer of ATC coding, S k represents the k-level clinical similarity between drug A and drug B;

[0094] Step 1.6.2: Obtain the single-layer clinical similarity between drugs using the method shown in formula (4);

[0095]

[0096] Among them, ATC k Represents all ATC codes on layer k, S atc (A, B) represents the clinical similarity between drug A and drug B, and n represents the number of coding levels in the ATC classification system;

[0097] In the embodiment, the Anatomical Therapeutic and Chemical Classification System (ATC) codes are used to quantitatively estimate the therapeutic similarity of drugs in organs and systems;

[0098] Step 1.7: Using the drug target sequence data, obtain the drug target sequence similarity using the method shown in formula (5), and construct a drug target sequence subnetwork that has been screened by target sequence similarity;

[0099]

[0100] Among them, targets a and b correspond to the targets of drugs A and B respectively. When a≠b, the overall sequence similarity S is obtained by averaging the protein sequence similarities of targets a (a∈A) and b (b∈B). p ;

[0101] Step 1.8: Using the drug target gene data, obtain the drug target gene similarity using the method shown in formula (6), and construct a drug target gene sub-network screened by target gene similarity;

[0102]

[0103] Among them, genes a and b correspond to the target genes of drugs A and B respectively. The gene ontology similarity S is obtained by averaging the semantic similarity of the gene pairs a (a∈A) and b (b∈B) using the graph. GO ;

[0104] Step 2: Perform representation learning and representation fusion on the sub-networks in the multimodal drug relocalization network;

[0105] Step 2.1: Representation learning of the sub-networks in the multimodal drug relocalization network;

[0106] Step 2.1.1: Use the scaled dot product attention method for capturing and learning inter-modal shared information as shown in Equation (7) to learn inter-modal representations for the sub-networks in the multimodal drug relocalization network;

[0107]

[0108] Where, P ij Indicates the degree of attention of the i-th modality to the j-th modality, τ is the scaling factor, is the transposition operator, exp is the exponential function, q i is the query matrix, k j is the bond matrix, M is the number of modes;

[0109] Step 2.1.2: Use the method shown in formula (8) to construct a representation for capturing and learning cross-modal shared information, and perform cross-modal representation learning on the sub-network in the multimodal drug relocation network;

[0110]

[0111] F sh =W h ·Vec(V^ T ) (8)

[0112] Among them, formula (8) adds shared cross-modal information to the value vector of each modality to represent the shared information of each modality value vector, P is the inter-modal attention matrix, I represents the identity matrix, and α is a hyperparameter that controls the degree of self-retention of the original modality information; is the projection matrix used to add V^; F sh Represents shared representations between modalities;

[0113] Step 2.1.3: Construct a representation for capturing and learning intra-modal shared information using the method shown in Equation (9), and perform intra-modal representation learning on the sub-networks in the multimodal drug relocalization network;

[0114]

[0115] Among them, Vec is the vectorized symbol, P is the inter-modal attention matrix, It is a modality-specific information representation;

[0116] Step 2.2: The sub-networks in the multimodal drug relocalization network are characterized by fusion using the method shown in formula (10);

[0117] h=Concat(h sh ,h sp ) (10)

[0118] Among them, h represents the overall representation, and Concat represents the feature vector concatenation operation;

[0119] Step 3: Adaptive graph structure learning of the multimodal drug relocalization network with fused representations;

[0120] Step 3.1: Construct a drug repositioning adjacency matrix for adaptive graph structure learning of a multimodal drug repositioning network using the method shown in formula (11);

[0121] A ij =Sim(h i , h j )=cos(W A h i ,W A h j ) (11)

[0122] Among them, A ij is the adjacency matrix, Sim is the similarity calculation function, h i and h jare the information of node i and node j respectively, cos is the cosine similarity calculation function, W A is a learnable weight matrix;

[0123] In the embodiment, an adaptive graph learning matrix module is constructed to model graph learning as a similarity matrix between learning nodes, and an adjacency matrix of the graph structure is obtained by calculating the weighted cosine similarity between different nodes;

[0124] Step 3.2: Optimize the drug repositioning adjacency matrix learned by the adaptive graph structure of the multimodal drug repositioning network;

[0125] Step 3.2.1: Optimize the sparsity of the drug relocation adjacency matrix using the method shown in formula (12);

[0126]

[0127] Among them, XX is the element in the original matrix, X.min is the minimum value of all elements in the matrix, X.max is the maximum value of all elements in the matrix, and X norm It is a scaled value used to process the element value interval of the adjacency matrix into a non-negative form and scale it;

[0128] Step 3.2.2: Optimize the connectivity of the drug relocation adjacency matrix using the method shown in formula (13);

[0129]

[0130] Among them, A is the adjacency matrix representing the graph structure, H is the modal representation, and N is the number of nodes.

[0131] Establish smoothness constraints on graph signals for connections between similar nodes;

[0132] Step 3.2.3: Optimize the connectivity of the drug relocation adjacency matrix using the method shown in formula (14);

[0133]

[0134] Where A is the adjacency matrix representing the graph structure, Represents a vector of all 1s, N is the number of nodes, and log is a logarithmic function is the connectivity constraint of the graph structure;

[0135] Step 3.2.4: Optimize the constraints of the drug relocation adjacency matrix using the method shown in formula (15);

[0136]

[0137] Among them, H is the modal representation, F is the Frobenius norm, γ is the hyperparameter, Used to regularize the graph, β and γ are hyperparameters that balance the regularization. Used to adjust the sparsity of graph structure;

[0138] Step 4: The adaptively learned multimodal drug relocation network is subjected to loss optimization using the method shown in formula (16) to obtain the drug relocation prediction result;

[0139]

[0140] Where, Y^ sp =f(H sp ) is the prediction result, is the regularization term, is the loss function of the GNN prediction module, γ and η are hyperparameters;

[0141] In one embodiment, a GNN is used to generate prediction results based on the learned embedding and adjacency matrix, and a branch prediction network is used to ensure the stability of the learning process and improve the learning of modality-specific information;

[0142] A computer-readable storage device stores a computer program, wherein the computer program, when executed, implements the drug repositioning method based on multimodal graph learning.

[0143] A drug repositioning system based on multimodal graph learning includes a storage device, a processor, and a computer program stored in the storage device and executable on the processor, wherein the processor executes the computer program to implement the drug repositioning method based on multimodal graph learning.

[0144] The above specific description further illustrates the purpose, technical solutions and beneficial effects of the invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A drug relocalization method based on multimodal graph learning, characterized by: The following steps are included: Step 1: Construct a multimodal drug relocalization network consisting of subnetworks with equal number of input modalities, so that the subnetworks can independently process single modality data; Step 1.1: Use drug data and disease data to construct a drug-disease association subnetwork; Step 1.2: Construct drug interaction subnetwork using drug interaction data; Step 1.3: Construct a drug-target interaction subnetwork using drug-target interaction data; Step 1.4: Using the side effect data between drugs, the similarity of side effects between drugs is obtained using the method shown in formula (1), and a drug side effect subnetwork screened by the side effect similarity is constructed; Where A and B represent drug A and drug B respectively, N A is the number of side effects associated with drug A; N B is the number of side effects associated with drug B; N AB is the number of side effects shared by drug A and drug B, and the range of Tanimoto coefficient is [0,1]; Step 1.5: Using medicinal chemistry data, obtain the drug molecule fingerprint similarity using the method shown in formula (2), and construct a medicinal chemistry subnetwork screened by the drug molecule fingerprint similarity; Where A and B represent drug A and drug B respectively, N′ A and N′ B Represents the binary bit strings of the MACCS fragments of the two drug molecules, N′ AB is a binary bit string that both have on the fragment; Step 1.6: Using drug data and clinical data, construct a drug clinical subnetwork that has been screened for drug clinical similarity; Step 1.7: Using the drug target sequence data, obtain the drug target sequence similarity using the method shown in formula (5), and construct a drug target sequence subnetwork that has been screened by target sequence similarity; Among them, targets a and b correspond to the targets of drugs A and B respectively. When a≠b, the overall sequence similarity S is obtained by averaging the protein sequence similarities of targets a (a∈A) and b (b∈B). p ; Step 1.8: Using the drug target gene data, obtain the drug target gene similarity using the method shown in formula (6), and construct a drug target gene sub-network screened by target gene similarity; Among them, genes a and b correspond to the target genes of drugs A and B respectively. The gene ontology similarity S is obtained by averaging the semantic similarity of the gene pairs a (a∈A) and b (b∈B) using the graph. GO ; Step 2: Perform representation learning and representation fusion on the sub-networks in the multimodal drug relocalization network; Step 2.1: Representation learning of the sub-networks in the multimodal drug relocalization network; Step 2.2: The sub-networks in the multimodal drug relocalization network are characterized by fusion using the method shown in formula (10); h=Concat(h sh ,h sp ) (10) Among them, h represents the overall representation, and Concat represents the feature vector concatenation operation; Step 3: Adaptive graph structure learning of the multimodal drug relocalization network with fused representations; Step 3.1: Construct a drug repositioning adjacency matrix for adaptive graph structure learning of a multimodal drug repositioning network using the method shown in formula (11); A ij =Sim(h i ,h j )=cos(W A h i ,W A h j ) (11) Among them, A ij is the adjacency matrix, Sim is the similarity calculation function, h i and h j are the information of node i and node j respectively, cos is the cosine similarity calculation function, W A is a learnable weight matrix; Step 3.2: Optimize the drug repositioning adjacency matrix learned by the adaptive graph structure of the multimodal drug repositioning network; Step 3.2.1: Optimize the sparsity of the drug relocation adjacency matrix using the method shown in formula (12); Among them, XX is the element in the original matrix, X.min is the minimum value of all elements in the matrix, X.max is the maximum value of all elements in the matrix, and X norm It is a scaled value used to process the element value interval of the adjacency matrix into a non-negative form and scale it; Step 3.2.2: Optimize the connectivity of the drug relocation adjacency matrix using the method shown in formula (13); Among them, A is the adjacency matrix representing the graph structure, H is the modal representation, and N is the number of nodes. Establish smoothness constraints on graph signals for connections between similar nodes; Step 3.2.3: Optimize the connectivity of the drug relocation adjacency matrix using the method shown in formula (14); Where A is the adjacency matrix representing the graph structure, Represents a vector of all 1s, N is the number of nodes, and log is a logarithmic function is the connectivity constraint of the graph structure; Step 3.2.4: Optimize the constraints of the drug relocation adjacency matrix using the method shown in formula (15); Among them, H is the modal representation, F is the Frobenius norm, γ is the hyperparameter, Used to regularize the graph, β and γ are hyperparameters that balance the regularization. Used to adjust the sparsity of graph structure; Step 4: The adaptively learned multimodal drug relocation network is subjected to loss optimization using the method shown in formula (16) to obtain the drug relocation prediction result; Where, Y^ sp =f(H sp ) is the prediction result, is the regularization term, is the loss function of the GNN prediction module, and γ and η are hyperparameters.

2. The drug relocalization method based on multimodal graph learning according to claim 1, wherein: Step 1.6 is implemented as follows: Step 1.6.1: Obtain the single-layer clinical similarity between drugs using the method shown in formula (3); Where k represents the kth layer of ATC coding, S k represents the k-level clinical similarity between drug A and drug B; Step 1.6.2: Obtain the single-layer clinical similarity between drugs using the method shown in formula (4); Among them, ATC k Represents all ATC codes on layer k, S atc (A, B) represents the clinical similarity between drug A and drug B, and n represents the number of coding levels in the ATC classification system.

3. The drug relocalization method based on multimodal graph learning according to claim 1, characterized in that: Step 2.1 is implemented as follows: Step 2.1.1: Use the scaled dot product attention method for capturing and learning inter-modal shared information as shown in Equation (7) to learn inter-modal representations for the sub-networks in the multimodal drug relocalization network; Where, P ij Indicates the degree of attention of the i-th modality to the j-th modality, τ is the scaling factor, is the transposition operator, exp is the exponential function, q i is the query matrix, k j is the bond matrix, M is the number of modes; Step 2.1.2: Use the method shown in formula (8) to construct a representation for capturing and learning cross-modal shared information, and perform cross-modal representation learning on the sub-network in the multimodal drug relocation network; F sh =W h ·Vec(V^ T ) (8) Among them, formula (8) adds shared cross-modal information to the value vector of each modality to represent the shared information of each modality value vector, P is the inter-modal attention matrix, I represents the identity matrix, and α is a hyperparameter that controls the degree of self-retention of the original modality information; is the projection matrix used to add V^; F sh Represents shared representations between modalities; Step 2.1.3: Construct a representation for capturing and learning intra-modal shared information using the method shown in Equation (9), and perform intra-modal representation learning on the sub-networks in the multimodal drug relocalization network; Among them, Vec is the vectorized symbol, P is the inter-modal attention matrix, It is a modality-specific information representation.

4. A computer-readable storage device storing a computer program, characterized in that: When the computer program is executed, the drug repositioning method based on multimodal graph learning according to any one of claims 1 to 3 is implemented.

5. A drug repositioning system based on multimodal graph learning, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that: The processor executes the computer program to implement the drug repositioning method based on multimodal graph learning as described in any one of claims 1 to 3.

Images