Drug-disease relationship prediction method and system based on dual-channel fusion knowledge graph

Abstract

The present invention relates to a drug-disease relationship prediction method and system based on a dual-channel fusion knowledge graph. The method comprises the following steps: S1, constructing a biomedical fusion knowledge graph; S2, constructing a drug-disease subgraph based on the biomedical fusion knowledge graph; S3, constructing a dual-channel adaptive fusion feature module and embedding the drug-disease subgraph; S4, performing enhanced splicing on the drug-disease subgraph pre-embedded in the dual-channel fusion knowledge graph network, introducing an edge-node iterative update learning mechanism, training the spliced ​​subgraph relationship perception learning network, and obtaining an enhanced subgraph feature embedding; S5, calculating and outputting the predicted probability of the drug-disease relationship using the trained enhanced subgraph feature embedding. The present invention integrates the semantic representation capability of the knowledge graph with the topological modeling advantages of the graph neural network, and realizes efficient and accurate prediction of the drug-disease relationship through multimodal feature interaction, subgraph enhancement, and dynamic subgraph learning mechanisms.

Description

Technical Field

[0001] The present invention relates to the field of medical big data technology, and in particular to a drug-disease relationship prediction method and system based on a dual-channel fusion knowledge graph. Background Art

[0002] The traditional drug development process is not only time-consuming and costly, but also carries a high risk of failure. This process encompasses five core phases: initial drug discovery, preclinical evaluation, clinical studies, review by the U.S. Food and Drug Administration (FDA), and post-market safety monitoring. Typically, drug development requires billions of dollars and spans a lengthy 10-15 years. Safety issues are a major cause of development failure. If a drug candidate fails Phase I due to safety concerns (e.g., toxicity), all investments from early drug discovery through Phase I are lost. To alleviate this burden, exploring new indications for approved drugs—drug repurposing or repositioning—has become a sensible option. According to statistics, approximately 30% of FDA-approved drugs are discovered to have at least one new indication after approval. These "old drugs for new uses" typically do not require further safety verification and can proceed directly to Phase II and Phase III clinical trials to evaluate their therapeutic efficacy for specific diseases. This strategy not only greatly shortens the time for traditional drug development, but also significantly reduces R&D costs. At the same time, it can improve the efficiency of drug use and maximize its potential therapeutic value.

[0003] However, traditional drug repurposing research relies primarily on manual screening based on experimental validation and expert experience, resulting in a lengthy and costly process. Limited by known drug mechanisms of action and accumulated clinical data, it lacks systematic methods for exploring potential new indications, leading to inefficient discovery. In recent years, artificial intelligence (AI) has demonstrated tremendous potential in assisting drug development. Knowledge graphs (KGs), as efficient AI tools, can integrate data from multiple sources, including complex, heterogeneous information from literature, clinical trials, and genomic data. They encompass a wide range of entities (e.g., drugs, diseases, proteins, or targets) and the intricate relationships between them (e.g., drug-drug interactions and drug-target pairings). These tools can deeply reveal the complex networks of entities within biological systems and intuitively display the complex relationships between biomedical entities such as drugs, diseases, genes, and proteins. Knowledge graph completion techniques can deeply mine the latent semantic features of triples and identify new potential connections between diseases and drugs. Traditionally, researchers have used machine learning methods to analyze drug chemical structures, gene expression data, or electronic medical record data. However, these methods require manual feature definition and struggle to integrate multimodal data. Similar deep learning methods have also been applied to drug repurposing research. Scholars at home and abroad have inputted knowledge graph triples into architectures such as convolutional neural networks (CNNs) and graph neural networks (GNNs) to extract entity embedding features and achieve their ultimate research goals. However, these research methods often treat neural networks as a black box, making it difficult to explain the biological basis of the predictions.

[0004] Current knowledge graph completion methods for drug repurposing still have some limitations. For example, the medical field lacks sufficient samples required for supervised learning and the neighborhood information of the graph structure is insufficient. Current knowledge graph completion methods find it difficult to achieve ideal training results. To overcome these problems, more and more researchers are trying to incorporate the rich information of external knowledge graphs into drug repurposing research to improve the performance of downstream tasks. However, the use of graph embedding alone often only utilizes topological structure information and does not use the entity's own attribute information. At the same time, as the scale of knowledge graphs continues to grow, the types of entities and relationships in the graph will continue to increase, and the heterogeneity of knowledge graphs will become more and more significant. This poses new challenges in how to effectively embed different types of relationships. In addition, although large biomedical knowledge graphs can utilize rich information, the embedding results are often poor due to noise when representing nodes. At the same time, passing the entire graph into the model also increases the computational burden.

[0005] Some researchers have tried to use subgraphs to precisely focus on key information for embedding learning. However, the current method of subgraph embedding learning relies on the topological structure of a fixed graph. The known topological structure of the knowledge graph may not be reliable. The explicit graph structure does not necessarily reflect the true dependency relationship. Not all relationship paths provide important association information. Excessive reliance may introduce noise.

[0006] Therefore, the field currently still lacks methods that can better utilize knowledge graphs to predict drug-disease relationships. Summary of the Invention

[0007] The present invention provides a drug-disease relationship prediction method and system based on a dual-channel fusion knowledge graph, the purpose of which is to fill the gap in the research on subgraph embedding learning and realize the effective use of the knowledge graph computing model in the drug-disease relationship prediction method.

[0008] To achieve the above objectives, the present invention provides a drug-disease relationship prediction method based on a dual-channel fusion knowledge graph, comprising the following steps:

[0009] S1. Obtain drug-disease association triples and external knowledge graph data to construct a biomedical fusion knowledge graph;

[0010] S2. Construct drug-disease subgraph based on biomedical fusion knowledge graph;

[0011] S3. Construct a dual-channel adaptive fusion feature module, embed the drug-disease subgraph, and obtain a dual-channel fusion knowledge graph pre-embedding network;

[0012] S4. Enhance the splicing of the drug-disease subgraph in the dual-channel fusion knowledge graph pre-embedded network, introduce the edge-node iterative update learning mechanism, train the spliced ​​subgraph relationship perception learning network, and obtain the enhanced subgraph feature embedding;

[0013] S5. Use the trained enhanced subgraph feature embedding to calculate the predicted probability of the output drug-disease relationship.

[0014] The dual-channel fusion knowledge graph network of the present invention consists of four key parts: external knowledge fusion module, subgraph construction module, dual-channel adaptive fusion feature embedding module and subgraph learning module.

[0015] For the dual-channel adaptive fusion feature embedding module, the present invention designs a dual-channel adaptive embedding module (Dual-Channel Embedding Module), which consists of a structural channel (UnionGIN) and a semantic channel (PubMed-BERT), which respectively encode the fused graph structure and the text description of the entity, thereby comprehensively capturing the multi-source feature information of the entity. The structural channel models the topological structure between entities, while the semantic channel mines the semantic relationship of the entity ontology, completing the multimodal representation of the graph nodes through adaptive feature fusion. This process first merges the original drug-disease graph with the external knowledge graph to introduce more common entities and relationships, thereby improving the representation capability of the graph.

[0016] For the subgraph construction module, the present invention designs a path-aware sampling strategy. Centered on the drug-disease pair to be predicted, the corresponding K-hop subgraph is generated from the fusion graph through path search. For the subgraph learning module, the present invention designs an end-to-end subgraph representation learning framework that employs an iterative edge-node collaborative update mechanism: edge weights are first calculated based on node embeddings to implement edge embedding updates, and then node representations are updated through neighbor aggregation.

[0017] For the subgraph learning module, in order to enhance the representation capabilities of the subgraph, the present invention introduces a relationship enhancement mechanism (Resemble Edge Construction). By constructing supplementary edges in the subgraph (such as drug-drug, disease-disease similarity connections, and drug-disease association connections), the structural associations between entities are explicitly enhanced, thereby improving the model's ability to model high-order relationships in complex paths. At the same time, the present invention designs an end-to-end subgraph representation learning framework that adopts an iterative edge-node collaborative update mechanism: first, edge embedding updates are implemented based on node embeddings to calculate edge weights, and then node representations are updated through neighbor aggregation, and iteratively implemented to achieve subgraph learning capabilities.

[0018] In addition, in the classification prediction stage, the present invention designs a triple embedding fusion strategy to splice the head entity embedding, tail entity embedding and subgraph embedding, and outputs the association type through the classifier. This strategy significantly improves the accuracy of association relationship judgment by complementing global context perception with local structural information.

[0019] Preferably, the acquisition of drug-disease association triples and external knowledge graph data in step S1 specifically includes: selecting drug-disease association triples provided by the Precision Medicine Knowledge Graph (PrimeKG) and the Drug Repositioning Database (repoDB) as training knowledge, wherein the node types include drug and disease, and the relationship type is a verified drug-disease association relationship; and selecting a biomedical heterogeneous knowledge graph (Heterogeneous Information Network, Hetionet) as the external knowledge graph;

[0020] The construction of the biomedical fusion knowledge graph specifically includes:

[0021] Remove isolated nodes from the external knowledge graph and filter out drug nodes and disease nodes;

[0022] Extract drug-disease association triples from the precision medicine knowledge graph and drug repositioning database;

[0023] The drug-disease association triple data is deeply integrated with drug nodes and disease nodes to construct a biomedical fusion knowledge graph; the drug-disease association triple data is divided into training set, validation set and test set.

[0024] Preferably, the step S2 of constructing the drug-disease subgraph specifically includes:

[0025] Target drug-disease pair As the center, a path-based K-hop subgraph sampling strategy is used to model the local context of the target pair. For a target triple ,in Indicates medicine, Indicates disease, Represented as the semantic relationship between them, Represents all target triplets; find the path through breadth-first traversal and extract the local subgraph within its jump range from the fusion graph, which is defined as:

[0026] ;

[0027] in, Indicates that drugs are captured by BFS and The K-hop local context of the departure includes nodes, relations, and triples. Indicates that from drugs and diseases Starting from, the set of nodes reachable within K hops; and are the relations and triple sets within the subgraph, respectively.

[0028] Preferably, the constructing of the dual-channel adaptive fusion feature module in step S3 includes constructing a dual-channel modeling including structural channel modeling and semantic channel modeling;

[0029] The structural channel modeling adopts the improved graph neural network model UnionGIN to process the structural characteristics of different types of nodes and edges through multi-relation aggregation mechanism and hierarchical message passing. In the layer, for the fusion knowledge graph Each node in The structural representation update formula is:

[0030] ;

[0031] ;

[0032] ;

[0033] in, is a node No. Layer embedding representation, is a node No. Layer embedding representation, It is The learnable parameters of the layer, Representation node The set of neighbor nodes of Selected from Node The set of neighbor nodes forms and Adjacent edges , and is a multilayer perceptron with different parameters, For the cross-edge type conversion module, adjust or normalize the edge information, It's the edge The original edge structural coefficient, which represents the edge strength, It's the edge The normalized structural coefficient of is used for attention allocation or weight calculation;

[0034] Using edges in the hierarchical message passing The path matrix is ​​subjected to singular value decomposition (SVD) to extract the singular values , where P represents the path matrix, and denote the left singular vector matrix and the right singular vector matrix obtained by decomposition, respectively. Represents the diagonal matrix composed of singular values, The sum of the singular values ​​of , as an edge The local structure coefficient of Represents a slave node arrive The path matrix reflects the connectivity strength or weight of the path, It's the edge The original side structure coefficient of quantifies the global importance of the path matrix, Represents a slave node arrive The singular value diagonal matrix of the path matrix;

[0035] Finally, the embedding representation of each node on the fusion graph is obtained as , is the number of nodes, D represents the embedding dimension;

[0036] for Each edge (i.e., relationship) in , is the edge set of the fusion graph, and its embedding is expressed as:

[0037] ;

[0038] in, is a learnable matrix, is the bias, is the type of edge, ; is the number of edges, The embedding dimension representing the edge type;

[0039] The semantic channel modeling is based on medical knowledge base resources and uses a large language model to extract the unstructured text description of the entity. For each entity node in the target triple , there is a medical text description , including but not limited to one or more of drug mechanism of action, indications, and disease symptoms; using PubMed-BERT as a text encoder Encode it to obtain a high-dimensional semantic vector:

[0040] ;

[0041] in, , is the number of nodes, Represents the embedding dimension of the text.

[0042] Preferably, the construction of the dual-channel adaptive fusion feature module in step S3 further includes introducing a cross-modal multi-head attention module to adaptively allocate weights of structural and semantic features, and dynamically fuse structural embedding and semantic embedding through the attention mechanism; which specifically includes:

[0043] Obtaining node embedding representation of target drug-disease subgraph from fused knowledge graph ,in is the number of nodes in the subgraph, and D represents the embedding dimension; Indicates that from drugs and diseases Starting from, the set of nodes reachable within K hops;

[0044] ;

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] in, Representation node The structural embedding of Representation node Text embedding, , , Respectively The projection parameter matrix of the heads, the value range of i is the number of heads, is the number of heads, represents the embedding dimension, is the projection matrix, used to map the multi-head results back to the original dimensional space; 、 、 is the input vector of the i-th attention head in the multi-head cross attention mechanism, representing the query, key, and value respectively;

[0050] Finally, the structural and semantic features are concatenated and linearly transformed to obtain a fused embedding representation:

[0051] ;

[0052] in, For the final fusion embedding representation, For normalization, Node embedding representation of the target drug-disease subgraph; represents the attention head output; is a multi-layer perceptron;

[0053] All nodes are aggregated through multiple layers to form a unified multimodal embedding representation, which serves as the basic input for subsequent subgraph learning and relationship classification.

[0054] Preferably, the enhanced splicing in step S4 specifically includes:

[0055] Assume that there are Target drug-disease pairs , then the spliced ​​subgraph is defined as:

[0056] ;

[0057] On the basis of preserving the structure of the atomic graph, enhanced connections are constructed on the spliced ​​subgraph, specifically including three types of virtual edges: drug-drug similarity edges, disease-disease similarity edges, and drug-disease virtual association edges.

[0058] Preferably, the introducing the iterative update learning mechanism in step S4 specifically includes:

[0059] An end-to-end subgraph learning mechanism based on edge-node collaborative updating dynamically updates subgraph edge weights and optimizes node embeddings in each iteration. After the spliced ​​subgraph is constructed, each edge is assigned an initialized connection strength value, which comprehensively considers the importance weight of known DDAs, the potential association probability between unknown drug-disease pairs, and the semantic similarity between drugs and drugs, and diseases and diseases.

[0060] For any two drugs in the batch graph and , whose initial edge weight is , use Jaccard similarity to calculate the neighborhood structure similarity of two drug nodes and assign edge weights:

[0061] ;

[0062] ;

[0063] in, Indicates drug and Jaccard similarity, and is the drug in the batch graph, is the indicator function, is the threshold, is the dimension index of the embedding vector, is the total dimension of the embedding vector, It's a drug The embedding vector of Indicates embedding the drug into the vector Passing the threshold The indicator vector formed after binarization is Indicates embedding the drug into the vector Passing the threshold The indicator vector formed after binarization is It's a drug The embedding vector of Is the weight balance parameter, which controls the degree of fusion between the old edge weight and the new edge similarity. To update the weights behind, Indicates updating the weight of all edges;

[0064] For any two disease nodes in the batch graph , calculate the cosine similarity in the embedding space and assign edge weights:

[0065] ;

[0066] ;

[0067] in, Indicates disease and Cosine similarity, and represents any two disease nodes in the batch graph, and is the embedding of the disease node, Disease nodes and The updated relationship weight between Disease nodes and The original relationship weight between is the weight balancing parameter, Represents the L2 norm of the vector;

[0068] For the drug node in the batch graph and disease nodes , the edge weights between them are:

[0069] ;

[0070] in, Represents a drug node and disease nodes The edge weights between and Represent drug nodes and disease nodes Embedded, Represents vector concatenation operation, is a multi-layer perceptron;

[0071] Afterwards, in each round In the iteration, considering the embedding difference between node u and node v in the previous iteration and the learnable embedding of relation r, the node representation of the previous round is used. and relation embedding ,use The multi-layer perceptron evaluates the semantic importance score of the edge:

[0072] ;

[0073] in Indicates the Node during round iteration and nodes The semantic importance score of Indicates the MLP multi-layer perceptron function during round iteration, represents the exponential function;

[0074] Current round rights A weighted combination of historical edge weights and scoring results:

[0075] ;

[0076] in is a hybrid hyperparameter that controls the New score for round iteration Hedi Side rights in the wheel The weight ratio of

[0077] In the subgraph Next, we perform neighbor message passing based on the updated edge weights and update the node representation:

[0078] ;

[0079] in, Representation node In the The updated node representation of the round, represents the activation function, Representation node The neighbor set of Represents an edge No. The dynamic weight of the round iteration, Represents neighbor nodes The embedding representation in the previous round.

[0080] Preferably, the outputting of the predicted probability of the drug-disease relationship in step S5 specifically includes: obtaining the local feature representation of the target drug finally learned by the subgraph , characteristic representation of target disease And the contextual feature representation of the target drug-disease pair:

[0081] ;

[0082] in Representation node All neighbor context nodes of are embedded;

[0083] Through feature concatenation, drug features, disease features, and context features are combined into a unified high-dimensional feature vector:

[0084] ;

[0085] The unified high-dimensional feature vector Input to the fully connected layer, which maps the feature space to the relation classification space through linear transformation and activation function mapping:

[0086] ;

[0087] in, and is the fully connected layer parameter; Indicates the number of predefined relationship categories; is a unified high-dimensional feature vector; Drug-disease pair The multi-class score vector.

[0088] Finally, the score vector is normalized into a probability distribution through the normalization function Softmax for classification prediction:

[0089] ;

[0090] in, Indicates that the sample belongs to The predicted probability of the class relationship, Indicates the Drug-disease pairs The multi-class scoring vector of the relationship is used for subsequent drug-disease relationship decision-making.

[0091] Under the same technical concept, the present invention also provides a drug-disease relationship prediction system based on a dual-channel fusion knowledge graph, the prediction system comprising:

[0092] External knowledge fusion module: used to obtain drug-disease association triples and external knowledge graph data to build a biomedical fusion knowledge graph;

[0093] Subgraph construction module: used to construct drug-disease subgraph based on biomedical fusion knowledge graph;

[0094] Dual-channel adaptive fusion feature embedding module: used to build a dual-channel adaptive fusion feature module, encode the drug-disease subgraph into a vector embedding, and obtain a dual-channel fusion knowledge graph pre-embedding network;

[0095] Subgraph learning module: This module is used to enhance the splicing of the drug-disease subgraphs pre-embedded in the dual-channel fusion knowledge graph network. It introduces an edge-node iterative update learning mechanism, trains the spliced ​​subgraph relationship perception learning network, and obtains enhanced subgraph feature embedding.

[0096] Prediction probability output module: used to calculate and output the predicted probability of the drug-disease relationship using the trained enhanced subgraph feature embedding.

[0097] The above solution of the present invention has the following beneficial effects:

[0098] This paper proposes a drug-disease relationship prediction method based on a dual-channel fusion knowledge graph. It systematically integrates the semantic representation capabilities of the knowledge graph with the topological modeling advantages of graph neural networks. Through multimodal feature interaction and dynamic subgraph learning mechanisms, it achieves efficient and accurate prediction of drug-disease associations.

[0099] Specifically, the present invention first supplements the original DDA triple data with multi-source heterogeneous knowledge, integrating biomedical entity descriptions to construct a fused biomedical knowledge graph. Furthermore, a structural semantic dual-channel graph embedding module, consisting of UnionGIN and PubMed-BERT, is employed to capture global and local features of the graph from the perspectives of graph structure and textual semantics, respectively. Adaptive deep fusion of features is achieved through a cross-modal attention alignment mechanism. The attention weights of noisy features are automatically attenuated, while the weights of key features are increased, achieving multimodal feature complementarity and improving representational integrity. During training, the present invention employs a subgraph learning paradigm. Using a path-aware heuristic sampling algorithm, breadth-first search (BFS) is used to extract K-hop paths between target drug-disease pairs. The method selects the most meaningful subgraph nodes and relationships, focusing on key nodes to reduce noise interference from irrelevant nodes. Furthermore, virtual connection enhancement is used to increase the average node degree of sparse subgraphs, alleviating the problem of poor correlation. During the subgraph learning phase, an end-to-end iterative update mechanism is employed. Through bidirectional, alternating message passing between nodes and edges, the edge weights are dynamically adjusted to adjust the subgraph structure, alleviating over-reliance on topological structures and exploring deep-seated association patterns between drugs and diseases. Ultimately, to achieve classification prediction output, the present invention combines the target drug, target disease, and their contextual feature representations into a unified high-dimensional feature vector through feature concatenation. This vector is then fed into a fully connected layer. Through linear transformation and activation function mapping, the feature space is mapped to the relational classification space, completing multi-classification predictions of the drug-disease relationship.

[0100] Experimental results show that DualSubNet significantly outperforms existing baseline models on multiple benchmark datasets, demonstrating its advancement and superiority in drug repurposing prediction tasks.

[0101] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] Figure 1 This is a diagram of the DualSubNet architecture of the present invention;

[0103] Figure 2 A schematic diagram of the process of obtaining entity text description in the semantic channel modeling of the present invention;

[0104] Figure 3 Schematic diagram of the dual-channel adaptive fusion feature module structure of the present invention;

[0105] Figure 4 Schematic diagram of a process of drug-disease relationship prediction method based on dual-channel fusion knowledge graph of the present invention; DETAILED DESCRIPTION

[0106] To make the technical problems, technical solutions, and advantages to be solved by the present invention more clear, the following is a detailed description with reference to the accompanying drawings and specific embodiments. It is obvious that the embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0107] In the description of the present invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0108] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood broadly. For example, they may refer to a locking connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0109] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0110] Example 1:

[0111] This embodiment provides a drug-disease relationship prediction method based on a dual-channel fusion knowledge graph, the flow chart of which is as follows: Figure 4 As shown, the following steps are included:

[0112] Step 1: Data acquisition and processing.

[0113] This embodiment selects the drug-disease association (DDA) triple data provided by the Precision Medicine Knowledge Graph (PrimeKG) and the Drug Repositioning Database (repoDB) as training knowledge. The node types include drugs and diseases, and the relationship types are verified drug-disease association relationships. The biomedical heterogeneous knowledge graph (HeterogeneousInformation Network, Hetionet) is selected as the external knowledge graph to supplement the structured background knowledge of the basic graph. Hetionet contains a wealth of entity types (such as proteins, genes, pathways, etc.) and multi-source relationship types, which can provide structured biomedical background knowledge for drug-disease association prediction. The graph data are all stored in triple format, and the triple form is represented as ,in It is a collection of entities, including drugs, diseases, genes, etc. It is a set of relationships, including drug-disease association (DDA), protein-protein interaction (PPI), gene regulation (Regulates) and other relationships. Represents a directed knowledge triple, h is the head entity, t is the tail entity, and r is the semantic relationship between them.

[0114] The datasets were then preprocessed. First, isolated nodes (nodes without connections) were removed from the external knowledge graph Hetionet, and drug and disease nodes were filtered. Next, the PrimeKG and repoDB datasets were preprocessed to extract drug-disease relationship triplets. Table 1 shows the analysis of the PrimeKG and repoDB datasets, and Table 2 shows the analysis of the Hetionet dataset.

[0115] Table 1 Data analysis of PrimeKG and repoDB

[0116]

[0117] Table 2 Hetionet data analysis

[0118]

[0119] The drug-disease subgraph is then deeply fused with the triple data in Hetionet to construct a unified comprehensive knowledge graph. Entities and relationships are encoded using unified ID naming rules to construct PrimeKG-Hetionet fusion graphs and repoDB-Hetionet fusion graphs. The entity alignment relationships across graphs are retained and the triple structure of the fusion graphs (e.g., <drug, association, disease>, <gene, regulation, drug>, etc.) is extracted. Expressed as:

[0120] ;

[0121] in, It indicates that there are N nodes, M edges, and T triplets in the fusion graph, including entity alignment and cross-graph association information.

[0122] The data analysis of PrimeKG-Hetionet fusion map and repoDB-Hetionet fusion map is shown in Table 3:

[0123] Table 3 Data analysis of PrimeKG-Hetionet and repoDB-Hetionet

[0124]

[0125] Finally, the processed triplet dataset is divided into 10-fold cross-validation methods: 80% of the data is used as a training set; 10% of the data is used as a validation set; and 10% of the data is used as a test set.

[0126] Step 2: Construction of drug-disease subgraph based on biomedical fusion knowledge graph path.

[0127] Integrating knowledge graphs In this example, the target drug-disease pair As the center, a path-based K-hop subgraph sampling strategy is used to model the local context of the target pair. ( are all target triplets), where Indicates medicine, Indicates disease, Represented as the semantic relationship between them; find the path through breadth-first traversal and extract its The local subgraph within the hop range is defined as:

[0128] ;

[0129] in, Indicates that drugs are captured by BFS and The K-hop local context of the departure includes nodes, relations, and triples. Indicates that from drugs and diseases Starting from, the set of nodes reachable within K hops; and are the relations and triple sets within the subgraph, respectively.

[0130] Constructing subgraphs allows us to focus on specific patterns or fragments of information that are of practical value, thereby reducing the impact of noise. At the same time, the path-based construction approach preserves the structural paths and multi-hop dependency information between entity pairs, helping to reveal potential high-order drug-disease associations.

[0131] Step 3: Two-channel adaptive graph embedding.

[0132] To capture richer semantic features, this embodiment models the network from two dimensions: structural channel and semantic channel.

[0133] 1. Structural channel modeling: In order to break through the limitations of traditional GNN on homogeneous graph structures, this embodiment adopts an improved graph neural network model UnionGIN to effectively process the structural characteristics of different types of nodes and edges through a multi-relationship aggregation mechanism and hierarchical message passing for highly heterogeneous entities and relationships in the knowledge graph. In the layer, for the fusion knowledge graph Each node in The structural representation update formula is:

[0134] ;

[0135] ;

[0136] ;

[0137] in, is a node No. Layer embedding representation, is a node No. Layer embedding representation, It is The learnable parameters of the layer, Representation node The set of neighbor nodes of Selected from Node The set of neighbor nodes forms and Adjacent edges , and is a multilayer perceptron with different parameters, For the cross-edge type conversion module, adjust or normalize the edge information, It's the edge The original edge structural coefficient, which represents the edge strength, It's the edge The normalized structural coefficient of is used for attention allocation or weight calculation;

[0138] Using edges in the hierarchical message passing The path matrix is ​​decomposed into singular values ​​and the singular values ​​are extracted. , where P represents the path matrix, and denote the left singular vector matrix and the right singular vector matrix obtained by decomposition, respectively. Represents the diagonal matrix composed of singular values, The sum of the singular values ​​of , as an edge The local structure coefficient of Represents a slave node arrive The path matrix reflects the connectivity strength or weight of the path, It's the edge The original side structure coefficient of quantifies the global importance of the path matrix, Represents a slave node arrive The singular value diagonal matrix of the path matrix;

[0139] Finally, the embedding representation of each node on the fusion graph is obtained as , is the number of nodes, H is the embedding representation of all nodes, and D is the embedding dimension.

[0140] for Each edge (i.e., relationship) in , is the edge set of the fusion graph, and its embedding is expressed as:

[0141] ;

[0142] in, is a learnable matrix, is the bias, is the type of edge, ; is the number of edges, The embedding dimension representing the edge type;

[0143] 2. Semantic channel modeling: To improve the richness and domain adaptability of semantic features, this embodiment uses authoritative medical knowledge base resources such as the DrugBank database and the U.S. National Library of Medicine document retrieval system PubMed, and uses a large language model to extract unstructured text descriptions of entities (such as drug indications, drug mechanisms of action, disease symptoms, clinical trial results, etc.). The flowchart for obtaining entity text descriptions is shown in the following figure. Figure 2 As shown. For each entity node in the target triple , there is a medical text description , including drug mechanism of action, indications, disease symptoms, etc. This example uses PubMed-BERT as the text encoder , thus obtaining a high-dimensional semantic vector:

[0144] ;

[0145] in, , is the number of nodes, Represents the embedding dimension of the text.

[0146] At the same time, this embodiment designs a cross-modal multi-head attention module to adaptively allocate the weights of structural and semantic features, and dynamically fuses structural embedding and semantic embedding through the attention mechanism, which not only retains the topological structure information of the knowledge graph, but also incorporates the semantic context of the entity, providing more comprehensive feature support for subsequent tasks (such as drug repositioning and disease mechanism prediction). The dual-channel adaptive fusion feature module is as follows: Figure 3 As shown. We first obtain the node embedding representation of the target drug-disease subgraph from the fusion knowledge graph ,in is the number of nodes in the subgraph, D represents the embedding dimension, Indicates that from drugs and diseases Starting from, the set of nodes reachable within K hops;

[0147] ;

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] in, Representation node The structural embedding of Representation node Text embedding, , , Respectively The projection parameter matrix of the heads, the value range of i is the number of heads, is the number of heads, represents the embedding dimension, is the projection matrix, used to map the multi-head results back to the original dimensional space; 、 、 is the input vector of the i-th attention head in the multi-head cross attention mechanism, representing the query, key, and value respectively;

[0153] Finally, the structural and semantic features are concatenated and linearly transformed to obtain a fused embedding representation:

[0154] ;

[0155] in, For the final fusion embedding representation, For normalization, Node embedding representation of the target drug-disease subgraph; represents the attention head output; is a multi-layer perceptron;

[0156] All nodes are aggregated through multiple layers to form a unified multimodal embedding representation, which serves as the basic input for subsequent subgraph learning and relationship classification.

[0157] Step 4: Subgraph enhancement splicing and subgraph iterative update learning mechanism based on end-to-end architecture.

[0158] Since drug-disease association data is usually highly sparse (i.e., a large number of drug-disease pairs lack direct association evidence), the information contained in a single subgraph may not be sufficient to support effective association prediction. In order to improve the information density and generalization ability of local subgraphs, this embodiment further designs a subgraph splicing mechanism based on structural and semantic similarity, and enhances the association ability of sparse networks through virtual connections. This embodiment splices the subgraphs of all target DDAs in the same batch (Batch), and constructs a larger-scale spliced ​​subgraph while retaining the atomic graph structure. Assume that there are a total of Target drug-disease pairs , then the spliced ​​subgraph is defined as:

[0159] ;

[0160] On the basis of preserving the structure of the atomic graph, enhanced connections are constructed on the spliced ​​subgraph, specifically including three types of virtual edges: drug-drug similarity edges, disease-disease similarity edges, and drug-disease virtual association edges.

[0161] In order to achieve high efficiency and accuracy in drug-disease association prediction, this embodiment designs an end-to-end subgraph learning mechanism based on edge-node collaborative updating, which dynamically updates the subgraph edge weights and optimizes node embeddings in each round of iteration to explore the deep association patterns between drugs and diseases. After the splicing subgraph is constructed, each edge is assigned an initialized connection strength value, which comprehensively considers the importance weight of known DDA, the potential association probability between unknown drug-disease pairs, and the semantic similarity between drug-drug and disease-disease. By introducing an iterative optimization strategy, the edge weights in the subgraph are dynamically adjusted in each iteration according to their relevance to the DDA prediction task, so as to gradually reduce the weights of redundant or noisy edges and thus streamline the subgraph structure, focusing on the association information that contributes most to the prediction task.

[0162] For any two drugs in the batch graph and , whose initial edge weight is , use Jaccard similarity to calculate the neighborhood structure similarity of two drug nodes and assign edge weights:

[0163] ;

[0164] ;

[0165] in, Indicates drug and Jaccard similarity, and is the drug in the batch graph, is the indicator function, is the threshold, is the dimension index of the embedding vector, is the total dimension of the embedding vector, It's a drug The embedding vector of Indicates embedding the drug into the vector Passing the threshold The indicator vector formed after binarization is Indicates embedding the drug into the vector Passing the threshold The indicator vector formed after binarization is It's a drug The embedding vector of Is the weight balance parameter, which controls the degree of fusion between the old edge weight and the new edge similarity. To update the weights behind, Indicates updating the weight of all edges;

[0166] For any two disease nodes in the batch graph , calculate the cosine similarity in the embedding space and assign edge weights:

[0167] ;

[0168] ;

[0169] in, Indicates disease and Cosine similarity, and Represents any two disease nodes in the batch graph, and is the embedding of the disease node, Disease nodes and The updated relationship weight between Disease nodes and The original relationship weight between is the weight balancing parameter, Represents the L2 norm of the vector;

[0170] For the drug node in the batch graph and disease nodes , the edge weights between them are:

[0171] ;

[0172] in, Represents a drug node and disease nodes The edge weights between and Represent drug nodes and disease nodes Embedded, Represents vector concatenation operation, is a multi-layer perceptron;

[0173] At this point we have completed the initial update of all edge weights in the subgraph.

[0174] Afterwards, in each round In the iteration, considering the embedding difference between node u and node v in the previous iteration and the learnable embedding of relation r, the node representation of the previous round is used. and relation embedding ,use The multi-layer perceptron evaluates the semantic importance score of the edge:

[0175] ;

[0176] in Indicates the Node during round iteration and nodes The semantic importance score of Indicates the MLP multi-layer perceptron function during round iteration, represents the exponential function;

[0177] Current round rights A weighted combination of historical edge weights and scoring results:

[0178] ;

[0179] in is a hybrid hyperparameter that controls the New score for round iteration Hedi Side rights in the wheel The weight ratio of

[0180] In the subgraph Next, we perform neighbor message passing based on the updated edge weights and update the node representation:

[0181] ;

[0182] in, Representation node In the The updated node representation of the round, represents the activation function, Representation node The neighbor set of Represents an edge No. The dynamic weight of the round iteration, Represents neighbor nodes The embedding representation in the previous round.

[0183] After multiple rounds of iteration, the node representations of drugs and diseases are gradually mapped from a general feature space to a specialized representation space for DDA prediction tasks, significantly improving prediction performance. Furthermore, this mechanism effectively and dynamically prunes redundant edges and enhances potential associations, allowing important semantic information to be conveyed even under sparse or missing conditions, thereby improving the robustness and accuracy of drug-disease association prediction.

[0184] Step 5: The classification layer outputs the predicted probability.

[0185] After multiple rounds of iterative optimization and learning, both the drug node and the disease node are embedded in the optimal latent space vector representation, ensuring efficient encoding and semantic expression of node features. This embodiment uses the subgraph learning mechanism to extract the local feature representation of the target drug. (focusing on the drug's own attributes and directly related node information), characteristic representation of the target disease (covering disease-related features and topological structure information), as well as contextual feature representation of target drug-disease pairs (By calculating the average embedding vector of all nodes in the subgraph, the interaction pattern between drugs and diseases in the global network is comprehensively reflected), where Representation node All neighbor context nodes of are embedded;

[0186] Then, the above three types of features are deeply fused: first, through feature splicing operation, the drug features, disease features and context features are combined into a unified high-dimensional feature vector in terms of dimension. ; Then the concatenated feature vector is input into the fully connected layer, which maps the feature space to the relation classification space through linear transformation and activation function mapping:

[0187] ;

[0188] in, and is the fully connected layer parameter; Represents the number of predefined relationship categories, is a unified high-dimensional feature vector, Drug-disease pair The multi-class score vector.

[0189] Finally, the score vector is normalized into a probability distribution through the normalization function Softmax for classification prediction:

[0190] ;

[0191] in, Indicates that the sample belongs to The predicted probability of the class relationship, Indicates the Drug-disease pairs The multi-class scoring vector of the relationship is used for subsequent drug-disease relationship decision-making.

[0192] A drug-disease relationship prediction system based on a dual-channel fusion knowledge graph, including:

[0193] External knowledge fusion module: used to obtain drug-disease association triples and external knowledge graph data to build a biomedical fusion knowledge graph;

[0194] Subgraph construction module: used to construct drug-disease subgraph based on biomedical fusion knowledge graph;

[0195] Dual-channel adaptive fusion feature embedding module: used to build a dual-channel adaptive fusion feature module, encode the drug-disease subgraph into a vector embedding, and obtain a dual-channel fusion knowledge graph pre-embedding network;

[0196] Subgraph learning module: This module is used to enhance the splicing of the drug-disease subgraphs pre-embedded in the dual-channel fusion knowledge graph network. It introduces an edge-node iterative update learning mechanism, trains the spliced ​​subgraph relationship perception learning network, and obtains enhanced subgraph feature embedding.

[0197] Prediction probability output module: used to calculate and output the predicted probability of the drug-disease relationship using the trained enhanced subgraph feature embedding.

[0198] The DualSubNet architecture of this embodiment is shown in the figure below: Figure 1 As shown, it is advanced in handling drug-disease relationship prediction and classification tasks. It can simultaneously capture the local characteristics and global interaction patterns of drugs and diseases through a dual-channel subgraph learning mechanism, and achieve efficient relationship classification with the help of feature splicing and linear classification layers. This example conducts drug-disease prediction experiments on two datasets: the fusion data of the precision medicine knowledge graph and the biomedical heterogeneous knowledge graph (PrimeKG-Hetionet), and the fusion data of the drug repositioning database and the biomedical heterogeneous knowledge graph (repoDB-Hetionet). For example, given the drug fosinopril and the disease hypertensive, the relationship between them needs to be predicted as an indication.

[0199] Furthermore, this example selected drug and disease pairs from the DrugBank data to test and validate the DualSubNet model. The specific approach is as follows: The DualSubNet model inputs drug and disease pairs, and the model outputs predicted relationships. In the outputted predicted relationships, a positive sign indicates treatment / prevention, an ineffective sign indicates ineffectiveness, and a negative sign indicates side effects or worsening of the disease. The specific predicted and actual relationships are shown in Table 4 below:

[0200] Table 4 Performance test of the DualSubNet model using DrugBank data

[0201]

[0202] The results show that the model and method of this embodiment have great application potential in terms of prediction accuracy, relationship category discrimination and generalization ability, especially showing significant advantages in the scenarios of drug repositioning and disease mechanism analysis in complex biological networks.

[0203] Comparative Example 1:

[0204] In order to comprehensively study the performance of DualSubNet, the present invention introduces some other existing advanced methods and compares the performance of these methods with the method proposed in the present invention in the experiments below.

[0205] AMDGT proposes a multimodal fusion method based on a dual-graph transformer and an attention mechanism. By integrating the similarity network of drugs and diseases with heterogeneous association networks, it uses a modal interaction module to capture deep features and improve the performance and generalization ability of drug-disease association prediction.

[0206] AdaDR significantly improves the performance of drug repurposing tasks by combining information from feature space and topological space, using the attention mechanism to dynamically adjust embedding weights, and enhancing the model's generalization ability through consistency constraints.

[0207] DSE-HNGCN proposes a deep learning framework based on heterogeneous networks and graph convolutional networks. It constructs a heterogeneous network by integrating multi-source information (drug similarity, side effect semantic similarity, and frequency information), and uses a multi-layer graph convolutional network and layer importance combination strategy to effectively predict the frequency of drug side effects.

[0208] SUMGNN reduces noise and improves efficiency by leveraging external biomedical knowledge and extracting local subgraphs related to target drug pairs from the knowledge graph. At the same time, it generates reasoning paths through the self-attention mechanism, extracts the most useful information for prediction, efficiently predicts drug-drug interactions, and generates explainable reasoning paths.

[0209] NAGTLDA proposes a graph Transformer model that combines local feature learning, global feature learning and structural encoding. Through an adaptive feature fusion mechanism, it efficiently predicts the unknown associations between lncRNAs and diseases, significantly improving the performance of the model on large-scale sparse datasets.

[0210] LaGAT proposes a link-aware graph attention mechanism that can generate different attention paths according to different drug pairs. At the same time, it dynamically generates attention paths to capture the semantic diversity of drug nodes in different drug pairs, thereby improving the accuracy and interpretability of predictions.

[0211] HGTDR proposes an end-to-end method based on heterogeneous graph transformer (HGT), which uses HGT for drug reuse for the first time. It can automatically process complex relationships in heterogeneous graphs and avoid information loss. By automatically processing large-scale heterogeneous knowledge graphs, it achieves efficient prediction of drug-disease relationships and demonstrates excellent performance in multiple tasks.

[0212] LDAGM proposes a model based on deep topological feature extraction and graph convolutional autoencoder, which can fuse functional similarity and Gaussian kernel similarity to solve the sparsity problem. It predicts the relationship between lncRNA and disease through multi-view heterogeneous network fusion and multi-layer perceptron. At the same time, it introduces an aggregation layer in the MLP and optimizes feature extraction through a gating mechanism, significantly improving the accuracy and stability of prediction.

[0213] Since the data and data processing methods used by these methods may differ, in order to achieve data uniformity and model fairness, eliminate bias caused by differences in data input or classification layers, and focus on the performance differences of the model architecture itself, the specific implementation methods of this comparison include:

[0214] (1) Standardization of data input:

[0215] The same training set, validation set, and test set as this model are used for all comparison models to ensure consistent data division.

[0216] This comparative example also employed the same PrimeKG-Hetionet and repoDB-Hetionet datasets as used in the examples above for the remaining data models. Comprehensive comparative experiments systematically evaluated the performance of DualSubNet against other advanced methods in drug-disease association prediction. The PrimeKG dataset contains more complex and rich biomedical background knowledge, while the repoDB dataset offers greater real-world applicability, demonstrating the model's generalization capabilities in drug repurposing scenarios.

[0217] (2) Adaptability modification of the classification layer:

[0218] The final classification layer of all comparison models was replaced with the same structure to output the drug-disease association probability.

[0219] In order to make a fair comparison on the classification performance of the models, this comparison standardizes the data input of all models and modifies the corresponding classification layers.

[0220] Table 5 shows that on the PrimeKG-Hetionet dataset, although the proposed DualSubNet model is slightly lower than the LDAGM model by 0.32% in terms of the AUROC indicator, it achieves optimal performance in both the ACC and Weighted_F1 indicators, reaching an accuracy rate (ACC) of 92.76% and a weighted F1 score of 92.52%, respectively, which is superior to the comparison method. This shows that although the DualSubNet model is slightly inferior to the LDAGM in ranking, it improves the recognition ability of small sample categories through subgraph construction and information enhancement mechanisms, and is better optimized under the same classification threshold.

[0221] Experimental results on the repoDB-Hetionet dataset are shown in Table 6. DualSubNet achieves 95.68% in the AUROC metric, the highest among all models, demonstrating strong discriminative capabilities. Although the LDAGM model is slightly more accurate than the present invention, this may be due to LDAGM's preference for high-confidence predictions of the primary class, which performs well in repoDB, where the class distribution is uneven. However, LDAGM neglects minority class recall, resulting in a decrease in the weighted F1 score. DualSubNet outperforms LDAGM in the weighted F1 score, achieving more balanced overall performance and strong generalization. In particular, compared to weaker methods such as LaGAT and SUMGNN, DualSubNet achieves significant improvements across multiple metrics.

[0222] The experimental results fully demonstrate that DualSubNet outperforms existing advanced methods in multiple key performance indicators, has stronger accuracy, robustness and generalization capabilities, and provides a more effective technical path for the prediction task of drug-disease relationship.

[0223] Table 5 Performance comparison of DualSubNet and other advanced methods on PrimeKG-Hetionet dataset

[0224]

[0225] Table 6 Performance comparison of DualSubNet and other advanced methods on the repoDB-Hetionet dataset

[0226]

[0227] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A drug-disease relationship prediction method based on dual-channel fusion knowledge graph, characterized by: The following steps are involved: S1. Obtain drug-disease association triples and external knowledge graph data to construct a biomedical fusion knowledge graph; S2. Construct drug-disease subgraph based on biomedical fusion knowledge graph; S3. Constructing a dual-channel adaptive fusion feature module, embedding the drug-disease subgraph, and obtaining a dual-channel fusion knowledge graph pre-embedding network; constructing the dual-channel adaptive fusion feature module includes constructing a dual-channel model including structural channel modeling and semantic channel modeling; The structural channel modeling adopts the improved graph neural network model UnionGIN through multi-relation aggregation mechanism and hierarchical message passing to process the structural characteristics of different types of nodes and edges; The semantic channel modeling is based on medical knowledge base resources and uses a large language model to extract the unstructured text description of the entity. For each entity node in the target triple , there is a medical text description , including one or more of the drug's mechanism of action, indications, and disease symptoms; The construction of the dual-channel adaptive fusion feature module also includes introducing a cross-modal multi-head attention module to adaptively allocate weights of structural and semantic features, and dynamically fuse structural embedding and semantic embedding through the attention mechanism; S4. Enhance the splicing of the drug-disease subgraph in the dual-channel fusion knowledge graph pre-embedded network, introduce the edge-node iterative update learning mechanism, train the spliced ​​subgraph relationship perception learning network, and obtain the enhanced subgraph feature embedding; S5. Use the trained enhanced subgraph feature embedding to calculate the predicted probability of the output drug-disease relationship.

2. The prediction method according to claim 1, wherein: The step S1 of obtaining drug-disease association triples and external knowledge graph data specifically includes: selecting drug-disease association triples provided by the precision medicine knowledge graph and the drug repositioning database as training knowledge, wherein the node types include drug and disease, and the relationship type is a verified drug-disease association relationship; selecting a biomedical heterogeneous knowledge graph as the external knowledge graph; The construction of the biomedical fusion knowledge graph specifically includes: Remove isolated nodes from the external knowledge graph and filter out drug nodes and disease nodes; Extract drug-disease association triples from the precision medicine knowledge graph and drug repositioning database; The drug-disease association triple data is deeply integrated with drug nodes and disease nodes to construct a biomedical fusion knowledge graph; the drug-disease association triple data is divided into training set, validation set and test set.

3. The prediction method according to claim 1, wherein: The construction of the drug-disease subgraph in step S2 specifically includes: Target drug-disease pair As the center, a path-based K-hop subgraph sampling strategy is used to model the local context of the target pair. For a target triple ,in Indicates medicine, Indicates disease, Represented as the semantic relationship between them, Represents all target triplets; find the path through breadth-first traversal and extract the local subgraph within its jump range from the fusion graph, which is defined as: ； in, Indicates that drugs are captured by BFS and The K-hop local context of the departure includes nodes, relations, and triples. Indicates that from drugs and diseases Starting from, the set of nodes reachable within K hops; and are the relations and triple sets within the subgraph, respectively.

4. The prediction method according to claim 1, wherein: The structural channel modeling in step S3 uses the improved graph neural network model UnionGIN to process the structural features of different types of nodes and edges through multi-relation aggregation mechanism and hierarchical message passing. In the layer, for the fusion knowledge graph Each node in The structural representation update formula is: ； ； ； in, is a node No. Layer embedding representation, is a node No. Layer embedding representation, It is The learnable parameters of the layer, Representation node The set of neighbor nodes of Selected from Node The set of neighbor nodes forms and Adjacent edges , and is a multilayer perceptron with different parameters, For the cross-edge type conversion module, adjust or normalize the edge information, It's the edge The original edge structural coefficient, which represents the edge strength, It's the edge The normalized structural coefficient of is used for attention allocation or weight calculation; Using edges in the hierarchical message passing The path matrix is ​​decomposed into singular values ​​and the singular values ​​are extracted. , where P represents the path matrix, and denote the left singular vector matrix and the right singular vector matrix obtained by decomposition, respectively. Represents the diagonal matrix composed of singular values, The sum of the singular values ​​of , as an edge The local structure coefficient of Represents a slave node arrive The path matrix reflects the connectivity strength or weight of the path, It's the edge The original side structure coefficient of quantifies the global importance of the path matrix, Represents a slave node arrive The singular value diagonal matrix of the path matrix; Finally, the embedding representation of each node on the fusion graph is obtained as , is the number of nodes, H is the embedding representation of all nodes, and D is the embedding dimension; for Each edge in , is the edge set of the fusion graph, and its embedding is expressed as: ； in, is a learnable matrix, is the bias, is the type of edge, ; is the number of edges, Embedding dimension representing edge type; The semantic channel modeling uses PubMed-BERT as the text encoder Encode it to obtain a high-dimensional semantic vector: ； in, , is the number of nodes, Represents the embedding dimension of the text.

5. The prediction method according to claim 4, wherein: In step S3, a cross-modal multi-head attention module is introduced to adaptively allocate the weights of structural and semantic features, and dynamically fuse structural embedding and semantic embedding through the attention mechanism; Specifically include: Obtaining node embedding representation of target drug-disease subgraph from fused knowledge graph ,in is the number of nodes in the subgraph, and D represents the embedding dimension; Indicates that from drugs and diseases Starting from, the set of nodes reachable within K hops; ； ； ； ； ； in, Representation node The structural embedding of Representation node Text embedding, , , Respectively The projection parameter matrix of the heads, the value range of i is the number of heads, is the number of heads, represents the embedding dimension, is the projection matrix, used to map the multi-head results back to the original dimensional space; 、 、 is the input vector of the i-th attention head in the multi-head cross attention mechanism, representing the query, key, and value respectively; Finally, the structural and semantic features are concatenated and linearly transformed to obtain a fused embedding representation: ； in, For the final fusion embedding representation, For normalization, Node embedding representation of the target drug-disease subgraph; represents the attention head output; is a multi-layer perceptron; All nodes are aggregated through multiple layers to form a unified multimodal embedding representation, which serves as the basic input for subsequent subgraph learning and relationship classification.

6. The prediction method according to claim 1, wherein: The enhanced splicing in step S4 specifically includes: Assume that there are Target drug-disease pairs , then the spliced ​​subgraph is defined as: ； On the basis of preserving the structure of the atomic graph, enhanced connections are constructed on the spliced ​​subgraph, specifically including three types of virtual edges: drug-drug similarity edges, disease-disease similarity edges, and drug-disease virtual association edges.

7. The prediction method according to claim 1, wherein: The imported edge-node iterative update learning mechanism in step S4 specifically includes: An end-to-end subgraph learning mechanism based on edge-node collaborative updates dynamically updates subgraph edge weights and optimizes node embeddings in each iteration. After the subgraph is constructed, each edge is assigned an initial connection strength value that comprehensively considers the importance weight of known DDAs, the potential association probability between unknown drug-disease pairs, and the semantic similarity between drugs and diseases. For any two drugs in the batch graph and , whose initial edge weight is , use Jaccard similarity to calculate the neighborhood structure similarity of two drug nodes and assign edge weights: ； ； in, Indicates drug and Jaccard similarity, and is the drug in the batch graph, is the indicator function, is the threshold, is the dimension index of the embedding vector, is the total dimension of the embedding vector, It's a drug The embedding vector of Indicates embedding the drug into the vector Passing the threshold The indicator vector formed after binarization is Indicates embedding the drug into the vector Passing the threshold The indicator vector formed after binarization is It's a drug The embedding vector of Is the weight balance parameter, which controls the degree of fusion between the old edge weight and the new edge similarity. To update the weights behind, Indicates updating the weight of all edges; For any two disease nodes in the batch graph , calculate the cosine similarity in the embedding space and assign edge weights: ； ； in, Indicates disease and Cosine similarity, and represents any two disease nodes in the batch graph, and is the embedding of the disease node, Disease nodes and The updated relationship weight between Disease nodes and The original relationship weight between is the weight balancing parameter, Represents the L2 norm of the vector; For the drug node in the batch graph and disease nodes , the edge weights between them are: ； in, Represents a drug node and disease nodes The edge weights between and Represent drug nodes and disease nodes Embedded, Represents vector concatenation operation, is a multi-layer perceptron; Afterwards, in each round In the iteration, considering the embedding difference between node u and node v in the previous iteration and the learnable embedding of relation r, the node representation of the previous round is used. and relation embedding ,use The multi-layer perceptron evaluates the semantic importance score of the edge: ； in Indicates the Node during round iteration and nodes The semantic importance score of Indicates the MLP multi-layer perceptron function during round iteration, represents the exponential function; Current round rights A weighted combination of historical edge weights and scoring results: ； in is a hybrid hyperparameter that controls the New score for round iteration Hedi Side rights in the wheel The weight ratio of In the subgraph Next, we perform neighbor message passing based on the updated edge weights and update the node representation: ； in, Representation node In the The updated node representation of the round, represents the activation function, Representation node The neighbor set of Represents an edge No. The dynamic weight of the round iteration, Represents neighbor nodes The embedding representation in the previous round.

8. The prediction method according to claim 6, wherein: The output of the predicted probability of the drug-disease relationship in step S5 specifically includes: obtaining the local feature representation of the target drug finally learned by the subgraph , characteristic representation of target disease And the contextual feature representation of the target drug-disease pair: ； in Representation node All neighbor context nodes of are embedded; Through feature concatenation, drug features, disease features, and context features are combined into a unified high-dimensional feature vector: ； The unified high-dimensional feature vector Input to the fully connected layer, which maps the feature space to the relation classification space through linear transformation and activation function mapping: ； in, and is the fully connected layer parameter; Indicates the number of predefined relationship categories; is a unified high-dimensional feature vector; Drug-disease pair Multi-class scoring vector; Finally, the score vector is normalized into a probability distribution through the normalization function Softmax for classification prediction: ； in, Represents drug-disease pair Belong to The predicted probability of the class relationship, Indicates the Drug-disease pairs The multi-class scoring vector of the relationship is used for subsequent drug-disease relationship decision-making.

9. A drug-disease relationship prediction system based on dual-channel fusion knowledge graph, characterized by: The prediction system includes: External knowledge fusion module: used to obtain drug-disease association triples and external knowledge graph data to build a biomedical fusion knowledge graph; Subgraph construction module: used to construct drug-disease subgraph based on biomedical fusion knowledge graph; Dual-channel adaptive fusion feature embedding module: used to construct a dual-channel adaptive fusion feature module, encode the drug-disease subgraph into a vector embedding, and obtain a dual-channel fusion knowledge graph pre-embedding network; the construction of the dual-channel adaptive fusion feature module includes constructing a dual-channel model including structural channel modeling and semantic channel modeling; The structural channel modeling adopts the improved graph neural network model UnionGIN through multi-relation aggregation mechanism and hierarchical message passing to process the structural characteristics of different types of nodes and edges; The semantic channel modeling is based on medical knowledge base resources and uses a large language model to extract the unstructured text description of the entity. For each entity node in the target triple , there is a medical text description , including one or more of the drug's mechanism of action, indications, and disease symptoms; The construction of the dual-channel adaptive fusion feature module also includes introducing a cross-modal multi-head attention module to adaptively allocate weights of structural and semantic features, and dynamically fuse structural embedding and semantic embedding through the attention mechanism; Subgraph learning module: This module is used to enhance the splicing of the drug-disease subgraphs pre-embedded in the dual-channel fusion knowledge graph network. It introduces an edge-node iterative update learning mechanism, trains the spliced ​​subgraph relationship perception learning network, and obtains enhanced subgraph feature embedding. Prediction probability output module: used to calculate and output the predicted probability of the drug-disease relationship using the trained enhanced subgraph feature embedding.

Images