A Drug Repositioning Method Based on a Multilevel Cross-Domain Relationship Network

Abstract

The present invention discloses a drug repositioning method based on a multi-level cross-domain relationship network, belonging to the technical field of drug repositioning. First, a multi-level cross-domain relationship network is constructed, including a drug-disease association native network, a drug association network, and a disease association network. By integrating multiple network architectures, the network can reveal richer semantic relationships and enhance the expressive ability of the network. A cross-domain fusion encoder is designed to encode the multi-level cross-domain relationship network for drug repositioning tasks. By comprehensively encoding multi-source heterogeneous data, the accuracy of network encoding is significantly improved. The association between drugs and diseases is predicted by using a multi-layer perceptron as a decoder. The multi-level cross-domain relationship network, the cross-domain fusion encoder, and the multi-layer perceptron are combined to form a drug repositioning model. A classification task is constructed to train the model. The drugs and diseases for testing are input into the trained drug repositioning model, and the probability score of the test drugs having an effect on the test diseases is output.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug repositioning, and in particular relates to a drug repositioning method based on a multi-level cross-domain relationship network. Background Art

[0002] In the process of drug discovery, traditional drug development often requires decades of clinical research and clinical trials, and then it can be marketed after market approval, which consumes huge manpower and financial resources. However, studies have shown that marketed drugs exhibit therapeutic effects different from the original indications in cytology, animal models or patient treatment. For example, Remdesivir, used to treat Ebola virus, has been proven to provide a rapid response in the treatment of COVID-19, or Fluoxetine was originally used to treat depression and was later found to be effective for obsessive-compulsive disorder. This is the concept of drug repositioning, also known as the development of new drug indications, which refers to the process of analyzing the mechanism of action of drugs that have been approved for marketing or entered the clinical stage to find new uses beyond their original indications. This strategy can effectively shorten the cycle of new drug development, reduce R&D costs, and reduce safety risks.

[0003] With the rapid development of the information age, a large amount of biomedical data has been rapidly accumulated, including genomic data, transcriptome data, etc., which provides research support for drug repositioning. In recent years, the rise of deep learning technology has significantly promoted the processing of biomedical data. Through advanced algorithms such as deep neural networks, researchers can more effectively mine potential patterns and features in complex data.

[0004] At present, many studies at home and abroad have conducted effective research in the field of drug repositioning based on a large amount of biomedical data. By analyzing the similarity of molecular structure or biochemical characteristics between drugs and targets to predict the potential application of drugs in different diseases, the core is to assume that drugs with similar molecular structures may also have similar therapeutic effects in biological mechanisms, which provides a simple and scalable strategy for drug repositioning, especially under the premise of known drug targets or characteristics, by calculating the similarity of molecular characteristics, it is expected to predict and verify new drug-disease pairings. Napolitano et al. integrated gene expression profiles and target features, combined multiple data sources through machine learning, and successfully predicted new uses for multiple drugs. However, the molecular similarity calculation method relies too much on the assumption of similarity of molecular structure, and in fact many potential therapeutic effects may not be discovered due to differences in molecular structure. In addition, this method is based on the static structural characteristics of molecules and it is difficult to capture the dynamic and complex mechanism of action of drugs in biological systems.

[0005] To overcome the limitations of molecular similarity calculation methods, more complex network biology methods have been proposed. By integrating multi-source biological information to construct a multi-source association network of diseases and drugs, the potential relationships between drugs and diseases can be further revealed. Ren Z H et al. constructed a heterogeneous network through a simplified graph convolutional network, combined with an adaptive information diffusion distance to dynamically adjust the transmission range of network information, and strengthened the representation of drug-disease associations, achieving satisfactory results in the drug repositioning task. Although the methods based on multi-source association networks have shown remarkable performance, the complexity of their networks also brings non-negligible limitations. For example, the data sources are diverse and of uneven quality, and the diversity of different types of nodes and relationships makes network representation learning complex. Therefore, researchers have explored using simpler network structures and focused on the direct associations between drugs and diseases. Sun X et al. extracted the partner-specific subgraphs of target drug-disease pairs in the direct association network and used a layer attention mechanism to integrate multi-scale layer information. Jin S et al. generated multiple views of nodes in the direct association network through data augmentation methods such as node random inactivation, edge random inactivation, and random walks, and used a graph convolutional network to learn node embeddings. At the same time, the learning process was further optimized through contrastive learning and adaptive denoising training. However, the above methods always deploy feature encoders in separate networks and cannot effectively guarantee the integrity of heterogeneous networks. Summary of the Invention

[0006] In view of the above problems existing in the prior art, the present invention proposes a drug repositioning method based on a multi-level cross-domain relationship network, which is reasonably designed, solves the deficiencies of the prior art, and has good effects.

[0007] A drug repositioning method based on a multi-level cross-domain relationship network includes the following steps:

[0008] S1. Construct a multi-level cross-domain relationship network , including a drug-disease association native network, a drug association network, and a disease association network;

[0009] S2. Design a cross-domain fusion encoder to encode the multi-level cross-domain relationship network for the drug repositioning task;

[0010] S3. Predict the association between drugs and diseases through a multi-layer perceptron as a decoder;

[0011] S4. Construct a drug repositioning model with the multi-level cross-domain relationship network, the cross-domain fusion encoder, and the multi-layer perceptron, and construct a classification task to train the model to obtain a trained drug repositioning model;

[0012] S5. Input the drugs and diseases for testing into the trained drug repositioning model, and output the probability score that the test drug has an effect on the test disease.

[0013] Furthermore, for the drug-disease association native network , first, drug-disease information is collected from public databases, including PubChem, DrugCentral, and DrugBank, to construct direct associations between drugs and diseases. If the th drug has an effect on the th disease, they are considered to be associated in ; otherwise, they are not.

[0014] Furthermore, for the drug association network , let be defined as , where is the set of drug nodes in , is the initial embedding set for each drug, and this embedding not only comes from the basic properties of the drug but also incorporates the DDI prediction score; is the association set between drugs in ; the k-nearest neighbor graph idea is used to construct , that is, the top k drugs are selected according to the DDI prediction score from largest to smallest to construct so as to achieve the purpose of association construction. The expression is:

[0015] (1);

[0016] Among them, is the association representation between the th drug and the th drug, is the DDI score between the th drug and the th drug, , represents the number of drugs, is the extended k-nearest neighbor set containing the th drug constructed according to the k-nearest neighbor graph idea and the DDI prediction score.

[0017] Furthermore, for the disease association network , the semantic information of each disease is represented based on the directed acyclic graph structure of MeSH. Each disease is mapped to a DAG. Let be defined as , is the set of diseases in , is the semantic similarity of different disease DAGs, is Set of associations between diseases.

[0018] Furthermore, for the cross - domain fusion encoder, is defined as , where is the set of drug nodes in , is the set of disease nodes in , represents the number of diseases, is in the set of associations between drugs and diseases. If the rd drug is effective against the th disease, then , otherwise . ; As the input of the cross - domain fusion encoder, its initial embedding is defined as the one - hot index that maps the entity index in to a learnable embedding and the entity association mapping in

[0019] The information aggregation process of the cross - domain fusion encoder is as shown in the formula:

[0020] (2);

[0021] Where represents the initial input of the network, represents the adjacency relationship between nodes in the network, is a learnable weight matrix, is the complete convolution process, including two forms, the association domain convolution and the collaboration domain convolution , is the node type mapping function, is the th node in is the th node, ;

[0022] Obtain the embeddings of the th drug and the th disease in and respectively, where the specific information transfer process using the th drug introduction is:

[0023] First, the The initial association information of each drug node is provided in , and expanding in formula (2), the expression is:

[0024] (3);

[0025] (4);

[0026] (5);

[0027] where is the pseudo - association feature corresponding to the rdrug, is the real - association feature corresponding to the th drug, is the association - domain feature corresponding to the th drug, is the tanh activation function, and are both learnable weight matrices, is the element - wise multiplication aggregation function; is the pseudo - association feature corresponding to the th drug, , is the real - association feature corresponding to the th drug, , and are respectively the neighborhood directly associated with the th drug and the neighborhood not directly associated;

[0028] Then, the node features of the drug association network are updated through , and the expression is:

[0029] (6);

[0030] (7);

[0031] where is the drug cooperation - domain feature corresponding to the th drug, is the perception - domain depth of , is the th drug's neighbor set obtained according to the k - nearest neighborhood in , is the learnable weight matrix of the th drug when the perception - domain depth is ;​ is the feature embedding of the rd drug after layers of perception, and is the embedding of the th drug after update. is the Hadamard product aggregation function;

[0032] Next, is fed into Equation (2) for deeper exploration. Through deep feature encoding, it not only captures the direct relationship between drugs and diseases in but also incorporates the collaboration information in

[0033] Finally, in the multi-layer architecture of the graph neural network, a layer-by-layer weighted fusion method is adopted to accumulate the features of each layer into the final output according to decreasing weights. The expression is:

[0034] (8);

[0035] where is the final embedding of the th drug obtained in , is the feature obtained by the th drug in the th layer of the cross-domain fusion encoder, and is the depth of the perception field of the cross-domain fusion encoder.

[0036] Furthermore, in step S3, and are respectively assigned to the attributes of the corresponding drug nodes and disease nodes in . Through the update operation of the edge attributes in , the features between drugs and diseases are integrated to generate the edge features between the target drugs and diseases. The expression is:

[0037] (9);

[0038] where is the feature integration function used to integrate drug and disease features into the network association;

[0039] Subsequently, the edge features are input into a multi-layer perceptron to obtain high-order features layer by layer, and finally the prediction score of whether there is a therapeutic effect relationship between the target drugs and diseases is generated. The expression is:

[0040] (10);

[0041] where is the prediction score made by the model, and is a learnable weight matrix, and is a bias parameter, is the ReLu activation function.

[0042] Furthermore, in step S4, during the model training process, the drug repositioning problem is modeled as a binary classification task, that is, predicting whether a drug is effective against a disease. To optimize the model performance, binary cross-entropy is used as the loss function for the classification task, and the expression is:

[0043] (11);

[0044] wherein, is the true label.

[0045] The beneficial technical effects brought by the present invention:

[0046] By constructing a multi-level cross-domain relationship network and integrating multiple network architectures, the present invention can deeply mine more diverse semantic associations, thereby comprehensively improving the expression and reasoning capabilities of the network. To overcome the limitations of traditional similarity-based drug association networks, the DDI concept is introduced to more fully reveal the deep associations between drugs at the biological level, and a cross-domain fusion encoder is proposed. By comprehensively encoding multi-source heterogeneous data, the accuracy of network encoding is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic diagram of a traditional dual-network architecture.

[0048] Figure 2 is a schematic diagram of the multi-level cross-domain relationship network architecture in the present invention.

[0049] Figure 3 is a schematic diagram of the process of the cross-domain fusion encoder in the present invention.

[0050] Figure 4 is a schematic diagram of the principle of associated domain convolution in the present invention.

[0051] Figure 5 is a schematic diagram of the principle of collaborative domain convolution in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0052] The following further describes the specific implementation manners of the present invention with reference to specific embodiments:

[0053] A drug repositioning method based on a multi-level cross-domain relationship network includes the following steps:

[0054] S1. Construct a multi-level cross-domain relationship network , including a drug-disease association native network, a drug association network, and a disease association network;

[0055] Traditional dual-network architectures usually treat the drug-disease association native network as an independent architecture, while treating the entity association network, i.e., the drug association network and the disease association network , as separate architectures for modeling, as shown in Figure 1 . Although it can capture the direct associations between drugs and diseases and their respective internal relationships to a certain extent, due to the lack of deep interaction between the networks, the ability to mine potential associations is often insufficient. Therefore, the present invention proposes a multi-level cross-domain relationship network, as shown in Figure 2 , which, while retaining the direct associations between network nodes, further integrates the information between drugs from and the information between diseases from , thereby enriching the feature expression of the native network, enabling it not only to effectively express the direct associations between drugs and diseases but also to fully exploit the enhancing effect of the internal relationships between drugs and diseases on the overall network, significantly improving the mapping ability of drug-disease potential associations and achieving a high degree of integration of the network structure.

[0056] For the drug-disease association native network , first, drug-disease information is collected from public databases. The public databases include PubChem, DrugCentral, and DrugBank to construct the direct associations between drugs and diseases. If the th drug has an effect on the th disease, it is considered that there is an association between the two in , otherwise not.

[0057] For the drug association network , in existing methods, the construction of the drug association network is usually based on the similarity between drugs, such as using the similarity of structure, function, or chemical properties. However, this method has significant limitations: the structural similarity of drugs does not always reflect the relevance of their mechanisms of action in actual treatment. The activities of some structurally similar drugs at specific targets may show significant differences, and even completely different therapeutic effects may occur in the treatment of diseases. Therefore, relying solely on similarity may lead to the construction of the drug association network lacking certain biological significance and affecting the accuracy and reliability of entity associations. To solve this problem, the present invention pre-trains a drug interaction prediction model from a large amount of drug pair data collected from the DrugBank database as a discriminator for the interaction strength between drug pairs in , and then statistically analyzes the prediction scores between each drug pair using Construction. This method is different from the traditional strategy for constructing association networks based on drug similarity (such as structure or chemical properties). Instead, by predicting the actual functional associations between drugs, it provides a more biologically meaningful basis for network construction.

[0058] Let be defined as , where is the set of drug nodes in , and is the set of initial embeddings for each drug. This embedding not only comes from the basic properties of the drug but also incorporates the DDI prediction score, which can more comprehensively reflect the associations between drugs. For the DDI prediction score, first, a drug interaction prediction model is built using the drug structure and drug fingerprints, and then a large number of drug pair data are collected from the DrugBank database to pre-train it as the discriminator for the DDI prediction score in ; is the set of associations between drugs in . The k-nearest neighbor graph idea is used to construct , that is, the top k drug entities are selected according to the DDI prediction score from large to small to construct

[0059] (1);

[0060] where is the association representation between the th drug and the th drug, is the DDI score between the th drug and the th drug, , represents the number of drugs, is the extended k-nearest neighbor set containing the th drug constructed according to the k-nearest neighbor graph idea and the DDI prediction score.

[0061] For the disease association network , the semantic information of each disease is represented based on the directed acyclic graph structure of MeSH (Medical Subject Headings). As a hierarchical medical subject vocabulary, MeSH can effectively reveal the hierarchical structure and relevance among diseases. Each disease is mapped to a DAG, where nodes represent the subject entries of the disease and their upper-level concepts, and edges represent the parent-child relationships between different levels. By calculating the semantic similarity of two disease DAGs, the degree of association between diseases can be quantified. The calculation of the similarity score is based on the co-occurrence information and hierarchical structure among nodes, which can comprehensively consider the specific characteristics of diseases and higher-level semantic associations, and is used for construction. is defined as , is the set of diseases in is the semantic similarity of different disease DAGs, is the set of associations between diseases in

[0062] By using differentiated strategies, and are constructed respectively, so as to more accurately depict the complex relationship between drugs and diseases. Without relying on similarity assumptions, it is closer to the actual performance of drugs in disease treatment, integrating the drug network based on functional associations and the disease network based on semantic associations, breaking through the limitations of traditional methods, and achieving a comprehensive and accurate depiction of the association between drugs and diseases.

[0063] Traditional dual-network architectures usually take as an independent architecture, and at the same time take and as separate architectures for modeling. Although it can capture the direct associations between drugs and diseases and their respective internal relationships to a certain extent, due to the lack of deep interaction between networks, it often leads to insufficient ability to mine potential associations. Therefore, the present invention proposes a multi-level cross-domain relationship network , where each drug node and disease node, while retaining their direct associations, further integrates the interaction information between drugs from and the disease similarity information in . It can not only effectively express the direct associations between drugs and diseases, but also fully explore the enhancement effect of the internal relationships between drugs and diseases on the overall network, significantly improving the mapping ability of drug-disease potential associations and achieving a high degree of integration of the network structure.

[0064] S2. Design a cross-domain fusion encoder to encode for the drug repositioning task;

[0065] For the cross - domain fusion encoder, as Figures 3 - 5 shown, define as , where is the set of drug nodes in , is the set of disease nodes in , represents the number of diseases, is the association set of drugs and diseases in . If the th drug is effective against the th disease, then , otherwise , . ; As the input of the cross - domain fusion encoder, its initial embedding is defined as the one - hot index that maps the entity index in to the learnable embedding and the entity association mapping in and ;

[0066] The information aggregation process of the cross - domain fusion encoder is as shown in the formula:

[0067] (2);

[0068] Where represents the initial input of the network, represents the adjacency relationship between nodes in the network, is the learnable weight matrix, is the complete convolution process, including two forms, the association domain convolution and the collaboration domain convolution , is the node type mapping function, is the th node in , is the th node,

[0069] Obtain the embeddings of the th drug and the th disease in and respectively, where the specific information transfer process using the th drug introduction is:

[0070] First, the initial association information of the th drug node in has been provided. Expanding the in formula (2), the expression is:

[0071] (3);

[0072] (4);

[0073] (5);

[0074] Among them, is the pseudo - association feature corresponding to the th drug, is the real - association feature corresponding to the th drug, is the association - domain feature corresponding to the th drug, is the tanh activation function, and are both learnable weight matrices, is the element - wise multiplication aggregation function, is the pseudo - association feature corresponding to the th drug, , is the real - association feature corresponding to the th drug, , and are respectively the neighborhood with direct association and the neighborhood without direct association of with the th drug; the purpose is to comprehensively capture the diverse association information of the drug node in the basic network through the neighbor feature aggregation of two types of edge relationships, providing an initial understanding for subsequent deeper - level feature representation;

[0075] Then, the node features of the drug - association network are updated through , and the expression is:

[0076] (6);

[0077] (7);

[0078] Among them, is the drug - collaboration - domain feature corresponding to the th drug, is the perception - domain depth of , is the th drug's neighbor set obtained according to the k - nearest neighborhood in , is the The learnable weight matrix of a drug when the depth of the perception field is , and is the feature embedding of the th drug after layers of perception in . is the updated embedding of the th drug, and

[0079] is the Hadamard product aggregation function; Next, is fed into Equation (2) for deeper exploration. Through deep feature encoding, not only the direct relationship between drugs and diseases in and is captured, but also the collaborative information in

[0080] is incorporated into the encoding process, enhancing the potential relationship mapping ability between drug and disease nodes;

[0081] (8);

[0082] where is the final embedding of the th drug obtained in , is the feature obtained by the th drug in the -layer cross-domain fusion encoder, is the depth of the perception field of the cross-domain fusion encoder.

[0083] The specific information transfer process using the th disease introduction is as follows:

[0084] First, the initial association information of the th disease is already provided in . Expanding in Equation (2), the expression is:

[0085] (9);

[0086] (10);

[0087] (11);

[0088] Among them, is the pseudo - associated feature corresponding to the th disease, is the real - associated feature corresponding to the th disease, is the associated domain feature corresponding to the th disease, and are respectively the neighborhood with direct association and the neighborhood without direct association corresponding to the th disease, and are both learnable weight matrices; is the pseudo - associated feature corresponding to the th disease, is the real - associated feature corresponding to the th disease;

[0089] Then, the node features of the disease association network are updated through The expression is:

[0090] (12);

[0091] (13);

[0092] Among them, is the disease cooperation domain feature corresponding to the th disease, is the perception domain depth of , is the neighbor set obtained according to the k - nearest neighborhood in for the th disease, is the learnable weight matrix when the perception domain depth of the th disease in is , is the feature embedding of the nd disease after layers of perception in , is the embedding of the th disease after update, is the Hadamard product aggregation function, is the associated domain feature;

[0093] Finally, in the multi-layer architecture of the cross-domain fusion encoder, a layer-by-layer weighted fusion method is adopted to accumulate the features of each layer into the final output according to decreasing weights. The expression is:

[0094] (14);

[0095] where, is the final embedding obtained for the th disease in , is the th disease, and is the feature obtained by the

[0096] S3. Predict the association between drugs and diseases through a multi-layer perceptron as a decoder;

[0097] and are respectively assigned to the attributes of the corresponding drug nodes and disease nodes in . Through the update operation of the edge attributes in , the features between drugs and diseases are integrated to generate the edge features between the target drug and disease. The expression is:

[0098] (15);

[0099] where, is a feature integration function used to integrate drug and disease features into the network association;

[0100] Subsequently, the edge features are input into the multi-layer perceptron to obtain high-order features layer by layer, and finally the prediction score of whether there is a therapeutic effect relationship between the target drug and disease is generated. The expression is:

[0101] (16);

[0102] where, is the prediction score made by the model, and are learnable weight matrices, and are bias parameters, is the ReLu activation function.

[0103] S4. Construct a drug repositioning model by combining the multi-level cross-domain relationship network, cross-domain fusion encoder, and multi-layer perceptron. Build a classification task to train the model and obtain a trained drug repositioning model;

[0104] There are already many drug repurposing datasets available for research, such as Gdataset, Cdataset, and LRSSL. Each dataset covers drug and disease information collected from multiple reliable databases. The known efficacy relationships are randomly divided into 10 subsets of equal size. Each subset is used as the test set in turn, and the remaining subsets are used as the training set. This process is repeated 10 times.

[0105] During the model training process, the drug repurposing problem is modeled as a binary classification task, that is, predicting whether a drug has efficacy on a disease. To optimize the model performance, binary cross-entropy is used as the loss function for the classification task, and the expression is:

[0106] (17);

[0107] where, is the true label.

[0108] S5. Input the drugs and diseases used for testing into the trained drug repurposing model, and output the probability score that the test drug has efficacy on the test disease.

[0109] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions, or substitutions made by those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope of the present invention.

Claims

1. A drug repositioning method based on a multi-level cross-domain relationship network, characterized in that: The following steps are involved: S1. Build a multi-level cross-domain relationship network , including drug-disease association native network, drug association network and disease association network; S2. Design a cross-domain fusion encoder to encode a multi-level cross-domain relationship network for drug repositioning tasks; S3, predicting the association between drugs and diseases using a multi-layer perceptron as a decoder; S4, a multi-level cross-domain relationship network, a cross-domain fusion encoder and a multi-layer perceptron are used to form a drug relocation model, and a classification task is constructed to train the model to obtain a trained drug relocation model; S5. Input the drug and disease to be tested into the trained drug repositioning model, and output the probability score of the test drug being effective for the test disease; For drug association networks ,Will Defined as ,in yes The set of drug nodes in The initial embedding set for each drug, which is derived not only from the basic properties of the drug but also incorporates the DDI prediction score; for The association set between Chinese medicines; constructed using the k-nearest neighbor graph idea , that is, select the top k drugs from large to small according to the DDI prediction score to construct To achieve The purpose of association construction is expressed as: (1)； in, For the Drugs and The association between drugs indicates that For the Drugs and The DDI scores between drugs , Indicates the amount of medicine, It includes The extended k-nearest neighbor set of the drug is constructed based on the k-nearest neighbor graph idea and DDI prediction score.

2. A drug repositioning method based on a multi-level cross-domain relationship network according to claim 1, characterized in that: For drug-disease association native network First, drug-disease information was collected through public databases, including PubChem, DrugCentral, and DrugBank, to build a direct relationship between drugs and diseases. Drugs for If the two diseases are cured, it is considered that There is a correlation, otherwise there is no correlation.

3. A drug repositioning method based on a multi-level cross-domain relationship network according to claim 2, characterized in that: For disease association networks , based on the directed acyclic graph structure of MeSH, the semantic information of each disease is represented. Each disease is mapped into a DAG. Defined as , for The disease collection in is the semantic similarity of different disease DAGs, for The set of associations between diseases.

4. A drug repositioning method based on a multi-level cross-domain relationship network according to claim 3, characterized in that: For the cross-domain fusion encoder, Defined as ,in, is The collection of drug nodes in is The set of disease nodes in Indicates the number of diseases, is The association set between Chinese medicine and disease, if Drugs for If the disease is effective, ,otherwise , ; As the input of the cross-domain fusion encoder, its initial embedding is defined as The entity index in is mapped to a one-hot index that can be learned to embed and and Entity association mapping in ; The information aggregation process of the cross-domain fusion encoder is shown in the formula: (2)； in, express The initial input to the network, Represents the adjacency relationship between nodes in the network. is a learnable weight matrix, It is a complete convolution process, including two forms, correlation domain convolution and collaborative domain convolution , is the node type mapping function, for Middle nodes, For the nodes, ; Get the Drugs and The diseases are Embedding in and , where the first Drug introduction The specific information transmission process is: First, The initial association information of drug nodes is It has been provided in the formula (2). , the expression is: (3)； (4)； (5)； in, For the The pseudo-correlation features corresponding to each drug, For the The real correlation features corresponding to each drug, For the The associated domain features corresponding to each drug, is the tanh activation function, and are all learnable weight matrices, is the element-wise multiplication aggregation function; For the The pseudo-correlation features corresponding to each drug, , For the The real correlation features corresponding to each drug, , and They are With Each drug has directly related neighbors and non-directly related neighbors; Then, the node features of the drug association network are obtained by Update, the expression is: (6)； (7)； in, For the The drug collaboration domain features corresponding to each drug, Yes The depth of the receptive field, It is Drugs in The neighbor set obtained according to the k-nearest domain, It is Drugs in The receptive field depth is The learnable weight matrix at It is Drugs in After Feature embedding after layer perception, After the update The insertion of a drug, is the Hadamard product aggregation function; Next, is fed into formula (2) for deeper exploration. Through deep feature encoding, not only The direct relationship between Chinese medicine and disease will also and Incorporate collaborative information into the encoding process; Finally, in the multi-layer architecture of the graph neural network, a layer-by-layer weighted fusion method is adopted to accumulate the features of each layer into the final output according to decreasing weights. The expression is: (8)； in, It is Drugs in The final embedding obtained in For the The drug in The features obtained by the encoder are fused across layers. is the receptive domain depth of the cross-domain fusion encoder.

5. A drug repositioning method based on a multi-level cross-domain relationship network according to claim 4, characterized in that: In S3, and Assigned to The corresponding drug node and disease node attributes are obtained by The update operation of the edge attribute integrates the features between drugs and diseases to generate the edge features between the target drug and disease. The expression is: (9)； in, It is a feature integration function, which is used to integrate drug and disease features into network associations; Subsequently, the edge features are input into a multilayer perceptron to obtain high-order features layer by layer, and finally generate a prediction score for whether there is a therapeutic effect relationship between the target drug and the disease, expressed as: (10)； in, is the prediction score made by the model, and is a learnable weight matrix, and is the bias parameter, is the ReLu activation function.

6. A drug repositioning method based on a multi-level cross-domain relationship network according to claim 5, characterized in that: In S4, during the model training process, the drug repositioning problem is modeled as a binary classification task, that is, predicting whether the drug is effective for the disease. In order to optimize the model performance, binary cross entropy is used as the loss function of the classification task, and the expression is: (11)； in, is the true label.

Images