Drug repositioning method and system based on heterogeneous knowledge and structural fusion

Abstract

The application discloses a drug repositioning method and system based on heterogeneous knowledge and structure fusion, and belongs to the technical field of medical care informatics. The method comprises the following steps: constructing a biomedical field knowledge heterogeneous graph; based on the biomedical field knowledge heterogeneous graph, generating disease knowledge embedding and drug knowledge embedding corresponding to a target drug-disease pair; by constructing a drug-drug similarity network, a disease-disease similarity network and a drug-disease association network, generating disease structure embedding and drug structure embedding corresponding to the target drug-disease pair; and based on the disease knowledge embedding, the drug knowledge embedding, the disease structure embedding and the drug structure embedding, obtaining a drug repositioning result. The application accurately captures complex biological network characteristics and finely models complex entity information through an innovative drug repositioning model, so that accurate drug repositioning is realized.

Description

TECHNICAL FIELD

[0001] The system belongs to the technical field of medical care informatics, and particularly relates to a drug repositioning method and system based on heterogeneous knowledge and structure fusion. BACKGROUND

[0002] In the field of drug research and development, there has been a technical problem of long research and development cycle, high research and development cost and low research and development success rate for a long time. In view of the problem, the drug repositioning technology has become a key technical means for breaking through the above-mentioned new drug research and development bottleneck because it can significantly shorten the drug development cycle and reduce the drug research and development cost. Early drug repositioning methods project drugs and diseases into a shared latent space using matrix decomposition, use latent vectors for representation, and integrate global drug-drug and disease-disease similarity matrices to improve prediction accuracy, such as the DRRS and SCMFDD methods.

[0003] In recent years, methods based on network and graph neural network (GNN) have become the mainstream paradigm because of their strong topological structure modeling capability. For example, LAGCN applies a graph convolution network to fuse embeddings from multiple graph convolution layers using an attention mechanism; DRWBNCF applies a new weighted bilinear graph convolution to the drug-disease association network, drug-drug similarity and disease-disease similarity network, and integrates all information in a unified representation.

[0004] In the construction of drug features and disease features, a variety of representation learning methods have been used in existing drug repositioning related technologies. For example, MGATRx integrates drug target associations, chemical substructures, side effect annotations and MeSH categories to construct drug features, and constructs disease features through disease gene associations, and then converts these features into binary adjacency matrices to construct multi-view heterogeneous graphs, and then uses a graph attention network (GAT) to learn node representations by selectively aggregating neighbor information.

[0005] In addition, protein information related to drugs or diseases is often introduced into the network to improve the prediction effect of drug repositioning. Some early methods model proteins as a type of abstract node in the network, such as the RLFDDA method. Subsequently, some methods introduce protein amino acid sequence information, for example, DRTerHGAT uses a protein pre-training model ESM-1b to extract original features from amino acid sequences, and constructs a multi-task autoencoder to simultaneously perform feature reconstruction and protein interaction network prediction tasks to obtain high-quality feature representations that integrate protein interaction function information.

[0006] Although existing technologies have made progress in integrating drug, disease and protein information, there are still significant limitations in accurately capturing complex biological network features and fine modeling of complex entity information.

[0007] 1) Defects in capturing complex biological network features: Existing technologies fail to integrate domain knowledge in an effective way to fully capture the complex associations of biological networks. Existing technologies mostly integrate global similarity to capture the overall structure of the network and integrate prior knowledge, but only integrate domain knowledge through similarity matrices, which limits the understanding of complex biological processes and relationships and fails to fully capture the complex features of biological networks and entities.

[0008] 2) Limitations in fine-grained modeling of complex biomedical entities: Existing technologies fail to model the high-order semantic features of disease entities. Existing technologies mainly focus on the optimization design of drug chemical structures, target associations, and other features, while for diseases, most methods still rely on similarity or binary feature encoding, making it difficult to fully reveal the complex hierarchical structure and pathological associations of diseases, and there are limitations in capturing the high-order semantics of diseases.

[0009] 3) Existing methods fail to fully utilize the deep biological and functional information contained in proteins. Existing methods mostly treat proteins as abstract nodes or only introduce sequence information of proteins, but lack modeling of three-dimensional spatial conformation and key functional sites, which cannot accurately describe the complex three-dimensional structure of proteins and fail to effectively focus on functional local domains directly related to drug binding or disease pathways.

[0010] 4) Most existing methods have obvious node type limitations in heterogeneous graph construction, often only containing three core nodes of drugs, proteins, and diseases, and fail to include more important biological entity types. This simplified setting severely fragments the multidimensional associations of biological systems, fails to capture key biological context information and direct associations between entities, and ultimately makes it difficult for the model to fully analyze drug action and disease mechanisms from multiple perspectives, limiting its modeling ability and analysis accuracy for complex biomedical problems. SUMMARY

[0011] The present application aims to provide a drug repositioning method and system based on heterogeneous knowledge and structure fusion, which accurately captures complex biological network features and fine-grained modeling of complex entity information through an innovative drug repositioning model, thereby achieving precise drug repositioning.

[0012] To achieve the above purpose, the technical solution of the present application includes the following contents.

[0013] A drug repositioning method based on heterogeneous knowledge and structure fusion, the method comprising:

[0014] Constructing a biomedical field knowledge heterogeneous graph , the biomedical field knowledge heterogeneous graph The node types in the biomedical domain knowledge heterogeneous graph include: a drug node, a disease node, a protein node, a molecular function node, a biological process node, a pathway node, a disease symptom node, and a side effect node The edge types in the biomedical domain knowledge heterogeneous graph include: binding, activation, inhibition, and regulation of expression between a drug-protein node pair; treatment and contraindication between a drug-disease node pair; induction between a drug-side effect node pair; execution between a protein-molecular function node pair; participation and regulation between a protein-biological process node pair; composition and regulation of a pathway between a protein-pathway node pair; association between a protein-disease node pair; support between a molecular function-biological process node pair; correspondence between a biological process-pathway node pair; abnormal association between a pathway-disease node pair; and manifestation and accompaniment between a disease-symptom node pair;

[0015] Based on the biomedical domain knowledge heterogeneous graph , a disease knowledge embedding and a drug knowledge embedding corresponding to a target drug-disease pair are generated;

[0016] By constructing a drug-drug similarity network, a disease-disease similarity network, and a drug-disease association network, a disease structure embedding and a drug structure embedding corresponding to the target drug-disease pair are generated;

[0017] Based on the disease knowledge embedding, the drug knowledge embedding, the disease structure embedding, and the drug structure embedding, a drug repositioning result is obtained.

[0018] Further, an initial embedding representation of the disease node is generated by the following steps:

[0019] A MeSH code of the disease is obtained;

[0020] The MeSH code is parsed into a sequence containing at least one token ; Then, the token is mapped to a semantic space through a trainable embedding layer to obtain an embedding vector of the token ;

[0021] Based on the embedding vector of the token , an initial feature to the disease is generated;

[0022] A hypergraph is constructed, where is a node set representing all disease entities; is a hyperedge set representing a MeSH category system, and an initial embedding representation of the hypergraph is the initial feature of the disease.

[0023] Using hypergraph convolutional networks to study hypergraphs Feature propagation is performed to obtain a heterogeneous graph of knowledge in the biomedical field. The initial embedding representation of disease nodes in the middle.

[0024] Furthermore, based on tokens Embedded vector The initial characteristics generated for the disease include:

[0025] For a disease containing only one MeSH code, the initial characteristics of the disease are determined by analyzing the sequences that make up that sequence. All tokens Embedded vector Obtain the mean pooling result;

[0026] For diseases containing multiple MeSH codes, the process of obtaining the initial features of the disease includes:

[0027] By analyzing the components of this sequence All tokens Embedded vector Perform mean pooling to obtain the features corresponding to the MeSH encoding. ;

[0028] Features of all MeSH codes involved in this disease Mean pooling is performed to obtain the initial characteristics of the disease.

[0029] Further, the initial embedding representation of the protein node is generated through the following steps:

[0030] Obtain the PDB structure file corresponding to the protein, and use the Saberi-Fathi algorithm to identify the pocket diagram structure of the protein. ,in Represents a set of residue nodes. Represents the set of edges and residue nodes. Features include: residue type, residue self-distance, and residue dihedral angle, edge Features include: whether residues are connected, residue CA distance, residue center distance, and maximum residue distance;

[0031] In pocket diagram structure Above, a Graph Transformer network is used to learn the embedding representation of proteins, resulting in a heterogeneous knowledge graph in the biomedical field. The initial embedding representation of protein nodes.

[0032] Further, the initial embedding representation of the drug node is generated through the following steps:

[0033] The drug molecule diagram is constructed using the atoms of the drug as nodes and the chemical bonds between atoms as edges. ;in, Represents a set of atomic nodes. A set of chemical bond edges is represented by the following characteristics of atomic nodes: atomic type, formal charge, number of unpaired electrons, isotopic labeling, chiral structure, minimum ring size, hybridization type, and aromaticity labeling. The characteristics of chemical bond edges include: bond type, conjugation labeling, intracyclic bond labeling, double bond configuration, and isotopic labeling.

[0034] The feature linear transformation of each atomic node is mapped to the hidden space to obtain the atomic node. initial embedding representation ;

[0035] Based on graph isomorphic networks for atomic nodes initial embedding representation Update to obtain atomic nodes Embedded representation , Indicates the number of layers in a graph isomorphic network;

[0036] All atomic nodes are treated by global average pooling. Embedded representation Aggregation yields a heterogeneous knowledge graph in the biomedical field. The initial embedding representation of the drug node in the middle.

[0037] Furthermore, based on the knowledge heterogeneity graph in this biomedical field Generate the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair, including:

[0038] Knowledge Heterogeneity Graph in the Biomedical Field In the process, obtain the drug node corresponding to the target drug-disease pair. and disease nodes ;

[0039] Acquisition and Drugs and disease The set of directly connected protein nodes is selected, and the nodes with the highest degree are chosen according to the degree priority principle. protein nodes ;

[0040] With each protein node Starting from this point, recursive sampling The protein's neighbor nodes are hopped, and the highest degree node in each hop is selected based on the degree priority principle. protein nodes ;

[0041] based on the drug node , the disease node , the protein node , the protein node and the molecular function node, the biological process node, the pathway node, the disease symptom node and the side effect node directly connected with the drug node , the disease node , the protein node or the protein node , construct a biomedical field knowledge heterogeneous graph related to the target drug-disease pair ;

[0042] use the HGT model to fuse the knowledge of the biomedical field knowledge heterogeneous graph composed of different node pairs to obtain the disease knowledge embedding and the drug knowledge embedding corresponding to the target drug-disease pair.

[0043] Further, by constructing a drug-drug similarity network, a disease-disease similarity network and a drug-disease association network, the disease structure embedding and the drug structure embedding corresponding to the target drug-disease pair are generated, including:

[0044] construct a drug-drug similarity network by calculating the target point combination information similarity of any two drugs ; wherein the target point combination information similarity is the ratio of the number of intersection elements to the number of union elements of the target point set;

[0045] convert each disease into a directed acyclic graph corresponding to the MeSH descriptor, and calculate the similarity of two diseases based on the directed acyclic graph to obtain a disease-disease similarity network ;

[0046] based on the association between drugs and diseases, construct a drug-disease association network, and represent the drug-disease association network as a binary adjacency matrix ;

[0047] splice the drug-drug similarity network , the disease-disease similarity network and the binary adjacency matrix to obtain an input graph ;

[0048] use a hierarchical GCN network on the input graph to learn the global structure representation of the node, and obtain the global structure embedding of the drug and the global structure embedding of the disease .

[0049] Further, based on the disease knowledge embedding, the drug knowledge embedding, the disease structure embedding and the drug structure embedding, a drug repositioning result is obtained, including:

[0050] The disease knowledge embedding and the disease structure embedding are spliced to obtain a disease final feature;

[0051] The drug knowledge embedding and the drug structure embedding are spliced to obtain a drug final feature;

[0052] The disease final feature and the drug final feature are spliced, and a drug repositioning result is obtained through a multi-layer perception machine.

[0053] A drug repositioning system based on heterogeneous knowledge and structure fusion, the system comprising:

[0054] A heterogeneous graph construction module for constructing a biomedical field knowledge heterogeneous graph The node types in the biomedical field knowledge heterogeneous graph include drug nodes, disease nodes, protein nodes, molecular function nodes, biological process nodes, pathway nodes, disease symptom nodes and side effect nodes, the edge types in the biomedical field knowledge heterogeneous graph include binding, activation, inhibition, regulation of expression between drug-protein node pairs; treatment, contraindication between drug-disease node pairs; induction between drug-side effect node pairs; execution between protein-molecular function node pairs; participation, regulation between protein-biological process node pairs; composition, regulation of pathways between protein-pathway node pairs; association between protein-disease node pairs; support between molecular function-biological process node pairs; correspondence between biological process-pathway node pairs; abnormal association between pathway-disease node pairs; and, manifestation, accompaniment between disease-symptom node pairs;

[0055] A heterogeneous graph semantic fusion module for generating disease knowledge embedding and drug knowledge embedding corresponding to a target drug-disease pair based on the biomedical field knowledge heterogeneous graph

[0056] A global-local fusion structure embedding module for generating disease structure embedding and drug structure embedding corresponding to the target drug-disease pair by constructing a drug-drug similarity network, a disease-disease similarity network and a drug-disease association network;

[0057] A prediction module for obtaining a drug repositioning result based on the disease knowledge embedding, the drug knowledge embedding, the disease structure embedding and the drug structure embedding.

[0058] ​An electronic device, characterized in that the electronic device comprises a processor and a memory storing computer program instructions; the processor implements the drug repositioning method based on heterogeneous knowledge and structure fusion of any one of the above when executing the computer program instructions.

[0059] Compared with the prior art, the present application has at least the following beneficial effects.

[0060] (1) Heterogeneous knowledge and global-local structure integration: The present application proposes a novel integration framework that systematically integrates heterogeneous biomedical knowledge and global-local structure information of the network, enhancing the comprehensive modeling capability of the drug repositioning model.

[0061] (2) Disease high-order semantic modeling: The present application uses a hypergraph convolution network to effectively capture the complex hierarchical relationships and pathological associations between diseases, achieving high-order semantic representation of diseases.

[0062] (3) Fine modeling of protein key functional regions and three-dimensional structure: The present application introduces protein three-dimensional structure information and models protein structure graphs based on binding pockets to accurately represent key functional regions involved in drug-target interaction and disease-protein association in proteins, achieving fine modeling of proteins.

[0063] (4) Introduction of multi-dimensional biological entity nodes to build a comprehensive heterogeneous network model: Breaking through the limitation of existing biomedical heterogeneous graphs containing only three core nodes of drugs, proteins, and diseases, the present application innovatively introduces entity nodes with key biological significance such as molecular function, biological process, pathway, disease symptom, and side effect, to build a more comprehensive multi-dimensional heterogeneous network model. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 The method flowchart of the present system. DETAILED DESCRIPTION

[0065] The present system will be further described in detail below in conjunction with the accompanying drawings, and the examples are only used to explain the present system and not to limit the scope of the present system.

[0066] The present application not only integrates biomedical field knowledge, global semantic and pharmacological information, and local substructure information to enhance the comprehensive modeling capability of the model, but also systematically represents disease associations in the disease classification system to achieve high-order semantic representation of diseases, and fine models the key regions of proteins by using three-dimensional information, thereby achieving precise drug repositioning. Specifically, as shown in Figure 1 the drug repositioning method based on heterogeneous knowledge and structure fusion of the present application includes the following steps.

[0067] Step S1: Construct a knowledge heterogeneous graph in the biomedical field .

[0068] This invention constructs heterogeneous networks , where the node set It includes eight types: in addition to drugs, diseases, and protein nodes, it also introduces molecular functions, biological processes, pathways, disease symptoms, and side effects. (Border sets) Eighteen types of biological relationships were defined, as shown in Table 1.

[0069] Node pair combination Edge type Biological implication Drug - Protein Binding Drug and protein specific domains undergo non-covalent interaction Drug - Protein Activation Drug enhances biological activity or function of protein Drug - Protein Inhibition Drug reduces or blocks activity of protein Drug - Protein Regulation of expression Drug affects transcription or translation of protein coding gene Drug - Disease Treatment Drug used to improve or cure disease Drug - Disease Contraindication Drug has usage contraindication for specific disease Drug - Side effect Induction Adverse clinical response induced after drug use Protein - Molecular function Perform Protein performs specific molecular level function Protein - Biological process Involved in Protein is a component involved in specific biological process Protein - Biological process Regulation Protein regulates rate or direction of biological process Protein - Pathway Component of Protein is a structural or functional component of a certain signaling pathway Protein - Pathway Regulation of pathway Protein activates or inhibits signaling transmission of pathway Protein - Disease Association Protein functional abnormality or mutation is associated with disease occurrence Molecular function - Biological process Support Molecular function provides basis for biological process to proceed Biological process - Pathway Correspondence Pathway is the specific molecular implementation path of biological process Pathway - Disease Abnormal association Abnormal activity (over-activation or inhibition) of pathway is associated with disease Disease - Symptom Manifestation Observable symptoms of disease presented in clinic Disease - Symptom Accompanying Non-specific symptoms that are likely to occur simultaneously with disease occurrence

[0070] Table 1

[0071] By introducing multiple types of nodes and associated edges, the association information between drugs and diseases is enriched from multiple dimensions, including molecular mechanisms, pathological logic, and clinical characteristics, constructing a complex association network that is more in line with biomedical realities. This not only breaks through the limitations of traditional drug repositioning that relies solely on direct associations, but also accurately captures the suitability of drugs to the potential mechanisms of disease action. Furthermore, it can use multi-dimensional information to screen for lower-risk candidate drugs, while clearly supporting the internal association logic of drug repositioning, ultimately significantly improving the reliability of candidate drugs and the efficiency of subsequent research and development.

[0072] The following section explains the entity embedding of each node.

[0073] (1) Disease coding module.

[0074] Based on the tree-like structure of disease MeSH (Medical Subject Headings) encoding, this invention designs a hypergraph embedding framework to capture semantic associations between diseases. First, the MeSH encoding is initialized for embedding; each MeSH encoding is parsed into a hierarchical semantic unit token sequence. Through trainable embedding layers Mapping to semantic space:

[0075]

[0076] in, Represents token Embedded vector, For the size of the token vocabulary, For the embedding dimension. Embedding of a single MeSH encoding. Obtained through mean pooling of its constituent tokens:

[0077]

[0078] For diseases containing multiple MeSH codes , initial features of which are generated by mean pooling with MeSH encoding.

[0079] Obtain initial embeddings of disease nodes After that, construct hypergraph , where is a set of nodes, representing all disease entities; is a set of hyperedges, representing MeSH category system. The association relationship of the hypergraph is represented by the association matrix :

[0080]

[0081] Feature propagation is performed using an L-layer hypergraph convolutional network (HGCN):

[0082]

[0083] wherein, is a node degree matrix, is a hyperedge degree matrix, and is a learnable weight matrix, represents an activation function, represents the disease node embedding representation of the i-th layer in the hypergraph .

[0084] (2) Protein encoding module.

[0085] The three-dimensional structure of a protein plays an important role in drug-target binding and disease-protein interaction. High-quality protein structures predicted by AlphaFold2 can effectively represent proteins, so the present application obtains the PDB structure file corresponding to the protein from the AlphaFold database through the Uniprot ID of the protein. Subsequently, the protein pocket map structure is identified based on the Saberi-Fathi method, focusing on key domains related to drug binding and disease pathways.

[0086] In terms of protein graph representation, a protein pocket map is constructed at the residue level, and a connection is established with a distance threshold of 10.0 Å. For a protein , the protein graph can be represented as , wherein represents a set of residue nodes, and the node feature , represents a set of edges, and the edge feature , wherein and represent the number of nodes and edges, respectively, and ​These represent the embedding dimensions of protein graph nodes and edges, respectively. For residue nodes, the following features are extracted: residue type, residue self-distance, and residue dihedral angle; for edges, the following features are extracted: whether residues are connected, residue CA distance, residue center distance, and residue maximum distance.

[0087] After obtaining the protein pocket graph structure, this invention employs a Graph Transformer network to learn the protein's embedding representation. This model captures the complex interactions between protein residues through a multi-layer attention mechanism, while also considering the influence of edge features.

[0088] For residue nodes , its node features With its Laplace position code The initial features are obtained by splicing. For nodes With edge initial features and Perform a linear transformation to map to the hidden space:

[0089]

[0090]

[0091] in and This is the weight matrix. and This is the bias vector.

[0092] For a multi-head graph transformer layer, the first... The formula for calculating the multi-head attention of a layer is:

[0093] , ,

[0094]

[0095]

[0096]

[0097]

[0098] in , It is the dimension of each attention head. and These are learnable parameters. the number of attention heads, representing a node set of neighbor nodes. Subsequently, and are input into the feedforward network:

[0099]

[0100]

[0101] Through the above-mentioned Graph Transformer module, the features of the residual nodes are obtained, and finally all node features are aggregated into a graph-level representation of the protein by global average pooling:

[0102]

[0103] Since the protein acts as an auxiliary information carrier in drug repositioning, its role is to enhance the contextual awareness of the drug and disease representation, so the initial embedding representation of the protein is obtained through the pre-training strategy. Specifically, the protein encoding module is trained on the entire protein-protein association network. The pre-training task is defined as protein association prediction - for any protein pair , the model predicts the interaction probability through the inner product operation of its embedding vector and the Sigmoid activation function:

[0104]

[0105] After obtaining the interaction probability, the cross-entropy loss function is used for training, and after pre-training, the model is used to generate the initial embedding for each protein in the dataset.

[0106] (3) Drug encoding module.

[0107] In order to accurately model the biological and structural properties of drug molecules, the present application uses InChI representation as the input source of drug molecules, and extracts key features of atoms and chemical bonds. The drug molecule graph is constructed with atoms as nodes and chemical bonds between atoms as edges. For a drug , the drug molecule graph can be represented as , where represents the set of atomic nodes, represents the set of chemical bond encodings. For atoms, the following features are extracted: atomic type, formal charge, number of unpaired electrons, isotope label, chiral structure, minimum ring size, hybridization type, and aromaticity label, which contain the basic chemical properties, stereochemical features, ring structure features, and chemical environment features of the node. For edges, the following features of the edge are extracted: bond type, conjugation label, intra-ring bond label, double bond configuration, and isotope label.

[0108] The present application adopts the InChI representation method of the International Union of Pure and Applied Chemistry (IUPAC) standard as the input source of drug molecules. InChI is a linear representation method based on molecular structure, which can accurately encode the atomic composition, bond connection relationship, stereochemical configuration and isotope information of the molecule. By parsing the InChI string into a molecular object through the RDKit chemical information toolkit, the structure and chemical characteristics of atoms and chemical bonds are extracted. The drug molecule graph is constructed by taking the atoms of the drug as nodes and the chemical bonds between the atoms as edges. For the drug , the drug molecule graph can be represented as , wherein represents the set of atomic nodes, and the node features , represents the set of chemical bond edges, and the edge features , wherein and represent the number of nodes and edges, respectively, and represent the embedding dimensions of the drug molecule graph nodes and edges, respectively. For atoms, the following features are extracted: atomic type, formal charge, number of unpaired electrons, isotope label, chiral structure, minimum ring size, hybridization type and aromaticity label, which contain the basic chemical properties, stereochemical features, ring structure features and chemical environment features of the nodes. For edges, the following features of the edges are extracted: bond type, conjugation label, intra-ring bond label, double bond configuration and isotope label.

[0109] The graph isomorphism network (GIN) with fused edge features is used for the drug molecule graph, and for the nodes in the drug molecule graph, the initial features are linearly transformed and mapped to the hidden space:

[0110]

[0111] wherein is the weight matrix, is the bias vector, is the hidden layer dimension. For edges , the neighbor node features and the edge features are fused to generate the message :

[0112]

[0113] wherein represents the concatenation operation, is a two-layer nonlinear transformation.

[0114] The additive aggregation integrates neighbor messages and updates itself:

[0115]

[0116] where is the neighbor set of node , is the learnable parameter of the th layer, The structure of is similar to that of , both using two layers of nonlinear functions.

[0117] After L-layer GIN, the representation of node is obtained The last layer , all node features are aggregated into the graph-level representation of drug by global average pooling:

[0118]

[0119] (4) Other biomedical node embeddings.

[0120] For molecular function, biological process, pathway, disease symptom, and side effect biomedical nodes, first obtain the corresponding text description of each node: the description of molecular function and biological process is taken from the Gene Ontology (GO) database, the pathway description is from the KEGG Pathway database, the disease symptom description is from the MeSH medical subject headings, and the side effect description is from the DrugBank database; If there is no public text description for the corresponding node, the node name is directly used as the text input.

[0121] For the above biomedical nodes , input the corresponding text description (or node name) of the biomedical pre-trained language model , and obtain the hidden layer features of the sequence after tokenization and encoding. Take the hidden layer output corresponding to the tag in the sequence as the embedding vector of the node , that is:

[0122]

[0123] where is the node name or text description of node , represents the hidden layer features of the tag in the output.

[0124] Through the above method, the node embeddings of molecular function, biological process, pathway, disease symptom, and side effect are obtained .

[0125] Step S2: generating disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair based on the biomedical field knowledge heterogeneous graph .

[0126] In order to capture the high-order semantic relationship of drug-disease pairs in the heterogeneous graph, the present application adopts an embedding learning method based on heterogeneous graph Transformer (HGT). This method first acquires a local interaction environment through subgraph sampling, and then applies an HGT model on the sampled subgraph.

[0127] Given a target drug-disease pair , the subgraph sampling process is as follows:

[0128] Direct neighbor sampling: acquire the protein set directly connected with the drug and the disease , and select the highest degree number of proteins according to the degree priority principle:

[0129]

[0130] Multi-hop neighbor expansion: take the proteins obtained by direct neighbor sampling in the above step as the starting point, recursively sample hop neighbors, and select the highest degree number of proteins according to the degree priority principle. For the hop proteins , sample their neighbors:

[0131]

[0132] For other nodes, acquire the directly connected molecular functions, biological processes, pathways, disease symptoms, and side effects through the drugs, diseases, and proteins obtained by the above steps.

[0133] Through the above sampling steps, the biomedical field knowledge heterogeneous graph related to the given target drug-disease pair is obtained. For this heterogeneous graph, an HGT model is used to fuse the knowledge of the (target node, edge type, neighbor node) triple composed of different biomedical pairs, and the various biomedical node embeddings obtained by the previous module are used as the initial embedding of the HGT node. For the target node and the neighbor node , the meta-relation triple is used to parameterize the attention weight:

[0134]

[0135]

[0136] wherein, , denote the node type of the neighbor node and the target node , denote the edge type of the edge , , denote the embedding of the neighbor node and the target node at layer , denote the query vector of the target node at layer , the th attention head, denote the key vector of the neighbor node at layer , the th attention head, denote the value vector of the neighbor node at layer , the th attention head, denote the projection matrix of the node type at layer , the th attention head for generating the query vector, denote the projection matrix of the node type at layer , the th attention head for generating the key vector, denote the projection matrix of the node type at layer , the th attention head for generating the value vector, denote the attention weight of the neighbor node to the target node via the edge at layer , the th attention head, denote the dimension size of the query vector / key vector , denote the total number of attention heads, denote the neighbor node set of the target node , denote the hidden output of the target node at layer .

[0137] Finally, update the node embedding by residual connection and layer normalization:

[0138]

[0139] Stacking After layer HGT, get semantic embedding of drug and disease as and .

[0140] Step S3: generate the disease structure embedding and the drug structure embedding corresponding to the target drug-disease pair by constructing a drug-drug similarity network, a disease-disease similarity network and a drug-disease association network.

[0141] The multi-scale structure embedding aims to obtain an embedding representation that can comprehensively reflect the interaction mode by integrating the global and local association structure information of drugs and diseases. First, a drug-drug similarity network and a disease-disease similarity network are constructed. For drug-drug similarity, based on the target binding information of drugs, Jaccard similarity is used to calculate the drug similarity matrix . For any two drugs, the similarity is the ratio of the number of intersection elements to the number of union elements of the target set combined by the two drugs. For disease-disease similarity, MeSH tree structure is used for calculation. Each disease is corresponded to a directed acyclic graph (DAG) by MeSH descriptor, which contains the hierarchical relationship and association characteristics of the disease in the classification system. Then the similarity of two diseases is calculated by the generalized Jaccard formula, and the disease similarity matrix is obtained. For drug-disease association, the present application represents it as a binary adjacency matrix . First, the drug-drug similarity network , the disease-disease similarity network and the adjacency matrix are spliced together as an input graph:

[0142]

[0143] where ~ represents normalizing the similarity network.

[0144] GCN is used on the input graph to learn the global structure representation of the node:

[0145]

[0146] where is the degree matrix of the input graph , is the node embedding of the l-th layer, is the trainable weight matrix of the l-th layer, is a nonlinear activation function. The embedding is initialized as:

[0147]

[0148] Stacking After the GCN layer, different embedding layers are combined using an attention mechanism to obtain the global structure embedding of drugs and diseases is and :

[0149]

[0150] where is the attention weight learned automatically by the neural network.

[0151] The global structure embedding of drugs and diseases and is reconstructed by the inner product decoder and the Sigmod activation function:

[0152]

[0153] In addition, in the training stage, the weighted cross-entropy loss function can be used for optimization:

[0154]

[0155] where and are the number of samples in the positive sample set and the negative sample set , respectively, is the number of drugs, is the number of diseases.

[0156] Step S4: Based on the disease knowledge embedding, drug knowledge embedding, disease structure embedding and drug structure embedding, the drug repositioning result is obtained.

[0157] After obtaining the semantic embedding and global structure embedding of drugs and diseases, the final features of drugs and diseases are obtained by splicing:

[0158]

[0159] Finally, the features of drugs and diseases are spliced and the prediction result is obtained by MLP:

[0160]

[0161] wherein is a weight matrix, is a bias term, is a Sigmod activation function.

[0162] During the training process, the present application can use a cross-entropy loss function to optimize after obtaining a prediction result:

[0163]

[0164] wherein is a total number of samples, is a true label of the th sample, is a prediction value of the th sample.

[0165] Although specific embodiments of the system are disclosed herein for illustrative purposes, and are to be considered merely illustrative and not restrictive of the scope of the system, it will be readily apparent to one of ordinary skill in the art that other embodiments can be practiced without departing from the spirit and scope of the system and the claims appended hereto. Accordingly, the system should not be limited to the disclosed best mode, but should be given the broadest scope in the claims appended hereto.

Claims

1. A drug relocation method based on heterogeneous knowledge and structural fusion, characterized in that, The method includes: Constructing a knowledge heterogeneous graph in the biomedical field The biomedical knowledge heterogeneity graph The node types in the graph include: drug nodes, disease nodes, protein nodes, molecular function nodes, biological process nodes, pathway nodes, disease symptom nodes, and side effect nodes. This is a heterogeneous knowledge graph in the biomedical field. The types of edges in the data include: binding, activation, inhibition, and expression regulation between drug-protein node pairs; treatment and contraindications between drug-disease node pairs; initiation of drug-side effects node pairs; execution between protein-molecular function node pairs; participation and regulation between protein-biological process node pairs; composition and regulatory pathways between protein-pathway node pairs; association between protein-disease node pairs; support between molecular function-biological process node pairs; correspondence between biological process-pathway node pairs; abnormal association between pathway-disease node pairs; and manifestations and accompaniments between disease-symptom node pairs. Based on the knowledge heterogeneity graph in the biomedical field Generate the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair; By constructing drug-drug similarity networks, disease-disease similarity networks, and drug-disease association networks, the disease structure embedding and drug structure embedding corresponding to the target drug-disease pair are generated. Based on disease knowledge embedding, drug knowledge embedding, disease structure embedding, and drug structure embedding, the drug relocation results are obtained. Among them, the knowledge heterogeneity graph based on this biomedical field Generate the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair, including: Knowledge Heterogeneity Graph in the Biomedical Field In the process, obtain the drug node corresponding to the target drug-disease pair. and disease nodes ; Acquisition and Drugs and disease The set of directly connected protein nodes is selected, and the nodes with the highest degree are chosen according to the degree priority principle. protein nodes ; With each protein node Starting from this point, recursive sampling The protein's neighbor nodes are hopped, and the highest degree node in each hop is selected based on the degree priority principle. protein nodes ; Based on drug nodes Disease nodes Protein nodes Protein nodes and the drug node Disease nodes Protein nodes or protein nodes By constructing a heterogeneous biomedical knowledge graph related to the target drug-disease pair, we can identify directly connected molecular functional nodes, biological process nodes, pathway nodes, disease symptom nodes, and side effect nodes. ; Using the HGT model to analyze the heterogeneity graph of knowledge in the biomedical field Knowledge fusion is performed on triples composed of different node pairs to obtain the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair.

2. The method according to claim 1, characterized in that, The initial embedding representation of the disease node is generated through the following steps: Obtain the MeSH code of the disease; MeSH encoding is parsed to contain at least one token. sequence Then, the token is embedded through a trainable embedding layer. Mapping to the semantic space yields the token. Embedded vector ; Based on token Embedded vector This generates the initial characteristics of the disease; Building a hypergraph ,in It is a set of nodes, representing all disease entities; It is a hyperedge set, representing the MeSH category system, the hypergraph. The initial embedding of the nodes is represented as the initial features of the disease; Using hypergraph convolutional networks to study hypergraphs Feature propagation is performed to obtain a heterogeneous graph of knowledge in the biomedical field. The initial embedding representation of disease nodes in the middle.

3. The method according to claim 2, characterized in that, Based on token Embedded vector The initial characteristics generated for the disease include: For a disease containing only one MeSH code, the initial characteristics of the disease are determined by analyzing the sequences that make up that sequence. All tokens Embedded vector Obtain the mean pooling result; For diseases containing multiple MeSH codes, the process of obtaining the initial features of the disease includes: By analyzing the components of this sequence All tokens Embedded vector Perform mean pooling to obtain the features corresponding to the MeSH encoding. ; Features of all MeSH codes involved in this disease Mean pooling is performed to obtain the initial characteristics of the disease.

4. The method according to claim 1, characterized in that, The initial embedding representation of the protein node is generated through the following steps: Obtain the PDB structure file corresponding to the protein, and use the Saberi-Fathi algorithm to identify the pocket diagram structure of the protein. ,in Represents a set of residue nodes. Represents the set of edges and residue nodes. Features include: residue type, residue self-distance, and residue dihedral angle, edge Features include: whether residues are connected, residue CA distance, residue center distance, and maximum residue distance; In pocket diagram structure Above, a Graph Transformer network is used to learn the embedding representation of proteins, resulting in a heterogeneous knowledge graph in the biomedical field. The initial embedding representation of protein nodes.

5. The method according to claim 1, characterized in that, The initial embedding representation of the drug node is generated through the following steps: The drug molecule diagram is constructed using the atoms of the drug as nodes and the chemical bonds between atoms as edges. ;in, Represents a set of atomic nodes. A set of chemical bond edges is represented by the following characteristics of atomic nodes: atomic type, formal charge, number of unpaired electrons, isotopic labeling, chiral structure, minimum ring size, hybridization type, and aromaticity labeling. The characteristics of chemical bond edges include: bond type, conjugation labeling, intracyclic bond labeling, double bond configuration, and isotopic labeling. The feature linear transformation of each atomic node is mapped to the hidden space to obtain the atomic node. initial embedding representation ; Based on graph isomorphic networks for atomic nodes initial embedding representation Update to obtain atomic nodes Embedded representation , Indicates the number of layers in a graph isomorphic network; All atomic nodes are processed using global average pooling. Embedded representation Aggregation yields a heterogeneous knowledge graph in the biomedical field. The initial embedding representation of the drug node in the middle.

6. The method according to claim 1, characterized in that, By constructing drug-drug similarity networks, disease-disease similarity networks, and drug-disease association networks, the disease structure embedding and drug structure embedding corresponding to the target drug-disease pair are generated, including: A drug-drug similarity network is constructed by calculating the similarity of target binding information between any two drugs. Wherein, the target-point combination information similarity is the ratio of the number of intersection elements to the number of union elements of the target point sets; Each disease is mapped to a directed acyclic graph (DAG) using a MeSH descriptor, and the similarity between two diseases is calculated based on this DAG to obtain a disease-disease similarity network. ; Based on the association between drugs and diseases, a drug-disease association network is constructed and represented as a binary adjacency matrix. ; Drug-drug similarity networks Disease-disease similarity network With binary adjacency matrix By splicing the images together, we obtain the input image. ; In the input image A hierarchical GCN network is then used to learn the global structural representation of nodes, resulting in the global structural embedding of the drug. Embedded in the global structure of disease .

7. The method according to claim 1, characterized in that, Based on disease knowledge embedding, drug knowledge embedding, disease structure embedding, and drug structure embedding, the drug relocation results are obtained, including: By combining disease knowledge embeddings and disease structure embeddings, the final disease features are obtained. By combining drug knowledge embeddings and drug structure embeddings, the final characteristics of the drug are obtained. The final characteristics of the disease and the final characteristics of the drug are spliced ​​together, and the drug relocation result is obtained through a multilayer perceptron.

8. A drug relocation system based on heterogeneous knowledge and structural fusion, characterized in that, The system includes: The heterogeneous graph construction module is used to construct heterogeneous graphs of knowledge in the biomedical field. The biomedical knowledge heterogeneity graph The node types in the graph include: drug nodes, disease nodes, protein nodes, molecular function nodes, biological process nodes, pathway nodes, disease symptom nodes, and side effect nodes. This is a heterogeneous knowledge graph in the biomedical field. The types of edges in the data include: binding, activation, inhibition, and expression regulation between drug-protein node pairs; treatment and contraindications between drug-disease node pairs; initiation of drug-side effects node pairs; execution between protein-molecular function node pairs; participation and regulation between protein-biological process node pairs; composition and regulatory pathways between protein-pathway node pairs; association between protein-disease node pairs; support between molecular function-biological process node pairs; correspondence between biological process-pathway node pairs; abnormal association between pathway-disease node pairs; and manifestations and accompaniments between disease-symptom node pairs. The heterogeneous graph semantic fusion module is used for knowledge heterogeneous graphs in this biomedical field. Generate the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair; The global-local fusion structure embedding module is used to generate the disease structure embedding and drug structure embedding corresponding to the target drug-disease pair by constructing a drug-disease similarity network, a disease-disease similarity network, and a drug-disease association network. The prediction module is used to obtain drug relocation results based on disease knowledge embedding, drug knowledge embedding, disease structure embedding, and drug structure embedding. Among them, the knowledge heterogeneity graph based on this biomedical field Generate the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair, including: Knowledge Heterogeneity Graph in the Biomedical Field In the process, obtain the drug node corresponding to the target drug-disease pair. and disease nodes ; Acquisition and Drugs and disease The set of directly connected protein nodes is selected, and the nodes with the highest degree are chosen according to the degree priority principle. protein nodes ; With each protein node Starting from this point, recursive sampling The protein's neighbor nodes are hopped, and the highest degree node in each hop is selected based on the degree priority principle. protein nodes ; Based on drug nodes Disease nodes Protein nodes Protein nodes and the drug node Disease nodes Protein nodes or protein nodes By constructing a heterogeneous biomedical knowledge graph related to the target drug-disease pair, we can identify directly connected molecular functional nodes, biological process nodes, pathway nodes, disease symptom nodes, and side effect nodes. ; Using the HGT model to analyze the heterogeneity graph of knowledge in the biomedical field Knowledge fusion is performed on triples composed of different node pairs to obtain the disease knowledge embedding and drug knowledge embedding corresponding to the target drug-disease pair.

9. An electronic device, characterized in that, The electronic device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the drug relocation method based on heterogeneous knowledge and structural fusion as described in any one of claims 1-7.

Images