Drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution

Abstract

The application discloses a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, comprising the following steps: constructing a first and a second drug-microorganism heterogeneous graph, establishing a locally neighbor-enhanced attribute embedding; combining a multi-view variational graph autoencoder to learn the attribute distribution of nodes; combining a cross-attention to establish an enhanced hyperedge attribute; constructing a dynamically evolving hypergraph topology; establishing a multi-scale hypergraph convolution framework, adaptively fusing multi-scale node attributes to obtain a dynamic hypergraph attribute representation of the nodes; and through comparative learning of multiple graphs, the consistency between the features of the same node in multiple graphs is enhanced. The prediction method can infer the association tendency between real-time obtained drugs and microorganisms. In summary, the application can learn multi-view node attribute distribution, multi-scale neighbor topology and dynamically encode the connection between multiple nodes of the hypergraph, enhance the consistency of node representation, and improve the performance of drug-related microorganism candidate prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and particularly relates to a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution. BACKGROUND

[0002] Human microbiome is a complex community containing bacteria, archaea, viruses and microbial eukaryotes. Studies have shown that microorganisms can enhance or inhibit the efficacy of drugs by changing the structure of drugs. Gut microbiota can affect the immune surveillance of colorectal cancer and the response to immunotherapy. Penicillin can treat pneumonia, meningitis, endocarditis, etc. However, the widespread use of antibiotics has led to drug resistance in human microorganisms such as Staphylococcus aureus and Escherichia coli. Therefore, predicting drug-microorganism associations can help to screen relevant microorganism candidates for a given drug and help to explore the mechanism of action of the drug.

[0003] Machine learning methods have been widely used to develop new drug-microorganism association prediction methods. For example, the association prediction method based on KATZ measure infers the association between drugs and microorganisms by calculating the path length between them; the Laplace regularization least squares algorithm infers the tendency of drug-microorganism association by evaluating the association score between drugs and microorganisms. However, these methods fail to learn the deep features of drugs and microorganisms. With the development of deep learning technology, prediction methods based on deep learning have been gradually introduced into the task of drug-microorganism association prediction. For example, graph autoencoder and variational graph autoencoder have been used to learn the feature distribution of nodes, and prediction methods that use drug and microorganism data and combine graph attention network, meta-path-oriented graph convolution network, graph neural network, graph neural network in hyperbolic space and relationship-sensitive transformer to infer drug-related microorganism candidates have also been used. However, these methods do not integrate the connections between multiple drugs and microorganism nodes. To solve this problem, some researchers have tried to integrate such connections based on hypergraph convolution methods, but traditional hypergraph convolution methods usually only focus on the aggregation of single-scale node neighbor topological information, ignoring the information of multiple-scale neighbor topologies. In addition, the consistency between the same nodes in multiple graphs is also ignored, which may limit the prediction performance of the model.

[0004] Therefore, it is necessary to propose a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution for the task of drug-microorganism association prediction, which can learn the feature distribution of multi-view drug and microorganism nodes, multi-scale neighbor topologies and dynamically encode the connections between multiple nodes from hypergraphs, enhance the consistency of node representation, and thus improve the performance of the drug-related candidate task. SUMMARY

[0005] Therefore, the present application provides a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, to solve the technical problems that the current drug-microorganism association prediction method does not integrate the connection between multiple drug and microorganism nodes, and only focuses on the aggregation of single-scale node neighbor topological information, ignoring the multi-scale neighbor topological information, resulting in low accuracy of association prediction results.

[0006] The present application provides a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, comprising:

[0007] Obtain a drug-microorganism association database;

[0008] According to the drug-microorganism association database, construct a first heterogeneous graph based on drug function similarity and a second heterogeneous graph based on drug attribute similarity containing drug and microorganism nodes, establish a local neighbor enhanced attribute embedding, and based on the local neighbor enhanced attribute embedding and the multi-view variational graph autoencoder, obtain the heterogeneous graph attribute representation of the node;

[0009] According to the drug-microorganism association database, generate a drug-microorganism hypergraph, adopt a cross-attention strategy, construct a dynamic hypergraph topology based on the joint hyperedge attribute matrix and the joint node attribute matrix of the hypergraph;

[0010] Introduce a hypergraph Laplacian operator in the dynamic hypergraph topology to form a multi-order neighbor topology embedding, adaptively fuse to obtain multi-scale node attributes, and supplement the multi-scale node attributes in the hypergraph convolution layer to obtain the dynamic hypergraph attribute representation of the node;

[0011] According to the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation, perform graph comparison learning on any two graphs in the first heterogeneous graph, the second heterogeneous graph and the multi-scale topology perspective hypergraph, and train to obtain a drug-microorganism association prediction model;

[0012] Real-time acquire the feature vectors of drugs and microorganisms, and analyze the association between the drugs and microorganisms through the drug-microorganism association prediction model.

[0013] Further, the method for establishing a local neighbor enhanced attribute embedding comprises:

[0014] In the first heterogeneous graph and the second heterogeneous graph, respectively, take a drug or a microorganism as a target node, and screen a plurality of neighbor nodes most similar to the attribute of the target node; wherein, in the first heterogeneous graph, cosine similarity is used to calculate the similarity of node attributes, and in the second heterogeneous graph, heat kernel similarity is used to calculate the similarity of node attributes;

[0015] concatenate the attribute vector of the target node with the attribute vectors of the neighbor nodes to obtain a neighbor node attribute matrix of the target node;

[0016] perform weighted aggregation on the neighbor node attribute matrix using a convolution filter, reserve original attribute information through a residual connection, and obtain a locally neighbor-enhanced attribute embedding.

[0017] Further, the node heterogeneous graph attribute representation is obtained based on the locally neighbor-enhanced attribute embedding and the multi-view variational graph autoencoder, and includes:

[0018] the locally neighbor-enhanced attribute embedding is taken as an input of a shared parameter graph convolutional encoding layer, common information is obtained according to the shared parameter graph convolutional encoding layer, and the common information is respectively propagated and supplemented into the first heterogeneous graph and the second heterogeneous graph to form attribute matrices;

[0019] the variational graph autoencoder is used to model a node feature distribution according to the attribute matrices, and the feature distribution of the nodes of the heterogeneous graph is calculated through a reparameterization technique.

[0020] Further, the locally neighbor-enhanced attribute embedding is taken as an input of a shared parameter graph convolutional encoding layer, common information is obtained according to the shared parameter graph convolutional encoding layer, and the common information is respectively propagated and supplemented into the first heterogeneous graph and the second heterogeneous graph to form attribute matrices, and includes:

[0021] the shared parameter graph convolutional network is used to encode topology and node attributes common to the first heterogeneous graph and the second heterogeneous graph to obtain a common information feature map;

[0022] the common information feature map obtained from the drug function perspective of the first heterogeneous graph is propagated and supplemented into the drug attribute perspective of the second heterogeneous graph to form an attribute matrix of the nodes of the second heterogeneous graph;

[0023] the common information feature map obtained from the drug attribute perspective of the second heterogeneous graph is propagated and supplemented into the drug function perspective of the first heterogeneous graph to form an attribute matrix of the nodes of the first heterogeneous graph.

[0024] Further, the variational graph autoencoder is used to model a node feature distribution according to the attribute matrices, and the feature distribution of the nodes of the heterogeneous graph is calculated through a reparameterization technique, and includes:

[0025] the first graph convolutional network and the second graph convolutional network are used to respectively calculate a mean and a variance of the attribute matrices of the first heterogeneous graph and the second heterogeneous graph;

[0026] the latent feature distribution of the nodes is calculated based on a reparameterization technique, and the latent feature distribution is sampled from a normal distribution.

[0027] The common information features are concatenated with the potential feature distribution to obtain heterogeneous graph attribute representations of all nodes in the first heterogeneous graph and the second heterogeneous graph.

[0028] Furthermore, generating a drug-microbe hypergraph based on the drug-microbe association database, adopting a cross-attention strategy, and constructing a dynamic hypergraph topology based on the joint hyperedge attribute matrix and the joint node attribute matrix of the hypergraph include:

[0029] The hyperedge attribute matrices of the drug-microorganism hypergraph are sampled respectively according to a learnable Gaussian distribution, and are concatenated to form a joint hyperedge attribute matrix; at the same time, the node attribute matrices of the hypergraph are concatenated to form a joint node attribute matrix;

[0030] Aggregating all attributes of nodes in the hypergraph through a cross-attention strategy to generate enhanced hyperedge attributes;

[0031] A dynamic hypergraph topology is constructed according to the enhanced hyperedge attributes and the joint node attribute matrix.

[0032] Furthermore, all attributes of the nodes in the hypergraph are aggregated through a cross-attention strategy to generate enhanced hyperedge attributes, including:

[0033] Linearly transforming the joint hyperedge attribute matrix to obtain a query matrix;

[0034] Linearly transforming the joint node attribute matrix to obtain a key matrix and a value matrix respectively;

[0035] Enhanced hyperedge properties are obtained according to the query matrix, the transpose of the key matrix and the value matrix.

[0036] Furthermore, a hypergraph Laplacian operator is introduced into the dynamic hypergraph topology to form a multi-order neighbor topology embedding, and multi-scale node attributes are obtained by adaptively fusing the neighbor topology embedding. The multi-scale node attributes are supplemented in the hypergraph convolution layer to obtain a dynamic hypergraph attribute representation of the node, including:

[0037] Normalizing the dynamic hypergraph topology according to the hypergraph Laplacian operator to form a multi-order neighbor topology embedding;

[0038] Generate a topological attribute matrix of multi-order neighbor nodes using the multi-order neighbor topology embedding;

[0039] Adaptively fuse the topological attribute matrices of all scales to form a multi-scale node attribute matrix from a multi-scale topological perspective;

[0040] For any hypergraph convolution layer in the hypergraph, the topological attribute matrix of the current convolution layer and the multi-scale node attribute representation of the previous convolution layer are spliced until the splicing of all hypergraph convolution layers in the hypergraph is completed, and the dynamic hypergraph attribute representation of all nodes is obtained.

[0041] Further, the first heterogeneous graph, the second heterogeneous graph and any two graphs in the multi-scale topological perspective hypergraph are subjected to graph contrast learning according to the heterogeneous graph attribute representation of the nodes and the dynamic hypergraph attribute representation of the nodes, and a drug-microorganism association prediction model is trained, comprising:

[0042] According to the heterogeneous graph attribute representation of the nodes and the dynamic hypergraph attribute representation of the nodes, a first loss function of the first heterogeneous graph and the dynamic hypergraph and a second loss function of the second heterogeneous graph and the dynamic hypergraph are calculated;

[0043] According to the heterogeneous graph attribute representation of the nodes in the first heterogeneous graph and the attribute representation of the nodes in the second heterogeneous graph, a third loss function is calculated;

[0044] According to the first loss function, the second loss function and the third loss function, a joint contrast loss function is calculated, and a drug-microorganism association prediction model is obtained according to the joint contrast loss function.

[0045] The application also provides a drug-microorganism association prediction system based on a multi-perspective variational graph autoencoder and a multi-scale hypergraph convolution, comprising:

[0046] A data acquisition module is configured to acquire a drug-microorganism association database;

[0047] A multi-perspective variational graph autoencoder module is configured to construct a first heterogeneous graph based on drug function similarity and a second heterogeneous graph based on drug attribute similarity containing drug and microorganism nodes according to the drug-microorganism association database, establish a locally neighbor-enhanced attribute embedding, and obtain a heterogeneous graph attribute representation of nodes based on the locally neighbor-enhanced attribute embedding and the multi-perspective variational graph autoencoder;

[0048] A dynamic hypergraph module is configured to generate a drug-microorganism hypergraph according to the drug-microorganism association database, adopt a cross-attention strategy, and construct a dynamic hypergraph topology based on a joint hyperedge attribute matrix and a joint node attribute matrix of the hypergraph.

[0049] A multi-scale convolution module is configured to introduce a hypergraph Laplacian operator in the dynamic hypergraph topology, form a multi-order neighbor topology embedding, adaptively fuse to obtain a multi-scale node attribute, supplement the multi-scale node attribute in a hypergraph convolution layer, and obtain a dynamic hypergraph attribute representation of nodes.

[0050] The contrast learning module is used for graph contrast learning of any two graphs in the first heterogeneous graph, the second heterogeneous graph and the hypergraph of the multi-scale topological perspective according to the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation, and a drug-microorganism association prediction model is trained.

[0051] The association prediction module is used for real-time acquisition of feature vectors of drugs and microorganisms, and analysis of the association between the drugs and the microorganisms by the drug-microorganism association prediction model.

[0052] Compared with the prior art, the advantages of the present application are that:

[0053] (1) A drug-microorganism heterogeneous graph containing unique information of a drug function perspective and a drug attribute perspective is respectively constructed, and a variational graph autoencoder that propagates the unique information in the two perspectives is designed to model the feature distribution of each drug and microorganism node. This helps to reveal the feature distribution of each node in the entire graph. Unlike traditional techniques that only model a single perspective, the present method can flexibly adapt to multiple different data sets and problem scenarios, rather than being limited to a single or specific scenario.

[0054] (2) The close relationship between multiple drugs and microorganisms is revealed from multiple biological perspectives by a hypergraph composed of hyperedges. This method introduces hyperedge attributes for dynamic learning, which can adaptively capture the association of drug and microorganism nodes in different biological backgrounds and form the topological structure of the dynamic evolving hypergraph, which is scalable.

[0055] (3) A multi-scale topological hypergraph convolution framework is used to generate drug-microorganism node neighbor topology embeddings covering multiple scales, and to adaptively integrate different scale neighbor information. This method solves the defect of traditional hypergraph convolution that only targets single neighbor topology.

[0056] (4) Multiple graph contrast learning strategies are used to maximize the similarity of node feature representations in heterogeneous graphs and hypergraphs, effectively maintaining the consistency of node representations under multiple graph perspectives, which improves the generalization ability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 A flowchart of the drug-microorganism association prediction method based on the multi-perspective variational graph autoencoder and multi-scale hypergraph convolution provided by the present application is shown in the figure;

[0058] Figure 2 A graphical data flow diagram provided by the present application is shown in the figure;

[0059] Figure 3 A data processing diagram of the multi-perspective variational graph autoencoder provided by the present application is shown in the figure;

[0060] Figure 4 A dynamic hyperedge attribute learning schematic diagram provided by the present application;

[0061] Figure 5 A multi-scale hypergraph convolution process schematic diagram provided by the present application;

[0062] Figure 6 A multi-graph comparison process schematic diagram provided by the present application. DETAILED DESCRIPTION

[0063] The preferred embodiments of the present application are specifically described below in conjunction with the accompanying drawings, which form a part of this application, and together with the embodiments of the present application serve to explain the principles of the present application, and are not used to limit the scope of the present application.

[0064] The present application provides a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, which is specifically described below.

[0065] Please refer to Figure 1 The present embodiment provides a drug-microorganism association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, which comprises:

[0066] Step S101: Obtain a drug-microorganism association database;

[0067] Step S102: According to the drug-microorganism association database, construct a first heterogeneous graph based on drug function similarity and a second heterogeneous graph based on drug attribute similarity containing drug and microorganism nodes, establish a locally neighbor-enhanced attribute embedding, and based on the locally neighbor-enhanced attribute embedding and a multi-view variational graph autoencoder, obtain a heterogeneous graph attribute representation of the nodes;

[0068] Step S103: Generate a drug-microorganism hypergraph according to the drug-microorganism association database, adopt a cross-attention strategy, construct a dynamic hypergraph topology based on a joint hyperedge attribute matrix and a joint node attribute matrix of the hypergraph;

[0069] Step S104: Introduce a hypergraph Laplacian operator in the dynamic hypergraph topology to form a multi-order neighbor topology embedding, adaptively fuse node attributes under multi-scale topology perspectives according to the neighbor topology embedding, and construct a hypergraph under multi-scale topology perspectives to obtain a hypergraph attribute representation of the nodes;

[0070] Step S105: According to the heterogeneous graph attribute representation and the hypergraph attribute representation, perform graph comparison learning on any two graphs in the first heterogeneous graph, the second heterogeneous graph and the hypergraph under multi-scale topology perspectives, and train to obtain a drug-microorganism association prediction model;

[0071] Step S106: Real-time acquisition of the feature vectors of the drug and the microorganism, and analysis of the correlation between the drug and the microorganism by the drug-microorganism correlation prediction model.

[0072] The method of the embodiment dynamically encodes the connection between multiple nodes by learning multi-view node feature distribution in multiple heterogeneous graphs and hypergraphs, enhances the consistency of node representation, and helps to screen potential microorganism candidates for drugs.

[0073] As a specific embodiment, in step S101, the drug-microorganism correlation database includes: drug-microorganism correlation, drug chemical substructure similarity, drug Gaussian kernel similarity and microorganism similarity; wherein the drug-microorganism correlation includes 2470 pairs of original drug-microorganism correlations supported by clinical or experimental data from the MDAD database (Sun et al., 2018), and these correlations contain 1373 drugs and 173 microorganisms. Drug chemical substructure similarity is calculated based on the maximum common subgraph of drug structure by SIMCOMP2 (Hattori et al., 2010), from previous work of EGATMDA (Long et al., 2020a). Drug Gaussian kernel similarity is calculated based on the drugs interacting with two drugs. The similarity of microorganisms is calculated by the cosine similarity of the sequence features of the microorganisms. The gene sequence of the microorganism is obtained from the National Center for Biotechnology Information (NCBI), and the key sequence features are extracted by principal component analysis (PCA) (Chen et al., 2002).

[0074] As a specific embodiment, in step S102, the first heterogeneous graph is constructed based on the functional perspective of drug-drug, the nodes in the graph include drug and microorganism nodes, and the edges include drug-drug edges and drug-microorganism edges. The drug-drug edge represents the functional similarity between drugs, and these edges can be based on information such as shared targets, mechanisms of action, therapeutic effects, or pharmacological categories of drugs. The drug-microorganism edge represents the effect of the drug on the microorganism, such as inhibition, promotion or no effect, and these edges can be based on experimental data and information from clinical studies.

[0075] The second heterogeneous graph is constructed based on the attribute perspective of drug-drug, the nodes in the graph include drug and microorganism nodes, and the relationship between the edges is that the drug-drug edge represents the attribute similarity between drugs, and these edges can be based on information such as chemical structure, physicochemical properties (such as solubility, stability), molecular function of drugs, and the drug-microorganism edge represents the effect of the drug on the microorganism, such as inhibition, promotion or no effect.

[0076] In order to learn the multi-view node feature distribution, the common information needs to be determined first, and the neighbor-enhanced attribute embedding is taken as the input of the shared parameter graph convolutional encoding layer. Figure 2 As a preferred embodiment, please refer to the part (a) in the description, the method for establishing the local neighbor-enhanced attribute embedding comprises the following steps:

[0077] In the first and second heterogeneous graphs, the top-k neighbor nodes with the most similar attributes to the target node are screened out, wherein the cosine similarity is used to calculate the similarity of the node attributes in the first heterogeneous graph, and the heat kernel similarity is used to calculate the similarity of the node attributes in the second heterogeneous graph.

[0078] The attribute vector of the target node is spliced with the attribute vector of the neighbor node to obtain a neighbor node attribute matrix of the target node.

[0079] A 1*1 convolution filter is used to aggregate the neighbor node attribute matrix, and the residual connection is used to retain the original attribute information, so as to obtain the local neighbor-enhanced attribute embedding.

[0080] Through the above method, the local neighbor-enhanced attribute embedding is established, and the target node can obtain more attribute information from different neighbors, which helps to establish a more rich multi-level feature representation. In the drug-microorganism association prediction task, the attributes of the nodes are not only determined by a single feature, but also obtained by interacting with the features of multiple adjacent nodes. The enhanced embedding can better capture this complex feature interaction, so as to mine the key information hidden in the complex interaction between nodes.

[0081] As a preferred embodiment, in step S102, the heterogeneous graph attribute representation of the node is obtained based on the local neighbor-enhanced attribute embedding and the multi-view variational graph autoencoder, comprising:

[0082] The local neighbor-enhanced attribute embedding is taken as the input of the shared parameter graph convolutional encoding layer, the common information is obtained according to the shared parameter graph convolutional encoding layer, and the common information is propagated and supplemented into the first and second heterogeneous graphs to form attribute matrices.

[0083] The variational graph autoencoder is used to model the node feature distribution according to the attribute matrix, and the feature distribution of the heterogeneous graph node is calculated through the reparameterization technique.

[0084] As a preferred embodiment, the local neighbor-enhanced attribute embedding is taken as the input of the shared parameter graph convolutional encoding layer, the common information is obtained according to the shared parameter graph convolutional encoding layer, and the common information is propagated and supplemented into the first and second heterogeneous graphs to form attribute matrices, comprising:

[0085] The shared topology and node attributes of the first heterogeneous graph and the second heterogeneous graph are encoded by a shared parameter graph convolutional network to obtain a shared information feature graph;

[0086] The shared information feature graph obtained from the drug function perspective of the first heterogeneous graph is propagated and supplemented to the drug attribute perspective of the second heterogeneous graph to form an attribute matrix of the nodes of the second heterogeneous graph;

[0087] The common information feature graph obtained from the perspective of drug attributes of the second heterogeneous graph is propagated and supplemented to the drug function perspective of the first heterogeneous graph to form the attribute matrix of the nodes of the first heterogeneous graph.

[0088] As a preferred embodiment, the variational graph autoencoder is used to model the node feature distribution according to the attribute matrix, and the heterogeneous graph attribute representation of the node is calculated by reparameterization technique, including:

[0089] Calculate the mean and variance of the attribute matrix of the first heterogeneous graph and the second heterogeneous graph respectively using the first graph convolutional network and the second graph convolutional network;

[0090] calculating a latent feature distribution of the node based on a reparameterization technique, wherein the latent feature distribution is sampled from a normal distribution;

[0091] The common information features are concatenated with the potential feature distribution to obtain heterogeneous graph attribute representations of all nodes in the first heterogeneous graph and the second heterogeneous graph.

[0092] The following combination Figure 2 A specific example is used to explain in detail the process of learning node feature distribution using the multi-view variational graph autoencoder method. Figure 2 Part (b) of

[0093] Step S201: Setting model parameters of the first isograph and the second isograph, including the drug-microbe isograph The adjacency matrix Node attribute matrix Attribute embedding for local neighborhood enhancement Independent first graph convolutional network and the second graph convolutional network And parameter-sharing graph convolutional network GCN co ;in, fu and he represent two different types of drug-drug similarities, respectively.

[0094] Step S202: Initialize parameters, including the weight matrices of the independent first and second graph convolutional networks The weight matrix W of the parameter-sharing graph convolutional network co .

[0095] Step S203: embed the attributes of local neighbor enhancement into Y fu ,Y he Graph convolution GCN as parameter sharing co Input, get the common information feature graph C containing drug and microbial nodes fu ,C he .

[0096] Step S204: Feature map C he and the node attribute matrix F fu Spread and replenish G fu Perspective, forming an attribute matrix

[0097] Step S205: The feature map C fu and the node attribute matrix F he Spread and replenish G he Perspective, forming an attribute matrix

[0098] Step S206: Using graph convolution and calculate The mean and variance And calculate the characteristic distribution of nodes based on reparameterization technology in, From the normal distribution Medium sampling.

[0099] Step S207: Splicing feature maps and node feature distribution Get all node attribute representations

[0100] like Figure 3 As shown, Figure 3The data processing flowchart of the multi-view variational graph autoencoder is shown. Through the above steps, the information in the two heterogeneous graphs is fused by using the shared graph convolutional network, so that each graph can obtain supplementary information from the other graph, and the understanding of the node features by the model is enhanced. In addition, by combining the output of the graph convolutional network with the variational autoencoder, the features of the nodes are extracted by the graph convolutional network, and the features of the nodes are mapped to the mean and variance of the latent space through the fully connected layer, and the parameters of the latent space distribution are learned (similar to the encoder in VAE, except that here the distribution parameters of the latent space are generated from the node features output by the GCN), which can learn the low-dimensional representation of the nodes to capture the latent features of the nodes, which is helpful for subsequent node classification and link prediction tasks. By propagating the information in one view to another view, the lack of information in different views can be supplemented, and the understanding ability of the model for complex relationships can be improved. The reparameterization trick enables the model to be optimized by stochastic gradient descent, improving the efficiency of training.

[0101] In one specific embodiment, the specific process of propagating and supplementing the feature graph C he (C fu ) and the attribute matrix F fu (F he ) of the node to the G fu (G he ) view in steps S204 and S205 is described as follows:

[0102]

[0103] Where P he ,P fu are learnable parameter matrices, and W sf ,W sh are weight matrices.

[0104] As can be seen, the parameter matrices P he ,P fu are learnable, and the model can automatically adjust these parameters to optimize performance. By combining features and attributes from different heterogeneous graphs, the expressiveness of node features can be enhanced, enabling the model to capture more complex patterns and relationships. Through cross-graph attribute propagation, the model can better generalize to unseen data and be suitable for complex tasks that require the integration of multiple types of information.

[0105] As a preferred embodiment, in step S103, a drug-microorganism supergraph is generated according to the drug-microorganism association database, and a cross-attention strategy is used to construct a dynamic supergraph topology based on the joint superedge attribute matrix and the joint node attribute matrix of the supergraph, including:

[0106] According to the learnable Gaussian distribution, the superedge attribute matrix of the drug-microorganism supergraph is sampled respectively, and is spliced to form a joint superedge attribute matrix; meanwhile, the node attribute matrix of the supergraph is spliced to form a joint node attribute matrix;

[0107] All attributes of nodes in the supergraph are aggregated by a cross-attention strategy to generate enhanced superedge attributes;

[0108] According to the enhanced superedge attributes and the joint node attribute matrix, a dynamic supergraph topology is constructed.

[0109] As a preferred embodiment, aggregating all attributes of nodes in the supergraph by a cross-attention strategy to generate enhanced superedge attributes comprises:

[0110] Linearly transforming the joint superedge attribute matrix to obtain a query matrix;

[0111] Linearly transforming the joint node attribute matrix to obtain a key matrix and a value matrix respectively;

[0112] According to the query matrix, the transpose of the key matrix and the value matrix, enhanced superedge attributes are obtained.

[0113] As a specific embodiment, the above process is described in combination with the (c) part of Figure 2 Superedge attribute learning comprises the following steps:

[0114] Step S301: inputting a joint superedge attribute matrix A joint node attribute matrix F mn ;

[0115] Step S302: initializing parameters;

[0116] Step S303: linearly transforming the joint superedge attribute matrix to obtain a query matrix Q;

[0117] Step S304: linearly transforming the joint node attribute matrix F mn to obtain a key matrix K and a value matrix V respectively;

[0118] Step S305: according to Q, K T and V, enhanced superedge attributes

[0119] Step S306: according to the product of the enhanced superedge attributes and the transpose of the joint node attribute matrix F mn , a dynamic supergraph topology is formed.

[0120] As shown in Figure 4 , Figure 4A dynamic hyperedge attribute learning diagram is shown. Through the above steps, the attribute matrices of all hyperedges are spliced to form a joint hyperedge attribute matrix. This helps to uniformly process the attribute information of all hyperedges in the model, thereby improving the efficiency and performance of the model. Similarly, the attribute matrices of the nodes can also be spliced to form a joint node attribute matrix, so that the model is more concentrated and efficient in processing node information. The cross-attention mechanism allows the model to interact information between different inputs, which can enhance the aggregation ability of the model to node attributes, thereby improving the expression ability of the hyperedge attributes. Based on the enhanced hyperedge attributes and the joint node attribute matrix, a dynamic hypergraph topology can be constructed, which can be dynamically adjusted according to the attributes and relationships of the data, thereby better reflecting the internal structure and relationship of the data.

[0121] As a specific embodiment, in step S305, the enhanced hyperedge attribute T and V are calculated according to Q, K The specific method is as follows:

[0122]

[0123] Where N is the total number of nodes.

[0124] Since two nodes with similar attribute distribution are more likely to belong to the same hyperedge, the hyperedge attribute is learned in combination with the attribute distribution of each node.

[0125] As a preferred embodiment, in step S104, a hypergraph Laplacian operator is introduced into the dynamic hypergraph topology to form a multi-order neighbor topology embedding, a multi-scale node attribute matrix is adaptively fused according to the neighbor topology embedding, and the multi-scale node attribute matrix is supplemented in the hypergraph convolution layer to obtain the dynamic hypergraph attribute representation of the node, including:

[0126] The dynamic hypergraph topology is standardized according to the hypergraph Laplacian operator to form a multi-order neighbor topology embedding;

[0127] The multi-order neighbor topology embedding is used to generate a topology attribute matrix of multi-order neighbor nodes;

[0128] All scale topology attribute matrices are adaptively fused to form a multi-scale node attribute matrix under a multi-scale topology perspective;

[0129] For any hypergraph convolution layer in the hypergraph, the topology attribute matrix of the current convolution layer and the multi-scale node attribute representation of the previous convolution layer are spliced until the splicing of all hypergraph convolution layers in the hypergraph is completed, and the dynamic hypergraph attribute representation of all nodes is obtained.

[0130] The specific processing process of the above method will be described below through a specific embodiment.

[0131] In actual processing, see Figure 2 Part (d) of the process consists of the following steps:

[0132] Step S401: Input enhanced hyperedge attributes Joint node attribute matrix F mn , dynamic hypergraph topology H dy and the hypergraph Laplacian θ;

[0133] Step S402: Initialize parameters, including the number of hypergraph convolution layers T and the order of neighbor nodes s;

[0134] Step S403: The input dynamic hypergraph topology H is calculated based on the hypergraph Laplacian operator θ. dy Normalize and form an s-order neighbor topology embedding A s ;

[0135] Step S404: In the first hypergraph convolution layer, the attributes of the hyperedges are aggregated based on the hypergraph topology to generate the attribute matrix of the first-order neighbor nodes.

[0136] Step S405: Embed A using the s-order neighbor topology s Generate the attribute matrix of the s-order neighbor node as

[0137] Step S406: Adaptively fuse the neighbor topology attribute matrices of all scales to form a node attribute matrix K under the multi-scale topology perspective ms ;

[0138] Step S407: In the lth hypergraph convolutional layer, concatenate the node attribute matrix K ms and the node attribute representation obtained by the l-1th graph convolutional layer;

[0139] Step S408: End the loop after T times and get the node attribute representation K jo .

[0140] like Figure 5 As shown, Figure 5 A schematic diagram of the multi-scale hypergraph convolution process is presented. This method utilizes a multi-scale hypergraph convolution framework to generate a topological embedding of drug-microbe node neighbors across multiple scales. It also adaptively integrates information from neighbors at different scales, overcoming the limitation of traditional hypergraph convolution, which only targets a single neighbor topology, and can encode connections between multiple drug-microbe nodes.

[0141] Furthermore, in step S405, the hypergraph Laplacian operator is introduced to form an s-order neighbor topology embedding, which is generated as follows:

[0142]

[0143] Among them, W e is the weight matrix, and the node degree matrix (D v ) ii =∑ j (H dy ) ij , hyperedge degree matrix (D e ) jj =∑ i (H dy ) ij .

[0144] By generating s-order neighbor topological embeddings, it helps to form topological embeddings for each drug and microorganism node in the hypergraph, and these embeddings cover the neighbor regions with multiple ranges.

[0145] As a preferred embodiment, in step S105, graph comparison learning is performed on any two graphs among the first heterogeneous graph, the second heterogeneous graph, and the hypergraph from a multi-scale topological perspective according to the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation, and a drug-microorganism association prediction model is trained, including:

[0146] A first loss function of the first heterogeneous graph and the dynamic hypergraph, and a second loss function of the second heterogeneous graph and the dynamic hypergraph are calculated based on the heterogeneous graph attribute representation of the node and the dynamic hypergraph attribute representation of the node;

[0147] Calculating a third loss function based on the heterogeneous graph attribute representation of the nodes in the first heterogeneous graph and the attribute representation of the nodes in the second heterogeneous graph;

[0148] A joint contrast loss function is calculated according to the first loss function, the second loss function and the third loss function, and a drug-microorganism association prediction model is obtained according to the joint contrast loss function.

[0149] By comparing multiple graphs, the model can comprehensively consider different graph structures and attribute representations and enhance the consistency of node representation, thereby more effectively integrating feature information from different graph structures and more accurately predicting the association between drugs and microorganisms.

[0150] As a specific embodiment, the specific processing steps of multi-image comparison include:

[0151] Step S501: Input the heterogeneous graph G fu ,G he Learn to get node attribute representation K fu and K he , from the hypergraph The learned node attribute matrix K jo, loss function L fc ,L hc ,L fh ;

[0152] Step S502: Initialize parameters, including temperature parameter τ;

[0153] Step S503: For heterogeneous graph G fu (G he ) and hypergraph Perform graph comparison learning: Represent K according to node attributes fu (K he ) and node attribute matrix K jo Calculate the loss function L fc (L hc );

[0154] Step S504: For heterogeneous graph G fu and heterogeneous graph G he Perform graph comparison learning: Represent K according to node attributes fu and K he Calculate the loss function L fh ;

[0155] Step S505: Calculate the joint contrast loss function.

[0156] like Figure 6 As shown, Figure 6 A schematic diagram of the multi-graph comparison process is shown. Learning through multi-graph comparison can maintain consistency in node representation across heterogeneous graphs and hypergraphs from different graph perspectives, effectively improving the model's ability to predict drug-microbe associations.

[0157] As a specific example, in order to improve the model's ability to predict drug-microorganism associations and maintain consistency in node representation between heterogeneous graphs and hypergraphs from different graph perspectives, the loss function L fc ,L hc ,L fh The calculation method is as follows:

[0158]

[0159] Where sim(·) is the cosine similarity and τ is the temperature parameter.

[0160] It should be noted that in step S106, when the drug-microorganism association prediction model provided by the present invention is used to calculate the association between the drug and the microorganism, since it is oriented to the heterogeneous graph G fu , G he and Hypergraph The corresponding learned attribute matrix of all drugs and microorganisms is K β The i-th row and the N-th rowd + jth column is the attribute vector of drug node d i and microbe node m j and d i and m j β (β∈{fu,he,jo}) are formed,

[0161]

[0162] B fu , B he and B jo are stacked at channel level to form

[0163] B = [B fu ; B he ; B jo ],

[0164] where “;” denotes the stacking operation. To preserve the marginal information of B, we perform zero-padding operation in each convolutional layer, denoted as After two convolutional and pooling layers, we get d i -m j attribute representation J. J is flattened into a vector denoted as J f and fed into a fully connected layer to get d i -m j correlation score,

[0165] η = softmax(W f J f + b f ),

[0166] where W f and b f are weight matrix and bias vector respectively. η = [(η)0, (η)1], (η)0and (η)1represent the probability of drug d i and microbe m j being irrelevant or relevant respectively.

[0167] The cross-entropy between the true labels of drug-microbe node pairs and their predicted correlation scores is computed as the loss function of the model, i.e.,

[0168]

[0169] where N t is the number of training sample set and C = 2. If there is a true correlation between drug and microbe, y j = 1, otherwise 0. ​​

[0170] To validate the practical effect of the present application, the prediction performance of the present method and other comparative methods is evaluated by five-fold cross-validation. The positive samples consist of all known drug-microbe associations, and they are randomly divided into 5 subsets of equal size. All drug-microbe node pairs that have not been found to have an association are considered negative samples. In each fold, we select 4 positive sample subsets and an equal amount of negative samples randomly as the training set, and the remaining positive sample subset and all remaining negative samples are used to test the model.

[0171] The evaluation indicators include the area under the receiver operating characteristic curve (ROC) (AUC), the area under the precision-recall (PR) curve (AUPR), and the recall rate of the top-k drug-microbe candidates. We calculate the AUC and AUPR of each fold respectively when cross-validation, and take their average values as the final prediction performance of the model.

[0172] The other 8 models used for comparison are DHDMP, MFLP, NGMDA, SCSMDA, GACNNMDA, GSAMDA, EGATMDA and GCNMDA. Among them, the number of encoding layers P in hypergraph convolution is 2, and the order s of the neighbor node is 2. The temperature parameter τ in multiple graph contrastive learning is 0.1. The AUC and AUPR of the present method and the other eight models for drug-microbe association prediction are shown in Table 1.

[0173] Table 1 Comparison of association prediction results of each model

[0174]

[0175] From Table 1, it can be found that the present method has the highest average AUC of 96.3%, which is 0.4% higher than DHDMP, 1.7% higher than MFLP, 1.9% higher than NGMDA, 4.7% higher than SCSMDA, 12.0% higher than GACNNMDA, 6.1% higher than GSAMDA, 2.3% higher than EGATMDA, and 6.0% higher than GCNMDA. The average AUPR of MMDMA is 88.0%, which is 5.7%, 11.3%, 15.2%, 54.0%, 68.4%, 63.3%, 57.3% and 56.5% higher than DHDMP, MFLP, NGMDA, SCSMDA, CACNNMDA, GSAMDA, EGATMDA and GCNMDA, respectively.

[0176] The performance of GACNNMDA and GSAMDA is poor, which may be due to the fact that they only learn the node features based on GAT without considering the feature distribution of the nodes in the heterogeneous graph. GCNMDA and EGATMDA respectively aggregate the neighbor node information of the target node based on graph convolution and node-level attention, and they both achieve good performance. This shows that it is necessary to aggregate the neighbor node information. SCSMDA only focuses on the learning of meta-path node features, ignoring the learning of node features in the entire heterogeneous graph. NGMDA achieves better prediction performance than SCSMDA, which learns the features of drug and microbial nodes from the perspective of neighborhood and the entire heterogeneous graph respectively. MFLP mainly focuses on learning the hierarchical features of multiple classes of nodes in hyperbolic space. However, these methods ignore the close relationship between multiple drug and microbial nodes. DHDMP predicts the association between drugs and microorganisms based on class-sensitive hypergraph convolution, but ignores the information propagation from nodes to hyperedges. The main reason for the best performance of our method is that it learns the relationship between multiple nodes and the distribution of node features.

[0177] The embodiment of the application further provides a drug-microorganism association prediction system based on a multi-view variational graph autoencoder and a multi-scale hypergraph convolution, comprising:

[0178] A data acquisition module is configured to acquire a drug-microorganism association database.

[0179] A multi-view variational graph autoencoder module is configured to construct a first heterogeneous graph based on drug function similarity and a second heterogeneous graph based on drug attribute similarity containing drug and microbial nodes according to the drug-microorganism association database, establish a locally neighbor-enhanced attribute embedding, and obtain a heterogeneous graph attribute representation of a node based on the locally neighbor-enhanced attribute embedding and a multi-view variational graph autoencoder.

[0180] A dynamic hypergraph module is configured to generate a drug-microorganism hypergraph according to the drug-microorganism association database, adopt a cross-attention strategy, and construct a dynamic hypergraph topology based on a joint hyperedge attribute matrix and a joint node attribute matrix of the hypergraph.

[0181] A multi-scale convolution module is configured to introduce a hypergraph Laplacian operator in the dynamic hypergraph topology, form a multi-order neighbor topology embedding, adaptively fuse to obtain a multi-scale node attribute, supplement the multi-scale node attribute in a hypergraph convolution layer, and obtain a dynamic hypergraph attribute representation of a node.

[0182] A contrastive learning module is configured to perform graph contrastive learning on any two graphs in the first heterogeneous graph, the second heterogeneous graph, and the multi-scale topology perspective hypergraph based on the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation, and train to obtain a drug-microorganism association prediction model.

[0183] The association prediction module is configured to acquire feature vectors of the drugs and the microorganisms in real time, and analyze the association between the drugs and the microorganisms by using the drug-microorganism association prediction model.

[0184] To sum up, the drug-microorganism association prediction method based on the multi-view variational graph autoencoder and the multi-scale hypergraph convolution provided by the embodiment can learn the multi-view node feature distribution, the multi-view drug and microorganism node feature distribution, the multi-scale neighbor topology, and dynamically encode the multiple node interconnections from the hypergraph, so as to enhance the consistency of the node representation and improve the performance of the drug-related candidate task.

[0185] The above merely describes the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, characterized by: include: Access to drug-microbe association databases; Based on the drug-microbe association database, a first heterogeneous graph based on drug function similarity and a second heterogeneous graph based on drug attribute similarity, comprising drug and microbe nodes, are constructed, a local neighbor-enhanced attribute embedding is established, and based on the local neighbor-enhanced attribute embedding and a multi-view variational graph autoencoder, a heterogeneous graph attribute representation of the node is obtained, including: using the local neighbor-enhanced attribute embedding as input to a shared parameter graph convolutional coding layer, obtaining common information based on the shared parameter graph convolutional coding layer, propagating and supplementing the common information into the first heterogeneous graph and the second heterogeneous graph to form an attribute matrix; using a variational graph autoencoder to model node feature distribution according to the attribute matrix, and calculating the heterogeneous graph node feature distribution through a reparameterization technique; Among them, the shared information is obtained according to the shared parameter graph convolution coding layer, and the shared information is propagated and supplemented to the first heterogeneous graph and the second heterogeneous graph to form an attribute matrix, which specifically includes: encoding the topology and node attributes shared by the first heterogeneous graph and the second heterogeneous graph through the shared parameter graph convolution network to obtain a shared information feature graph; propagating and supplementing the shared information feature graph obtained from the drug function perspective of the first heterogeneous graph to the drug attribute perspective of the second heterogeneous graph to form an attribute matrix of the nodes of the second heterogeneous graph; propagating and supplementing the shared information feature graph obtained from the drug attribute perspective of the second heterogeneous graph to the drug function perspective of the first heterogeneous graph to form an attribute matrix of the nodes of the first heterogeneous graph; A variational graph autoencoder is used to model the node feature distribution according to the attribute matrix, and the heterogeneous graph node feature distribution is calculated by reparameterization techniques, specifically comprising: using a first graph convolutional network and a second graph convolutional network to respectively calculate the mean and variance of the attribute matrix of the first heterogeneous graph and the second heterogeneous graph; calculating the potential feature distribution of the node based on the reparameterization technique, wherein the potential feature distribution is sampled from a normal distribution; splicing the common information feature with the potential feature distribution to obtain heterogeneous graph attribute representations of all nodes in the first heterogeneous graph and the second heterogeneous graph; generating a drug-microbe hypergraph according to the drug-microbe association database, and constructing a dynamic hypergraph topology based on a joint hyperedge attribute matrix and a joint node attribute matrix of the hypergraph using a cross-attention strategy; Introducing a hypergraph Laplacian operator into the dynamic hypergraph topology to form a multi-order neighbor topology embedding, adaptively fusing the neighbor topology embedding to obtain multi-scale node attributes, and supplementing the multi-scale node attributes in the hypergraph convolution layer to obtain a dynamic hypergraph attribute representation of the node; Performing graph comparison learning on any two graphs among the first heterogeneous graph, the second heterogeneous graph, and the multi-scale topological perspective hypergraph according to the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation of the nodes, and training to obtain a drug-microorganism association prediction model; The characteristic vectors of the drug and the microorganism are acquired in real time, and the correlation between the drug and the microorganism is analyzed using the drug-microorganism correlation prediction model.

2. The drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution according to claim 1 is characterized in that: The method for establishing attribute embedding for local neighborhood enhancement includes: In each of the first and second heterogeneous graphs, a drug or a microorganism is used as a target node, and several neighboring nodes with the most similar attributes to the target node are screened; wherein, in the first heterogeneous graph, cosine similarity is used to calculate the similarity of the node attributes, and in the second heterogeneous graph, heat kernel similarity is used to calculate the similarity of the node attributes; Concatenate the attribute vector of the target node with the attribute vector of the neighboring node to obtain the neighboring node attribute matrix of the target node; The convolution filter is used to perform weighted aggregation on the attribute matrix of the neighbor nodes, and the original attribute information is retained through residual connection to obtain the attribute embedding of local neighbor enhancement.

3. The drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution according to claim 1 is characterized in that: The method generates a drug-microorganism hypergraph according to the drug-microorganism association database, adopts a cross-attention strategy, and constructs a dynamic hypergraph topology based on a joint hyperedge attribute matrix and a joint node attribute matrix of the hypergraph, including: The hyperedge attribute matrices of the drug-microorganism hypergraph are sampled respectively according to a learnable Gaussian distribution, and are concatenated to form a joint hyperedge attribute matrix; at the same time, the node attribute matrices of the hypergraph are concatenated to form a joint node attribute matrix; Aggregating all attributes of nodes in the hypergraph through a cross-attention strategy to generate enhanced hyperedge attributes; A dynamic hypergraph topology is constructed according to the enhanced hyperedge attributes and the joint node attribute matrix.

4. The drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution according to claim 3 is characterized in that: All attributes of nodes in the hypergraph are aggregated through a cross-attention strategy to generate enhanced hyperedge attributes, including: Linearly transforming the joint hyperedge attribute matrix to obtain a query matrix; Linearly transforming the joint node attribute matrix to obtain a key matrix and a value matrix respectively; Enhanced hyperedge properties are obtained according to the query matrix, the transpose of the key matrix and the value matrix.

5. The drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution according to claim 1 is characterized in that: A hypergraph Laplacian operator is introduced into the dynamic hypergraph topology to form a multi-order neighbor topology embedding, multi-scale node attributes are obtained by adaptively fusing the neighbor topology embedding, and the multi-scale node attributes are supplemented in the hypergraph convolution layer to obtain a dynamic hypergraph attribute representation of the node, including: Normalizing the dynamic hypergraph topology according to the hypergraph Laplacian operator to form a multi-order neighbor topology embedding; Generate a topological attribute matrix of multi-order neighbor nodes using the multi-order neighbor topology embedding; Adaptively fuse the topological attribute matrices of all scales to form a multi-scale node attribute matrix from a multi-scale topological perspective; For any hypergraph convolutional layer in the hypergraph, the topological attribute matrix of the current convolutional layer and the multi-scale node attribute representation of the previous convolutional layer are spliced ​​together until all hypergraph convolutional layers in the hypergraph are spliced ​​together to obtain the dynamic hypergraph attribute representation of all nodes.

6. The drug-microbe association prediction method based on multi-view variational graph autoencoder and multi-scale hypergraph convolution according to claim 1 is characterized in that: Performing graph comparison learning on any two graphs among the first heterogeneous graph, the second heterogeneous graph, and the multi-scale topological perspective hypergraph according to the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation, and training a drug-microorganism association prediction model, including: A first loss function of the first heterogeneous graph and the dynamic hypergraph, and a second loss function of the second heterogeneous graph and the dynamic hypergraph are calculated based on the heterogeneous graph attribute representation of the node and the dynamic hypergraph attribute representation of the node; Calculating a third loss function based on the heterogeneous graph attribute representation of the nodes in the first heterogeneous graph and the attribute representation of the nodes in the second heterogeneous graph; A joint contrast loss function is calculated according to the first loss function, the second loss function and the third loss function, and a drug-microorganism association prediction model is obtained according to the joint contrast loss function.

7. A drug-microbe association prediction system based on multi-view variational graph autoencoder and multi-scale hypergraph convolution, characterized by: include: Data acquisition module, used to obtain drug-microorganism association database; A multi-view variational graph autoencoder module is used to construct, based on the drug-microbe association database, a first heterogeneous graph based on drug function similarity and a second heterogeneous graph based on drug attribute similarity, comprising drug and microbe nodes, establish a local neighbor-enhanced attribute embedding, and obtain a heterogeneous graph attribute representation of the node based on the local neighbor-enhanced attribute embedding and the multi-view variational graph autoencoder, including: using the local neighbor-enhanced attribute embedding as input to a shared parameter graph convolutional coding layer, obtaining common information based on the shared parameter graph convolutional coding layer, propagating and supplementing the common information into the first heterogeneous graph and the second heterogeneous graph to form an attribute matrix; using a variational graph autoencoder to model the node feature distribution based on the attribute matrix, and calculating the heterogeneous graph node feature distribution through a reparameterization technique; Among them, the shared information is obtained according to the shared parameter graph convolution coding layer, and the shared information is propagated and supplemented to the first heterogeneous graph and the second heterogeneous graph to form an attribute matrix, which specifically includes: encoding the topology and node attributes shared by the first heterogeneous graph and the second heterogeneous graph through the shared parameter graph convolution network to obtain a shared information feature graph; propagating and supplementing the shared information feature graph obtained from the drug function perspective of the first heterogeneous graph to the drug attribute perspective of the second heterogeneous graph to form an attribute matrix of the nodes of the second heterogeneous graph; propagating and supplementing the shared information feature graph obtained from the drug attribute perspective of the second heterogeneous graph to the drug function perspective of the first heterogeneous graph to form an attribute matrix of the nodes of the first heterogeneous graph; A variational graph autoencoder is used to model the node feature distribution according to the attribute matrix, and the heterogeneous graph node feature distribution is calculated by reparameterization techniques, specifically comprising: using a first graph convolutional network and a second graph convolutional network to respectively calculate the mean and variance of the attribute matrix of the first heterogeneous graph and the second heterogeneous graph; calculating the potential feature distribution of the node based on the reparameterization technique, wherein the potential feature distribution is sampled from a normal distribution; splicing the common information feature with the potential feature distribution to obtain heterogeneous graph attribute representations of all nodes in the first heterogeneous graph and the second heterogeneous graph; A dynamic hypergraph module is used to generate a drug-microbe hypergraph based on the drug-microbe association database, and adopt a cross-attention strategy to construct a dynamic hypergraph topology based on the joint hyperedge attribute matrix and the joint node attribute matrix of the hypergraph; A multi-scale convolution module is used to introduce a hypergraph Laplacian operator into the dynamic hypergraph topology to form a multi-order neighbor topology embedding, adaptively fuse the neighbor topology embedding to obtain multi-scale node attributes, and supplement the multi-scale node attributes in the hypergraph convolution layer to obtain a dynamic hypergraph attribute representation of the node; a comparative learning module, configured to perform graph comparative learning on any two graphs among the first heterogeneous graph, the second heterogeneous graph, and the multi-scale topological perspective hypergraph based on the heterogeneous graph attribute representation and the dynamic hypergraph attribute representation, and train a drug-microorganism association prediction model; The association prediction module is used to obtain the characteristic vectors of drugs and microorganisms in real time, and analyze the association between the drugs and microorganisms through the drug-microorganism association prediction model.

Images