Topology-enhanced hypergraph learning and multi-graph converter driven drug-related side effect prediction method

Abstract

The invention discloses a topology-enhanced hypergraph learning and multi-graph converter driven drug-related side effect prediction method. The method comprises the following steps: constructing a multi-view heterogeneous graph, a drug-side effect hypergraph and a corresponding line graph; based on a cross-hypergraph and line graph learning strategy, node features of the hypergraph and the line graph are extracted and fused, and a first feature matrix is obtained; integrating node features of the multi-view heterogeneous graph by using an adaptive multi-graph Transform model to obtain a second feature matrix; performing multi-view feature fusion on the first feature matrix and the second feature matrix through a feature-level gating network; and splicing the fusion features and node features of the heterogeneous graph to form drug-side effect node pair feature representation, and performing model training based on a cross entropy loss function to obtain a drug-side effect association model. According to the method, the accuracy of predicting the relationship between the drug and the side effect is improved, and powerful technical support is provided for drug safety evaluation.

Description

Technical Field

[0001] This invention relates to the field of drug safety assessment technology, and in particular to a method for predicting drug-related side effects driven by topology-enhanced hypergraph learning and multigraph converter. Background Technology

[0002] Drug side effects refer to adverse reactions unrelated to the therapeutic purpose that occur after normal administration, often leading to drug failure in clinical trials. Identifying potential drug side effects is a crucial step in drug development and clinical application, and is of great significance for reducing the risk of development failure and improving drug safety. Identifying potential drug-related side effects through various computational prediction methods can support pharmacologists in screening reliable side effect candidates, reducing the cost and time of new drug design and development.

[0003] Existing methods for predicting the association between drugs and their side effects can be mainly divided into three categories. The first category focuses on using drug chemistry-protein interactions (CPIs) and target protein information. Since adverse reactions are often caused by off-target activities of drugs, these methods predict potential new uses and adverse reactions by constructing CPIs. However, these methods are limited by the lack of three-dimensional structural information of drug-related proteins, relying only on one-dimensional sequences or two-dimensional topological features. They cannot characterize atomic-level spatial complementarity and binding pocket geometry, and are prone to missing allosteric sites and non-classical binding modes. The second category infers the tendency of drug-side effect associations by constructing machine learning-based models. However, these methods are shallow predictive models and struggle to effectively extract deep features of drug and side effect nodes. The third category uses deep learning techniques to construct drug-side effect association prediction models, improving prediction accuracy by learning deep and representative features of drug and side effect nodes. However, these methods do not fully integrate the differential contributions of multiple heterogeneous graphs to association prediction. Furthermore, hypergraph-based prediction models have also been applied to drug-side effect prediction, but they neglect the feature learning of the topological structure composed of multiple hyperedges.

[0004] Therefore, there is a need to propose a drug-related side effect prediction method driven by topology-enhanced hypergraph learning and multigraph converter, which can encode the complex relationships between multiple drug and side effect nodes and deeply fuse the topology structure composed of all drugs, side effects and their attributes. It can learn deeply and adaptively fuse multi-perspective information from multiple hypergraphs, multi-line graphs and multi-heterogeneous graphs, providing technical support for improving drug safety and reducing the risk of R&D failure. Summary of the Invention

[0005] In view of this, the present invention provides a drug-related side effect prediction method driven by topology-enhanced hypergraph learning and multigraph converter, in order to solve the technical problem that the existing technology cannot effectively integrate multi-view heterogeneous graph structural features and make accurate predictions in the prediction of drug-side effect association.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting drug-related side effects driven by topology-enhanced hypergraph learning and multigraph converters, comprising: Obtain datasets related to drugs, diseases, and side effects; Based on the aforementioned drug, disease, and side effect related dataset, a multi-view heterogeneous graph and multiple drug-side effect hypergraphs are constructed, and a corresponding line graph is constructed based on each drug-side effect hypergraph. Based on the learning strategy of cross-hypergraph and line graph, a hypergraph convolutional neural network is used to extract node features from the hypergraph, and a line graph convolutional neural network is used to extract node features from the line graph. The two types of node features are enhanced bidirectionally through the feature cross-propagation mechanism, and the enhanced two types of features are fused using convolutional pooling layers to obtain the first feature matrix. By using an adaptive multi-graph Transformer model, node features of heterogeneous graphs from multiple perspectives are integrated to obtain a second feature matrix; The first and second feature matrices are fused from multiple perspectives using a feature-level gating network to obtain fused features. The fused features are concatenated with the node features of the heterogeneous graph to form a drug-side effect node pair feature representation, and the model is trained based on the cross-entropy loss function to obtain the drug-side effect association model. The drug chemical substructures and candidate side effect types whose associations are to be predicted are input into the drug-side effect association model, and the association prediction results between drugs and side effects are output.

[0007] Furthermore, a multi-view heterogeneous graph is constructed based on the aforementioned drug, disease, and side effect related dataset, including: Based on the aforementioned drug, disease, and side effect related dataset, extract the association between drugs and side effects, drug chemical substructure information, drug-related disease information, and side effect-related drug information; A drug-side effect correlation matrix is ​​generated based on the relationship between drugs and side effects; a drug similarity matrix is ​​generated based on drug chemical substructure information and drug-related disease information; and a side effect similarity matrix is ​​generated based on side effect-related drug information. The drug similarity matrix includes a structural similarity matrix and a related disease similarity matrix. Based on the same set of nodes, a first heterogeneous graph is constructed using the drug-side effect association matrix, side effect similarity matrix, and structural similarity matrix; a second heterogeneous graph is constructed using the drug-side effect association matrix, side effect similarity matrix, and related disease similarity matrix; wherein, the first heterogeneous graph and the second heterogeneous graph reflect the node connection relationships of drugs from the perspective of chemical structure and related diseases, respectively.

[0008] Furthermore, multiple drug-side effect hypergraphs are constructed based on the aforementioned drug, disease, and side effect related datasets, including: Based on the structural similarity matrix, the related disease similarity matrix, and the side effect similarity matrix, after setting similarity thresholds for filtering, a first hypergraph containing the hyperedge set from the perspective of drug structure and a second hypergraph containing the hyperedge set from the perspective of drug-related diseases are constructed respectively. Each hypergraph uses an adjacency matrix to describe the relationship between hyperedges and drugs and side effects. The i-th column of the adjacency matrix represents the i-th hyperedge of the corresponding hypergraph.

[0009] Furthermore, the construction of the corresponding line graph based on each drug-side effect hypergraph includes: Each hyperedge in each hypergraph is treated as a node to construct a corresponding line graph, which is used to represent the topological structure between hyperedges; If two hyperedges have at least one common node in the original hypergraph, then an edge is established between the corresponding nodes in the line graph. The weight of the edge between connected nodes is the ratio of the number of nodes shared by the two hyperedges to their total number of nodes. Based on the weight of the edge between connected nodes, the adjacency matrix and node feature matrix of the line graph are obtained.

[0010] Furthermore, a hypergraph convolutional neural network is used to extract node features from the hypergraph, including: The hypergraph convolutional neural network contains a multi-layer convolutional structure, with drugs and side effects as nodes, and hyperedges filtered by similarity thresholds as connection units. The features of the nodes connected to each hyperedge are aggregated to obtain the edge features of the hyperedge. The edge features of each hyperedge are distributed back to each node it is connected to, and fused with the original features of the nodes to complete the forward computation of one hypergraph convolutional layer. By stacking multiple layers of the hypergraph convolutional layers, the embedding representations of drug nodes and side effect nodes are updated layer by layer to obtain hypergraph node features that integrate the similarity and association of members within the hyperedge.

[0011] Furthermore, the step of using a line graph convolutional neural network to extract node features from the line graph includes: The line graph convolutional neural network contains a multi-layer convolutional structure. The network uses hyperedges as nodes and Jaccard similarity between hyperedges as edge weights. It extracts the topological features of hyperedges through Laplacian normalized graph convolution. By leveraging node-hyperedge associations, the topological features of hyperedges are backpropagated to the original drug and side effect nodes, thereby achieving joint enhancement of the hyperedge structural context and node attribute information.

[0012] Furthermore, the feature cross-propagation mechanism includes: In each layer of the online graph convolutional neural network and the hypergraph convolutional neural network, the node features output by the hypergraph convolutional neural network are used as query vectors, and the node features output by the online graph convolutional neural network are used as key and value vectors. At the same time, the node features output by the online graph convolutional neural network are used as query vectors, and the node features output by the hypergraph convolutional neural network are used as key and value vectors. The updated node features are generated through bidirectional cross-enhancement using a scaling dot product attention mechanism.

[0013] Furthermore, the second feature matrix is ​​obtained by integrating node features from multi-view heterogeneous graphs using an adaptive multi-graph Transformer model, including: The node features of the first heterogeneous graph and the second heterogeneous graph are mapped to corresponding query, key, and value vectors respectively through learnable weight matrices; Learnable scalar weights are introduced into the two sets of query, key, and value vectors. After softmax normalization, the dynamic contribution coefficients of the drug structure perspective and the drug-related disease perspective are obtained, thereby realizing the information ratio fusion of the first heterogeneous graph and the second heterogeneous graph. Multi-head scaled dot product attention is performed on the fused query, key, and value to capture the contextual dependencies between drug-side effect nodes and output attention features; The attention features are input into the feedforward neural network, and after residual connections and layer normalization, the updated node embeddings are obtained. After multi-layer stacking, the final node embeddings are output, which constitute the second feature matrix, and are used for subsequent multi-view fusion and association prediction.

[0014] Furthermore, the multi-view feature fusion of the first feature matrix and the second feature matrix through a feature-level gating network to obtain fused features includes: The first feature matrix and the second feature matrix are projected onto a unified dimension space through linear mapping; After performing the concatenation operation, the sigmoid activation function is input to generate the gated weight matrix; The two mapped feature matrices are then fused element-wise using the gate weights to obtain the fused features.

[0015] On the other hand, the present invention also provides a drug-related side effect prediction system driven by topology-enhanced hypergraph learning and multigraph converter, comprising: The data acquisition module is used to acquire datasets related to drugs, diseases, and side effects. The graph structure construction module is used to construct a multi-view heterogeneous graph and multiple drug-side effect hypergraphs based on the drug, disease and side effect related dataset, and to construct a corresponding line graph based on each drug-side effect hypergraph; The first feature extraction module is used to extract node features from the hypergraph and line graph based on the cross-hypergraph and line graph learning strategy. It uses a hypergraph convolutional neural network to extract node features from the hypergraph and a line graph convolutional neural network to extract node features from the line graph. It enhances the two types of node features in both directions through a feature cross-propagation mechanism and uses a convolutional pooling layer to fuse the enhanced two types of features to obtain the first feature matrix. The second feature extraction module is used to integrate the node features of multi-view heterogeneous graphs using an adaptive multi-graph Transformer model to obtain the second feature matrix. The fusion module is used to fuse the first feature matrix and the second feature matrix from multiple perspectives through a feature-level gating network to obtain fused features. The training module is used to concatenate the fused features with the node features of the heterogeneous graph to form a drug-side effect node pair feature representation, and to train the model based on the cross-entropy loss function to obtain a drug-side effect association model. The prediction module is used to input the drug chemical substructures and candidate side effect types of the association to be predicted into the drug-side effect association model, and output the association prediction results between drugs and side effects.

[0016] Compared with existing technologies, the method proposed in this invention has the following advantages: (1) Multiple hyperedges were constructed, each containing drug and side effect nodes unique to that edge and the relationships between them. Furthermore, we treated each hyperedge as a new node to build a line graph, better capturing the topological structure composed of multiple hyperedges. This method allows for the establishment of complex relationships between multiple drugs and multiple side effect nodes.

[0017] (2) A node feature learning strategy across hypergraphs and line graphs (FLHL) was designed, which can interactively encode the connections between multiple drug and side effect nodes and integrate the topology of hyperedges. The designed hypergraph convolutional neural networks (HGCNs) can model and propagate knowledge between multiple drug and side effect nodes, while the line graph convolutional neural networks (LGCNs) introduce the topology of hyperedges and learn the features of each node in the line graph. At the same time, a feature cross-propagation strategy (FCP) was also proposed, which can cross-propagate the node features learned from the hypergraph perspective and the line graph perspective.

[0018] (3) An adaptive multi-graph Transformer strategy (AMT) is proposed, which can dynamically learn the different contributions of multiple heterogeneous graphs, thereby encoding the contextual relationships between the features of multiple heterogeneous graph nodes.

[0019] (4) The importance of node feature representations from node feature learning strategies across hypergraphs and linegraphs and the importance of node feature representations from adaptive multigraph Transformer strategies are evaluated by designing a feature-level gated network (FLGN), and these node representations are then fused.

[0020] This invention can improve the accuracy of predicting the relationship between drugs and side effects, enhance the comprehensive capabilities and adaptability of the model, and provide strong technical support for drug safety assessment, personalized medicine, and side effect early warning. Attached Figure Description

[0021] Figure 1 A schematic diagram of the drug-related side effect prediction method driven by topology-enhanced hypergraph learning and multigraph converter provided by the present invention; Figure 2 This is a schematic diagram illustrating the feature cross-propagation between hypergraph and line graph learning provided by the present invention; Figure 3 This is a schematic diagram of a multi-view feature fusion process guided by a feature-level gating network provided by the present invention. Figure 4 This is a schematic diagram of the implementation process of the method provided by the present invention; Figure 5 A schematic diagram of the structure of the drug-related side effect prediction system driven by topology-enhanced hypergraph learning and multigraph converter provided by the present invention; Figure 6 The performance comparison of the drug-related side effect prediction system driven by topology-enhanced hypergraph learning and multigraph converter provided by this invention with the comparative method on ROC and PR curves. Detailed Implementation

[0022] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0023] Example 1 Please see Figure 1 This embodiment provides a method for predicting drug-related side effects driven by topology-enhanced hypergraph learning and multigraph converters, including: Step S101: Obtain datasets related to drugs, diseases, and side effects; Step S102: Construct a multi-view heterogeneous graph and multiple drug-side effect hypergraphs based on the drug, disease and side effect related dataset, and construct a corresponding line graph based on each drug-side effect hypergraph; Step S103: Based on the cross-hypergraph and line graph learning strategy, a hypergraph convolutional neural network is used to extract node features from the hypergraph, and a line graph convolutional neural network is used to extract node features from the line graph. The two types of node features are bidirectionally enhanced through a feature cross-propagation mechanism, and the enhanced two types of features are fused using a convolutional pooling layer to obtain the first feature matrix. Step S104: Using the adaptive multi-graph Transformer model, integrate the node features of multi-view heterogeneous graphs to obtain the second feature matrix; Step S105: The first feature matrix and the second feature matrix are fused from multiple perspectives using a feature-level gating network to obtain fused features; Step S106: Concatenate the fused features with the node features of the heterogeneous graph to form a drug-side effect node pair feature representation, and train the model based on the cross-entropy loss function to obtain the drug-side effect association model; Step S107: Input the drug information and side effect information of the relationship to be predicted into the drug-side effect association model, and output the association prediction results of drugs and side effects.

[0024] The method in this embodiment constructs a multi-view heterogeneous graph and a drug-side effect hypergraph. Based on a cross-hypergraph and line graph learning strategy, it extracts node features using a hypergraph convolutional neural network and a line graph convolutional neural network, then integrates these features using an adaptive multi-graph Transformer. Finally, it trains a model through feature fusion and a cross-entropy loss function to predict the association between drugs and side effects. This invention's method effectively improves the accuracy and robustness of drug-side effect relationship prediction through multi-view feature fusion, graph structure learning, and feature cross-propagation mechanisms.

[0025] As a specific embodiment, in step S101, raw datasets are obtained from the Comparative Toxicology-Genomics Database and the SIDER database. The datasets include drug-disease association data, drug-side effect association data, medicinal chemical substructure data, disease-related data, and side effect-related data. Further, based on existing similarity measurement methods, we calculate drug functional similarity, medicinal chemical substructure similarity, and side effect similarity based on drug-related diseases, medicinal chemical substructures, and drugs related to side effects, respectively. By establishing standardized raw data, a legitimate, reproducible, and interpretable data foundation can be provided for subsequent construction of heterogeneous graphs, hypergraphs, line graphs, etc.

[0026] In a preferred embodiment, step S102 involves constructing a multi-view heterogeneous graph based on the drug, disease, and side effect related dataset, including: Based on the aforementioned drug, disease, and side effect related dataset, extract the association between drugs and side effects, drug chemical substructure information, drug-related disease information, and side effect-related drug information; A drug-side effect correlation matrix is ​​generated based on the relationship between drugs and side effects; a drug similarity matrix is ​​generated based on drug chemical substructure information and drug-related disease information; and a side effect similarity matrix is ​​generated based on side effect-related drug information. The drug similarity matrix includes a structural similarity matrix and a related disease similarity matrix. Based on the same set of nodes, a first heterogeneous graph is constructed using the drug-side effect association matrix, side effect similarity matrix, and structural similarity matrix; a second heterogeneous graph is constructed using the drug-side effect association matrix, side effect similarity matrix, and related disease similarity matrix; wherein, the first heterogeneous graph and the second heterogeneous graph reflect the node connection relationships of drugs from the perspective of chemical structure and related diseases, respectively.

[0027] As a specific example, the construction process of a multi-view heterogeneous graph is as follows: First, use a set of nodes. To represent entities that represent drugs and side effects; among which, N represents d Drug nodes ; N represents s Side effect node .

[0028] Secondly, determine the connections between nodes, including connections between different types of drug-side effect nodes, and internal connections between nodes of the same type. Specifically: 1) The relationship between drugs and side effects (representation of relationships): Drug-Side Effect Relationship Matrix The definition is as follows: Wherein, given ,if =1, then it means and There is a correlation between them; otherwise =0.

[0029] 2) Internal drug relationships: The internal relationship matrix for drugs includes two internal similarity matrices and two drug similarity matrices. Represented as: in, This is a structural similarity matrix, also known as a medicinal chemical substructure similarity matrix. It assesses the degree of similarity between two drugs by calculating the chemical substructure similarity between them. It is a disease similarity matrix. Based on the biological premise that "if two drugs are more similar to the disease sets they are associated with, then the drugs usually exhibit more similar functions", it uses Jaccard similarity to calculate the overlap between the two sets and quantifies the functional similarity between drugs for related diseases. ( This describes the similarity between drug nodes, with values ​​ranging from 0 to 1.

[0030] 3) Internal relationships of side effects: Side effect similarity matrix The definition of is: When there are more associated drugs of the same class for two side effects, the side effects are generally more similar. We obtained data from the dataset related to side effects. Related drug collection and side effects Related drug collection By calculating the drug set and We can obtain the similarity between them. and The similarity between them.

[0031] Finally, based on the similarity between the two drugs, two heterogeneous graphs consisting of drugs and side effects were constructed respectively. and ;in, From the perspective of medicinal chemistry substructure, nodes A set of heterogeneous edges From the perspective of drug-related diseases, nodes A set of heterogeneous edges.

[0032] Heterogeneous graph Feature matrix It can be represented as: (4) in, This represents the total number of nodes representing drugs and their side effects. Representation of the correlation matrix The transpose of .

[0033] In a preferred embodiment, in step S102, multiple drug-side effect hypergraphs are constructed based on the drug, disease, and side effect related dataset, including: Based on the structural similarity matrix, the related disease similarity matrix, and the side effect similarity matrix, after setting similarity thresholds for filtering, a first hypergraph containing the hyperedge set from the perspective of drug structure and a second hypergraph containing the hyperedge set from the perspective of drug-related diseases are constructed respectively. Each hypergraph uses an adjacency matrix to describe the relationship between hyperedges and drugs and side effects. The i-th column of the adjacency matrix represents the i-th hyperedge of the corresponding hypergraph.

[0034] Furthermore, a corresponding line graph is constructed based on each drug-side effect hypergraph, including: Each hyperedge in each hypergraph is treated as a node to construct a corresponding line graph, which is used to represent the topological structure between hyperedges; If two hyperedges have at least one common node in the original hypergraph, then an edge is established between the corresponding nodes in the line graph. The weight of the edge between connected nodes is the ratio of the number of nodes shared by the two hyperedges to their total number of nodes. Based on the weight of the edge between connected nodes, the adjacency matrix and node feature matrix of the line graph are obtained.

[0035] It should be noted that, since a drug often has multiple side effects, and multiple different drugs may synergistically trigger a single side effect, complex relationships exist between multiple drugs and side effect nodes. Therefore, to represent this relationship, as a specific example, we construct two drug-side effect hypergraphs composed of multiple hyperedges. The implementation steps are as follows: First, for the similarity matrix We set a threshold. ,reserve The similarity values ​​greater than or equal to the corresponding threshold are used to obtain the filtered matrix. .

[0036] Secondly, based on Construct two drug-side effect hypergraphs and ;in, Represents a set of nodes ( (Entities used to represent drugs and side effects) It is a super-edge set. For the perspective of medicinal chemistry substructure, hyperedge set, It is a hyperedge set from the perspective of drug-related diseases.

[0037] Next, build a hypergraph. adjacency matrix To describe the relationship between hyperedges, drugs, and side effects, we define an adjacency matrix. for: in, The i-th column of the adjacency matrix represents the i-th hyperedge of the hypergraph. .

[0038] from From the perspective of nodes Similarities and associations with all drugs and side effects are included. Therefore, It can also be viewed as the feature matrix of drug and side effect nodes in the hypergraph. .

[0039] Similarly, hypergraph adjacency matrix It can also be obtained in the same way, which will not be elaborated here.

[0040] Since each hyperedge in a hypergraph can be considered a new node to build a line graph, this helps to reflect the topological structure between hyperedges, thus providing richer structural information for subsequent drug-side effect association prediction. Therefore, in some embodiments, a hypergraph is used... For example, its line chart Defined as ,in It is a set of nodes. It is an edge set.

[0041] like and In SuperGraph If there is at least one common node, then the corresponding node is considered to be... and Online chart The nodes are connected. The weight of the edge between two connected nodes... The calculation is based on the ratio of the number of nodes shared by the two corresponding hyperedges to their total number of nodes, expressed by the formula: Line chart The adjacency matrix can be represented as And satisfy: in, The i-th column contains the weight relationships between the i-th hyperedge node and all other nodes. By weighted aggregation of the features of multiple drug and side effect nodes in the hyperedge, we can obtain... Node feature matrix , is represented as: in, It is the characteristic matrix of the hypergraph nodes. It is an adjacency matrix, through Aggregate the features of multiple nodes connected by a single hyperedge.

[0042] The same method can be used to obtain a line graph. adjacency matrix and node feature matrix This will not be elaborated upon here.

[0043] In a preferred embodiment, step S103 involves using a hypergraph convolutional neural network to extract node features from the hypergraph, including: The hypergraph convolutional neural network contains a multi-layer convolutional structure, with drugs and side effects as nodes, and hyperedges filtered by similarity thresholds as connection units. The features of the nodes connected to each hyperedge are aggregated to obtain the edge features of the hyperedge. The edge features of each hyperedge are distributed back to each node it is connected to, and fused with the original features of the nodes to complete the forward computation of one hypergraph convolutional layer. By stacking multiple layers of the hypergraph convolutional layers, the embedding representations of drug nodes and side effect nodes are updated layer by layer to obtain hypergraph node features that integrate the similarity and association of members within the hyperedge.

[0044] In a preferred embodiment, the step of using a line graph convolutional neural network to extract node features from the line graph includes: The line graph convolutional neural network contains a multi-layer convolutional structure. The network uses hyperedges as nodes and Jaccard similarity between hyperedges as edge weights. It extracts the topological features of hyperedges through Laplacian normalized graph convolution. By leveraging node-hyperedge associations, the topological features of hyperedges are backpropagated to the original drug and side effect nodes, thereby achieving joint enhancement of the hyperedge structural context and node attribute information.

[0045] As a specific example, in drug-side effect association prediction, hypergraphs and line graphs establish the complex relationships between multiple nodes and the topological structure of hyperedges from different perspectives. We propose FLHL (a learning strategy that spans both hypergraphs and line graphs). This will be explained in detail below: To more accurately encode the complex relationships between multiple drugs and their side effects, we designed Hypergraph Convolutional Neural Networks (HGCNs) with multi-layer convolutional structures. HGCNs networks aggregate node features to hyperedges and then spread the hyperedge features back to the nodes, completing one convolution in two steps. Combined with layer-by-layer learnable weights and non-linear activations, this enables multi-order feature propagation on the hypergraph structure.

[0046] Specifically, for hypergraphs Adjacency matrix and characteristic matrix The input is fed into the HGCNs model. After l layers of encoding, the output feature matrix is ​​obtained. : in, and It is a weight matrix. This is the LeakyReLU activation function, where L is the number of HGCNs layers. (Diagonal matrix) and Let represent the degree matrix of the hypergraph nodes and the degree matrix of the hyperedges, respectively. .

[0047] For hypergraphs , its adjacency matrix and characteristic matrix The input is fed into another independent HGCNs to learn the feature matrices of all drug and side effect nodes. .

[0048] Each hyperedge connects multiple hypergraph nodes, and the multiple hyperedges together form the topology from the perspective of the hyperedge.

[0049] Traditional hypergraph convolutional neural networks often focus only on the aggregation of node features, ignoring the topological structure of hyperedges, resulting in the loss of important hyperedge structural information. To learn node features with hyperedge topology enhancement, we designed line graph convolutional neural networks (LGCNs) containing multi-layer convolutional structures.

[0050] Adjacency Matrix After Laplace normalization, we obtain: ;in, It is a line chart The degree matrix of a node. Node feature matrix Adjacency matrix of hypergraph The feature matrices of drug and side effect nodes are obtained at the l-th layer of the LGCNs network. , in, It is a weight matrix. = . and (l>1) is the feature matrix of the learned line graph nodes, which is related to the hypergraph adjacency matrix. Multiplication can pass the learned hyperedge topology to the drug and side effect node feature learning, thereby enhancing the expressive power of these node features.

[0051] Similarly, a line graph can be obtained through another independent LGCNs network. Node feature matrix .

[0052] Hypergraphs focus on the direct associations between drug nodes and side-effect nodes within specific hyperedges, while line graphs are better able to learn the topological structure formed by the nodes along the hyperedges. Therefore, it is necessary to simultaneously optimize node features in both graph spaces.

[0053] In a preferred embodiment, the feature cross-propagation mechanism in step S103 includes: In each layer of the online graph convolutional neural network (OBN) and the hypergraph convolutional neural network (HPN), the node features output by the HPN are used as the query vector, and the node features output by the online graph convolutional neural network (OBN) are used as the key-value vector. Simultaneously, the node features output by the online graph convolutional neural network (OBN) are used as the query vector, and the node features output by the hypergraph convolutional neural network (HPN) are used as the key-value vector. A bidirectional cross-enhancement mechanism using scaling dot product attention is employed to generate updated node features.

[0054] Specifically, such as Figure 2 As shown, we designed FCP (Feature Cross-Propagation Strategy). Figure 2 This diagram illustrates the process of feature cross-propagation between a hypergraph and a line graph.

[0055] Through the Feature Cross-Propagation (FCP) strategy, the input of the l-th layer (l > 1) of HGCNs includes not only the node feature matrix of its own (l-1)-th layer, but also... Furthermore, it incorporates the node feature matrix of the (l-1)th layer of LGCNs enhanced by the cross-attention mechanism. .

[0056] Similarly, the input of the l-th layer of LGCNs also incorporates the node feature matrix of the (l-1)-th layer of the enhanced HGCNs. After cross-propagation, we obtained... and , is represented as: (11) Cro is a cross-attention operation.

[0057] At once In contrast, the feature matrix of line graph topology enhancement The value matrix after linear transformation is considered as , (12) in, It is a weight matrix.

[0058] Feature matrix obtained by hypergraph convolution Feature matrix of line graph topology enhancement The query matrix is ​​obtained by performing linear transformations respectively. and key matrix , (13) in, It is a weight matrix.

[0059] Next, we calculate And scale it, where This indicates the importance of the j-th node feature vector learned by LGCNs to the i-th node feature vector learned by HGCNs.

[0060] By applying the softmax function to the scaled result, we can obtain the normalized attention weight matrix.

[0061] Finally, we multiply the attention weight matrix by the value matrix. Obtain the features of the superedge nodes , , (14) in, This is the feature dimension of each drug and side effect node in the (l-1)th layer of HGCNs. Because It contains more detailed information, which we will compare with The summation forms the hypergraph node features enhanced by cross attention. .

[0062] Using the same method, we can obtain cross-attention-enhanced line graph node features that contain both hyperedge topological information and node details from the perspective of the corresponding hypergraph drug structure. Hypergraph features enhanced by cross-attention, which are associated with nodes from both the perspective of drug-related diseases and the perspective of corresponding line graphs. And the cross-attention enhanced line graph node features, which contain both hyperedge topological information and node details from the perspective of drug-related diseases in the corresponding hypergraph. .

[0063] After passing through L layers of HGCNs and LGCNs, we obtain... . They are stacked at the channel level to form .

[0064] To preserve edge information, we... Padding with zeros yields the expanded matrix. Furthermore, convolution and pooling operations are employed to further integrate the features. In the convolutional layers, the filter size and number are set to [values ​​to be filled in]. And cnum. The average pooling layer uses a cnum×1×1 window to perform dimensionality reduction at the channel level. Based on convolution and pooling operations, a node feature matrix with complex connections and hyperedge topology enhancement can be obtained. This can be expressed as a formula: (15) Here, average indicates that average pooling is performed at the channel level. This represents the convolution operation. and These are the filters and bias terms of the convolutional layer, respectively.

[0065] In a preferred embodiment, in step S104, the process of integrating node features of multi-view heterogeneous graphs using an adaptive multi-graph Transformer model to obtain a second feature matrix includes: The node features of the first heterogeneous graph and the second heterogeneous graph are mapped to corresponding query, key, and value vectors respectively through learnable weight matrices; Learnable scalar weights are introduced into the two sets of query, key, and value vectors. After softmax normalization, the dynamic contribution coefficients of the drug structure perspective and the drug-related disease perspective are obtained, thereby realizing the information ratio fusion of the first heterogeneous graph and the second heterogeneous graph. Multi-head scaled dot product attention is performed on the fused query, key, and value to capture the contextual dependencies between drug-side effect nodes and output attention features; The attention features are input into the feedforward neural network, and after residual connections and layer normalization, the updated node embeddings are obtained. After multi-layer stacking, the final node embeddings are output, which constitute the second feature matrix, and are used for subsequent multi-view fusion and association prediction.

[0066] As a specific example, heterogeneous graphs and These respectively imply drug chemical substructure similarity and drug functional similarity, as well as shared side effect similarity and drug-side effect association.

[0067] Therefore, they reflect the relationship between drugs and side effects from different perspectives. The node features of each heterogeneous graph provide important auxiliary information for predicting drug side effect associations, but their contribution is difficult to determine in advance. In addition, there are contextual relationships between node features in multiple heterogeneous graphs, and capturing these relationships can effectively improve the expressive power of node representations.

[0068] We designed AMT (Adaptive Multi-Graph Transformer) to dynamically integrate information from multiple heterogeneous graphs and capture the contextual relationships between node features in multiple heterogeneous graphs.

[0069] In the first coding layer (d = 1), and Three pairs of learnable weight parameters were assigned. and They are independently normalized using the softmax function. The normalized weight parameters are... and , is represented as: (16) in, .

[0070] We introduce a multi-head attention mechanism (MHA) to reduce variance during the learning process by running multiple independent attention computation modules in parallel. For the m-th attention head (i.e., the last attention head), we use a weight matrix. , and For the characteristic matrix Perform a linear transformation.

[0071] Using weight parameters and The feature matrix after linear transformation is weighted and fused to obtain the query matrix. key matrix Sum matrix , (17) in, and It can adaptively adjust the contribution of multiple heterogeneous graph features to ensure that the features of each heterogeneous graph can be reasonably utilized according to their actual contribution to the prediction of drug-side effect associations.

[0072] After adaptively fusing features from multiple heterogeneous graphs, we calculate Form an attention score matrix, where Recorded all drug and side effect nodes Attention score.

[0073] Then we will After normalization and Multiplying them yields the feature matrix of the m-th attention node. , (18) in, This refers to the dimension of each attention head. The output of MHA is obtained by concatenating the results of all attention heads. : (19) The input is fed into a feedforward neural network (FFN) and subjected to residual connections and layer normalization to obtain the feature matrix of the first coding layer of the AMT. , (20) LN stands for layer normalization operation.

[0074] Since it has already been completed in the first coding layer and Adaptive feature fusion allows subsequent linear transformations to focus on enhancing the expressive power of the features. Therefore, in encoding layers with d ≥ 2, we only perform linear transformations on the node features obtained from the previous layer. (twenty one) in, These are the node features obtained from the (d-1)th encoding layer. , , It is a weight matrix. Multi-head attention is used to learn the contextual relationships between features of nodes in a heterogeneous graph.

[0075] After D encoding layers, we can obtain the feature representation of the multi-heterogeneous graph. .

[0076] In a preferred embodiment, in step S105, the first feature matrix and the second feature matrix are fused using a feature-level gating network to obtain fused features, including: The first feature matrix and the second feature matrix are projected onto a unified dimension space through linear mapping; After performing the concatenation operation, the sigmoid activation function is input to generate the gated weight matrix; The two mapped feature matrices are then fused element-wise using the gate weights to obtain the fused features.

[0077] Specifically, as mentioned earlier, FLHL (a learning strategy across hypergraphs and line graphs) can encode complex relationships between multiple drug and side effect nodes, as well as the topological structure between line graphs, and obtain a feature matrix. AMT (Adaptive Multi-Graph Transformer) adaptively integrates the features of multiple heterogeneous graph nodes to obtain a feature matrix. To fuse node features from both the hypergraph and multi-heterogeneous graph perspectives, we established FLGN (Feature-Level Gated Network). and Each feature matrix is ​​multiplied by a different weight matrix, thus projecting them onto a unified feature space. The resulting feature matrices are then summed element-wise to form the final feature space. , (twenty two) Where tanh is the activation function. , This is the weight matrix. The weight matrix of FLGN is represented as follows: , (twenty three) Where sigmoid is the activation function. and ∥ represents the weight matrix and bias terms, and ∥ represents the concatenation operation.

[0078] Finally, the gate weight matrix S is used to guide F in forming the final fusion matrix Z. (twenty four) In this context, ⊙ represents the Hadamarda complex. like Figure 3 As shown, Figure 3 The process of multi-view feature fusion guided by a feature-level gating network of d2 is demonstrated.

[0079] In some embodiments, nodes for drugs and side effects in a heterogeneous graph and This contains more detailed information. Therefore, in step S106, we add the two matrices element-wise and concatenate the result with the fusion matrix Z to obtain... , (25) in, The corresponding i-th drug node The feature vector is , This is a side effect node. The feature vectors are stacked to form... and The final expression : (26) In this context, ":" indicates a stacking operation.

[0080] We further integrate this through multi-layer convolution and pooling. After the l-th convolutional pooling layer, the node pair features are represented as , is represented as: (27) in, = `max` represents the max pooling operation. The activation function is indicated by *, which represents the convolution operation. and These are the filters and biases of the l-th convolutional layer.

[0081] The feature representation obtained from the last convolutional pooling layer is as follows: It generates probability scores for drug-side effect associations through fully connected layers. .

[0082] During model training, we use the cross-entropy function to calculate the model's prediction loss. (28) in, This represents the number of samples in each training batch. This indicates the true correlation between drugs and their side effects. ( ) represents a drug node With side effect nodes The predicted probability of no (existing) association.

[0083] To illustrate the above process more clearly, as follows: Figure 4 As shown, Figure 4 The figure shows a schematic diagram of the implementation process of the method in this embodiment. In the figure, (a) shows the construction process of multiple hypergraphs, multiple line graphs and multiple heterogeneous graphs, (b) shows the feature learning across hypergraphs and line graphs (FLHL), (c) shows the multi-heterogeneous graph node feature learning (AMT) process based on adaptive multi-graph Transformer, and (d) shows the multi-view feature fusion (FLGN) based on feature-level gating network.

[0084] Example 2 like Figure 5 As shown, this embodiment of the invention provides a drug-related side effect prediction system 500 driven by topology-enhanced hypergraph learning and multigraph converter, comprising: Data acquisition module 501 is used to acquire datasets related to drugs, diseases, and side effects; The graph structure construction module 502 is used to construct a multi-view heterogeneous graph and multiple drug-side effect hypergraphs based on the drug, disease and side effect related dataset, and to construct a corresponding line graph based on each drug-side effect hypergraph; The first feature extraction module 503 is used to extract node features of the hypergraph and line graph based on the cross-hypergraph and line graph learning strategy, using a hypergraph convolutional neural network to extract node features of the hypergraph and a line graph convolutional neural network to extract node features of the line graph, using a feature cross-propagation mechanism to enhance the two types of node features in both directions, and using a convolutional pooling layer to fuse the enhanced two types of features to obtain the first feature matrix; The second feature extraction module 504 is used to integrate the node features of multi-view heterogeneous graphs using an adaptive multi-graph Transformer model to obtain the second feature matrix. The fusion module 505 is used to fuse the first feature matrix and the second feature matrix through a feature-level gating network to obtain fused features; Training module 506 is used to concatenate the fused features with the node features of the heterogeneous graph to form a drug-side effect node pair feature representation, and to train the model based on the cross-entropy loss function to obtain a drug-side effect association model. The prediction module 507 is used to input the drug chemical substructures and candidate side effect types of the association to be predicted into the drug-side effect association model, and output the association prediction results between the drug and the side effect.

[0085] To demonstrate the effectiveness of this invention, the system proposed in this embodiment (hereinafter referred to as MGLG for ease of description) was implemented based on the PyTorch framework and trained on an Nvidia GeForce RTX 4090 GPU card. The model training used 40 epochs, with a batch size of 64, and a learning rate of 0.00001. and The values ​​are set to 0.4, 0.4, and 0.5 respectively. In FLHL, both hypergraph and linegraph convolutional layers have two layers, with output feature dimensions of 1024 and 256 for each layer. The number of filters (cnum) in the convolutional layers is set to 4, and the kernel size is 3×3. The average pooling layer uses a 4×1×1 window. In AMT, the feature dimension... The number of nodes is 1225, with 4 attention heads per layer, for a total of 2 layers. Node-to-node feature learning consists of two convolutional layers, each with a kernel size of 2×8, and two pooling layers with window sizes of 1×4 and 2×4, respectively.

[0086] A comparison with nine new predictive methods and a case study of five drugs demonstrates that MGLG achieves superior predictive performance and exhibits a strong ability to screen reliable drug-related side effect candidates. Ablation experiments were also performed to further demonstrate the impact of MGLG's innovations on overall performance.

[0087] MGLG was evaluated using a five-fold cross-validation method. Specifically, all known drug-side effect associations (i.e., positive samples) were randomly divided into five equal groups, while all unproven drug-side effect node pairs were treated as negative samples. In each fold of cross-validation, four positive samples were selected as training samples, and an equal number of negative samples were randomly selected and added to the training set. The remaining positive and negative samples not used for training were used for testing. The ratio of positive to negative samples was 1:36, indicating a class imbalance. Class imbalance makes evaluation metrics such as precision and recall difficult to accurately reflect the model's performance. The area under the receiver operating characteristic (AUC) can evaluate the overall performance of the model at different thresholds, while the area under the precision-recall curve (AUPR) focuses more on the ability to identify positive samples and is suitable for imbalanced datasets. Therefore, we used AUC and AUPR as evaluation metrics. In the five-fold cross-validation, we calculated the AUC and AUPR values ​​for each drug in each fold, and took the average of these five folds as the final AUC and AUPR score for each drug. In addition, considering that biologists tend to prioritize the top-ranked side effects for further validation, we calculated the recall rate of the top k drug-related side effects.

[0088] To validate the effectiveness of FLHL (Feature Learning Strategies Across Hypergraphs and Line Graphs), FCP (Feature Cross-Propagation Strategies), AMT (Adaptive Multi-Graph Transformer), and FLGN (Feature-Level Gated Network), we conducted ablation experiments. The complete model, including FLHL, FCP, AMT, and FLGN, achieved state-of-the-art performance in both mean AUC and AUPR, at 0.991 and 0.466, respectively. Compared to the complete model, removing FLHL resulted in a 0.2% decrease in AUC and a 4.2% decrease in AUPR. This demonstrates the crucial role of encoding multiple drug-side association relationships and integrating hyperedge topologies in improving the accuracy of drug-side association predictions. The model without FCP showed a 0.1% decrease in AUC and a 1.5% decrease in AUPR, indicating the necessity of cross-propagation hypergraph and line graph node features for improving predictive performance. When AMT was removed, the complete model's AUPR decreased by 1.5%, demonstrating the effectiveness of AMT in adaptively integrating features from different heterogeneous graphs. Compared to the model without FLGN, our full model improved AUC and AUPR by 0.3% and 4.2%, respectively. This demonstrates that FLGN significantly enhances the predictive performance of the model when fusing features from both hypergraph and multi-heterogeneous graph learning perspectives. Experimental results show that FLGN contributes the most to predicting drug-side effect combinations. This is mainly attributed to its gating mechanism's ability to effectively fuse key features from different perspectives.

[0089] Furthermore, nine advanced drug-side effect association prediction methods were compared with MGLG. These nine prediction methods include MVDSA, ICAL, Net, TCSD, GCRS, SDPred, Galeaon's method, Ding's method, and FGRMF. Each method uses the optimal parameters recommended in its original paper, and our model and the comparison methods are trained and tested using the same dataset and partitioning.

[0090] MVDSA: This method constructs a functional knowledge graph of drugs and side effects and a structural knowledge graph of drugs and side effects, and uses a relation-gated semantic encoder and MLP to predict potential associations between drugs and side effects; ICAL: This model builds two hypergraphs based on the similarity of two types of drugs, and uses hypergraph convolutional networks and convolutional neural networks to predict potential drug-side effect associations; Net: This method constructs a drug chemistry molecular graph and a side effect similarity graph. It uses a framework based on graph attention learning and self-attention interaction to infer the association between drugs and side effects; TCSD: This method constructs two heterogeneous graphs of drug side effects using two types of drug similarity, and predicts drug-related side effects through a heterogeneous graph transformer and capsule network; GCRS: This method forms two heterogeneous graphs covering drugs and side effects. It uses GCA with representation-level attention and a multilayer convolutional neural network with attribute-level attention to explore the association between drugs and side effects; SDPred: This method integrates multiple drug similarities and multiple side effect correlations, and learns the features of drug and side effect nodes based on MLP and convolutional neural networks; Galeaon's method: This method learns the signatures of drugs and side effects by constructing a matrix factorization model; Ding's method: This method constructs a model based on multi-kernel learning and Kronecker regularized least squares to predict candidate drug-related side effects; FGRMF: This method uses a feature-driven graph regularized matrix factorization model to infer the tendency of drug side effect associations.

[0091] The final prediction results show that MGLG achieved the highest average AUC of 99.1%, which is 0.2% higher than MVDSA, 0.6% higher than ICAL, and higher than... Net is 3.0% higher, 1.4% higher than TCSD, 3.4% higher than GCRS, 4.5% higher than SDPred, 7.9% higher than Galeaon's method, 4.6% higher than Ding's method, and 7.2% higher than FGRMF. MGLG's average AUPR is 46.6%, higher than MVDSA, ICAL, Net, TCSD, GCRS, SDPred, Galeaon's method, Ding's method, and FGRMF showed higher recall rates of 4.0%, 6.1%, 25.0%, 10.5%, 19.4%, 24.0%, 33.5%, 27.5%, and 28.7%, respectively. The higher recall rates among top-ranked candidates indicate that MGLG is able to retrieve more real-world associations between drugs and side effects among the candidates.

[0092] Five-fold crossover experiments showed that MGLG achieved an AUC of 0.991 and an AUPR of 0.466, outperforming nine state-of-the-art methods in predicting drug-side effect associations. Ablation studies demonstrated the effectiveness of FLHL, AMT, FLGN, and FCP in cross-propagation of features in hypergraphs and linegraphs.

[0093] To more intuitively present the results of the five-fold crossover experiment, such as Figure 6 As shown, Figure 6 The figure shows a performance comparison of MGLG with nine advanced methods in terms of ROC and PR curves. In the figure, MGLG's ROC curve is closest to the top left corner, and its PR curve is closest to the top right corner, indicating that its AUC and AUPR are superior to the other nine comparative methods in predicting drug-side effect associations.

[0094] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting drug-related side effects driven by topology-enhanced hypergraph learning and multigraph converters, characterized in that, include: Obtain datasets related to drugs, diseases, and side effects; Based on the aforementioned drug, disease, and side effect related dataset, a multi-view heterogeneous graph and multiple drug-side effect hypergraphs are constructed, and a corresponding line graph is constructed based on each drug-side effect hypergraph. Based on the learning strategy of cross-hypergraph and line graph, a hypergraph convolutional neural network is used to extract node features from the hypergraph, and a line graph convolutional neural network is used to extract node features from the line graph. The two types of node features are enhanced bidirectionally through the feature cross-propagation mechanism, and the enhanced two types of features are fused using convolutional pooling layers to obtain the first feature matrix. By using an adaptive multi-graph Transformer model, node features of heterogeneous graphs from multiple perspectives are integrated to obtain a second feature matrix; The first and second feature matrices are fused from multiple perspectives using a feature-level gating network to obtain fused features. The fused features are concatenated with the node features of the heterogeneous graph to form a drug-side effect node pair feature representation, and the model is trained based on the cross-entropy loss function to obtain the drug-side effect association model. The drug chemical substructures and candidate side effect types whose associations are to be predicted are input into the drug-side effect association model, and the association prediction results between drugs and side effects are output.

2. The method according to claim 1, characterized in that, A multi-perspective heterogeneous graph was constructed based on the aforementioned drug, disease, and side effect related datasets, including: Based on the aforementioned drug, disease, and side effect related dataset, extract the association between drugs and side effects, drug chemical substructure information, drug-related disease information, and side effect-related drug information; A drug-side effect correlation matrix is ​​generated based on the relationship between drugs and side effects; a drug similarity matrix is ​​generated based on drug chemical substructure information and drug-related disease information; and a side effect similarity matrix is ​​generated based on side effect-related drug information. The drug similarity matrix includes a structural similarity matrix and a related disease similarity matrix. Based on the same set of nodes, a first heterogeneous graph is constructed using the drug-side effect association matrix, side effect similarity matrix, and structural similarity matrix; a second heterogeneous graph is constructed using the drug-side effect association matrix, side effect similarity matrix, and related disease similarity matrix; wherein, the first heterogeneous graph and the second heterogeneous graph reflect the node connection relationships of drugs from the perspective of chemical structure and related diseases, respectively.

3. The method according to claim 2, characterized in that, Multiple drug-side effect hypergraphs were constructed based on the aforementioned drug, disease, and side effect related datasets, including: Based on the structural similarity matrix, the related disease similarity matrix, and the side effect similarity matrix, after setting similarity thresholds for filtering, a first hypergraph containing the hyperedge set from the perspective of drug structure and a second hypergraph containing the hyperedge set from the perspective of drug-related diseases are constructed respectively. Each hypergraph uses an adjacency matrix to describe the relationship between hyperedges and drugs and side effects. The i-th column of the adjacency matrix represents the i-th hyperedge of the corresponding hypergraph.

4. The method according to claim 3, characterized in that, The construction of the corresponding line graph based on each drug-side effect hypergraph includes: Each hyperedge in each hypergraph is treated as a node to construct a corresponding line graph, which is used to represent the topological structure between hyperedges; If two hyperedges have at least one common node in the original hypergraph, then an edge is established between the corresponding nodes in the line graph. The weight of the edge between connected nodes is the ratio of the number of nodes shared by the two hyperedges to their total number of nodes. Based on the weight of the edge between connected nodes, the adjacency matrix and node feature matrix of the line graph are obtained.

5. The method according to claim 1, characterized in that, Hypergraph convolutional neural networks are used to extract node features from the hypergraph, including: The hypergraph convolutional neural network contains a multi-layer convolutional structure, with drugs and side effects as nodes, and hyperedges filtered by similarity thresholds as connection units. The features of the nodes connected to each hyperedge are aggregated to obtain the edge features of the hyperedge. The edge features of each hyperedge are distributed back to each node it is connected to, and fused with the original features of the nodes to complete the forward computation of one hypergraph convolutional layer. By stacking multiple layers of the hypergraph convolutional layers, the embedding representations of drug nodes and side effect nodes are updated layer by layer to obtain hypergraph node features that integrate the similarity and association of members within the hyperedge.

6. The method according to claim 5, characterized in that, The step of using a line graph convolutional neural network to extract node features from the line graph includes: The line graph convolutional neural network contains a multi-layer convolutional structure. The network uses hyperedges as nodes and Jaccard similarity between hyperedges as edge weights. It extracts the topological features of hyperedges through Laplacian normalized graph convolution. By leveraging node-hyperedge associations, the topological features of hyperedges are backpropagated to the original drug and side effect nodes, thereby achieving joint enhancement of the hyperedge structural context and node attribute information.

7. The method according to claim 6, characterized in that, The feature cross-propagation mechanism includes: In each layer of the online graph convolutional neural network and the hypergraph convolutional neural network, the node features output by the hypergraph convolutional neural network are used as query vectors, and the node features output by the online graph convolutional neural network are used as key and value vectors. At the same time, the node features output by the online graph convolutional neural network are used as query vectors, and the node features output by the hypergraph convolutional neural network are used as key and value vectors. The updated node features are generated through bidirectional cross-enhancement using a scaling dot product attention mechanism.

8. The method according to claim 2, characterized in that, The method utilizes an adaptive multi-graph Transformer model to integrate node features from multi-view heterogeneous graphs, resulting in a second feature matrix, which includes: The node features of the first heterogeneous graph and the second heterogeneous graph are mapped to corresponding query, key, and value vectors respectively through learnable weight matrices; Learnable scalar weights are introduced into the two sets of query, key, and value vectors. After softmax normalization, the dynamic contribution coefficients of the drug structure perspective and the drug-related disease perspective are obtained, thereby realizing the information ratio fusion of the first heterogeneous graph and the second heterogeneous graph. Multi-head scaled dot product attention is performed on the fused query, key, and value to capture the contextual dependencies between drug-side effect nodes and output attention features; The attention features are input into the feedforward neural network, and after residual connections and layer normalization, the updated node embeddings are obtained. After multi-layer stacking, the final node embeddings are output, which constitute the second feature matrix, and are used for subsequent multi-view fusion and association prediction.

9. The method according to claim 1, characterized in that, The first feature matrix and the second feature matrix are fused using a feature-level gating network to obtain fused features, including: The first feature matrix and the second feature matrix are projected onto a unified dimension space through linear mapping; After performing the concatenation operation, the sigmoid activation function is input to generate the gated weight matrix; The two mapped feature matrices are then fused element-wise using the gate weights to obtain the fused features.

10. A drug-related side effect prediction system driven by topology-enhanced hypergraph learning and multigraph converter, characterized in that, include: The data acquisition module is used to acquire datasets related to drugs, diseases, and side effects. The graph structure construction module is used to construct a multi-view heterogeneous graph and multiple drug-side effect hypergraphs based on the drug, disease and side effect related dataset, and to construct a corresponding line graph based on each drug-side effect hypergraph; The first feature extraction module is used to extract node features from the hypergraph and line graph based on the cross-hypergraph and line graph learning strategy. It uses a hypergraph convolutional neural network to extract node features from the hypergraph and a line graph convolutional neural network to extract node features from the line graph. It enhances the two types of node features in both directions through a feature cross-propagation mechanism and uses a convolutional pooling layer to fuse the enhanced two types of features to obtain the first feature matrix. The second feature extraction module is used to integrate the node features of multi-view heterogeneous graphs using an adaptive multi-graph Transformer model to obtain the second feature matrix. The fusion module is used to fuse the first feature matrix and the second feature matrix from multiple perspectives through a feature-level gating network to obtain fused features. The training module is used to concatenate the fused features with the node features of the heterogeneous graph to form a drug-side effect node pair feature representation, and to train the model based on the cross-entropy loss function to obtain a drug-side effect association model. The prediction module is used to input the drug chemical substructures and candidate side effect types of the association to be predicted into the drug-side effect association model, and output the association prediction results between drugs and side effects.

Images