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

Abstract

The 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 sub-graph based on the biomedicine fusion knowledge graph; s3, constructing a dual-channel adaptive fusion feature module, and embedding a drug-disease sub-graph; s4, performing enhanced splicing on the drug-disease sub-graphs in the dual-channel fusion knowledge graph pre-embedded network, importing an edge-node iterative updating learning mechanism, training a spliced sub-graph relation perception learning network, and obtaining enhanced sub-graph feature embedding; and S5, utilizing the trained enhanced subgraph features to embed, calculate and output the prediction probability of the drug-disease relationship. According to the method, the semantic representation capability of the knowledge graph and the topological modeling advantages of the graph neural network are integrated, and efficient and accurate prediction of the drug-disease relationship is realized through multi-modal feature interaction, sub-graph enhancement and a dynamic sub-graph learning mechanism.

Description

Technical Field

[0001] The present invention relates to the technical field of medical big data, and particularly relates to a method and system for predicting drug-disease relationships 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 accompanied by extremely high failure risks. This process covers five core stages, namely, initial drug discovery, preclinical evaluation, clinical research, review by the US Food and Drug Administration (FDA), and post-marketing safety monitoring. Usually, the research and development of a drug often requires an investment of billions of dollars and takes about 10 to 15 years. Among them, safety issues are one of the main reasons for R & D failure. If a candidate drug fails to pass Phase I due to safety issues (for example, the candidate drug is toxic to the human body), then all investments from early drug discovery to Phase I will be in vain. To alleviate this heavy burden, exploring new indications of approved drugs, that is, drug repurposing or repositioning, has become a wise choice. According to statistics, about 30% of the drugs approved by the FDA are found to have at least one new indication after approval. "Old drugs with new uses" usually do not need to verify safety again and can directly enter Phase II and Phase III clinical trials to evaluate their therapeutic effects on certain specific diseases. This strategy not only significantly shortens the time of traditional drug development, but also significantly reduces the R & D cost. At the same time, it can also improve the utilization efficiency of drugs and maximize their potential therapeutic value.

[0003] However, traditional drug repurposing research mainly relies on experimental verification and expert experience for manual screening. The process has a long cycle, high costs, and is limited by known drug action mechanisms and clinical data accumulation. There is a lack of systematic methods for exploring potential new indications, resulting in low efficiency in discovering new indications. In recent years, artificial intelligence has shown great potential in the field of drug R & D assistance. As an efficient artificial intelligence tool, the Knowledge Graph (KG) can integrate data from multiple sources, including complex heterogeneous information such as literature, clinical trials, and genomic data, covering a wide range of entities (such as drugs, diseases, proteins, or targets, etc.) and the intricate relationships between them (such as drug-drug interactions, drug-target pairing relationships, etc.), deeply revealing the complex network between entities in the biological system and intuitively showing the complex relationships between biomedical entities such as drugs, diseases, genes, and proteins. Through knowledge graph completion technology, the potential semantic features of triples can be deeply explored to identify new potential connections between diseases and drugs. In traditional research, researchers often use machine learning methods to analyze drug chemical structures, gene expressions, or electronic medical record data. However, such methods require manual feature definition and are difficult to integrate multi-modal data. Similarly, deep learning methods have also been applied to drug repurposing research. Scholars at home and abroad have input knowledge graph triple data into architectures such as convolutional neural networks (CNNs) and graph neural networks (GNNs) to extract entity embedding features and achieve the final research goals. However, such research methods often regard the neural network as a black box and are difficult to explain the biological basis of the prediction results.

[0004] There are still some limitations in the current knowledge graph completion methods for drug repurposing. For example, there are insufficient samples for supervised learning in the medical field, and the neighborhood information of the graph structure is insufficient. It is difficult for current knowledge graph completion methods to achieve ideal training results. To overcome the above problems, more and more researchers have tried to integrate the rich information of external knowledge graphs into drug repurposing research to improve the performance of downstream tasks. However, simply using graph embeddings often only utilizes topological structure information and does not use the entity's own attribute information. At the same time, as the scale of the knowledge graph continues to grow, the types of entities and relationships in the graph will continue to increase, and the heterogeneity of the knowledge graph will become more prominent, which poses new challenges in terms of how to effectively perform embedding representations for different types of relationships. In addition, although large biomedical knowledge graphs can utilize rich information, when performing node representations, the embedding results are often poor due to noise, and inputting the entire graph into the model will also increase the computational burden.

[0005] Some researchers have attempted to use subgraphs to precisely focus on key information for embedding learning. However, the current method of subgraph embedding learning depends on the fixed topological structure of the graph. The known topological structures of knowledge graphs may not be reliable, and the explicit graph structure does not necessarily reflect the true dependence relationship. Not all relationship paths provide important association information, and over-reliance may even introduce noise.

[0006] Therefore, there is still a lack of a method in the current field that can effectively utilize knowledge graphs to predict drug-disease relationships. Summary of the Invention

[0007] The present invention provides a method and system for predicting drug-disease relationships based on a dual-channel fusion knowledge graph, aiming to fill the gap in subgraph embedding learning research and realize the effective utilization of the knowledge graph computing model in drug-disease relationship prediction methods.

[0008] To achieve the above objective, the present invention provides a method for predicting drug-disease relationships based on a dual-channel fusion knowledge graph, including the following steps: S1. Obtain drug-disease association triples and external knowledge graph data, and construct a biomedical fusion knowledge graph; S2. Construct a drug-disease subgraph based on the biomedical fusion knowledge graph; S3. Construct a dual-channel adaptive fusion feature module, embed the drug-disease subgraph, and obtain a pre-embedded network of the dual-channel fusion knowledge graph; S4. Enhance the splicing of the drug-disease subgraph in the pre-embedded network of the dual-channel fusion knowledge graph, import an edge-node iterative update learning mechanism, and train a splicing subgraph relationship perception learning network to obtain enhanced subgraph feature embeddings; S5. Use the trained enhanced subgraph feature embeddings to calculate and output the prediction probability of drug-disease relationships.

[0009] The dual-channel fusion knowledge graph network of the present invention consists of four key parts: an external knowledge fusion module, a subgraph construction module, a dual-channel adaptive fusion feature embedding module, and a subgraph learning module.

[0010] For the dual-channel adaptive fusion feature embedding module, the present invention designs a dual-channel embedding module (Dual-Channel Embedding Module), which consists of a structural channel (UnionGIN) and a semantic channel (PubMed-BERT). They respectively encode the fused graph structure and the text description of entities, so as to comprehensively capture the multi-source feature information of entities. The structural channel models the topological structure between entities, while the semantic channel mines the semantic relationships of entity ontologies, and completes the multi-modal representation of graph nodes through an adaptive feature fusion method. This process first merges the original drug-disease graph with an external knowledge graph to construct a graph, introducing more general entities and relationships to enhance the graph representation ability.

[0011] For the subgraph construction module, the present invention designs a path-aware sampling strategy, which takes the drug-disease pair to be predicted as the center and generates the corresponding K-hop subgraph from the fused graph through path search. For the subgraph learning module, the present invention designs an end-to-end subgraph representation learning framework, which adopts an iterative edge-node collaborative update mechanism: first, it calculates the edge weights based on node embeddings to update the edge embeddings, and then updates the node representations through neighbor aggregation.

[0012] For the subgraph learning module, in order to enhance the representation ability of subgraphs, the present invention introduces a relationship enhancement mechanism (Resemble Edge Construction). By constructing supplementary edges in the subgraph (such as similarity connections between drugs-drugs, diseases-diseases and association connections between drugs-diseases), it explicitly enhances the structural associations between entities, 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, which adopts an iterative edge-node collaborative update mechanism: first, it calculates the edge weights based on node embeddings to update the edge embeddings, and then updates the node representations through neighbor aggregation and iterates to realize the learning function of subgraphs.

[0013] In addition, in the classification prediction stage, the present invention designs a triple embedding fusion strategy, which concatenates the head entity embedding, the tail entity embedding and the subgraph embedding, and outputs the association type through a classifier. This strategy significantly improves the accuracy of association relationship discrimination through the complementarity of global context awareness and local structure information.

[0014] Preferably, the obtaining of the drug-disease association triples and the external knowledge graph data in step S1 specifically includes: selecting the drug-disease association triple data provided by the Precision Medicine Knowledge Graph (PrimeKG) and the Drug Repositioning Database (repoDB) as the training knowledge, where the node types include drugs and diseases, and the relationship type is the verified drug-disease association relationship; selecting the Heterogeneous Information Network (Hetionet) as the external knowledge graph; The construction of the biomedical fusion knowledge graph specifically includes: Removing the isolated nodes in the external knowledge graph and screening the drug nodes and disease nodes therein; Extracting the drug-disease association triple data from the Precision Medicine Knowledge Graph and the Drug Repositioning Database; Deeply fusing the drug-disease association triple data with the drug nodes and disease nodes to construct a biomedical fusion knowledge graph; dividing the drug-disease association triple data into a training set, a validation set, and a test set.

[0015] Preferably, the construction of the drug-disease subgraph in step S2 specifically includes: Taking the target drug-disease pair as the center, adopting a path-based K-hop subgraph sampling strategy to perform local context modeling for the target pair. For a target triple , where represents the drug, represents the disease, represents the semantic relationship between them, represents all target triples; finding paths through breadth-first traversal, and extracting the local subgraph within its hop range from the fusion graph, defined as: ; Among them, represents the K-hop local context captured by BFS starting from the drug and , including nodes, relationships, and triples, represents the set of nodes reachable within K hops starting from the drug and the disease ; and are the relationship and triple sets within the subgraph respectively.

[0016] Preferably, the construction of the dual-channel adaptive fusion feature module in step S3 includes constructing a dual-channel modeling that includes structural channel modeling and semantic channel modeling; The structural channel modeling uses an improved graph neural network model UnionGIN to process the structural features of different types of nodes and edges through a multi-relationship aggregation mechanism and hierarchical message passing. In the th layer, for each node in the fused knowledge graph ; ; ; Among them, is the th layer embedding representation of node is the th layer embedding representation of node is the learnable parameter of the th layer, represents the set of neighbor nodes of node selected from the set of neighbor nodes of node to form an edge adjacent to , and are multi-layer perceptrons with different parameters, is a cross-edge type conversion module that adjusts or normalizes the information of the edge, is the original edge structure coefficient of edge , representing the edge strength, is the normalized structure coefficient of edge , used for attention allocation or weight calculation; In the hierarchical message passing, the path matrix of edge is used, and the singular value decomposition (SVD) is performed on the path matrix to extract the singular value , where P represents the path matrix, and respectively represent the left singular vector matrix and the right singular vector matrix obtained by the decomposition, represents the diagonal matrix composed of singular values, and the sum of the singular values of is expressed as , as the local structure coefficient of edge , where represents the path from node to The path matrix, which reflects the connectivity strength or weight of the path, is the edge The original edge structure coefficient, which quantifies the global importance of the path matrix, represents the diagonal matrix of singular values of the path matrix from node to ; Finally, the embedding representation of each node on the fused graph is obtained as , where For each edge (i.e., relationship) in , is the edge set of the fused graph, and its embedding representation is: ; where is a learnable matrix, is the bias, is the type of the edge, ; is the number of edges, represents the embedding dimension of the edge type; The semantic channel modeling is based on medical knowledge base resources, and uses a large language model to extract the unstructured text description of entities. For each entity node in the target triple, there is a medical text description , including but not limited to one or more of the drug mechanism of action, indications, and disease symptoms; using PubMed-BERT as the text encoder to encode it, so as to obtain a high-dimensional semantic vector: ; where , is the number of nodes, represents the embedding dimension of the text.

[0017] 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 the weights of the structural and semantic features, and dynamically fuse the structural embedding and the semantic embedding through the attention mechanism; specifically, it includes: Obtaining the node embedding representation of the target drug-disease subgraph from the fused knowledge graph , where is the number of nodes in the subgraph, and D represents the embedding dimension; represents the set of nodes reachable within K hops starting from drug and disease ; ; ; ; ; ; Among them, represents the structural embedding of node . represents the text embedding of node . , , are the projection parameter matrices of the -th head respectively. 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; , , are the input vectors of the -th attention head in the multi-head cross-attention mechanism, representing query, key, and value respectively; ; Among them, is the final fused embedding representation, is normalization, is the node embedding representation of the target drug-disease subgraph; represents the output of the attention head; is the multi-layer perceptron; All nodes form a unified multi-modal embedding representation through multi-layer aggregation, serving as the basic input for subsequent subgraph learning and relation classification.

[0018] Preferably, the enhanced splicing described in step S4 specifically includes: Suppose there are in the same batch for the target drug-disease pair , then the spliced subgraph is defined as: ; On the basis of retaining the structure of the atomic graph, enhanced connections are constructed on this spliced subgraph, specifically including three types of virtual edges: drug-drug similarity edges, disease-disease similarity edges, and drug-disease virtual association edges.

[0019] Preferably, the introduction of the iterative update learning mechanism described in step S4 specifically includes: The end-to-end subgraph learning mechanism based on edge-node collaborative update dynamically updates the subgraph edge weights and optimizes the node embeddings in each iteration; after the splicing subgraph is constructed, each edge is assigned an initial connection strength value, which comprehensively considers the importance weights of known DDAs, the potential association probabilities between unknown drug-disease pairs, and the semantic similarities between drugs and drugs, and diseases and diseases.

[0020] For any two drugs in the batch graph and , its initial edge weight is , and the Jaccard similarity is used to calculate the neighborhood structure similarity of the two drug nodes and assign the edge weight: ; ; where represents the Jaccard similarity between drugs and , and are drugs in the batch graph, is an indicator function, is a threshold, is the dimension index of the embedding vector, is the total dimension of the embedding vector, is the embedding vector of drug , represents the indicator vector formed by binarizing the drug embedding vector after passing through the threshold , represents the indicator vector formed by binarizing the drug embedding vector after passing through the threshold , is the embedding vector of drug , is the weight balance parameter, which controls the fusion degree of the old edge weight and the new edge similarity, is the updated edge weight, represents updating the full edge weight; For any two disease nodes in the batch graph, the cosine similarity is calculated in the embedding space and the edge weight is assigned: ; ; where represents the cosine similarity between diseases and , and represent any two disease nodes in the batch graph, and is the embedding of the disease node, representing the disease node and the updated relationship weight between them, representing the disease node and the original relationship weight between them, is the weight balance parameter, representing the L2 norm of the vector; For the drug nodes and disease nodes in the batch graph and disease nodes , the edge weight between them is: ; where represents the edge weight between the drug node and the disease node , and respectively represent the embeddings of the drug node and the disease node , represents the vector concatenation operation, is the multi-layer perceptron; After that, in each iteration, considering the embedding difference between node u and node v in the previous iteration and the learnable embedding of relationship r, using the node representation and relationship embedding , use the multi-layer perceptron to evaluate the semantic importance score of the edge: ; where represents the semantic importance score of node and node in the iteration, represents the MLP multi-layer perceptron function in the iteration, represents the exponential function; The current round edge weight is weighted and combined by the historical edge weight and the scoring result: ; where is the mixing hyperparameter, controlling the weight ratio of the new score in the round and the edge weight in the round; In the subgraph On this basis, neighbor message passing is performed according to the updated edge weights to update the node representation: ; Among them, represents the updated node representation of node in the th round, represents the activation function, represents the neighbor set of node , represents the edge the dynamic weight of the th round of iteration, represents the embedding representation of the neighbor node in the previous round.

[0021] Preferably, the prediction probability of the output drug-disease relationship in step S5 specifically includes: obtaining the local feature representation of the target drug finally learned by the subgraph , the feature representation of the target disease and the context feature representation of the target drug-disease pair: ; Among them represents all neighbor context node embeddings of node ; Through the feature splicing operation, the drug feature, disease feature and context feature are combined into a unified high-dimensional feature vector in terms of dimension: ; Input the unified high-dimensional feature vector into the fully connected layer, and this layer maps the feature space to the relationship classification space through linear transformation and activation function: ; Among them, and are the parameters of the fully connected layer; represents the predefined number of relationship categories; is the unified high-dimensional feature vector; is the multi-class scoring vector of the drug-disease pair .

[0022] Finally, the score vector is normalized into a probability distribution through the normalization function Softmax for classification prediction: ; Among them, represents the prediction probability that the sample belongs to the th class of relationship, represents the Drug-disease pairs A multi-class scoring vector of the relationship for subsequent drug-disease relationship decision-making.

[0023] Under the same technical concept, the present invention also provides a drug-disease relationship prediction system based on a dual-channel fusion knowledge graph, and the prediction system includes: External knowledge fusion module: used to obtain drug-disease association triples and external knowledge graph data, and construct a biomedical fusion knowledge graph; Sub-graph construction module: used to construct a drug-disease sub-graph based on the 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 sub-graph into vector embeddings, and obtain a pre-embedded network of the dual-channel fusion knowledge graph; Sub-graph learning module: used to perform enhanced splicing on the drug-disease sub-graph in the pre-embedded network of the dual-channel fusion knowledge graph, import an edge-node iterative update learning mechanism, train a splicing sub-graph relationship perception learning network, and obtain enhanced sub-graph feature embeddings; Prediction probability output module: used to calculate and output the prediction probability of the drug-disease relationship by using the trained enhanced sub-graph feature embeddings.

[0024] The above solution of the present invention has the following beneficial effects: The present invention proposes a drug-disease relationship prediction method based on a dual-channel fusion knowledge graph, systematically integrates the semantic representation ability of the knowledge graph and the topological modeling advantages of the graph neural network, and realizes the efficient and accurate prediction of the drug-disease association relationship through a multi-modal feature interaction and dynamic sub-graph learning mechanism; Specifically, the present invention first supplements the original DDA triple data with multi-source heterogeneous knowledge, integrates biomedical entity descriptions, and constructs a fusion biomedical knowledge graph. On this basis, a structural semantic dual-channel graph embedding module composed of UnionGIN and PubMed-BERT is used to capture the global and local features of the graph from the two dimensions of graph structure and text semantics, and realizes adaptive deep fusion of features through a cross-modal attention alignment mechanism. The attention weight of noise features is automatically attenuated, and the weight of key features is increased to achieve multi-modal feature complementarity and improve representation integrity. During the training process, the present invention adopts a subgraph learning paradigm, and uses a path-aware heuristic sampling algorithm to extract the K-hop path between the target drug-disease pair using breadth-first search (BFS), select the most meaningful subgraph nodes and relationships, focus on key nodes, reduce the noise interference problem of irrelevant nodes, and at the same time, enhance the average node degree of sparse subgraphs through virtual connection enhancement to alleviate the problem of poor correlation. In the subgraph learning stage, an iterative update mechanism of an end-to-end architecture is adopted. Through bidirectional alternating message transmission between nodes and edges, the edge weights are dynamically adjusted to adjust the subgraph structure, alleviate the problem of over-reliance on topological structure, and mine the deep-level association pattern between drugs and diseases. Finally, in order to achieve classification prediction output, the present invention combines the target drug, target disease and its context feature representation into a unified high-dimensional feature vector through feature concatenation operation, and inputs it into the fully connected layer. Through linear transformation and activation function mapping, the feature space is mapped to the relational classification space, and the multi-classification prediction of the drug-disease relationship is completed.

[0025] Experimental results show that DualSubNet significantly outperforms existing baseline models on multiple benchmark datasets, demonstrating its advancement and superiority in drug repurposing prediction tasks.

[0026] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a DualSubNet architecture diagram of the present invention; Figure 2 A schematic diagram of the process of obtaining entity text description in the semantic channel modeling of the present invention; Figure 3 It is a schematic diagram of the structure of the dual-channel adaptive fusion feature module of the present invention; Figure 4 It is a schematic diagram of the process of a drug-disease relationship prediction method based on a dual-channel fusion knowledge graph of the present invention; DETAILED DESCRIPTION

[0028] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0029] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as a limitation of the present invention. In addition, the terms "first", "second", "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0030] In the description of the present invention, it should be noted that unless otherwise clearly defined and limited, the terms "installed", "connected", "connected" should be understood in a broad sense. For example, it can be a locking connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0031] In addition, the technical features involved in different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0032] Embodiment 1:

[0033] A drug-disease relationship prediction method based on a dual-channel fusion knowledge graph provided in this embodiment has a flow schematic diagram as Figure 4 shown, and includes the following steps: Step 1: Data acquisition and processing.

[0034] In this embodiment, the drug-disease association (DDA) triple data provided by the Precision Medicine Knowledge Graph (PrimeKG) and the Drug Repositioning Database (repoDB) is selected as the training knowledge. The node types include drugs and diseases, and the relationship type is the verified drug-disease association relationship. The Heterogeneous Information Network (Hetionet) is selected as the external knowledge graph to supplement the structured background knowledge of the basic graph. Hetionet contains rich entity types (such as proteins, genes, pathways, etc.) and multi-source relationship types, and can provide structured biomedical background knowledge for drug-disease association prediction. The graph data is stored in the triple format, and the triple form is expressed as , where is the entity set, including entities such as drugs, diseases, genes, etc., is the relationship set, including relationships such as drug-disease association (DDA), protein-protein interaction (PPI), gene regulation (Regulates), etc., represents a directed knowledge triple, where h is the head entity, t is the tail entity, and r is the semantic relationship between them.

[0035] After that, the data set is preprocessed. First, the isolated nodes (nodes without connection relationships) are removed from the external knowledge graph Hetionet, and the drug nodes and disease nodes are screened. Next, the PrimeKG and repoDB data sets are preprocessed separately to extract the drug-disease relationship triples. The analysis of the PrimeKG and repoDB data sets is shown in Table 1, and the analysis of the Hetionet data set is shown in Table 2: Table 1 Analysis of PrimeKG and repoDB Data

[0036] Table 2 Analysis of Hetionet Data

[0037] Then, the drug-disease subgraph is deeply integrated with the triple data in Hetionet to construct a unified comprehensive knowledge graph. The entity and relationship are encoded according to the unified ID naming rules to construct the PrimeKG-Hetionet fusion graph and the repoDB-Hetionet fusion graph, retaining the entity alignment relationships across graphs and extracting the triple structure of the fusion graph (such as <drug, associated with, disease>, <gene, regulates, drug>, etc.). The fused knowledge graph is represented as: ; where represents that there are N nodes, M edges, and T triples in the fusion graph, including entity alignment and cross-graph association information.

[0038] The data analysis of the PrimeKG-Hetionet fusion graph and the repoDB-Hetionet fusion graph is shown in Table 3: Table 3 Data analysis of PrimeKG-Hetionet and repoDB-Hetionet

[0039] Finally, the processed triple dataset is divided according to the 10-fold cross-validation method: 80% of the data is used as the training set; 10% of the data is used as the validation set; 10% of the data is used as the test set.

[0040] Step 2: Construction of the drug-disease subgraph based on the path of the biomedical fusion knowledge graph.

[0041] On the fused knowledge graph in this embodiment, taking the target drug-disease pair as the center, a path-based K-hop subgraph sampling strategy is adopted for local context modeling of the target pair. For each target triple ( are all target triples), where represents the drug, represents the disease, represents the semantic relationship between them; the path is found through breadth-first traversal, and the local subgraph within its hop range is extracted from the fusion graph and defined as: ; where represents the K-hop local context captured from the drug and departure, including nodes, relationships, and triples, represents from the drug and the disease The set of nodes reachable within K jumps starting from... and are the relationship and triple sets within the subgraph respectively.

[0042] Constructing a subgraph allows us to focus on specific patterns or fragmentary information with practical value, thereby reducing the impact of noise. At the same time, the path-based construction method preserves the structural paths and multi-hop dependency information between entity pairs, which helps to reveal potential high-order drug-disease associations.

[0043] Step 3: Dual-channel adaptive graph embedding.

[0044] To capture richer semantic features, this embodiment models from two dimensions: the structural channel and the semantic channel: 1. Structural channel modeling: To break through the limitations of traditional GNNs on homogeneous graph structures, for the highly heterogeneous entity and relationship types in the knowledge graph, this embodiment adopts an improved graph neural network model, UnionGIN, which effectively processes the structural features of different types of nodes and edges through a multi-relational aggregation mechanism and hierarchical message passing. In the th layer, for each node in the fused knowledge graph the update formula for the structural representation is: ; ; ; where is the th layer embedding representation of node , is the th layer embedding representation of node , is the learnable parameter of the th layer, represents the set of neighbor nodes of node , is selected from the set of neighbor nodes of node to form an edge adjacent to , and are multi-layer perceptrons with different parameters, is a cross-edge type conversion module that adjusts or normalizes the information of the edge, is the original edge structure coefficient of edge , representing the edge strength, is the normalized structure coefficient of edge , used for attention allocation or weight calculation; Using edges in the hierarchical message passing For the path matrix, perform singular value decomposition on the path matrix and extract the singular values , where P represents the path matrix, and represent the left singular vector matrix and the right singular vector matrix obtained from the decomposition respectively, represents the diagonal matrix composed of singular values. Denote the sum of the singular values of as , which is used as the local structure coefficient of the edge . Among them, represents the path matrix from node to , reflecting the connectivity strength or weight of the path, is the original edge structure coefficient of the edge , quantifying the global importance of the path matrix, represents the singular value diagonal matrix of the path matrix from node to ; Finally, the embedding representation of each node on the fused graph is obtained as , where N is the number of nodes, H is the embedding representation of all nodes, and D represents the embedding dimension.

[0045] For each edge (i.e., relationship) in , is the edge set of the fused graph, and its embedding representation is: ; Among them, is a learnable matrix, is the bias, is the type of the edge, ; is the number of edges, represents the embedding dimension of the edge type; 2. Semantic channel modeling: To improve the richness and domain adaptability of semantic features, in this embodiment, based on authoritative medical knowledge base resources such as the DrugBank database and the PubMed literature retrieval system of the National Library of Medicine of the United States, a large language model (Large Language) is used to extract the unstructured text descriptions of entities (such as drug indications, drug mechanisms of action, disease symptoms, clinical trial results, etc.). The schematic diagram of the process for obtaining entity text descriptions is as shown in Figure 2 . For each entity node in the target triple, there is a medical text description , including the mechanism of drug action, indications, disease symptoms, etc. In this embodiment, PubMed-BERT is used as the text encoder , so as to obtain high-dimensional semantic vectors: ; Among them, , is the number of nodes, represents the embedding dimension of the text.

[0046] 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 fuse the 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 entities, providing more comprehensive feature support for subsequent tasks (such as drug repositioning, disease mechanism prediction). The dual-channel adaptive fusion feature module is shown in Figure 3 . First, we obtain the node embedding representation of the target drug-disease subgraph from the fused knowledge graph , where is the number of nodes in the subgraph, D represents the embedding dimension, represents starting from the drug and the disease , the set of nodes reachable within K hops; ; ; ; ; ; Among them, represents the structural embedding of node , represents the text embedding of node , , , are the projection parameter matrices of the i-th head respectively, 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 result back to the original dimensional space; , , are the input vectors of the i-th attention head in the multi-head cross-attention mechanism, representing query, key, and value respectively; Finally, after concatenating the structural and semantic features, the fused embedding representation is obtained through a linear transformation: ; Among them, is the final fused embedding representation, is normalization, is the node embedding representation of the target drug-disease subgraph; represents the output of the attention head; is a multi-layer perceptron;

[0047] All nodes form a unified multi-modal embedding representation through multi-layer aggregation, which serves as the basic input for subsequent subgraph learning and relationship classification.

[0048] Step 4: Subgraph enhancement splicing and subgraph iterative update learning mechanism based on an end-to-end architecture.

[0049] Since drug-disease association data usually has high sparsity (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. 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 the sparse network through virtual connections. This embodiment splices the subgraphs of all target DDAs in the same batch, and constructs a larger-scale spliced subgraph on the basis of retaining the atomic graph structure. Suppose there are For the target drug-disease pair , the spliced subgraph is defined as: ; Enhanced connections are constructed on this spliced subgraph while retaining the structure of the atomic graph, including three types of virtual edges: drug-drug similarity edges, disease-disease similarity edges, and drug-disease virtual association edges.

[0050] To achieve the efficiency and accuracy of drug-disease association prediction, this embodiment designs an end-to-end subgraph learning mechanism based on edge-node collaborative update, dynamically updates the subgraph edge weights and optimizes the node embeddings in each iteration, and mines the deep association patterns between drugs and diseases. After the spliced subgraph is constructed, each edge is assigned an initial 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. By introducing an iterative optimization strategy, the edge weights in the subgraph are dynamically adjusted according to their relevance to the DDA prediction task in each iteration, so as to gradually reduce the weights of redundant or noisy edges to streamline the subgraph structure and focus on the association information that contributes the most to the prediction task.

[0051] For any two drugs in the batch graph and , whose initial edge weight is , calculate the neighborhood structure similarity between two drug nodes using Jaccard similarity and assign an edge weight: ; ; where represents the Jaccard similarity between drugs and , and are drugs 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, is the embedding vector of drug , represents the indicator vector formed after binarizing the drug embedding vector through the threshold , represents the indicator vector formed after binarizing the drug embedding vector through the threshold , is the embedding vector of drug , is the weight balance parameter, which controls the degree of fusion between the old edge weight and the new edge similarity, is the updated edge weight, represents updating the full edge weight; For any two disease nodes in the batch graph, calculate the cosine similarity in the embedding space and assign an edge weight: ; ; where represents the cosine similarity between diseases and , and represent any two disease nodes in the batch graph, and are the embeddings of the disease nodes, represents the updated relationship weight between disease nodes and , represents the original relationship weight between disease nodes and , is the weight balance parameter, represents the L2 norm of the vector; For the drug nodes in the batch graph and disease nodes , the edge weight between them is: ; Among them, represents the edge weight between the drug node and the disease node , and respectively represent the embeddings of the drug node and the disease node , represents the vector concatenation operation, is a multi-layer perceptron; So far, we have completed the initial update of all edge weights in the subgraph.

[0052] After that, in each th iteration, considering the embedding difference between node u and node v in the previous iteration and the learnable embedding of relationship r, using the node representation and relationship embedding in the previous round, use a multi-layer perceptron to evaluate the semantic importance score of the edge: ; Among them represents the semantic importance score of node and node in the th iteration, represents the MLP multi-layer perceptron function in the th iteration, represents the exponential function; The edge weight in the current round is weighted and combined by the historical edge weight and the scoring result: ; Among them is a hybrid hyperparameter that controls the weight ratio of the new score in the th iteration and the edge weight in the th round; On the subgraph , perform neighbor message passing according to the updated edge weight to update the node representation: ; Among them, represents the updated node representation of node in the th round, denotes the activation function, denotes the set of neighbors of node , denotes the edge the dynamic weight of the r-th round of iteration, denotes the neighbor node 's embedding representation in the previous round.

[0053] After multiple rounds of iteration, the node representations of drugs and diseases are gradually mapped from the general feature space to the dedicated representation space for the DDA prediction task, significantly improving the prediction performance. At the same time, the above mechanism can effectively and dynamically prune redundant edges, enhance potential associated connections, and still transmit important semantic information under sparse or missing conditions, improving the robustness and accuracy of drug-disease association prediction.

[0054] Step 5: The classification layer outputs the prediction probability.

[0055] After the model is optimized and learned through multiple rounds of iteration, both the drug nodes and the disease nodes are embedded into the optimal latent space vector representation, ensuring the efficient encoding and semantic expression of node features. In this embodiment, through the subgraph learning mechanism, the local feature representations of the target drug (focusing on the drug's own attributes and the node information directly associated with it), the feature representation of the target disease (covering disease-related features and topological structure information), and the context feature representation of the target drug-disease pair (by calculating the average embedding vector of all nodes in the subgraph, comprehensively reflecting the interaction pattern between the drug and the disease in the global network) are extracted respectively, where denotes all neighbor context node embeddings of node ; Subsequently, the above three types of features are deeply fused: First, through the feature concatenation 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 relationship classification space through linear transformation and activation function mapping: ; where and are the parameters of the fully connected layer; denotes the number of predefined relationship categories, is the unified high-dimensional feature vector, is the drug-disease pair 's multi-class scoring vector.

[0056] Finally, the score vector is normalized into a probability distribution through the normalization function Softmax for classification prediction: ; wherein, represents the predicted probability that the sample belongs to the th class relationship, represents the multi-class scoring vector of the th class drug-disease pair relationship for subsequent drug-disease relationship decision-making.

[0057] A drug-disease relationship prediction system based on a dual-channel fusion knowledge graph, comprising: External knowledge fusion module: used to obtain drug-disease association triples and external knowledge graph data, and construct a biomedical fusion knowledge graph; Subgraph construction module: used to construct a drug-disease subgraph based on the 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 pre-embedded network of the dual-channel fusion knowledge graph; Subgraph learning module: used to perform enhanced splicing on the drug-disease subgraph in the pre-embedded network of the dual-channel fusion knowledge graph, import an edge-node iterative update learning mechanism, train a spliced subgraph relationship perception learning network, and obtain enhanced subgraph feature embedding; Predicted probability output module: used to calculate and output the predicted probability of the drug-disease relationship by using the trained enhanced subgraph feature embedding.

[0058] The architecture diagram of the DualSubNet in this embodiment is as Figure 1 shown. It is advanced in dealing with drug-disease relationship prediction classification tasks, can simultaneously capture the local features 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 a linear classification layer. This embodiment conducts drug-disease prediction experiments on two datasets, namely 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, it is necessary to predict their relationship as indication).

[0059] Furthermore, in this embodiment, drug pairs and disease pairs in the DrugBank data were selected to test and validate the DualSubNet model in this embodiment. The specific scheme is as follows: drug pairs and disease pairs are input into the DualSubNet model, and the predicted relationship is output through the model. Among the output predicted relationships, positive indicates treatable / preventable, invalid indicates invalid, and negative indicates having side effects or causing the disease to worsen. The results of the specific predicted relationships and the true relationships are shown in Table 4 below: Table 4 Performance Test of Applying DrugBank Data to the DualSubNet Model

[0060] 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, etc., and particularly show significant advantages in the scenarios of drug repositioning and disease mechanism analysis in complex biological networks.

[0061] Comparative Example 1: To comprehensively study the performance of DualSubNet, the present invention introduces some other existing advanced methods and compares the performance between these methods and the method proposed by the present invention in the subsequent experiments.

[0062] AMDGT proposed a multi-modal fusion method based on dual graph transformers and attention mechanisms. By integrating the similarity networks of drugs and diseases and the heterogeneous association network, and using the modal interaction module to capture deep features, it improves the performance and generalization ability of drug-disease association prediction.

[0063] AdaDR combines the information in the feature space and the topological space, dynamically adjusts the embedding weights using the attention mechanism, and enhances the model generalization ability through consistency constraints, significantly improving the performance of the drug repurposing task.

[0064] DSE-HNGCN proposed a deep learning framework based on heterogeneous networks and graph convolutional networks. By integrating multi-source information (drug similarity, side effect semantic similarity, and frequency information) to construct a heterogeneous network, and using a multi-layer graph convolutional network and layer importance combination strategy, it effectively predicts the occurrence frequency of drug side effects.

[0065] SUMGNN reduces noise and improves efficiency by using external biomedical knowledge and extracting local subgraphs related to the target drug pair from the knowledge graph. At the same time, it generates inference paths through the self-attention mechanism, extracts the information most useful for prediction, efficiently predicts drug-drug interactions, and generates interpretable inference paths.

[0066] NAGTLDA proposes a graph Transformer model that combines local feature learning, global feature learning, and structural encoding. Through an adaptive feature fusion mechanism, it can efficiently predict the unknown associations between lncRNAs and diseases, significantly improving the performance of the model on large-scale sparse datasets.

[0067] LaGAT proposes a link-aware graph attention mechanism that can generate different attention paths according to different drug pairs, and dynamically generate attention paths to capture the semantic diversity of drug nodes in different drug pairs, thus improving the accuracy and interpretability of predictions.

[0068] HGTDR proposes an end-to-end method based on heterogeneous graph Transformer (HGT). It is the first to use HGT for drug repurposing, which can automatically process complex relationships in heterogeneous graphs, avoid information loss, and achieve efficient prediction of drug-disease relationships by automatically processing large-scale heterogeneous knowledge graphs, and shows excellent performance in multiple tasks.

[0069] LDAGM proposes a model based on deep topological feature extraction and graph convolutional autoencoders, which can fuse functional similarity and Gaussian kernel similarity to solve the sparsity problem, predict the relationship between lncRNAs and diseases through multi-view heterogeneous network fusion and multi-layer perceptron, and introduce an aggregation layer in the MLP to optimize feature extraction through a gating mechanism, significantly improving the accuracy and stability of predictions.

[0070] Due to the possible differences in the data and data processing methods used by these methods, to achieve data unity and model fairness, eliminate the biases caused by data input or classification layer differences, and focus on the performance differences of the model architecture itself, the specific implementation methods of this comparative example include: (1) Unification of data input: Use the same training set, validation set, and test set as this model for all comparison models to ensure consistent data partitioning.

[0071] In the remaining data models of this comparative example, the exactly same PrimeKG-Hetionet dataset and repoDB-Hetionet dataset in the embodiments of the present invention are also used to conduct comprehensive comparative experiments, and systematically evaluate the performance differences between DualSubNet and other advanced methods in the drug-disease association prediction task. Among them, the PrimeKG dataset contains more complex and rich biomedical background knowledge, while the repoDB dataset has stronger practical usability and can reflect the generalization ability of the model in the drug repurposing scenario.

[0072] (2) Adaptive modification of the classification layer: Replace the final classification layer of all comparison models with the same structure to output the drug-disease association probability.

[0073] In this comparative example, in order to make a fair comparison of the classification performance of the models, the data inputs of all models were unified and the corresponding classification layers were modified.

[0074] As can be seen from Table 5, on the PrimeKG-Hetionet dataset, although the DualSubNet model proposed in the present invention is slightly lower than the LDAGM model by 0.32% in terms of the AUROC metric, it achieves the best performance in terms of the ACC and Weighted_F1 metrics, reaching an accuracy (ACC) of 92.76% and a weighted F1 score of 92.52% respectively, which is better than the comparative methods. This shows that although the ranking of DualSubNet is slightly inferior to that of LDAGM, DualSubNet improves the recognition ability of small-sample categories through subgraph construction and information enhancement mechanisms, and is better optimized under the same classification threshold.

[0075] The experimental results on the repoDB-Hetionet dataset are shown in Table 6. DualSubNet reaches 95.68% in terms of the AUROC metric, which is the highest among all models, showing extremely strong discrimination ability. Although the LDAGM model is slightly higher than the present invention in terms of accuracy, this may be because LDAGM is biased towards high-confidence predictions of the main categories and performs well in predicting the main categories in repoDB with uneven class distributions, but LDAGM ignores the recall of minority classes, resulting in a decrease in the weighted F1 score. While DualSubNet outperforms LDAGM in terms of the weighted F1 score, with more balanced overall performance and strong generalization ability. Especially compared with methods with weaker performance such as LaGAT and SUMGNN, DualSubNet has achieved significant improvements in multiple metrics.

[0076] The experimental results fully demonstrate that DualSubNet is superior to existing advanced methods in multiple key performance metrics, has stronger accuracy, robustness and generalization ability, and provides a more effective technical path for the prediction task of drug-disease relationships.

[0077] Table 5 Performance comparison between DualSubNet and other advanced methods on the PrimeKG-Hetionet dataset

[0078] Table 6 Performance comparison between DualSubNet and other advanced methods on the repoDB-Hetionet dataset

[0079] The above are the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for predicting drug-disease relationships based on a dual-channel fusion knowledge graph, characterized in that, It includes the following steps: S1. Obtain drug-disease association triples and external knowledge graph data, and construct a biomedical fusion knowledge graph; S2. Construct a drug-disease subgraph based on the biomedical fusion knowledge graph; 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; S4. Enhance and splice the drug-disease subgraph in the dual-channel fusion knowledge graph pre-embedding network, import an edge-node iterative update learning mechanism, and train a spliced subgraph relationship perception learning network to obtain enhanced subgraph feature embeddings; S5. Use the trained enhanced subgraph feature embeddings to calculate and output the prediction probability of the drug-disease relationship.

2. The prediction method according to claim 1, wherein, The obtaining of the drug-disease association triples and external knowledge graph data in step S1 specifically includes: selecting the drug-disease association triple data provided by the precision medicine knowledge graph and the drug repositioning database as training knowledge, where the node types include drugs and diseases, and the relationship type is the verified drug-disease association relationship; selecting the biomedical heterogeneous knowledge graph as the external knowledge graph; The construction of the biomedical fusion knowledge graph specifically includes: Removing the isolated nodes in the external knowledge graph, and screening the drug nodes and disease nodes therein; Extracting the drug-disease association triple data from the precision medicine knowledge graph and the drug repositioning database; Deeply fusing the drug-disease association triple data with the drug nodes and disease nodes to construct a biomedical fusion knowledge graph; dividing the drug-disease association triple data into a training set, a validation set, and a test set.

3. The prediction method according to claim 1, wherein The construction of the drug-disease subgraph in step S2 specifically includes: Centered on the target drug-disease pair , a path-based K-hop subgraph sampling strategy is adopted to model the local context of the target pair. For a target triple , where represents the drug, represents the disease, represents the semantic relationship between them, represents all target triples; find paths through breadth-first traversal, extract the local subgraph within its hop range from the fusion graph, and define it as: ； Among them, represents capturing drugs through BFS and the K-hop local context starting from represents starting from drugs and diseases the set of nodes reachable within K hops; and are the sets of relationships and triples within the subgraph respectively.

4. The prediction method according to claim 1, wherein The construction 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; The structure channel modeling uses an improved graph neural network model, UnionGIN, which processes the structural features of different types of nodes and edges through a multi-relation aggregation mechanism and hierarchical message passing. In the th layer, for each node in the fused knowledge graph , the update formula for the structural representation is as follows: ； ； ； Among them, is the -th -layer embedding representation of node is the -th -layer embedding representation of node is the -layer learnable parameter, represents the set of neighbor nodes of node , selected from the set of neighbor nodes of node forms an edge adjacent to , and are multi-layer perceptrons with different parameters, is a cross-edge type conversion module that adjusts or normalizes the information of the edge, is the original edge structure coefficient of edge , representing the edge strength, is the normalized structure coefficient of edge , which is used for attention allocation or weight calculation; Using edges in the hierarchical message passing Perform singular value decomposition on the path matrix and extract the singular values , where P represents the path matrix, and represent the left singular vector matrix and the right singular vector matrix obtained from the decomposition respectively, represents the diagonal matrix composed of singular values. Denote the sum of the singular values of as , which serves as the local structure coefficient of edge . Among them, represents the path matrix from node to , reflecting the connectivity strength or weight of the path, is the original edge structure coefficient of edge , quantifying the global importance of the path matrix, represents the singular value diagonal matrix of the path matrix from node to ; Finally, the embedding representation of each node on the fusion graph is obtained as , where $N$ is the number of nodes, $H$ is the embedding representation of all nodes, and $D$ represents the embedding dimension; For each edge in , which is the edge set of the fusion graph and its embedding representation is: ； Among them, is a learnable matrix, is a bias, is the type of edge, ; is the number of edges, represents the embedding dimension of the edge type; The semantic channel modeling is based on medical knowledge base resources, and uses a large language model to extract unstructured text descriptions of entities. For each entity node in the target triple , there is a medical text description , including one or more of drug mechanism of action, indications, and disease symptoms; Use PubMed-BERT as a text encoder to encode it, so as to obtain a high-dimensional semantic vector: ； Among them, , is the number of nodes, represents the embedding dimension of the text.

5. The prediction method according to claim 4, wherein The construction of the dual-channel adaptive fusion feature module in step S3 also includes introducing a cross-modal multi-head attention module to adaptively allocate the weights of structural and semantic features, and dynamically fusing the structural embedding and the semantic embedding through an attention mechanism; Specifically, it includes: Obtain the node embedding representation of the target drug-disease subgraph from the integrated knowledge graph , where is the number of nodes in the subgraph, and D represents the embedding dimension; represents the set of nodes reachable within K hops starting from the drug and the disease ; ； ； ； ； ； Among them, represents the structural embedding of node , represents the text embedding of node , , , are the projection parameter matrices of the -th head respectively. 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 result back to the original dimensional space; , , are the input vectors of the -th attention head in the multi-head cross-attention mechanism, representing query, key and value respectively; Finally, after the structural and semantic features are spliced, a fusion embedding representation is obtained through a linear transformation: ； Among them, is the final fused embedding representation, is normalization, is the node embedding representation of the target drug-disease subgraph; represents the output of the attention head; is a multi-layer perceptron; All nodes form a unified multi-modal embedding representation through multi-layer aggregation, which is used 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: Suppose there are for the target drug-disease pair , then the spliced subgraph is defined as: ； Constructing enhanced connections on the spliced subgraph while retaining the structure of the atomic graph, 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, characterized in that The import of the edge-node iterative update learning mechanism in step S4 specifically includes: A end-to-end subgraph learning mechanism based on edge-node collaborative update, dynamically updating the subgraph edge weights and optimizing the node embeddings in each iteration; after the spliced subgraph is constructed, each edge is assigned an initial connection strength value, which comprehensively considers the importance weight of the known DDA, the potential association probability between unknown drug-disease pairs, and the semantic similarity between drugs and drugs, and diseases and diseases; For any two drugs in the batch graph and , their initial edge weight is . Calculate the neighborhood structure similarity of the two drug nodes using the Jaccard similarity and assign the edge weight: ； ； Among them, represents the Jaccard similarity of drugs and . and are drugs in the batch graph, is an indicator function, is a threshold, is the dimension index of the embedding vector, is the total dimension of the embedding vector, is the embedding vector of drug . represents the indicator vector formed after binarizing the drug embedding vector with the threshold . represents the indicator vector formed after binarizing the drug embedding vector with the threshold . is the embedding vector of drug . is a weight balance parameter that controls the fusion degree of the old edge weight and the new edge similarity, is the updated edge weight, represents updating the full edge weight; For any two disease nodes in the batch graph , calculate the cosine similarity in the embedding space and assign an edge weight: ； ； Among them, represents a disease and cosine similarity, and represent any two disease nodes in the batch graph, and are the embeddings of the disease nodes, represents the disease node and the updated relationship weight between them, represents the disease node and the original relationship weight between them, is the weight balance parameter, represents the L2 norm of the vector; For the drug nodes in the batch graph and disease nodes , the edge weight between them is: ； Among them, represents the edge weight between the drug node and the disease node, and respectively represent the embeddings of the drug node and the disease node, represents the vector concatenation operation, is a multi-layer perceptron; After that, in each iteration of each round, considering the embedding difference between node u and node v in the previous iteration and the learnable embedding of relation r, using the node representation from the previous round and the relation embedding , use a multi-layer perceptron to evaluate the semantic importance score of the edge: ； Among them represents the semantic importance scores of nodes and node at the -th iteration, represents the MLP (Multi-Layer Perceptron) function at the -th iteration, represents the exponential function; Current wheel side weight Weighted combination of historical edge weights and scoring results: ； Among them is a mixed hyperparameter that controls the new score in the round of iteration and the edge weight in the round, as well as their weight ratio; On the subgraph perform neighbor message passing according to the updated edge weights to update the node representation: ； Among them, represents the node after the update in the round of iteration, represents the activation function, represents the node 's neighbor set, represents the edge the dynamic weight in the round of iteration, represents the embedding representation of the neighbor node in the previous round.

8. The prediction method according to claim 6, wherein, The predicted probability of outputting the drug-disease relationship described in step S5 specifically includes: obtaining the local feature representation of the target drug finally learned by the subgraph , the feature representation of the target disease and the context feature representation of the target drug-disease pair: ； Among them represents all neighbor context node embeddings of the node ; Through the feature splicing operation, the drug features, disease features, and context features are combined into a unified high-dimensional feature vector in terms of dimensions: ； Input the unified high-dimensional feature vector into the fully connected layer, which maps the feature space to the relation classification space through linear transformation and activation function: ； Among them, and are fully connected layer parameters; represents the number of predefined relationship categories; is a unified high-dimensional feature vector; is a drug-disease pair 's multi-class scoring vector; Finally, the score vector is normalized into a probability distribution through the normalization function Softmax for classification prediction: ； Among them, indicates the predicted probability that the sample belongs to the category relationship, represents the multi-category scoring vector of the category drug-disease pair relationship for subsequent drug-disease relationship decision-making.

9. A drug-disease relationship prediction system based on a dual-channel fusion knowledge graph, characterized in that, The prediction system includes: External knowledge fusion module: used to obtain drug-disease association triples and external knowledge graph data to construct a biomedical fusion knowledge graph; Subgraph construction module: used to construct a drug-disease subgraph based on the 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 vector embeddings, and obtain a dual-channel fusion knowledge graph pre-embedding network; Subgraph learning module: used to perform enhanced splicing on the drug-disease subgraph in the dual-channel fusion knowledge graph pre-embedding network, introduce an edge-node iterative update learning mechanism, train a spliced subgraph relationship-aware learning network, and obtain enhanced subgraph feature embeddings; Prediction probability output module: used to calculate and output the prediction probability of the drug-disease relationship using the trained enhanced subgraph feature embeddings.

Images