A Two-View Drug Interaction Prediction Method Based on Multi-Source Feature Fusion and Hypergraph Enhancement

Abstract

This invention proposes a dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement, comprising the following steps: First, acquiring multi-source features of the drug: molecular structure features, network behavior features, similarity features, and semantic features, and combining them with a pre-trained knowledge graph to obtain a macroscopic embedding. Second, adaptive fusion of multi-source features is performed using a Transformer to generate an initial representation of the drug. Third, the initial representation is input into a dual-view iterative learning network. The inner view uses a graph attention network to extract atomic-level local features, while the outer view constructs a substructure hypergraph and a multi-attribute hypergraph. These are then fused using spectral weighting to form a dynamic hypergraph to model higher-order interactions between drugs. Learnable weights are introduced to fuse the dual-view representations, achieving view-level interpretability. Finally, the interaction probability is output through a predictor. This invention integrates multi-source information and multi-scale structures, introduces hypergraphs to model higher-order relationships, and possesses good predictive performance, generalization ability, and interpretability.

Description

Technical Field

[0001] This invention relates to the field of drug action prediction technology, specifically to a dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement. Background Technology

[0002] Drug-drug interaction (DDI) refers to the phenomenon where, when two or more drugs are used simultaneously or sequentially, the pharmacological effect or toxic reaction of one drug is altered by another. This alteration may lead to enhanced or diminished drug efficacy, or it may produce new adverse reactions.

[0003] With the widespread use of multidrug combination therapy in clinical practice, drug interactions have become a significant cause of adverse drug reactions. Statistics show that approximately 30% of adverse drug reactions are directly related to drug interactions, and this proportion is even more pronounced in elderly individuals with multiple comorbidities. Traditional drug interaction detection (DDI) methods rely primarily on in vitro experiments and clinical trials, which are not only costly and time-consuming but also struggle to provide comprehensive coverage given the exponentially growing pool of drug combinations.

[0004] Early DDI prediction methods were primarily based on traditional machine learning models, using manually designed drug features (such as molecular fingerprints and drug classification codes) to train classifiers like random forests, support vector machines, or logistic regression for prediction. The limitations of these methods are that feature engineering heavily relies on domain-specific knowledge, and manually designed features struggle to fully capture the complex structural and functional information of drug molecules, resulting in significant performance bottlenecks in prediction.

[0005] In recent years, graph neural networks (Graph Neural Networks) have been widely applied to drug interaction prediction tasks due to their ability to effectively represent graph structure data. However, existing DDI prediction methods based on Graph Neural Networks still have several shortcomings. Drug characterization sources are relatively singular; most methods rely solely on single information such as molecular structure diagrams or molecular fingerprints, neglecting complementary knowledge of drugs in different information spaces such as molecular structure, structural similarity, and pharmacological semantics, making it difficult to form a comprehensive and robust representation. Furthermore, there is insufficient modeling of higher-order interactions between drugs; most existing methods are based on ordinary graphs to characterize binary relationships between drugs, making it difficult to express the adverse effects of multiple drugs due to the same pathway or similar drug classifications. Meanwhile, multi-source information fusion methods lack adaptability; the few existing fusion studies mostly employ simple splicing or fixed weighting strategies, failing to effectively handle differences in semantic space, feature scale, and information density between different modalities, easily introducing redundant noise and semantic conflicts. Moreover, existing models often view drug molecules in isolation, lacking linkage with external authoritative knowledge graphs, making it difficult to further reveal the key molecular internal structural features or external biological factors leading to interactions, thus limiting their practical application value in drug development and clinical decision-making. Summary of the Invention

[0006] Objective: To overcome the shortcomings of existing technologies in terms of feature dimension, interaction modeling, and interpretability, this invention proposes a dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement. By introducing a Transformer architecture to fuse multi-source heterogeneous features and macroscopic knowledge, and combining a dual-view substructure network with a multi-scale hypergraph convolutional network, comprehensive modeling from the atomic level, substructure level to the view level is achieved, thereby improving the accuracy and clinical interpretability of DDI prediction.

[0007] Technical solution: This invention also discloses a dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement, comprising:

[0008] Step 1: Extract drug molecular graph structural features, drug behavior features, drug structural similarity features, and drug semantic features to obtain macroscopic knowledge graph features from the pre-trained knowledge graph;

[0009] Step 2: Adaptively fuse multi-source features with macro-knowledge graph features to form an enhanced initial embedding representation of the drug. After linear projection, the enhanced initial embedding is injected into the initial feature vector of each atom node in the molecular graph in a broadcast manner to form an optimized initial atom feature matrix.

[0010] Step 3: Optimize the initial atomic feature input graph attention network GAT, perform atomic-level local message passing and feature aggregation inside the drug molecule graph, capture local chemical bond associations within the molecule, and then extract substructure-level embeddings through hierarchical pooling to obtain the Intra-view embedding representation;

[0011] Step 4: Construct a substructure hypergraph based on substructure nodes, construct a multi-attribute hypergraph based on drug multi-dimensional attribute information, perform weight allocation and weighted fusion on multiple hypergraphs to determine the final dynamic hypergraph, and capture high-order interaction features between drugs through hypergraph convolutional networks to obtain the Inter-view embedding representation;

[0012] Step 5: Introduce learnable weight parameters to adaptively fuse the Intra-view embedding representation and the Inter-view embedding representation to construct a joint feature vector representation of the drug pair, and output the interaction probability through the predictor.

[0013] Furthermore, multi-source features specifically include:

[0014] The RDKit tool was used to parse the drug's SMILES string into a molecular graph. Using graph neural networks to aggregate feature vectors of atomic nodes and chemical bond edge characteristics Output 3D molecular structure embedding ;

[0015] In drug interaction networks Perform a biased random walk. For the set of all drug nodes, Given a set of known drug-drug interaction edges, generate a behavioral embedding that encodes drug topological location information. ;

[0016] Based on Morgan molecular fingerprinting, the Tanimoto similarity matrix between the target drug and all other drugs in the drug library is calculated. The similarity vector of the target drug is then nonlinearly mapped using a multilayer perceptron to extract similarity embeddings that encode the spatial distribution information of the drug's chemistry. ;

[0017] Input the drug text description into the pre-trained BioBERT model to extract... The semantic embedding is obtained by linearly projecting the position vector. ;

[0018] Loading drug entity embeddings pre-trained using the ComplEx method from the PrimeKG knowledge graph For drugs not included in the catalog, their Initialize as a vector of all zeros.

[0019] Furthermore, the specific method for adaptive fusion of multi-source features is as follows:

[0020] First, embed the behavior Similarity embedding Semantic embedding Knowledge graph embedding Each is mapped to a unified path through an independent linear projection layer. The 3D space is cascaded, and the side information is comprehensively embedded through multilayer perceptron (MLP) mapping. ;

[0021] Secondly, embedding molecular structures Integrated embedding with side information The Transformer encoder, acting as the sequence input, performs feature reconstruction through a multi-head self-attention mechanism; mean pooling is performed on the Transformer output sequence to generate the initial embedding for drug enhancement. Finally, the residual superposition mechanism is used to... Through linear projection matrix Mapping to obtain the projection vector The initial features are superimposed element by element to each atomic node in a broadcast manner. The optimized initial atomic features are obtained as follows: .

[0022] Furthermore, the specific method of Intra-view encoding in step 3 is as follows:

[0023] The optimized atomic feature matrix is ​​input into a multi-layer graph attention network. The importance weights of neighboring atoms are dynamically calculated through a self-attention mechanism. After Softmax normalization, the neighbor features are aggregated and the atomic representation is updated.

[0024] By stacking Layered GAT recursively aggregates chemical information from its multi-order neighborhood at each atomic node, capturing the local chemical microenvironment features determined by chemical bonds within the molecule, and obtaining an updated atomic feature matrix. ;

[0025] Based on the current atomic feature matrix Adjacency matrix of molecular graph The self-attention graph pooling algorithm SAGPool is used to learn the self-attention score vector of each atomic node. After sorting, the top-ranked atomic nodes are retained. The group of high-scoring atomic nodes retained represents the key substructures in the drug molecule, and the atomic-level graph is gradually shrunk into a substructure-level graph.

[0026] Update the pooled drug atom feature matrix and adjacency matrix, and perform a global readout operation on the pooled substructure-level graph to obtain the drug's Intra-view embedding representation. .

[0027] Furthermore, the final dynamic hypergraph construction process is as follows:

[0028] Substructure hypergraph construction: Using the substructure nodes retained after pooling of drugs A and B as vertices, node associations are calculated based on cosine similarity, and the nodes with the highest similarity are linked. Each node is associated with a substructure hyperedge;

[0029] Multi-attribute hypergraph construction: Calculating Jaccard similarity between drugs based on their ATC codes, target proteins, and biological pathway information. ,Will Drug vertices exceeding a preset threshold are classified as superedges of the same attribute;

[0030] Weighted fusion of spectral Hamming similarity: Calculating the normalized Laplacian matrix of two hypergraphs The eigenvalue spectrum is extracted and mapped to binary hash codes; a baseline DDI hypergraph is constructed, and spectral hash codes are calculated; the spectral Hamming similarity between the spectral hash codes of each hypergraph and the baseline DDI hypergraph is calculated. Determine the fusion weights of each hypergraph. ;

[0031] The association matrices of the two hypergraphs are weighted and fused based on the fusion weights to construct the final dynamic hypergraph association matrix.

[0032] Furthermore, the message passing mechanism for obtaining the Inter-view embedding representation by the hypergraph convolutional network to capture high-order interaction features between drugs is as follows:

[0033] The update formula for node features in a hypergraph convolutional network is:

[0034] ,

[0035] in, For dynamic hypergraph incidence matrix, The hyperedge weight diagonal matrix is ​​determined by the weighted fusion of the longitudinal spectrum. For the first The node feature matrix of a hypergraph convolution. and Based on The calculated node degree matrix and hyperedge degree matrix, For the first The trainable parameter matrix of the layer, It is a non-linear activation function;

[0036] After multiple hypergraph convolutions, a global pooling operation is performed on the output node feature matrix to obtain the Inter-view embedding representation. .

[0037] Furthermore, the specific method of adaptive fusion in step 5 is as follows:

[0038] Introducing view-level learnable scalars , The final fusion weights are calculated through normalization: , The final drug embedding vector is obtained by fusion. , This indicates that the Inter-view is embedded. Embedded representation for Intra-view.

[0039] Furthermore, a joint feature vector representation of the drug pair is constructed, and the interaction probability is output through a predictor, specifically as follows:

[0040] drug pair The fusion vectors are combined to construct joint features. ,in, and This represents the final embedding vector of the drug. Indicates vector concatenation. Represents element-wise product. Represents the absolute value of each element. Input to a multilayer perceptron decoder, output interaction probability. .

[0041] Furthermore, after outputting the interaction probabilities, a multi-level interpretability analysis is performed, including:

[0042] View-level interpretability: by outputting normalized weights of Intra-view embedded representations and Inter-view embedded representations. and The numerical value quantifies the contribution of internal molecular features and external higher-order interaction features to the prediction results.

[0043] Substructure-level interpretability: Extract the substructure pair interaction weight matrix from the decoder, filter high-weight substructure pairs and draw a heatmap;

[0044] Atomic-level interpretability: High-weighted substructures are mapped back to the original molecular diagram, and the atomic coordinates involved are highlighted.

[0045] Furthermore, the comprehensive loss function used during model training. Defined as: in, To balance the weighted binary cross-entropy loss of the sample distribution; To constrain the semantic consistency loss of feature distribution consistency between two views; This is the L2 regularization loss.

[0046] Beneficial effects:

[0047] 1. This invention comprehensively extracts multi-source heterogeneous features of drugs from five dimensions: molecular structure, network behavior, structural similarity, textual semantics, and knowledge graph. These five types of features cover the drug's microscopic chemical structure, network behavior patterns, chemical spatial distribution, literature semantic description, and macroscopic biomedical associations, forming a complete information chain from microscopic to macroscopic. This breaks through the limitations of traditional methods that rely solely on single molecular structural features, making drug characterization more comprehensive and robust. In particular, for new drugs not present in the training set, effective feature representations can be generated with the help of multi-dimensional information, significantly alleviating the cold start problem of new drugs and improving the model's generalization ability in inductive prediction scenarios.

[0048] 2. This invention does not employ simple vector concatenation or fixed-weight summation to fuse the aforementioned five types of heterogeneous features. Instead, it designs a four-step fusion chain: "dimensional unified mapping → cascaded compression → Transformer cross-modal self-attention fusion → residual superposition injection." The Transformer is applied to the cross-modal adaptive fusion between "molecular structure embedding" and "multi-source side information comprehensive embedding," and the fused drug-level global semantics are injected into each atomic node of the molecular graph through residual superposition, achieving "global guidance of local" information enhancement. This invention uses the multi-head self-attention mechanism in the Transformer encoder to adaptively calculate the cross-modal semantic association weights between molecular structure embedding and side information comprehensive embedding, realizing deep semantic interaction and information reconstruction between features from different sources and modalities, overcoming the shortcomings of simple concatenation or weighted summation methods that cannot bridge the semantic gap.

[0049] 3. This invention innovatively introduces a dual-view parallel learning framework, employing a hypergraph structure to replace the traditional bipartite graph structure for drug-drug interaction modeling. The hyperedges of the hypergraph can simultaneously connect multiple nodes, naturally suited for representing "many-to-many" group associations between multiple substructures or biological entities. The DDI prediction task is decomposed into two complementary views: the Intra-view uses GAT and SAGPool to hierarchically encode atomic-level interactions within individual drug molecules, extracting key substructures; the Inter-view constructs drug-drug hypergraphs at both the substructure and attribute levels, capturing high-order associations between drug groups through hypergraph convolution. The two views operate in parallel, each encoding information at different levels, and are ultimately fused at the view level through an attention mechanism. This dual-view parallel architecture of "fine-grained intramolecular encoding + drug-drug group association modeling" can simultaneously capture high-order synergistic and antagonistic mechanisms at both the molecular structure and macroscopic biomedical levels, significantly improving the modeling capability for complex DDI mechanisms.

[0050] 4. This invention proposes a weighted fusion strategy for hypergraphs based on spectral Hamming similarity. In Interview, this invention does not directly convert known drug relationships into hyperedges. Instead, it constructs hypergraphs at two different granularities: substructure level and attribute level. Then, it adaptively fuses the Laplacian spectral features of the two hypergraphs using spectral Hamming similarity to generate a unified hypergraph structure. This "dual-scale construction + spectral adaptive fusion" scheme is a data-driven dynamic process. By comparing the Laplacian spectral features of each heterogeneous hypergraph with the spectral similarity of the benchmark DDI hypergraph, it adaptively determines the weight of each hypergraph in the fusion process, avoiding the subjectivity of manually setting the fusion ratio. This allows the fused dynamic hypergraph to automatically balance the contributions of substructure association information and attribute association information according to data characteristics.

[0051] 5. This invention provides interpretable outputs at three levels above the prediction results: view level (which view contributes more), substructure level (which functional groups are involved in the interaction), and atomic level (which atoms are key sites), which has important practical application value for drug safety assessment.

[0052] The systematic selection of five types of heterogeneous features, the design of a four-step fusion link, the proposal of a dual-view parallel architecture, the spectral domain adaptive fusion of a dual-scale hypergraph, and the construction of a three-level interpretability mechanism in this invention are all closely related in terms of technical logic: multi-source features provide rich input information for the dual views; the fusion injection mechanism transfers global semantics to atomic-level representations; the dual-view architecture encodes from two complementary perspectives, internal and external; and multi-level interpretability depends on the attention weights at different levels within the dual-view architecture. The organic combination of these elements results in prediction accuracy and interpretability significantly superior to existing methods. Attached Figure Description

[0053] Figure 1 This is a flowchart illustrating a dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to the present invention.

[0054] Figure 2 This is a schematic diagram of the multi-source feature extraction and fusion structure of the present invention;

[0055] Figure 3 This is a schematic diagram of the dual-view learning network structure of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figure 1 As shown, this invention discloses a dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement, comprising the following steps:

[0058] Step 1: Multi-source feature extraction and characterization. Obtain the target drug's SMILES string and multi-source biomedical information; extract molecular-level structural features using a graph neural network; extract behavioral features on the drug interaction network using the Node2Vec algorithm; calculate structural similarity features based on Morgan's molecular fingerprint; extract textual semantic features using BioBERT; and obtain macroscopic knowledge graph features from a pre-trained knowledge graph.

[0059] Step 2: Adaptive fusion and injection of multi-source features. The multi-source lateral features extracted in Step 1 are subjected to adaptive semantic fusion through a Transformer encoder to generate an initial drug-enhanced embedding that incorporates global information. This initial enhanced embedding is then injected into each atomic node of the drug molecule graph using a residual stacking method to form an optimized initial atomic feature matrix.

[0060] Step 3: Intra-view local feature encoding within the molecule. Using the optimized initial atomic feature matrix generated in Step 2 as input, an atomic-level message passing is performed through a graph attention network to capture local chemical bond relationships within the molecule. Then, sub-structural embeddings are extracted using SAGPool hierarchical pooling to obtain the intra-view embedding representation.

[0061] Step 4: Inter-view multi-scale hypergraph high-order interaction modeling. A sub-structure-level hypergraph is constructed based on the extracted sub-structure nodes, and a multi-attribute hypergraph is constructed simultaneously based on multi-dimensional attributes such as drug ATC encoding and target points. The spectral Hamming similarity algorithm is used to perform weighted fusion of heterogeneous hypergraphs to construct a dynamic hypergraph. The high-order interaction features between drugs are captured through a hypergraph convolutional network to obtain the inter-view embedding representation.

[0062] Step 5: View Fusion and Multi-Level Interpretable Prediction. Learnable view-level weight parameters are introduced to perform adaptive fusion of intra-view and inter-view embeddings, constructing a joint feature vector for drug pairs. The interaction probabilities are output through a decoder, generating multi-level interpretability analysis results covering the view, substructure, and atomic levels.

[0063] The following detailed explanation of the implementation process of the method of the present invention is based on an example of simvastatin and itraconazole. The relevant parameters and values ​​are for illustrative purposes only and do not constitute a limitation on the scope of protection of the present invention.

[0064] Step 1: Multi-source feature extraction and representation, such as Figure 2 The diagram illustrates the multi-source feature extraction and fusion structure. This step aims to comprehensively extract multi-source heterogeneous feature representations of drugs from multiple dimensions, including microscopic molecular structure and macroscopic biomedical knowledge (including molecular structure embedding, behavioral embedding, similarity embedding, semantic embedding, and knowledge graph embedding), providing rich input information for subsequent multimodal fusion and interaction modeling. Specifically, it includes the following sub-steps:

[0065] S11. Molecular graph structural feature extraction

[0066] The SMILES string of the drug was parsed and converted into a two-dimensional molecular graph structure using the cheminformatics tool RDKit. This molecular graph is defined as follows: ,in For a set of nodes, each node Corresponding to one atom in a drug molecule, This represents the total number of atoms in the drug molecule. As an edge set, in this embodiment, the molecular graph uses a representation that does not explicitly include hydrogen atoms; therefore, the number of nodes corresponds to the number of non-hydrogen atoms (heavy atoms) in the molecule. Taking simvastatin and itraconazole as examples, the molecular graphs obtained after parsing with RDKit are... and The number of atomic nodes , Number of chemical bond edges , Each edge Represents atoms With atoms The chemical bond connections between them.

[0067] For each atomic node in the molecular diagram Extract its predefined chemical property features to construct an initial atomic feature vector. ,in This represents the dimension of the atomic eigenvectors. The eigenvectors of all atoms form the initial atomic eigenma matrix. In this embodiment, the atomic feature vector dimension This encodes chemical properties such as atom type, number of bonded hydrogen atoms, and chiral labeling. The initial atomic feature matrix of simvastatin is... The initial atomic characteristic matrix of itraconazole is .

[0068] Similarly, for each chemical bond edge Extract edge attributes such as key type, conjugate, and whether it is in a cycle, and encode them into edge feature vectors. ,in Let be the dimension of the edge feature vector. In the subsequent message passing process of the graph neural network, the edge features... By concatenating features with corresponding nodes to participate in message computation, the GNN can distinguish different types of chemical bond connections. Subsequently, a graph neural network is used to encode the molecular graph. After multi-layer GNN encoding, a graph-level readout operation is performed on all node features to obtain the molecular-level structure embedding vector of the drug.

[0069]

[0070] in, This indicates the drug's 3D molecular structure embedding vector, For a unified embedding dimension, The molecular-level structure embedding vector of simvastatin is itraconazole is Dimension of edge feature vectors According to the molecular diagram The adjacency matrix is ​​constructed from the chemical bond connections in the matrix. ,in Represents atoms With atoms There are chemical bonds present; otherwise, the value is 0, which is used directly by SAGPool in step 3.

[0071] S12. Behavioral Feature Extraction

[0072] In existing drug interaction networks The network behavior characteristics of the extracted drugs, among which For the set of all drug nodes, Given the set of known drug-drug interaction edges, the Node2Vec graph walk algorithm is used. Node2Vec utilizes two hyperparameters... (Return parameters) and (Ingress / Exgress Parameters) Control Walking Strategy: The smaller the value, the more likely the roaming will return to the node visited in the previous step, capturing the local community structure; The smaller the value, the more the walk tends to move away from the source node, capturing global topological roles. A biased random walk is performed on each drug node to generate a node sequence. The transition probability of the random walk is defined as:

[0073]

[0074] in Indicates swimming to the first The nodes visited during the step. For parameters and Adjusted nonnormalized transition probabilities, This is the normalization constant.

[0075] The generated node sequences are input into the Skip-gram model for training, learning the topological distribution pattern of drug nodes in the DDI network, and finally outputting the behavior embedding vector of each drug. ,in This is the embedding dimension for behavioral features. This vector encodes the topological location information and connectivity patterns of the drug within a known interaction network. In this embodiment, the return parameters are set preferentially. , The walk length is 80 steps, each node generates 10 walk sequences, the Skip-gram window size is 5, and the embedding dimension is... The above parameters can be adjusted according to the actual network scale and task requirements, and are not limited to the values ​​listed in this embodiment. After training with Node2Vec, the behavioral embedding vector of simvastatin is obtained. behavioral embedding vectors of itraconazole .

[0076] S13. Structural Similarity Feature Extraction

[0077] Structural similarity features between drugs were calculated based on Morgan's molecular fingerprinting. First, the RDKit tool was used to calculate the structural similarity features between drugs by radius. Generate a fixed length of Morgan fingerprint bit vector .in The fingerprint bit length is specified. Subsequently, the target drug is calculated. The Tanimoto similarity coefficient between the drug and all other drugs in the drug library. For drugs and medicine Its Tanimoto coefficient is defined as:

[0078]

[0079] in This represents the element count of the set. From this, we can derive the drug. Similarity vector The similarity vector is nonlinearly mapped using a multilayer perceptron to extract a similarity embedding vector encoding the spatial distribution information of medicinal chemistry.

[0080]

[0081] in For similarity embedding vectors, For the embedding dimension of similarity features, It is a multilayer perceptron consisting of several fully connected layers, ReLU activation functions, and Dropout layers.

[0082] The Tanimoto similarity coefficient between simvastatin and itraconazole is This reflects the significant differences between the two in their chemical structure space. (Simvastatin similarity vector) After mapping by a multilayer perceptron, the following is obtained: Similarly, the similarity embedding vector for itraconazole is .

[0083] S14. Text Semantic Feature Extraction

[0084] Input the textual description of the drug into the pre-trained BioBERT model to extract... The hidden state vector corresponding to the labeled bit is then mapped to the target dimension through a linear projection layer to obtain the semantic embedding vector:

[0085]

[0086] in For semantic embedding vectors, For the embedding dimension of semantic features, , The weight matrix and bias vector are for linear projection. For drugs The text description.

[0087] Extracted by BioBERT Hide state and pass After projection, the semantic embedding vector of simvastatin is obtained. semantic embedding vectors of itraconazole .

[0088] S15. Macro-level knowledge graph feature extraction

[0089] Embedded representations of drug entities are obtained from a pre-trained biomedical knowledge graph. The PrimeKG knowledge graph is pre-trained using the ComplEx knowledge graph embedding method. The ComplEx method utilizes a bilinear scoring function in complex space. For triples Head entity, relation, tail entity Modeling is performed, in which To perform the real part operation, This is a complex conjugate operation. In this embodiment, the knowledge graph embedding dimension is set. .

[0090] Load the entity embedding vector corresponding to the target drug directly from the pre-trained knowledge graph embedding: ,in This refers to the dimensions of knowledge graph embedding. Taking simvastatin and itraconazole as examples, their knowledge graph embeddings are loaded respectively. and The study found that both drugs share the key entity node CYP3A4, providing a priori knowledge foundation for capturing higher-order associations in subsequent hypergraph modeling. For drugs not included in PrimeKG, their knowledge graph embeddings are initialized as all-zero vectors. Simvastatin and itraconazole are both included in PrimeKG and pre-trained embedding vectors can be loaded directly.

[0091] Step 2: Adaptive Fusion and Injection of Multi-Source Features. This step aims to map the four types of lateral information embeddings extracted in Step 1 to the same embedding dimension space through their respective independent linear projection layers. After cascading, they are compressed into a comprehensive lateral information embedding using a multilayer perceptron. This comprehensive lateral information embedding and the molecular structure embedding form the input sequence, which is fed into a Transformer encoder for adaptive semantic fusion. A multi-head self-attention mechanism is used to dynamically calculate cross-modal semantic association weights to generate an enhanced initial embedding for the drug. Subsequently, the enhanced initial embedding is injected into the initial feature vector of each atom node in the molecular graph in a broadcast manner through a residual superposition mechanism. Specifically, it includes the following sub-steps:

[0092] S21. Unified Dimensional Mapping and Cascading of Side Information Features

[0093] First, the four types of side information embedding vectors extracted in step 1 are uniformly mapped to the same path through their respective independent linear projection layers. In 3D space:

[0094] ,

[0095] ,

[0096] in, , , , For their respective projection weight matrices, , , , These are their respective bias vectors.

[0097] In this embodiment, the original dimension of the four types of side information embedding is 128, and the dimension of the projection target is also 128. Each projection weight matrix .

[0098] The four types of side information, after being unified in dimension, are concatenated and then subjected to nonlinear mapping and dimensionality compression using a multilayer perceptron to obtain the comprehensive side information embedding: ,in This represents a vector concatenation operation. To form a comprehensive embedding vector that integrates four types of side information, It is a multilayer perceptron consisting of a linear transformation layer, a ReLU activation function, and a Dropout layer, with an input dimension of . The output dimension is .

[0099] Taking simvastatin as an example, after concatenating four types of embeddings, a 512-dimensional vector is obtained. After mapping through a multilayer perceptron, the side information integrated embedding is obtained. Similarly, itraconazole was obtained. .

[0100] S22. Transformer Adaptive Semantic Fusion

[0101] Embed the molecular structure obtained in step S11 Integrated with the side information obtained in step S21 Construct an input sequence of length 2 The input sequence is then fed into the Transformer encoder for adaptive semantic fusion. The Transformer encoder achieves adaptive semantic fusion through a multi-head self-attention mechanism. Specifically, the input sequence is first processed... The query matrix is ​​generated through three sets of independent linear transformations. Key matrix Sum matrix : , , ,in The projection matrix is ​​learnable. Then, the scaled dot product attention is calculated:

[0102]

[0103] in, The feature space is defined as the feature dimension for each attention head. To enhance the model's expressive power, the feature space is divided into... Attention is computed in parallel across subspaces; in some implementations, the number of attention heads... The value can be set to 4, 8, or 16; in this embodiment, 8 is preferred. The outputs of each head are cascaded and linearly projected to obtain the final multi-head attention output. The Transformer encoder also includes standard components such as residual connections, layer normalization, and feedforward networks. The FFN consists of two layers of linear transformations and a GELU activation function, with the middle dimension being... .

[0104] Perform mean pooling on the sequence output by the Transformer encoder to generate an enhanced initial embedding for the drug: ,in To enhance the initial embedding of drugs that integrate molecular structure information and multi-source side information, This means taking the element-wise mean of the output vector at all positions in the sequence.

[0105] Taking simvastatin as an example, and Construct the input sequence The input is fed into the Transformer encoder and, after mean pooling, yields the enhanced initial embedding. Similarly, itraconazole was obtained. .

[0106] S23. Residual superposition injection mechanism

[0107] To inject global-level fusion semantics into each atom node in the molecular graph while preserving the local chemical properties of the atom nodes themselves, this invention employs a residual superposition mechanism. First, a learnable linear projection matrix is ​​used... Will enhance the initial embedding Mapped to the same dimensional space as the atomic features : ,in, This is the projected global embedding vector.

[0108] Then, this global embedding is broadcast element-wise to each atom node in the molecular graph. On the initial eigenvectors: ,in Atoms after injecting global information Optimize the initial feature vector, This refers to the original atomic feature vectors extracted in step S11. The optimized initial atomic feature matrix is ​​composed of the optimized initial feature vectors of all atomic nodes. Taking simvastatin as an example, Through projection matrix Mapping The optimized initial atomic feature matrix is ​​obtained by superimposing the features onto the initial features of all 30 heavy atom nodes in a broadcast manner. Similarly, the optimized initial atomic characteristic matrix of itraconazole is: .

[0109] Step 3: Intra-view local feature encoding within the molecule, such as... Figure 3 The Intra-view portion of the dual-view learning network structure diagram shows a step that focuses on the internal structure of a single drug molecule. It aims to extract key pharmacophores (substructures) that trigger drug interactions from optimized atomic features, generating an Intra-view embedding representation. This step achieves feature abstraction from microscopic atoms to mesoscopic substructures through a two-stage encoding process: "atomic-level message passing - substructure-level hierarchical pooling." Specifically, it includes the following sub-steps:

[0110] S31. Atomic Message Passing Based on Graph Attention Networks

[0111] The optimized initial atomic feature matrix generated in step 2 As input to the graph attention network. For the central atom node in the drug molecule graph. and its neighboring atomic nodes (in Represents a node Calculate the nonnormalized attention coefficients between the first-order neighbor nodes in the molecular graph. : ,in For nodes In the Hidden feature vectors of layer (when) hour, , ), For the first The learnable linear transformation weight matrix of the layer, This is a learnable attention parameter vector. This represents a vector concatenation operation. It is a nonlinear activation function with a negative slope. In this embodiment, the first layer... , 2nd and 3rd floors Attention parameter vector The LeakyReLU algorithm has a negative slope of 0.2, and each layer uses four attention heads. The outputs of each attention head are concatenated and then fused using linear projection to maintain the output dimension of each layer. The feature dimension of each attention head is In this embodiment Each head has a dimension of The outputs from each head are spliced ​​together and then linearly projected back to their original form. .

[0112] Subsequently, the nodes are processed using the Softmax function. Calculate normalized attention weights for all neighbors :

[0113] in Indicates the first Neighbor nodes in the layer For the central node Importance weights. The transformation features of neighboring nodes are weighted and summed based on normalized attention weights to update the node. Feature representation: ,in It is a non-linear activation function. For nodes In the Updated features of the layer.

[0114] By stacking Layered GATs, where each atomic node recursively aggregates chemical information within its multi-order neighborhood, captures the characteristics of the local chemical microenvironment within the molecule determined by chemical bond connections. After layer GAT encoding, the updated atomic feature matrix is ​​obtained. Output dimensions of each layer exist Time average equals uniform embedding dimension ,Right now GAT layer number The number of layers can be set from 2 to 5 depending on the molecular complexity; in this embodiment, 3 layers are preferred. After 3-layer GAT encoding, simvastatin obtains an updated atomic feature matrix. Itraconazole obtained .

[0115] S32. Substructure-level hierarchical pooling based on SAGPool

[0116] To extract higher-level substructure representations from atomic-level features, this invention employs a self-attention-based graph pooling algorithm, SAGPool. SAGPool learns the importance score of each node, adaptively selecting and retaining key atomic nodes, gradually shrinking the atomic-level graph into a substructure-level graph, thereby achieving hierarchical compression of structural information.

[0117] Specifically, SAGPool first utilizes an auxiliary graph neural network (denoted as...). Based on the current atomic feature matrix Adjacency matrix of molecular graph Calculate the self-attention score vector for each atomic node: ,in This is the importance score vector for all atomic nodes. Represents atoms Importance score. In this embodiment, A single-layer GCN is used, with a weight matrix. Taking simvastatin as an example, the lactone ring region atoms (such as the carbonyl carbon) received higher scores. ester bond oxygen ), while the methyl terminal carbon atom, which is far from the active center, has a lower score ( ).

[0118] Then, according to the score Sort by size and select the top scorers. The atomic nodes are retained, among which The preset pooling ratio, Indicates rounding up, pooling ratio The pooling ratio can be set between 0.3 and 0.7 depending on the molecular complexity. In this embodiment, the pooling ratio is preferably set. The node feature matrix and adjacency matrix after pooling are as follows: , ,in This represents element-wise multiplication (i.e., using the attention score as a gating signal to weight the node features). The hyperbolic tangent activation function is used. The node feature matrix retained after pooling. This is the adjacency matrix after pooling. The Travelogue That is, the first The feature vectors of the substructure nodes after pooling.

[0119] The retained high-scoring atomic node groups represent key substructures (pharmacodynamic groups) in drug molecules, such as carboxyl, amino, and aromatic rings, which are functional fragments that have a decisive influence on drug activity and interactions. Taking simvastatin as an example, pooling retains 15 high-scoring nodes, corresponding to key substructures including: lactone ring substructure (…). (5 atoms), decahydronaphthalene skeleton substructure ( (4 atoms), 2,2-dimethylbutyrate side chain substructure ( (4 atoms) and hydroxyl substituent substructure ( (2 atoms), pooled feature matrix Adjacency matrix after pooling After pooling, itraconazole retains 24 nodes, corresponding to key substructures including: 1,2,4-triazole ring substructure ( (5 atoms), dichlorophenyl substructure ( 8 atoms), piperazine ring substructure ( (5 atoms), dioxolane substructure ( (4 atoms) and sec-butyl side chain substructure ( (2 atoms), pooled feature matrix Adjacency matrix after pooling .

[0120] Finally, a global readout operation is performed on the pooled substructure-level graph to obtain the Intra-view embedding representation of the drug: ,in In this embodiment, the vector is embedded in the Intra-view. The operation uses a combination of mean pooling and max pooling: ,in The concatenation of mean pooling and max pooling can extract richer semantic information. This embedding encodes the local chemical features of key substructures within the drug molecule. After global readout, the Intra-view embedding of simvastatin is... The Intra-view of itraconazole is embedded as .

[0121] Step 4: Inter-view multi-scale hypergraph high-order interaction modeling, such as... Figure 3 The Inter-view portion of the dual-view learning network structure diagram illustrates how this step overcomes the limitations of binary relation modeling by introducing a hypergraph structure. It constructs heterogeneous hypergraphs at both the substructure and attribute levels, and adaptively fuses them using the spectral Hamming similarity algorithm. Finally, a hypergraph convolutional network captures high-order interaction features between drugs. Specifically, it includes the following sub-steps:

[0122] S41. Substructure Hypergraph Construction

[0123] Using the substructure nodes retained after SAGPool pooling in step 3 as the basic units, a substructure-level hypergraph is constructed. Among them, vertex set Includes all substructure nodes retained after pooling of drug A and drug B; superedge set Build it in the following way:

[0124] Calculate any two substructure nodes and Cosine similarity between them:

[0125] ,in and Substructure nodes and The feature vector after pooling (i.e. The row and number OK), This represents the L2 norm of a vector.

[0126] Based on cosine similarity, a K-nearest neighbor strategy is adopted for each substructure node. Take the one with the highest similarity. 1 node, including these nodes (including A connection between itself and a superedge is a hyperedge. ,in Represents nodes Top cosine similarity ranking A set of nodes. The value can be set to 3 to 10, and is preferred in this embodiment. This yields the correlation matrix of the substructure hypergraph. ,in Represents vertices Overtake Includes, otherwise 0. Taking the simvastatin-itraconazole drug pair as an example, the substructure hypergraph... vertex set It includes 15 pooled substructure nodes for simvastatin and 24 pooled substructure nodes for itraconazole, for a total of 39 vertices. Cosine similarity between the simvastatin lactone site and the itraconazole triazole site Both are ring structures containing heteroatoms. Under the K-nearest neighbor hyperedge construction strategy, each substructure node generates a hyperedge, therefore the number of hyperedges is... Final association matrix ,in Represents vertices Overtake Includes, otherwise 0.

[0127] S42. Construction of Multi-Attribute Hypergraphs

[0128] Constructing a multi-attribute hypergraph based on the multidimensional external attribute information of drugs. Among them, vertex set Includes the drug pairs to be predicted and their associated drugs; hyperedgeset Constructed based on the following multiple drug properties:

[0129] ATC-encoded hyperedges: Drugs with the same ATC level 1 to level 3 codes are grouped into the same hyperedge, so that drugs with similar therapeutic effects form high-order associations in the hypergraph.

[0130] Target protein superedge: Based on drug-target data, drugs that act on the same target protein are associated with the same superedge.

[0131] Biological pathway hyperedge: Based on drug-pathway data, drugs involved in the same biological pathway are associated with the same hyperedge.

[0132] For each of the above attributes, calculate the Jaccard similarity between drugs, and classify drugs with similarity exceeding a preset threshold. Drug vertices are grouped into hyperedges with the same attribute. Jaccard similarity is defined as:

[0133]

[0134] in and Drugs and The associated set of attributes, This represents the cardinality of the set. From this, we obtain the association matrix of the multi-attribute hypergraph. Similarity thresholds for each attribute type These thresholds can be set according to the characteristics of the data. In this embodiment, the ATC encoding superedge threshold is set to 0.3, the target protein superedge threshold is set to 0.1, and the biological pathway superedge threshold is set to 0.1. Taking simvastatin and itraconazole as examples, their ATC encodings are different and they are not connected by the same ATC superedge; however, they share the target CYP3A4, and their Jaccard similarity is... Exceeding the preset threshold of the target protein superedge They are connected by the same target hyperedge; they share the "drug metabolism-cytochrome P450" pathway and are also connected by the same pathway hyperedge.

[0135] S43. Hypergraph Weighted Fusion Based on Spectral Hamming Similarity

[0136] Due to the substructure hypergraph and multi-attribute hypergraph The higher-order relationships between drugs are characterized at both the molecular structure and macroscopic property levels, and these two approaches are complementary. To achieve adaptive fusion of the two types of hypergraphs, this invention proposes a weighted fusion strategy based on spectral Hamming similarity.

[0137] First, calculate the normalized Laplacian matrices of the two hypergraphs respectively:

[0138]

[0139] in For Hypergraph The vertex degree matrix, It is the hypermarginality matrix. This is a diagonal matrix with hyperedge weights. It is the identity matrix. For Hypergraph The correlation matrix.

[0140] Laplace matrix Perform eigenvalue decomposition and extract its antecedents. The smallest non-zero eigenvalues ​​form the eigenvalue spectrum vector. The spectral vector is mapped to a binary hash code through median thresholding. :

[0141]

[0142] in Represents spectral vector of the median.

[0143] Simultaneously, a baseline DDI hypergraph is constructed, and its spectral hash code is calculated. Calculate the spectral Hamming similarity between the spectral hash codes of each heterogeneous hypergraph and the spectral hash codes of the baseline hypergraph:

[0144]

[0145] in Hamming distance (the number of bits that differ at corresponding positions in two binary strings of equal length). Where is the hash code length. Spectral Hamming similarity is used to measure how closely the topology of a candidate hypergraph is similar to that of a baseline DDI hypergraph. The larger the value, the more likely it is to be a hypergraph. The more similar the topology of a hypergraph is to the baseline DDI hypergraph, the higher the weight it should be given in the fusion process.

[0146] The fusion weights of each hypergraph are obtained through normalization calculation:

[0147]

[0148] Based on the fusion weights, the association matrices of the two hypergraphs are weighted and fused to construct the final dynamic hypergraph association matrix:

[0149]

[0150] in This is the fused dynamic hypergraph association matrix. This dynamic hypergraph integrates substructure-level chemical structural association information with attribute-level macroscopic biomedical association information. (Hash code length) The value can be set from 20 to 100; in this embodiment, the value is preferably set to 20. The spectral Hamming similarity of the substructure hypergraph was calculated. Multi-attribute hypergraph Normalized fusion weights , The higher weighting of the multi-attribute hypergraph is consistent with the mechanism by which the drug's DDI is mediated by the common metabolic enzyme CYP3A4.

[0151] S44. Hypergraph Convolutional Network Coding

[0152] Hypergraph convolutional network operations are performed on the constructed dynamic hypergraph to capture high-order interaction features between drugs. The message passing of hypergraph convolution follows a two-stage aggregation pattern of "node → hyperedge → node": first, the features of all nodes contained in the hyperedge are aggregated to the hyperedge, and then the aggregated features of the hyperedge are distributed back to all the nodes it contains, thereby realizing the synchronous exchange of information between multiple nodes within the hyperedge.

[0153] The node feature update formula for hypergraph convolution is:

[0154]

[0155] in For the first The node feature matrix of the hypergraph convolution (where the first layer is the node feature matrix of the hypergraph convolution) Initial node feature matrix of the layer The substructure node feature matrix retained after SAGPool pooling in step 3 Composition, that is ), For dynamic hypergraph-based association matrix Recalculated vertex degree diagonal matrix (its diagonal elements) ), The corresponding hyperedge degree diagonal matrix (its diagonal elements) ), This is the super-edge weighted diagonal matrix determined by the weighted fusion of the longitudinal spectrum. For the first The trainable parameter matrix of the layer, It is a non-linear activation function.

[0156] After multiple hypergraph convolutions, a global pooling operation is performed on the output node feature matrix to obtain the Inter-view embedding representation:

[0157]

[0158] in Embed vectors for Inter-view. This represents the total number of hypergraph convolutional layers. This embedding encodes high-order synergistic and antagonistic interactions between drugs based on multiple substructures and various macroscopic properties. (Number of hypergraph convolutional layers) It can be set to 1 to 4 layers, and is preferably set in this embodiment. Initial node feature matrix (Simvastatin has 15 nodes, itraconazole has 24 nodes). After hypergraph convolution, the simvastatin [data] is obtained. and itraconazole .

[0159] Step 5: View Fusion and Multi-Level Interpretable Prediction

[0160] This step integrates the embedded representations of two views: Intra-view (intramolecular features) and Inter-view (higher-order interaction features between drugs). Through adaptive fusion, a joint feature vector of the drug pair is constructed. The decoder outputs the interaction prediction probability, generating a three-layer interpretable analysis result covering the view level, substructure level, and atomic level. Specifically, it includes the following sub-steps:

[0161] S51. View-level adaptive weight fusion

[0162] Introducing two view-level learnable scalar parameters and (These are all model parameters that need to be optimized through backpropagation), and the final view fusion weights are calculated using Softmax normalization:

[0163] ,

[0164] in , and These represent the contribution weights of Intra-view and Inter-view in the final prediction, respectively.

[0165] For a single drug, its final embedding after fusion is represented as:

[0166]

[0167] in, This is the final embedding vector for the drug. This represents the dimension of the drug embedding after fusion. In the prediction of the simvastatin-itraconazole drug pair, after training convergence... , This indicates that the DDI is primarily driven by external higher-order interaction factors (the co-metabolic enzyme CYP3A4). The final embedding of simvastatin... and the final insertion of itraconazole .

[0168] S52. Construction and Prediction of Drug Pair Combination Characteristics

[0169] For the drug pair to be predicted (drug A and drug B), calculate their final embeddings respectively. and The two are combined to construct a joint feature vector for the drug pair:

[0170]

[0171] in , Indicates vector concatenation. Represents element-wise product. This indicates taking the absolute value element by element. Among them, cascading operations... contribute Dimension, element-wise product contribute Dimension, element-wise absolute difference contribute Dimensions, total In this embodiment, is The cascade operation preserves the complete feature information of each of the two drugs, the element-wise product captures the synergy of the two drugs in the feature dimensions, and the element-wise absolute difference captures the differences between the two drugs.

[0172] The joint feature vector is input into the multilayer perceptron (MLP) decoder. After nonlinear transformation through several hidden layers, the probability of a specific type of drug interaction between the drug pair is output through the sigmoid activation function.

[0173]

[0174] in For the predicted DDI probability, It is the Sigmoid activation function. For a decoder, a multilayer perceptron. When When the threshold value exceeds a preset threshold (e.g., 0.5), the drug pair is considered to have an interaction. Taking simvastatin-itraconazole as an example, the joint feature vector... ( (Dimension) Input It contains several fully connected hidden layers, and the dimensions of each layer can be adjusted according to task requirements. In this embodiment, the dimensions of the hidden layers are set sequentially to 512→256→128→1, and the final output layer dimension is 1. Output prediction probability. The value was much higher than the threshold of 0.5, and the model determined that there was a drug interaction between the two, which is consistent with known clinical facts.

[0175] S53. Model Training Loss Function

[0176] The comprehensive loss function used during model training is defined as follows:

[0177]

[0178] in The weighted binary cross-entropy loss, used to balance the sample distribution, measures the difference between the predicted probability and the true label:

[0179]

[0180] in For the first The true label of each sample To predict probabilities for the model, The number of training samples. and These are the weighting coefficients for positive and negative samples, used to alleviate the problem of class imbalance. For semantic consistency loss, constrain Intra-view features With Inter-view features Consistent distribution in semantic space prevents excessive shifts in the feature representations of two views:

[0181]

[0182] in L2 regularization loss is used to prevent model overfitting, where, and These are the embedding vectors of the same drug in Intra-view and Inter-view, respectively. This loss is calculated by averaging all drug samples in the training batch and is used to constrain the feature representations of the two views to maintain consistency in the semantic space, preventing the two views from shifting too much. and This is a hyperparameter used to balance the contributions of each loss term.

[0183] The embodiment is set , The weights of positive and negative samples are calculated based on the reciprocal of the ratio of positive to negative samples in the training set. , ,in and These represent the number of positive samples and the number of negative samples, respectively.

[0184] S54. Multi-level interpretability analysis

[0185] This invention outputs three levels of interpretability analysis results, covering the view level, substructure level, and atomic level:

[0186] View-level interpretability: by extracting the normalized weights after model training convergence. and The magnitude of the value can intuitively show whether the predicted interaction is mainly dominated by internal molecular structural factors (Intra-view) or by external higher-order interaction factors (Inter-view). For example, when This suggests that the interaction is more driven by macroscopic property similarity or higher-order synergy among multiple groups. Taking the simvastatin-itraconazole drug pair as an example... This suggests that the DDI is primarily driven by external higher-order interaction factors.

[0187] Substructure-level interpretability: Based on the feature response strength and fusion weights of nodes after propagation through hyperedges during the dynamic hypergraph convolution process in step 4, construct the substructure pair interaction weight matrix. ,in and These represent the number of substructure nodes retained after SAGPool treatment for drugs A and B, respectively. Indicates the first of drug A The first substructure and drug B The interaction strength between substructures. (Filtering) The substructure pairs with the highest weights are identified and visualized using heatmaps to pinpoint the core substructure combinations that significantly contribute to DDI. Taking the simvastatin-itraconazole drug pair as an example, the interaction weight matrix... ( , In the ), the substructure pair with the highest weight is the simvastatin lactone ring ( ) and itraconazole triazole ring ( ), This aligns with the drug's chemical mechanism—simvastatin lactone ring is a substrate for CYP3A4 enzyme-catalyzed hydrolysis, and itraconazole triazole ring is a key pharmacophore that binds to the active site of CYP3A4 enzyme.

[0188] Atomic-level interpretability: Tracing back the Top-K sampling index of SAGPool in step 3, the substructure pairs with high interaction weights are mapped back to the specific sets of atoms in the original molecular diagram. Using the RDKit molecular visualization tool, the involved atomic regions are highlighted on the original two-dimensional molecular structure diagram, identifying reactive chemical groups with potential pharmacological risks. Taking the simvastatin-itraconazole drug pair as an example, the carbonyl carbon in the simvastatin lactone ring (importance score) ) and ester bond oxygen ( ), the N-3 nitrogen atom in the itraconazole triazole ring ( ) and N-4 nitrogen atoms ( () was identified as a key atom, identifying the key chemical group at which the two drugs competitively bind to the CYP3A4 enzyme active site.

[0189] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement, characterized in that, include: Step 1: Extract drug molecular graph structural features, drug behavior features, drug structural similarity features, and drug semantic features to obtain macroscopic knowledge graph features from the pre-trained knowledge graph; Step 2: Adaptively fuse multi-source features with macro-knowledge graph features to form an enhanced initial embedding representation of the drug. After linear projection, the enhanced initial embedding is injected into the initial feature vector of each atom node in the molecular graph in a broadcast manner to form an optimized initial atom feature matrix. Step 3: Optimize the initial atomic feature input graph attention network GAT, perform atomic-level local message passing and feature aggregation inside the drug molecule graph, capture local chemical bond associations within the molecule, and then extract substructure-level embeddings through hierarchical pooling to obtain the Intra-view embedding representation; Step 4: Construct a substructure hypergraph based on substructure nodes, construct a multi-attribute hypergraph based on drug multi-dimensional attribute information, perform weight allocation and weighted fusion on multiple hypergraphs to determine the final dynamic hypergraph, and capture high-order interaction features between drugs through hypergraph convolutional networks to obtain the Inter-view embedding representation; Step 5: Introduce learnable weight parameters to adaptively fuse the Intra-view embedding representation and the Inter-view embedding representation to construct a joint feature vector representation of the drug pair, and output the interaction probability through the predictor.

2. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, Multi-source features specifically include: The RDKit tool was used to parse the drug's SMILES string into a molecular graph. Using graph neural networks to aggregate feature vectors of atomic nodes and chemical bond edge characteristics Output 3D molecular structure embedding ; In drug interaction networks Perform a biased random walk. For the set of all drug nodes, Given a set of known drug-drug interaction edges, generate a behavioral embedding that encodes drug topological location information. ; Based on Morgan molecular fingerprinting, the Tanimoto similarity matrix between the target drug and all other drugs in the drug library is calculated. The similarity vector of the target drug is then nonlinearly mapped using a multilayer perceptron to extract similarity embeddings that encode the spatial distribution information of the drug's chemistry. ; Input the drug text description into the pre-trained BioBERT model to extract... The semantic embedding is obtained by linearly projecting the position vector. ; Loading drug entity embeddings pre-trained using the ComplEx method from the PrimeKG knowledge graph For drugs not included in the catalog, their Initialize as a vector of all zeros.

3. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, The specific method of adaptive fusion of multi-source features is as follows: First, embed the behavior Similarity embedding Semantic embedding Knowledge graph embedding Each is mapped to a unified path through an independent linear projection layer. The 3D space is cascaded, and the side information is comprehensively embedded through multilayer perceptron (MLP) mapping. ; Secondly, embedding molecular structures Integrated embedding with side information As a sequence input Transformer encoder, feature reconstruction is performed through a multi-head self-attention mechanism; Perform mean pooling on the Transformer output sequence to generate the initial embedding for drug enhancement. Finally, the residual superposition mechanism is used to... Through linear projection matrix Mapping to obtain the projection vector The initial features are superimposed element by element to each atomic node in a broadcast manner. The above yields optimized initial atoms. feature: .

4. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, The specific method of Intra-view encoding in step 3 is as follows: The optimized atomic feature matrix is ​​input into a multi-layer graph attention network. The importance weights of neighboring atoms are dynamically calculated through a self-attention mechanism. After Softmax normalization, the neighbor features are aggregated and the atomic representation is updated. By stacking Layered GAT recursively aggregates chemical information from its multi-order neighborhood at each atomic node, capturing the local chemical microenvironment features determined by chemical bonds within the molecule, and obtaining an updated atomic feature matrix. ; Based on the current atomic feature matrix Adjacency matrix of molecular graph The self-attention graph pooling algorithm SAGPool is used to learn the self-attention score vector of each atomic node. After sorting, the top-ranked atomic nodes are retained. The group of high-scoring atomic nodes retained represents the key substructures in the drug molecule, and the atomic-level graph is gradually shrunk into a substructure-level graph. Update the pooled drug atom feature matrix and adjacency matrix, and perform a global readout operation on the pooled substructure-level graph to obtain the drug's Intra-view embedding representation. .

5. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, The final dynamic hypergraph construction process is as follows: Substructure hypergraph construction: Using the substructure nodes retained after pooling of drugs A and B as vertices, node associations are calculated based on cosine similarity, and the nodes with the highest similarity are linked. Each node is associated with a substructure hyperedge; Multi-attribute hypergraph construction: Calculating Jaccard similarity between drugs based on their ATC codes, target proteins, and biological pathway information. ,Will Drug vertices exceeding a preset threshold are classified as superedges of the same attribute; Weighted fusion of spectral Hamming similarity: Calculating the normalized Laplacian matrix of two hypergraphs Extract the feature value spectrum and map it into a binary hash code; Construct a baseline DDI hypergraph and calculate the spectral hash code; The spectral Hamming similarity between each hypergraph hash code and the baseline DDI hypergraph hash code was calculated. Determine the fusion weights of each hypergraph. ; The association matrices of the two hypergraphs are weighted and fused based on the fusion weights to construct the final dynamic hypergraph association matrix.

6. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, The message passing mechanism for obtaining the Inter-view embedding representation by capturing high-order interaction features between drugs using hypergraph convolutional networks is as follows: The update formula for node features in a hypergraph convolutional network is: ， in, For dynamic hypergraph incidence matrix, The hyperedge weight diagonal matrix is ​​determined by the weighted fusion of the longitudinal spectrum. For the first The node feature matrix of a hypergraph convolution. and Based on The calculated node degree matrix and hyperedge degree matrix, For the first The trainable parameter matrix of the layer, It is a non-linear activation function; After multiple hypergraph convolutions, a global pooling operation is performed on the output node feature matrix to obtain the Inter-view embedding representation. .

7. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, The specific method of adaptive fusion in step 5 is as follows: Introducing view-level learnable scalars , The final fusion weights are calculated through normalization: , The final drug embedding vector is obtained by fusion. , This indicates that the Inter-view is embedded. Embedded representation for Intra-view.

8. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, Construct a joint feature vector representation of the drug pair, and output the interaction probability through a predictor, specifically as follows: drug pair The fusion vectors are combined to construct joint features. ,in, and This represents the final embedding vector of the drug. Indicates vector concatenation. Represents element-wise product. Represents the absolute value of each element. Input to a multilayer perceptron decoder, output interaction probability. .

9. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, After outputting the interaction probabilities, a multi-level interpretability analysis is performed, including: View-level interpretability: by outputting normalized weights of Intra-view embedded representations and Inter-view embedded representations. and The numerical value quantifies the contribution of internal molecular features and external higher-order interaction features to the prediction results. Substructure-level interpretability: Extract the substructure pair interaction weight matrix from the decoder, filter high-weight substructure pairs and draw a heatmap; Atomic-level interpretability: High-weighted substructures are mapped back to the original molecular diagram, and the atomic coordinates involved are highlighted.

10. The dual-view drug interaction prediction method based on multi-source feature fusion and hypergraph enhancement according to claim 1, characterized in that, The comprehensive loss function used during model training Defined as: in, To balance the weighted binary cross-entropy loss of the sample distribution; To constrain the semantic consistency loss of feature distribution consistency between two views; This is the L2 regularization loss.

Images