A method for discovering disease-symptom associations using disease knowledge graphs

By decomposing the disease knowledge graph into multiple binary networks and fusion of similarity matrix, using graph autoencoder and non-negative matrix decomposition, the problem of professional knowledge required for the mining of the correlation between diseases and symptoms in the prior art is solved, and more accurate and reliable disease-symptom association prediction.

CN117747125BActive Publication Date: 2024-11-15CHONGQING UNIV OF POSTS & TELECOMM
View PDF 8 Cites 0 Cited by

Patent Information

Application Number
CN202311778674.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-12-22
Publication Date
2024-11-15
Estimated Expiration
2043-12-22

AI Technical Summary

Technical Problem

The prior art requires the use of professional medical knowledge and similarity network calculation methods in the mining of the correlation between diseases and symptoms, and the calculation methods are relatively single, and there is a lack of multi-view information and comprehensive similarity network calculation methods.

Method used

By decomposing the disease knowledge graph into multiple binary networks, the similarity matrix of diseases and symptoms is calculated, and nonlinear representations are extracted using graph autoencoder and non-negative matrix decomposition, combining with fully connected neural networks for correlation prediction, using SimRank and DeepWalk The algorithm performs similarity matrix fusion and proposes a multi-view data and nonlinear similarity network fusion method.

Benefits of technology

Without professional medical knowledge, it improves the performance of disease-symptom association prediction, obtains a more comprehensive and reliable similarity network, and enhances the accuracy of prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117747125B_ABST
    Figure CN117747125B_ABST
Patent Text Reader

Abstract

The present invention relates to the fields of medical care and data mining technology, and in particular to a method for discovering disease-symptom associations using a disease knowledge graph, comprising decomposing the disease knowledge graph to obtain a bipartite network related to diseases and symptoms, and saving the network in the form of an adjacency matrix; calculating similarity matrices of disease nodes and symptom nodes based on the bipartite network, and fusing the similarity matrices into a comprehensive disease similarity matrix and a comprehensive symptom similarity matrix; constructing a graph autoencoder to extract nonlinear representations of nodes; using non-negative matrix decomposition to extract linear representations of nodes; combining the linear representations with the nonlinear representations, and predicting disease-symptom associations through a fully connected neural network; and proposing a more comprehensive and effective similarity network calculation method in the similarity network calculation stage, combining graph structure context information with node sequence information, and obtaining a more reliable similarity network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of medical treatment and data mining, and in particular to a method for discovering disease-symptom associations using a disease knowledge graph. Background Art

[0002] Mining the relationship between diseases and symptoms helps provide a more accurate basis for the diagnosis and treatment of diseases. At present, the relationship between diseases and symptoms mainly comes from the experience of doctors or knowledge in medical books. In the computer field, the past research on biomedical bipartite network association mining has mainly focused on genetics, proteomics and other fields, and the mining of disease-symptom associations is mainly based on simple statistical analysis methods.

[0003] Existing association mining of biomedical bipartite networks first constructs a bipartite graph with nodes representing the two types of entities whose relationships are to be mined, such as disease and symptom. Graph embedding technology is then used to characterize the nodes, and finally the prediction scores between all entity pairs are calculated.

[0004] However, existing methods usually need to rely on professional medical knowledge to assist in association mining. This knowledge is usually presented in the form of similarity networks, such as the functional similarity of MiRNA, the semantic similarity of diseases, and the structural similarity of drugs. At the same time, most existing methods use Gaussian kernel similarity to calculate similarity networks, and the calculation method is relatively simple. Summary of the invention

[0005] In view of the problem that the existing technology usually requires the use of professional medical knowledge and a single similarity network calculation method, the present invention proposes a method for discovering disease-symptom associations using a disease knowledge graph, which specifically includes the following steps:

[0006] S100, decomposing the disease knowledge graph into a plurality of bipartite networks including disease nodes and / or symptom nodes, and saving the bipartite networks in the form of an adjacency matrix;

[0007] S200, calculating similarity matrices of disease nodes and symptom nodes according to all obtained bipartite networks, and merging them into a comprehensive disease similarity matrix and a comprehensive symptom similarity matrix;

[0008] S300, constructing a graph autoencoder, using the graph autoencoder to extract a nonlinear representation of the disease node based on a comprehensive disease similarity matrix and a vector representation of a disease node, and extracting a nonlinear representation of the disease node based on a comprehensive symptom similarity matrix and a vector representation of a symptom node;

[0009] S400, extracting linear representations of diseases and symptoms using non-negative matrix decomposition, that is, decomposing the disease-symptom association matrix into a linear representation matrix U of diseases and a linear representation matrix V of symptoms, wherein U and V are low-rank non-negative matrices;

[0010] S500, combining linear representation and nonlinear representation to predict disease-symptom association through a fully connected neural network.

[0011] Furthermore, the disease knowledge graph is decomposed into multiple bipartite networks containing disease nodes, and the process of obtaining multi-view data includes:

[0012] S110, extracting disease nodes and other nodes from the disease knowledge graph to form a bipartite network of the disease node and any other node, and all bipartite networks constitute a disease-related bipartite network;

[0013] S120, extracting symptom nodes and other nodes from the disease knowledge graph to form a bipartite network of the symptom node and any other node, and all bipartite networks constitute a symptom-related bipartite network;

[0014] S130, cyclically deleting nodes of one type whose node degree is less than p and nodes of another type whose node degree is less than q in the bipartite network, to obtain a simplified bipartite network;

[0015] S140, saving the simplified bipartite network in the form of an adjacency matrix.

[0016] Furthermore, the process of calculating the similarity matrix of the disease nodes and the symptom nodes according to all the obtained bipartite networks and fusing them into a comprehensive disease similarity matrix and a comprehensive symptom similarity matrix includes the following steps:

[0017] S210, in the set of disease-related bipartite networks, respectively calculating the SimRank similarity of each disease for each bipartite network based on the SimRank algorithm to obtain a SimRank similarity matrix for each disease;

[0018] S220, in the set of bipartite networks related to the symptoms, calculating the SimRank similarity of each symptom for each bipartite network based on the SimRank algorithm to obtain a SimRank similarity matrix for each symptom;

[0019] S230, in the set of symptom-related bipartite networks, obtain the feature representation of each symptom based on the DeepWalk algorithm for each bipartite network, calculate the cosine similarity between two symptom feature representations, and construct a DeepWalk similarity matrix of the symptoms;

[0020] S240, in the set of disease-related bipartite networks, obtaining a feature representation of each disease based on the DeepWalk algorithm for each bipartite network, calculating the cosine similarity between two disease feature representations, and constructing a DeepWalk similarity matrix of the disease;

[0021] S250, based on the nonlinear similarity network fusion method SNF, the SimRank similarity matrix of the disease and the DeepWalk similarity matrix are fused to obtain a comprehensive disease similarity matrix;

[0022] S260. Based on the nonlinear similarity network fusion method SNF, the SimRank similarity matrix of symptoms and the DeepWalk similarity matrix are fused to obtain a comprehensive symptom similarity matrix.

[0023] Furthermore, the SimRank similarity calculation between two nodes includes:

[0024]

[0025] Among them, s k+1 (a, b) represents the SimRank similarity between nodes a and b in the k+1th iteration, and C is the damping coefficient; I(a) represents the set of other types of nodes in the bipartite network connected to node a, that is, if node a is a disease node, then I(a) represents the set of all other nodes connected to node a; s k (I i (a),I j (b)) is the i-th node I in the node set I(a) i (a) and the jth node I in the node set I (b) j (b) The similarity between them; |I(a)| represents the number of nodes in set I(a).

[0026] Furthermore, the element in the i-th row and j-th column of the disease SimRank similarity matrix represents the SimRank similarity between disease node i and disease node j. The SimRank similarity matrix of the disease at the k-th iteration is expressed as:

[0027]

[0028] Among them, S k represents the SimRank similarity matrix obtained at the kth iteration; P is the transition probability matrix, which is obtained by normalizing the rows of the bipartite network in the present invention; I n is an n×n unit matrix, where n is the number of nodes; diag(M) means taking the matrix M to obtain a diagonal vector, and Diag(V) means converting the vector V into a diagonal matrix.

[0029] Furthermore, based on the nonlinear similarity network fusion method, the process of fusing the SimRank similarity matrix and the DeepWalk similarity matrix includes:

[0030] The SimRank similarity matrix and the DeepWalk similarity matrix are normalized respectively, and the local affinity of each similarity matrix is ​​calculated based on the K-nearest neighbor algorithm;

[0031] The normalized similarity matrix is ​​iteratively updated using the local affinity, and the data after each iterative update is normalized. When the iterative result converges, the iteration ends. Otherwise, the local affinity of each similarity matrix is ​​continued to be calculated and iteratively updated according to the local affinity.

[0032] The mean matrix of the converged similarity matrix is ​​calculated, the mean matrix is ​​converted into a symmetric matrix, and the symmetric matrix is ​​used as the comprehensive similarity matrix.

[0033] Furthermore, the process of normalizing the similarity matrix includes:

[0034]

[0035] The calculation process of local affinity includes:

[0036]

[0037] Among them, S 1 ′(i,j) represents the pair S 1 (i, j) is the similarity between node i and node j after normalization; S 1 (i, j) represents the similarity between node i and node j in a similarity matrix; represents the local affinity between node i and node j; N i is the set of k nearest neighbor nodes of node i.

[0038] Furthermore, the process of iteratively updating the similarity matrix according to the local affinity includes:

[0039]

[0040] in, represents the state matrix of the vth similarity matrix after t iterations, represents the local affinity kernel of the vth similarity matrix, and m is the number of similarity matrices.

[0041] Furthermore, the graph autoencoder is used to obtain nonlinear representations of diseases and symptoms. The training process of the graph autoencoder includes:

[0042] The comprehensive similarity matrix and feature matrix of a node are taken as the output of the encoder, and the encoder obtains the potential representation of the node;

[0043] The decoder reconstructs the comprehensive similarity matrix of the nodes according to the potential representation, and calculates the cross entropy of the comprehensive similarity matrix of the nodes based on the reconstruction result of the encoder as the loss function to train the encoder and decoder.

[0044] Furthermore, non-negative matrix factorization (NMF) is used to extract nonlinear representations of diseases and symptoms, that is, the disease-symptom association relationship is projected into the disease subspace and the symptom subspace. The disease-symptom association matrix A is decomposed into a low-rank non-negative matrix U and a low-rank non-negative matrix V as nonlinear representations of diseases and symptoms respectively. The specific decomposition process includes:

[0045] Define the indicator weighting matrix W = w ij ∈R m×n , the value of W is the same as A, for the i-th row and j-th column u in the low-rank non-negative matrix U ik and the kth row and jth column of the low-rank non-negative matrix V kj The update is performed, and when the update result converges, the currently updated matrix is ​​output as a nonlinear representation of the disease and symptoms. The update process includes:

[0046]

[0047]

[0048] in, represents the value of the i-th row and j-th column in the t-th updated low-rank non-negative matrix U; 1 and λ 2 is the regularization coefficient, ⊙ is the Hadamard product; (·) ik Represents the value of the i-th row and j-th column.

[0049] On the one hand, the present invention utilizes the multi-view information in the disease knowledge graph to improve the prediction performance, that is, decomposing the disease knowledge graph into multiple disease-related binary networks, thereby eliminating the need for professional medical knowledge; on the other hand, in the similarity network calculation stage, a more comprehensive and effective similarity network calculation method is proposed, which combines the graph structure context information with the node sequence information to obtain a more reliable similarity network. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a method for discovering disease-symptom association relationships using a disease knowledge graph according to the present invention;

[0051] Figure 2 It is a flow chart for decomposing the disease knowledge graph of the present invention;

[0052] Figure 3 This is a similarity network calculation flow chart of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The present invention proposes a method for discovering disease-symptom associations using disease knowledge graphs, such as Figure 1 , specifically including the following steps:

[0055] S100, decomposing the disease knowledge graph into a plurality of bipartite networks including disease nodes and / or symptom nodes, and saving the bipartite networks in the form of an adjacency matrix;

[0056] S200, calculating similarity matrices of disease nodes and symptom nodes according to all obtained bipartite networks, and merging them into a comprehensive disease similarity matrix and a comprehensive symptom similarity matrix;

[0057] S300, constructing a graph autoencoder, using the graph autoencoder to extract a nonlinear representation of the disease node based on a comprehensive disease similarity matrix and a vector representation of a disease node, and extracting a nonlinear representation of the disease node based on a comprehensive symptom similarity matrix and a vector representation of a symptom node;

[0058] S400, extracting linear representations of diseases and symptoms using non-negative matrix decomposition, that is, decomposing the disease-symptom association matrix into a linear representation matrix U of diseases and a linear representation matrix V of symptoms, wherein U and V are low-rank non-negative matrices;

[0059] S500, combining linear representation and nonlinear representation to predict disease-symptom association through a fully connected neural network.

[0060] In order to solve the problem that the existing technology usually requires specific medical knowledge and a single similarity network calculation method, this embodiment decomposes the disease knowledge graph into multiple binary networks containing disease nodes to obtain multi-view data, and proposes a more comprehensive and effective similarity network calculation method to calculate network similarity. The flow chart is as follows Figure 2 As shown in the figure, the specific process of disease knowledge graph decomposition is as follows:

[0061] S110. The data used is a disease knowledge graph. Each time a disease node and another type of node and the associated edges between them are extracted, a disease-related bipartite network is obtained. All bipartite networks extracted from the disease knowledge graph constitute a disease-related bipartite network set, and the set includes a bipartite network composed of disease nodes and symptom nodes. In this embodiment, the network is called a disease-symptom association network.

[0062] Similarly, a set of symptom-related bipartite networks can be obtained, which will not be described in detail in this embodiment.

[0063] S120. For the decomposed network, since there are many relationships in the disease knowledge graph, it is easy to observe that the associated edges are sparse and there are many isolated nodes, so the bipartite network needs to be simplified. This embodiment takes the disease-symptom association network as an example, and deletes the disease nodes with a degree less than p and the symptom nodes with a degree less than q in the loop, and finds the appropriate p and q so that the number of disease nodes, symptom nodes and associated edges is more appropriate. The values ​​of p and q can be obtained by those skilled in the art based on experience or experiments, and the present invention does not perform linear analysis on this.

[0064] S130, after simplifying all the bipartite networks, save all the bipartite networks in the form of adjacency matrices.

[0065] The similarity matrix of the disease nodes and symptom nodes is calculated based on all the obtained bipartite networks, and is merged into a comprehensive disease similarity matrix and a comprehensive symptom similarity matrix, wherein the comprehensive disease similarity matrix is ​​calculated based on the disease-related bipartite network, and the comprehensive symptom similarity matrix is ​​calculated based on the symptom-related bipartite network. The two calculation processes are the same, but the calculations are mainly based on different types of nodes. This embodiment is described by taking the disease node as an example. If n disease-related bipartite networks are obtained after simplification in S130, the process of obtaining the comprehensive disease similarity matrix is ​​as follows: Figure 3 As shown, the specific steps include:

[0066] S210. In the set of disease-related bipartite networks, the SimRank similarity of each disease is calculated for each bipartite network based on the SimRank algorithm to obtain the SimRank similarity matrix of each disease.

[0067] After simplification of S130, n disease-related bipartite networks are obtained. The SimRank similarity matrix of n diseases can be obtained by using the SimRank algorithm for each bipartite network;

[0068] Specifically, in a bipartite network, the SimRank similarity calculation between two types of nodes includes:

[0069]

[0070] Among them, s k+1 (a,b) represents the SimRank similarity between node a and node b in the k+1th iteration calculation, and C is the damping coefficient; I(a) represents the set of nodes of another type in the bipartite network connected to node a. If node a is a disease node, the set of nodes of another type is the set of symptom nodes; if node a is a symptom node, the set of nodes of another type is the set of disease nodes; s k (I i (a),I j (b)) is the i-th node I in the node set I(a) i (a) and the jth node I in the node set I (b) j (b) The similarity between them; |I(a)| represents the number of nodes in set I(a).

[0071] The element in the i-th row and j-th column of the disease SimRank similarity matrix represents the SimRank similarity between disease node i and disease node. The SimRank similarity matrix of the disease at the k-th iteration is expressed as:

[0072]

[0073] Among them, S k represents the SimRank similarity matrix obtained at the kth iteration; P is the transition probability matrix; I n is an n×n unit matrix, where n is the number of nodes; diag(M) means taking the matrix M to obtain a diagonal vector, and Diag(V) means converting the vector V into a diagonal matrix.

[0074] Perform the above operation on each bipartite network in the set of disease-related bipartite networks to obtain the SimRank similarity matrix of all diseases, expressed as SD 1 , SD 2 ,......,SD n , S.D. n Represents the SimRank similarity matrix of the disease in the nth bipartite network in the set of disease-related bipartite networks.

[0075] Similarly, by performing the above operation on each bipartite network in the set of symptom-related bipartite networks, the SimRank similarity matrix of all symptoms can be obtained.

[0076] S220. At the same time, in the set of disease-related bipartite networks, the feature representation of each disease is obtained based on the DeepWalk algorithm for each bipartite network, the cosine similarity between the two disease feature representations is calculated, and the DeepWalk similarity matrix of the disease is constructed.

[0077] When calculating the DeepWalk similarity matrix of nodes in a bipartite network, the following steps are specifically included:

[0078] Randomly walk and sample node sequences in the bipartite network, perform SkipGram training on the sampled node sequences, and learn low-dimensional vector representations of nodes;

[0079] Calculate the cosine similarity of all disease nodes pairwise to obtain the DeepWalk similarity matrix of the disease;

[0080] Performing the above operations on all bipartite networks can obtain the DeepWalk similarity matrix of all diseases, which are represented as SD n+1 , SD n+2 ,......,SD m , S.D. n+1 Represents the DeepWalk similarity matrix of the disease in the first bipartite network in the set of disease-related bipartite networks, m=2n.

[0081] Similarly, by performing the above operations on each bipartite network in the set of symptom-related bipartite networks, we can obtain the DeepWalk similarity matrix of all symptoms.

[0082] S230, based on the nonlinear similarity network fusion method SNF, the SimRank similarity matrix of the disease and the DeepWalk similarity matrix are fused to obtain a comprehensive disease similarity matrix

[0083] The process of the similarity network fusion method SNF includes the following steps:

[0084] Each type of similarity network is normalized to obtain the SimRank similarity matrix SD in the disease-symptom association network. 1 The SimRank similarity matrix obtained for the disease-symptom association network, the normalized calculation process is as follows:

[0085]

[0086] Among them, SD 1 ′(i,j) represents the SD 1 (i, j) is the similarity between node i and node j after normalization; S 1 (i,j) represents the similarity between disease node i and symptom node j.

[0087] In this embodiment, K nearest neighbors (KNN) is used to measure the local affinity. The calculation process of the local affinity includes:

[0088]

[0089] in, represents the local affinity between node i and node j; N i is the set of k nearest neighbor nodes of node i.

[0090] Based on the assumption that local similarity (high value) is more reliable than remote similarity, the similarity between non-adjacent nodes is set to 0. For n disease bipartite networks, after obtaining the normalized matrix and local affinity kernel of all disease similarity matrices, each similarity matrix is ​​iteratively updated, and the update process is expressed as:

[0091]

[0092] in, represents the state matrix of the vth disease similarity matrix after t iterations, Sd v represents the local affinity kernel of the vth similarity matrix of the disease, and m is the number of similarity matrices.

[0093] After each iteration, the state matrix Normalization is performed and the iteration stops when the convergence condition is reached. The convergence condition is:

[0094]

[0095] Here, ||·|| represents the Frobenius norm of the matrix.

[0096] After the iteration is stopped (this embodiment assumes that t iterations are performed), the average value of all similarity matrices is obtained to obtain the disease comprehensive similarity matrix obtained by the disease symptom network:

[0097]

[0098] According to the above calculation rules, the disease comprehensive similarity matrix SD is not a symmetric matrix, so we use The final comprehensive disease similarity matrix is ​​obtained.

[0099] like Figure 3 When obtaining the comprehensive symptom similarity matrix, the above operations are performed from the symptom-related bipartite network to obtain the comprehensive symptom similarity matrix. In this embodiment, Figure 3 Process, the bipartite network of symptoms can only select a network consisting of a disease node and a symptom node as its bipartite network, and the remaining calculations are the same as the calculation of the comprehensive symptom similarity matrix of the disease, which will not be repeated in the present invention.

[0100] This embodiment still takes obtaining a linear representation of a disease as an example for explanation. The process of obtaining a linear representation of a disease includes:

[0101] Construct a graph autoencoder, which includes an encoder and a decoder. The graph autoencoder takes the comprehensive disease similarity matrix and the vector representation of a disease node as input, and outputs the potential expression Z of the disease node through the graph autoencoder, where one row of the disease-symptom association matrix is ​​the vector representation of a disease node.

[0102] In the decoding part, GAE adopts inner product to reconstruct the comprehensive disease similarity matrix.

[0103] The entire encoder contains two layers of graph convolutional neural networks. Each layer uses the first-order approximation of Chebyshev polynomials as the convolution kernel to process data. Each layer of graph convolutional neural network can be expressed by a simple formula:

[0104] Z=GCN(X,SD)

[0105] Among them, X represents the vector representation of a disease node, and SD represents the comprehensive disease similarity matrix.

[0106] The definition of the graph convolutional neural network function is as follows:

[0107] GCN(X,SD)=SDReLU(SDXW 0 )W 1

[0108] in, W0 and W1 are the parameters to be learned; ReLU(·) represents the ReLU activation function, and D represents the degree matrix of the matrix SD.

[0109] The adjacency matrix determines the structure of the graph. The reconstructed adjacency matrix should be as similar to the original matrix as possible. Therefore, GAE uses cross entropy as the loss function during training:

[0110]

[0111] Where y represents the value of an element in the adjacency matrix SD (0 or 1), and y represents the value of the corresponding element in the reconstructed adjacency matrix SD (between 0 and 1).

[0112] Similarly, according to the above steps, based on the comprehensive symptom similarity matrix and the vector representation of a symptom node as input, the potential expression of the disease node can be output through the graph autoencoder, where one column of the disease-symptom association matrix is ​​the vector representation of a symptom node.

[0113] Non-negative matrix factorization (NMF) is used to extract nonlinear representations of diseases and symptoms, that is, the disease-symptom association relationship is projected into the disease subspace and symptom subspace. The disease-symptom association matrix A is decomposed into a low-rank non-negative matrix U and a low-rank non-negative matrix V as nonlinear representations of diseases and symptoms respectively. The specific decomposition process includes:

[0114] Define the indicator weighting matrix W = w ij ∈R m×n , the value of W is the same as A, for the i-th row and j-th column u in the low-rank non-negative matrix U ik and the kth row and jth column of the low-rank non-negative matrix V kj The update is performed, and when the update result converges, the currently updated matrix is ​​output as a nonlinear representation of the disease and symptoms. The update process includes:

[0115]

[0116]

[0117] in, represents the value of the i-th row and j-th column in the t-th updated low-rank non-negative matrix U; 1 and λ 2 is the regularization coefficient, ⊙ is the Hadamard product; (·) ik Represents the value of the i-th row and j-th column.

[0118] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for discovering disease-symptom associations using disease knowledge graphs, It is characterized in that The specific steps include: S100, decomposing the disease knowledge graph into a plurality of bipartite networks including disease nodes and / or symptom nodes, and saving the bipartite networks in the form of an adjacency matrix; The specific steps include: S110, extracting disease nodes and other nodes except disease nodes from the disease knowledge graph, forming a bipartite network of the disease node and any other node except disease nodes extracted from the disease knowledge graph, and all bipartite networks constitute a disease-related bipartite network; S120, extracting symptom nodes and other nodes except symptom nodes from the disease knowledge graph, forming a bipartite network of the symptom nodes and any other nodes except symptom nodes extracted from the disease knowledge graph, and all bipartite networks constitute a symptom-related bipartite network; S130, cyclically deleting nodes of one type whose node degree is less than p and nodes of another type whose node degree is less than q in the bipartite network, to obtain a simplified bipartite network; S140, saving the simplified bipartite network in the form of an adjacency matrix; S200, calculating similarity matrices of disease nodes and symptom nodes according to all obtained bipartite networks, and merging them into a comprehensive disease similarity matrix and a comprehensive symptom similarity matrix; specifically comprising the following steps: S210, in the set of disease-related bipartite networks, respectively calculating the SimRank similarity of each disease for each bipartite network based on the SimRank algorithm to obtain a SimRank similarity matrix for each disease; S220, in the set of bipartite networks related to the symptoms, calculating the SimRank similarity of each symptom for each bipartite network based on the SimRank algorithm to obtain a SimRank similarity matrix for each symptom; S230, in the set of symptom-related bipartite networks, obtain the feature representation of each symptom based on the DeepWalk algorithm for each bipartite network, calculate the cosine similarity between two symptom feature representations, and construct a DeepWalk similarity matrix of the symptoms; S240, in the set of disease-related bipartite networks, obtaining a feature representation of each disease based on the DeepWalk algorithm for each bipartite network, calculating the cosine similarity between two disease feature representations, and constructing a DeepWalk similarity matrix of the disease; S250, based on the nonlinear similarity network fusion method SNF, the SimRank similarity matrix of the disease and the DeepWalk similarity matrix are fused to obtain a comprehensive disease similarity matrix; S260, based on the nonlinear similarity network fusion method SNF, the SimRank similarity matrix of symptoms and the DeepWalk similarity matrix are fused to obtain a comprehensive symptom similarity matrix; S300, constructing a graph autoencoder, using the graph autoencoder to extract a nonlinear representation of the disease node based on a comprehensive disease similarity matrix and a vector representation of a disease node, and extracting a nonlinear representation of the symptom node based on a comprehensive symptom similarity matrix and a vector representation of a symptom node; S400, extracting linear representations of diseases and symptoms using non-negative matrix decomposition, that is, decomposing a disease-symptom association matrix into a linear representation matrix U of diseases and a linear representation matrix V of symptoms, wherein U and V are low-rank non-negative matrices, and a row of the disease-symptom association matrix is ​​a vector representation of a disease node, and a column is a vector representation of a symptom node; S500, combining linear representation and nonlinear representation to predict disease-symptom association through a fully connected neural network.

2. According to claim 1, a method for discovering disease-symptom associations using a disease knowledge graph, It is characterized in that The SimRank similarity calculation between two nodes includes: Among them, s k+1 (a,b) represents the SimRank similarity between node a and node b in the k+1th iteration calculation, and C is the damping coefficient; I(a) represents the set of other types of nodes in the bipartite network connected to node a that are different from the type of node a; I(b) represents the set of other types of nodes in the bipartite network connected to node b that are different from the type of node b; s k (I i (a),I j (b)) is the i-th node I in the node set I(a) i (a) and the jth node I in the node set I (b) j (b); |I(a)| represents the number of nodes in set I(a), and |I(b) represents the number of nodes in set I(b).

3. According to claim 1, a method for discovering disease-symptom associations using a disease knowledge graph, It is characterized in that The element in the i-th row and j-th column of the disease SimRank similarity matrix represents the SimRank similarity between disease node i and disease node j. The SimRank similarity matrix of the disease at the k-th iteration is expressed as: Among them, S k represents the SimRank similarity matrix obtained at the kth iteration; P is the transition probability matrix; I n is an n×n unit matrix, where n is the number of nodes; diag(M) means taking the diagonal vector of the matrix M, and Diag(X) means converting the vector X into a diagonal matrix.

4. According to claim 1, a method for discovering disease-symptom associations using a disease knowledge graph, It is characterized in that Based on the nonlinear similarity network fusion method SNF, the process of fusing the SimRank similarity matrix and the DeepWalk similarity matrix includes: The SimRank similarity matrix and the DeepWalk similarity matrix are normalized respectively, and the local affinity of each similarity matrix is ​​calculated based on the K-nearest neighbor algorithm; The normalized similarity matrix is ​​iteratively updated using the local affinity, and the data after each iterative update is normalized. When the iterative result converges, the iteration ends. Otherwise, the local affinity of each similarity matrix is ​​continued to be calculated and iteratively updated according to the local affinity. The mean matrix of the converged similarity matrix is ​​calculated, the mean matrix is ​​converted into a symmetric matrix, and the symmetric matrix is ​​used as the comprehensive similarity matrix.

5. According to claim 4, a method for discovering disease-symptom associations using a disease knowledge graph, It is characterized in that The process of normalizing the similarity matrix includes: The calculation process of local affinity includes: Among them, S 1 ′(i,j) represents the pair S 1 (i, j) is the similarity between node i and node j after normalization; S 1 (i, j) represents the similarity between node i and node j in a similarity matrix; represents the local affinity between node i and node j; N i is the set of k nearest neighbor nodes of node i.

6. A method for discovering disease-symptom associations using a disease knowledge graph according to claim 5, It is characterized in that The process of iteratively updating the similarity matrix based on local affinity includes: in, represents the state matrix of the vth similarity matrix after t iterations, represents the local affinity kernel of the vth similarity matrix, and m is the number of similarity matrices.

7. The method of claim 1 for discovering disease-symptom associations using a disease knowledge graph, It is characterized in that The graph autoencoder is used to obtain nonlinear representations of diseases and symptoms. The training process of the graph autoencoder includes: The comprehensive similarity matrix and feature matrix of a node are taken as the output of the encoder, and the encoder obtains the potential representation of the node; The decoder reconstructs the comprehensive similarity matrix of the nodes according to the potential representation, and calculates the cross entropy of the comprehensive similarity matrix of the nodes based on the reconstruction result of the encoder as the loss function to train the encoder and decoder.

8. According to claim 1, a method for discovering disease-symptom associations using a disease knowledge graph, It is characterized in that Non-negative matrix factorization (NMF) is used to extract nonlinear representations of diseases and symptoms, that is, the disease-symptom association relationship is projected into the disease subspace and symptom subspace. The disease-symptom association matrix A is decomposed into a low-rank non-negative matrix U and a low-rank non-negative matrix V as nonlinear representations of diseases and symptoms respectively. The specific decomposition process includes: Define the indicator weighting matrix W = w ij ∈R m×n , the value of W is the same as A, for the i-th row and j-th column u in the low-rank non-negative matrix U ik and the kth row and jth column of the low-rank non-negative matrix V kj The update is performed, and when the update result converges, the currently updated matrix is ​​output as a nonlinear representation of the disease and symptoms. The update process includes: in, represents the value of the i-th row and j-th column in the low-rank non-negative matrix U updated for the tth time; λ 1 and λ 2 is the regularization coefficient, ⊙ is the Hadamard product; (·) ik Represents the value of the i-th row and j-th column.

Citation Information

Patent Citations

  • MiRNA and disease incidence relation prediction method based on self-encoding model

    CN112183837A

  • Disease prediction method and system based on multi-granularity feature fusion

    CN112331332A

  • MiRNA-disease association prediction method and device based on graph neural network fusion multi-view information

    CN112784913A

  • Graph neural network model construction method, system and device, diagnosis and treatment scheme recommendation method, system and device

    CN113434692A

  • Network-based deep learning technology for target identification and drug repurposing

    US20210142173A1