Method, device and storage medium for predicting miRNA-disease association

Abstract

The application discloses a miRNA and disease association prediction method and device and a storage medium, and relates to the field of miRNA and disease association prediction.The method comprises the following steps: determining a miRNA and disease association matrix, a miRNA similarity matrix and a disease similarity matrix; constructing a heterogeneous network graph about the miRNA and disease association; calculating a Laplacian graph regularization matrix of the miRNA / disease similarity; constructing a graph autoencoder, wherein the encoder and the decoder of the graph autoencoder both adopt two layers of GAT+GCN; taking the heterogeneous association matrix, the feature space matrix and the Laplacian graph regularization matrix as inputs of the graph autoencoder to obtain an output matrix; calculating a loss value according to the output matrix and the miRNA and disease association matrix; training the graph autoencoder according to the loss value, and obtaining a miRNA and disease association prediction result.The application improves the performance of association prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of disease association prediction, and in particular relates to a miRNA and disease association prediction method, device and storage medium based on a multi-source information improved graph autoencoder. Background Art

[0002] In recent years, microRNA (miRNA) has become a hot topic of research in biomedical research. In the early 1990s, the Ambros and Ruvkun laboratories discovered a short ribonucleotide that controls the timing of developmental transitions in the nematode Caenorhabditis elegans, and named it miRNA. This miRNA is a short, non-coding ribonucleotide, only 22 in length, that regulates gene expression at the post-transcriptional level. Initially, researchers conducted detailed explorations of miRNAs from synthesis to biogenesis. miRNAs regulate gene expression within cells primarily by interacting with the 3' untranslated region of target gene mRNAs, thereby inhibiting translation of the target gene or inducing degradation of the target mRNA.

[0003] miRNAs have a wide range of effects. They not only play important roles in fundamental biological processes such as metabolism, cell differentiation, development, cell cycle, and apoptosis, but also play a crucial role in the development and progression of various human diseases. Therefore, the study of miRNAs is of great significance for understanding basic biological processes and developing new treatments. Targeting miRNAs for therapeutic purposes has become a new research direction. In this direction, scientists are working hard to develop miRNA-targeted drugs to treat various diseases. These drugs are designed to target miRNAs, modulating their expression or blocking their effects, thereby regulating gene expression and cellular biological processes, ultimately achieving the goal of treating diseases.

[0004] As research on miRNAs deepens, more and more experimental evidence indicates the important role of miRNAs in disease treatment. Consequently, a growing number of computational methods have been proposed to study the relationship between diseases and miRNAs. Among these, some major computational methods include traditional matrix-based methods and machine learning-based methods, such as a miRNA-disease association prediction method, system, terminal, and storage medium (application publication number CN 111681705 A) and a heterogeneous graph-based miRNA-disease association prediction method and system (CN 114613437 A). However, these methods all have significant drawbacks, namely, the incomplete modeling of association relationships in existing miRNA-disease association prediction models, which results in a lack of information about the association relationship and, consequently, reduced prediction performance. Summary of the Invention

[0005] The purpose of the present invention is to provide a miRNA-disease association prediction method, device and storage medium to solve the problem that the association relationship modeling in traditional miRNA-disease association prediction models is incomplete, resulting in missing association relationship information and thus resulting in low prediction performance.

[0006] The present invention solves the above technical problems through the following technical solutions: a method for predicting the association between miRNA and disease, comprising the following steps:

[0007] Acquire miRNA and disease data, obtain a miRNA and disease association matrix based on the miRNA and disease data, calculate a miRNA similarity matrix based on the miRNA data, and calculate a disease similarity matrix based on the disease data;

[0008] constructing a heterogeneous association matrix based on the miRNA-disease association matrix, the miRNA similarity matrix, and the disease similarity matrix, and defining a feature space matrix using the miRNA similarity matrix and the disease similarity matrix; and constructing a heterogeneous network diagram regarding the association between miRNA and disease based on the heterogeneous association matrix and the feature space matrix;

[0009] Calculating a Laplacian graph regularization matrix of miRNA similarity based on the miRNA similarity matrix, and calculating a Laplacian graph regularization matrix of disease similarity based on the disease similarity matrix;

[0010] Constructing a graph autoencoder, the graph autoencoder comprising an encoder and a decoder, the encoder comprising a first graph attention mechanism, a first graph convolution layer, a first splicing layer, a second graph attention mechanism, a second graph convolution layer, and a second splicing layer connected in sequence, and the decoder comprising a third graph attention mechanism, a third graph convolution layer, a third splicing layer, a fourth graph attention mechanism, and a fourth graph convolution layer connected in sequence;

[0011] Using the heterogeneous association matrix, the feature space matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity as inputs of the graph autoencoder to obtain an output matrix of the graph autoencoder;

[0012] Calculating a loss value based on the output matrix and the miRNA-disease association matrix;

[0013] The training parameters of the graph autoencoder are adjusted according to the loss value, and when the loss value is stable, the association prediction score matrix between miRNA and disease is calculated according to the output matrix of the graph autoencoder.

[0014] Furthermore, the specific calculation process of the miRNA similarity matrix is:

[0015] Taking each miRNA as a string, the sequence similarity between different miRNAs is calculated using the Levenshtein distance. The specific calculation formula is:

[0016]

[0017] Among them, m i indicates miRNAi, m j miRNAj, Seq(m i ,m j ) represents m i With m j The sequence similarity between s(m i ) represents the sequence information of miRNAi, s(m j ) represents the sequence information of miRNAj, dis represents the sequence information of m i Convert to m j Minimum editing costs required;

[0018] Calculate the functional similarity between different miRNAs. The specific calculation formula is:

[0019]

[0020]

[0021] Among them, MF(m i ,m j ) represents m i With m j The functional similarity between i Indicates that m i Related disease collection, D j Indicates that m j Related disease collection, d ir Denotes the disease set D i Diseases in jr Denotes the disease set D j Diseases in Sim(d ir ,D j ) indicates disease d ir and disease set D j The similarity of Sim(d jr ,D i ) indicates disease d jr and disease set D i The similarity of |D i | represents the disease set D i The number of diseases in |D j | represents the disease set D j The number of diseases in Sim(d,d r) represents the relationship between disease d and a disease d in the disease set D r The similarity between them, n represents the number of diseases in the disease set D;

[0022] Calculate the Gaussian kernel similarity between different miRNAs. The specific calculation formula is:

[0023] GM(m i ,m j )=exp(-υ m ||IP(m i )-IP(m j )|| 2 )

[0024]

[0025] Among them, GM(m i ,m j ) represents m i With m j Gaussian kernel similarity between IP(m i ) represents m i The interaction spectrum, IP(m j ) represents m j The interaction spectrum, υ m represents the variable that controls the bandwidth of the miRNA kernel, υ m ' represents the normalized miRNA kernel bandwidth variable, Num m represents the number of miRNAs;

[0026] The sequence similarity, functional similarity and Gaussian kernel similarity between different miRNAs are integrated to obtain miRNA similarity, and the miRNA similarity matrix is ​​constructed by the similarities of all miRNAs. The specific calculation formula of the miRNA similarity is:

[0027]

[0028]

[0029] Among them, MM (m i ,m j ) represents m i With m j The similarity between IM(m i ,m j ) represents the average of sequence similarity and functional similarity.

[0030] Furthermore, the specific calculation process of the disease similarity matrix is:

[0031] Construct disease d iDirected acyclic graph of The directed acyclic graph Including diseases i The node and the set of all its ancestor disease nodes and the set of all direct edges from the parent disease node to the child disease nodes

[0032] Calculate the two disease nodes d i d j The semantic similarity DL (d i ,d j ), the specific calculation formula is:

[0033]

[0034]

[0035]

[0036]

[0037] in, Indicates disease i The semantic value of the disease d, Indicates disease j The semantic value of the disease d, DValue(d i ) indicates disease d i The semantic value of DValue(d j ) indicates disease d j The semantic value of Indicates disease d versus disease d i The semantic contribution of Indicates disease d versus disease d j The semantic contribution of disease d, η represents the semantic contribution factor, and d' represents the sub-disease of disease d;

[0038] Calculate the Gaussian kernel similarity between different diseases. The specific calculation formula is:

[0039] GD(d i ,d j )=exp(-υ d ||IP(d i )-IP(d j )|| 2 )

[0040]

[0041] Among them, GD(d i ,d j) indicates disease d i and disease j Gaussian kernel similarity between IP(d i ) indicates disease d i The interaction spectrum, IP(d j ) indicates disease d j The interaction spectrum, υ d represents the variable that controls the bandwidth of the disease kernel, υ d ' represents the normalized disease kernel bandwidth variable, Num d Indicates the number of diseases;

[0042] The semantic similarity and Gaussian kernel similarity between different diseases are integrated to obtain disease similarity, and a disease similarity matrix is ​​constructed by all disease similarities. The specific calculation formula of the disease similarity is:

[0043]

[0044] Among them, DD(d i ,d j ) indicates disease d i and disease j The similarities between them.

[0045] Furthermore, the specific implementation process of constructing a heterogeneous association matrix based on the miRNA-disease association matrix, the miRNA similarity matrix, and the disease similarity matrix is ​​as follows:

[0046] Construct the first zero matrix O m and the second zero matrix O d , the first zero matrix O m The miRNA similarity matrix is ​​a homotype matrix, and the second zero matrix O d The disease similarity matrix is ​​a homotypic matrix;

[0047] The miRNA and disease association matrix, the first zero matrix O m and the second zero matrix O d The heterogeneous correlation matrix is ​​formed, and the specific expression is:

[0048]

[0049] Among them, HN is the heterogeneous association matrix, and H is the miRNA-disease association matrix.

[0050] Furthermore, the specific expression of the feature space matrix is:

[0051]

[0052] Among them, Y0 is the feature space matrix, Num m Indicates the number of miRNA, Num d Indicates the number of diseases; W mm and W dd represents the transformation matrix; MM represents the miRNA similarity matrix, DD represents the disease similarity matrix; Y mm Represents the characteristics of miRNA nodes, Y dd represents the characteristics of the disease node; cat(·) represents the splicing function.

[0053] Furthermore, the specific calculation formula of the Laplace graph regularization matrix of the miRNA similarity is:

[0054]

[0055] Among them, M m represents the Laplacian graph regularization matrix of miRNA similarity, MM represents the miRNA similarity matrix, and the diagonal matrix B m represents the degree matrix of MM;

[0056] The specific calculation formula of the Laplacian graph regularization matrix of the disease similarity is:

[0057]

[0058] Among them, M d represents the Laplacian graph regularization matrix of disease similarity, DD represents the disease similarity matrix, and the diagonal matrix B d represents the degree matrix of DD.

[0059] Furthermore, the specific process of obtaining the output matrix is ​​as follows:

[0060] The heterogeneous correlation matrix, feature space matrix, Laplacian graph regularization matrix of miRNA similarity, and Laplacian graph regularization matrix of disease similarity are encoded for the first time using the first graph attention mechanism and the first graph convolution layer to obtain the first miRNA output feature matrix and the first disease output feature matrix. The specific calculation formula is:

[0061]

[0062]

[0063] in, represents the first miRNA output feature matrix, Represents the first disease output feature matrix, M m Laplacian graph regularization matrix representing miRNA similarity, M dThe Laplace graph regularization matrix representing disease similarity, HN is the heterogeneous correlation matrix, Y0 is the feature space matrix, W1 and W2 are both weight matrices, represents the miRNA node feature matrix obtained by the first graph attention mechanism, represents the disease node feature matrix obtained by the first graph attention mechanism;

[0064] Using the first splicing layer to splice the first miRNA output feature matrix and the first disease output feature matrix to obtain a first spliced ​​feature matrix;

[0065] The second graph attention mechanism and the second graph convolution layer are used to encode the heterogeneous correlation matrix, the first splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity for the second time to obtain the second miRNA output feature matrix and the second disease output feature matrix. The specific calculation formula is:

[0066]

[0067]

[0068] in, represents the second miRNA output feature matrix, represents the second disease output feature matrix, Y1 is the first splicing feature matrix, W3 and W4 are both weight matrices, represents the miRNA node feature matrix obtained by the second graph attention mechanism, represents the disease node feature matrix obtained by the second graph attention mechanism;

[0069] Using the second splicing layer to splice the second miRNA output feature matrix and the second disease output feature matrix to obtain a second splicing feature matrix;

[0070] The heterogeneous correlation matrix, the second splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity are decoded for the first time using the third graph attention mechanism and the third graph convolution layer to obtain the third miRNA output feature matrix and the third disease output feature matrix. The specific calculation formula is:

[0071]

[0072]

[0073] in, represents the third miRNA output feature matrix, represents the third disease output feature matrix, Y2 is the second splicing feature matrix, W5 and W6 are both weight matrices, represents the miRNA node feature matrix obtained by the third graph attention mechanism, represents the disease node feature matrix obtained by the third graph attention mechanism;

[0074] Using the third splicing layer to splice the third miRNA output feature matrix and the third disease output feature matrix to obtain a third splicing feature matrix;

[0075] The heterogeneous association matrix, the third splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity are decoded for the second time using the fourth graph attention mechanism and the fourth graph convolution layer to obtain a fourth miRNA output feature matrix and a fourth disease output feature matrix. The fourth miRNA output feature matrix and the fourth disease output feature matrix are the output matrices of the graph autoencoder. The specific calculation formula is:

[0076]

[0077]

[0078] in, represents the fourth miRNA output feature matrix, represents the fourth disease output feature matrix, Y3 is the third splicing feature matrix, W7 and W8 are both weight matrices, represents the miRNA node feature matrix obtained by the attention mechanism in the fourth figure, Represents the disease node feature matrix obtained by the fourth graph attention mechanism.

[0079] Furthermore, the miRNA node feature matrix is ​​obtained by the first graph attention mechanism and disease node feature matrix The specific implementation process is:

[0080] The calculation coefficient of the attention mechanism is calculated according to the feature space matrix. The specific calculation formula is:

[0081]

[0082] in, Represents the parameterized attention mechanism; Represents the calculation coefficient of attention; W represents the weight matrix; Represents the i-th miRNA node m in the heterogeneous network graph i The eigenvector of Represents the jth disease node d in the heterogeneous network graphj The eigenvector of Represents the rth disease node d in the heterogeneous network graph r The eigenvector of r Represents node m i Neighbor nodes; || represents the concatenation operation of two matrices; Represents miRNA node m in the heterogeneous network graph i The set of all neighbor nodes of

[0083] Multi-head attention is used to aggregate the features of the neighboring nodes of the disease node to obtain a new feature space matrix GAT(HN,Y0). The specific aggregation formula is:

[0084]

[0085] Among them, ζ represents the nonlinear LeakyReLU activation function, K represents the number of attention heads, represents the disease node d under the k-th attention head j Neighbor aggregation information, W k represents the weight matrix under the k-th attention head, Represents the calculation coefficient under the k-th attention;

[0086] Determine the miRNA node feature matrix according to the new feature space matrix GAT(HN,Y0) and disease node feature matrix The specific formula is:

[0087]

[0088]

[0089] Among them, GAT(HN,Y0)[1:Num m ,:] represents the first to Num of the new feature space matrix GAT(HN,Y0) m Row, GAT(HN,Y0)[(Num m +1):(Num m +Num d ),:] represents the Num-th matrix of the new feature space matrix GAT(HN,Y0) m +1~Num m +Num d OK.

[0090] Furthermore, the specific calculation formula of the loss value is:

[0091]

[0092] Among them, Γ represents the loss value, h ij represents the element in row i and column j of the miRNA-disease association matrix H, represents the fourth miRNA output feature matrix, Represents the fourth disease output feature matrix, r m ij Represents the matrix R m The element in row i and column j in r d ij Represents the matrix R d T The element at row i and column j in .

[0093] Furthermore, the specific calculation formula of the association prediction score matrix is:

[0094]

[0095] Among them, R represents the association prediction score matrix, represents the fourth miRNA output feature matrix, represents the fourth disease output feature matrix, represents the equilibrium parameter.

[0096] Based on the same concept, the present invention also provides an electronic device, which includes: a processor and a memory storing a computer program, wherein the processor is configured to implement the above-mentioned miRNA and disease association prediction method when executing the computer program.

[0097] Based on the same concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the miRNA-disease association prediction method described above is implemented.

[0098] Beneficial effects

[0099] Compared with the prior art, the advantages of the present invention are:

[0100] The present invention utilizes the known association between miRNA and disease to construct the architecture of heterogeneous network graph, uses similarity matrix to define the characteristics of nodes in the heterogeneous network graph, and propagates information through the nodes in the heterogeneous network graph to achieve feature aggregation, thereby achieving the purpose of modeling the association relationship between nodes, making the association relationship modeling more complete, improving the problem of missing association relationship information, and improving the model prediction performance.

[0101] The present invention adopts graph attention mechanism + graph convolution to replace the graph convolution layer in the traditional graph autoencoder, uses the improved graph autoencoder to predict potential disease-related miRNAs, captures the association relationship between neighbor nodes in the heterogeneous network graph, effectively aggregates the features between nodes, and finally obtains the prediction score matrix by calculating the aggregated features; the present invention is robust and effectively improves the performance of association prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only one embodiment of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0103] Figure 1 is a flow chart of the method for predicting the association between miRNA and disease in an embodiment of the present invention, wherein M q represents the Laplacian graph regularization matrix, q∈{m,d}, Y p Represents the feature space matrix or concatenated feature matrix, p∈{0,1,2,3}. DETAILED DESCRIPTION

[0104] The following is a clear and complete description of the technical solutions of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts are within the scope of protection of the present invention.

[0105] The following specific embodiments are used to describe the technical solution of the present application in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0106] like Figure 1 As shown, a method for predicting the association between miRNA and disease provided by an embodiment of the present invention includes the following steps:

[0107] Step 1: Obtain miRNA and disease data, obtain the miRNA and disease association matrix based on the miRNA and disease data, calculate the miRNA similarity matrix based on the miRNA data, and calculate the disease similarity matrix based on the disease data.

[0108] This example obtains miRNA and disease data from HMD2.0 and determines the miRNA and disease association matrix H. If the i-th miRNA (i.e., miRNAi or m i) and the jth disease (ie d j ) association, the value of the i-th row and j-th column of the miRNA-disease association matrix H is 1, otherwise it is 0. The cases where it is 0 include no association and unknown association. The present invention uses the known association between miRNA and disease to predict unknown association.

[0109] In this embodiment, miRNA sequence similarity, miRNA function similarity and miRNA Gaussian kernel similarity are calculated and integrated to obtain a miRNA similarity matrix. The specific calculation process of the miRNA similarity matrix is:

[0110] Step 1.11: Treat each miRNA as a string and calculate the sequence similarity between different miRNAs using Levenshtein distance. This is the minimum number of insertions, deletions, and substitutions required to convert one miRNA string into another. Therefore, the specific formula for calculating sequence similarity between different miRNAs is:

[0111]

[0112] Among them, m i indicates miRNAi, m j miRNAj, Seq(m i ,m j ) represents m i With m j The sequence similarity between s(m i ) represents the sequence information of miRNAi, s(m j ) represents the sequence information of miRNAj, dis represents the sequence information of m i Convert to m j The minimum editing cost required (i.e., the minimum number of insertion, deletion, and substitution operations). As can be seen from formula (1), the smaller the editing cost between two miRNAs, the greater the sequence similarity between the two miRNAs.

[0113] Step 1.12: Calculate the functional similarity between different miRNAs. The specific calculation formula is:

[0114]

[0115]

[0116] Among them, MF(m i ,m j ) represents m i With m j The functional similarity between i Indicates that m iRelated disease collection, D j Indicates that m j Related disease collection, d ir Denotes the disease set D i Diseases in jr Denotes the disease set D j Diseases in Sim(d ir ,D j ) indicates disease d ir and disease set D j The similarity of Sim(d jr ,D i ) indicates disease d jr and disease set D i The similarity of |D i | represents the disease set D i The number of diseases in |D j | represents the disease set D j The number of diseases in Sim(d,d r ) represents the disease d and the disease set D(D={d1,d2,...,d n}) a disease r The similarity between them is n, and n represents the number of diseases in the disease set D.

[0117] Formula (3) defines the similarity between diseases and disease sets based on the principle that miRNAs with similar functions are usually associated with similar diseases. i ,m j )>0, then the two miRNAs are connected and MF(m i ,m j ) score is set as the functional similarity between miRNAs.

[0118] Step 1.13: Based on the principle that miRNAs with similar functions are often associated with similar diseases, we observed i Whether there is a known association with each disease, the vector IP (m i ) to represent m i The interaction spectrum between different miRNAs is calculated using the Gaussian kernel similarity. The specific calculation formula is:

[0119] GM(m i ,m j )=exp(-υ m ||IP(m i )-IP(m j )|| 2 ) (4)

[0120]

[0121] Among them, GM(m i ,m j ) represents m i With m j Gaussian kernel similarity between IP(m i ) represents m i The interaction spectrum, IP(m j ) represents m j The interaction spectrum, υ m represents the variable that controls the bandwidth of the miRNA kernel, υ m ' represents the normalized miRNA kernel bandwidth variable, Num m Represents the number of miRNA. From formula (5), we can see that υ m is obtained by normalizing the new kernel bandwidth variable υ m 'Average number of diseases associated with all miRNAs.

[0122] Step 1.14: Integrate the sequence similarity, functional similarity, and Gaussian kernel similarity between different miRNAs to obtain miRNA similarity. The miRNA similarity matrix is ​​constructed from all miRNA similarities. The specific calculation formula for miRNA similarity is:

[0123]

[0124]

[0125] Among them, MM (m i ,m j ) represents m i With m j The similarity between IM(m i ,m j ) represents the average value of sequence similarity and functional similarity. As shown in formula (7), the sequence similarity and functional similarity of miRNA are equally important, so the two similarities are averaged to integrate them.

[0126] In this embodiment, the disease semantic similarity and disease Gaussian kernel similarity are calculated and integrated to obtain a disease similarity matrix. The specific calculation process of the disease similarity matrix is:

[0127] Step 1.21: Construct disease d i Directed acyclic graph of The directed acyclic graph Including diseases i The node and the set of all its ancestor disease nodes and the set of all direct edges from the parent disease node to the child disease nodes

[0128] Disease semantic similarity is constructed based on the disease ontology identification number (DOID) in the Disease Ontology database. In disease ontology, diseases are usually modeled as a directed acyclic graph (DAG), where each node represents a disease or symptom and each edge represents a relationship, such as etiology, pathophysiology, or treatment.

[0129] Diseased i Directed acyclic graph of is a disease node d i (or disease d i ) and all its ancestral disease nodes and the set of all direct edges from the parent disease node to the child disease nodes Specifically expressed as exist In the disease d vs disease d i Semantic contribution The calculation formula is as follows:

[0130]

[0131] Where η represents the semantic contribution factor, and d' represents the sub-disease of disease d. In the example, it is assumed that the more distant ancestral disease node is the disease node d i The contribution of disease d to the specific semantic value of is smaller, so η∈(0,1), which is set to 0.5 in this embodiment. Similarly, disease d contributes less to disease d j Semantic contribution The calculation formula is:

[0132]

[0133] For disease d i The semantic value DValue(d i ) is calculated as follows:

[0134]

[0135] Similarly, for disease d j The semantic value DValue(d j ) is calculated as follows:

[0136]

[0137] Step 1.22: By considering two disease nodes d i d j The relative position in the DAG, assuming that diseases that share most of the DAG tend to have higher semantic similarity, therefore, the distance between two disease nodes d is calculated.i d j The semantic similarity DL (d i ,d j ), the specific calculation formula is:

[0138]

[0139] in, Indicates disease i The semantic value of the disease d, Indicates disease j The semantic value of the disease d.

[0140] Step 1.23: Based on the principle that miRNAs with similar functions are often associated with similar diseases, we observed the disease d i Whether there is a known association with each miRNA, the vector IP (d i ) to indicate disease d i The interaction spectrum of different diseases is used to calculate the Gaussian kernel similarity between them. The specific calculation formula is:

[0141] GD(d i ,d j )=exp(-υ d ||IP(d i )-IP(d j )|| 2 ) (13)

[0142]

[0143] Among them, GD(d i ,d j ) indicates disease d i and disease j Gaussian kernel similarity between IP(d i ) indicates disease d i The interaction spectrum, IP(d j ) indicates disease d j The interaction spectrum, υ d represents the variable that controls the bandwidth of the disease kernel, υ d ' represents the normalized disease kernel bandwidth variable, Num d Represents the number of diseases. From formula (14), we can see that υ d is obtained by normalizing the new kernel bandwidth variable υ d 'Average number of miRNAs associated with all diseases.

[0144] Step 1.24: Integrate the semantic similarity and Gaussian kernel similarity between different diseases to obtain disease similarity. The disease similarity matrix is ​​constructed from all disease similarities. The specific calculation formula for disease similarity is:

[0145]

[0146] Among them, DD(d i ,d j ) indicates disease d i and disease j The similarities between them.

[0147] Step 2: Construct a heterogeneous association matrix based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, and use the miRNA similarity matrix and disease similarity matrix to define the feature space matrix; construct a heterogeneous network diagram about the association between miRNA and disease based on the heterogeneous association matrix and feature space matrix.

[0148] First, construct the first zero matrix O m and the second zero matrix O d , the first zero matrix The miRNA similarity matrix is ​​the isotype matrix, the second zero matrix The disease similarity matrix is ​​a homotypic matrix;

[0149] Secondly, the miRNA-disease association matrix and the first zero matrix O m and the second zero matrix O d Construct a heterogeneous correlation matrix, the specific expression is:

[0150]

[0151] Among them, HN is the heterogeneous association matrix, and H is the miRNA-disease association matrix.

[0152] The feature space matrix is ​​defined using the miRNA similarity matrix and the disease similarity matrix. The specific expression is:

[0153]

[0154] Among them, Y0 is the feature space matrix, represents the real number space, f represents the number of features of each node; W mm and W dd Represents the transformation matrix, using and The feature information of miRNA nodes and disease nodes is projected into the same feature space matrix Y0; MM represents the miRNA similarity matrix, DD represents the disease similarity matrix; cat(·) represents the splicing function, which splices the two matrices with the same column number up and down; Y mm Represents the characteristics of miRNA nodes, Y mm The i-th row of Y represents the characteristics of the i-th miRNA node in the heterogeneous network graph; dd Represents the characteristics of the disease node, Y dd The i-th row of represents the features of the i-th disease node in the heterogeneous network graph.

[0155] Based on the heterogeneous association matrix and feature space matrix, the DGL library in Python is used to construct a heterogeneous network graph about the association between miRNA and disease. The heterogeneous network graph includes nodes and edges, where nodes include all miRNA nodes and all disease nodes. Edge represents the edges associated with all miRNAs, diseases, and diseases.

[0156] Step 3: Calculate the Laplacian graph regularization matrix of miRNA similarity based on the miRNA similarity matrix, and calculate the Laplacian graph regularization matrix of disease similarity based on the disease similarity matrix.

[0157] From the perspective of miRNA, according to the definition of the Laplace graph regularization matrix, the calculation formula of the Laplace graph regularization matrix of miRNA similarity is:

[0158]

[0159]

[0160] Among them, M m represents the Laplacian graph regularization matrix of miRNA similarity, MM represents the miRNA similarity matrix, and the diagonal matrix B m Represents the degree matrix of MM.

[0161] Similarly, from the perspective of disease, according to the definition of the Laplace graph regularization matrix, the calculation formula of the Laplace graph regularization matrix of disease similarity is:

[0162]

[0163]

[0164] Among them, M d represents the Laplacian graph regularization matrix of disease similarity, DD represents the disease similarity matrix, and the diagonal matrix B drepresents the degree matrix of DD.

[0165] Step 4: Build the graph autoencoder.

[0166] Traditional graph autoencoders mainly include an encoder and a decoder, each of which contains two layers of graph convolution. The present invention improves the graph autoencoder by replacing each layer of graph convolution in the encoder and decoder with a graph attention network (GAT) and a graph convolution network (GCN). Therefore, the graph autoencoder of the present invention includes an encoder and a decoder, wherein the encoder includes a first graph attention mechanism, a first graph convolution layer, a first splicing layer, a second graph attention mechanism, a second graph convolution layer, and a second splicing layer connected in sequence, and the decoder includes a third graph attention mechanism, a third graph convolution layer, a third splicing layer, a fourth graph attention mechanism, and a fourth graph convolution layer connected in sequence. That is, the encoder part of the graph autoencoder of the present invention includes two layers of GAT+GCN, and the decoder part also includes two layers of GAT+GCN, and the matrix is ​​encoded or decoded using the graph attention mechanism + graph convolution.

[0167] Taking the feature space matrix Y0 as an example, we will explain the specific process of obtaining the node feature matrix using the graph attention mechanism:

[0168] The feature space matrix can be expressed as:

[0169]

[0170] in, Represents the i-th miRNA node m in the heterogeneous network graph i The eigenvector of Represents the jth disease node d in the heterogeneous network graph j The feature vector of ; f represents the feature number of miRNA node or disease node.

[0171] To transform the features in the feature space matrix into higher-level features, a learnable linear transformation is required. Therefore, a weight matrix or feature projection parameter matrix is ​​set. As a shared linear transformation (where f represents the number of features of the miRNA node or disease node), and used on each miRNA node, disease node. A self-attention mechanism a is used on each miRNA node or disease node: To calculate the attention coefficient, the specific formula is as follows:

[0172]

[0173] in, Represents miRNA node mi The features of disease node d j Calculate the importance or attention coefficient of disease nodes of Represents miRNA node m in the heterogeneous network graph i The set of all neighbor nodes of . Using the softmax function in all diseases d j The attention coefficient on the neighborhood of Normalize so that the attention coefficient It is easy to compare between different nodes. The specific formula is as follows:

[0174]

[0175] in, Represents the calculation coefficient of attention. The attention mechanism a is a single-layer feedforward neural network, through the weight vector Parameterization, where the parameterized attention mechanism Applying the nonlinear activation function LeakyReLU as the linear transformation function in the attention mechanism can help improve the model's ability to learn the attention coefficient. By fully expanding formula (24), we can get the calculation coefficient of attention The specific formula is as follows:

[0176]

[0177] in, express The transpose of || represents the concatenation of two matrices. Represents the rth disease node d in the heterogeneous network graph r The eigenvector of r Represents node m i neighbor nodes.

[0178] In order to stabilize the learning process of self-attention and fully aggregate the neighborhood features of the node, the present invention uses multi-head attention to aggregate the disease node d j The characteristics of the neighbor nodes (or adjacent nodes) of the disease node d are obtained. j The new feature representation, that is, the new feature space matrix GAT(HN,Y0):

[0179]

[0180] Among them, ζ represents the nonlinear LeakyReLU activation function, K represents the number of attention heads, represents the disease node d under the k-th attention head j Neighbor aggregation information, W kRepresents the weight matrix under the k-th attention head.

[0181] Determine the miRNA node feature matrix based on the new feature space matrix GAT(HN,Y0) and disease node feature matrix The specific formula is:

[0182]

[0183]

[0184] Among them, GAT(HN,Y0)[1:Num m ,:] represents the first to Num of the new feature space matrix GAT(HN,Y0) m Row, GAT(HN,Y0)[(Num m +1):(Num m +Num d ),:] represents the Num-th matrix of the new feature space matrix GAT(HN,Y0) m +1~Num m +Num d The functions in formulas (27) and (28) are convenient for representing the features of miRNA nodes and disease nodes aggregated by the graph attention mechanism.

[0185] Step 5: The heterogeneous correlation matrix, the feature space matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity are used as the input of the graph autoencoder in step 4 to obtain the output matrix of the graph autoencoder.

[0186] In this embodiment, the specific process of obtaining the output matrix is ​​as follows:

[0187] Step 5.1: Use the first graph attention mechanism and the first graph convolution layer (i.e., the first layer GAT+GCN of the encoder) to perform the first encoding on the heterogeneous correlation matrix, feature space matrix, miRNA similarity Laplacian graph regularization matrix, and disease similarity Laplacian graph regularization matrix to obtain the first miRNA output feature matrix and the first disease output feature matrix. The specific calculation formula is:

[0188]

[0189]

[0190] in, represents the first miRNA output feature matrix, Represents the first disease output feature matrix, W1 and W2 are weight matrices (parameters to be learned), represents the miRNA node feature matrix obtained by the first graph attention mechanism (the specific acquisition process is shown in formulas (22) to (28)), Represents the disease node feature matrix obtained by the first graph attention mechanism (see formulas (22) to (28) for the specific acquisition process).

[0191] Step 5.2: Use the first splicing layer to splice the first miRNA output feature matrix and the first disease output feature matrix to obtain the first splicing feature matrix Y1. The specific formula is:

[0192]

[0193] Where cat(·) represents the concatenation function.

[0194] Step 5.3: Use the second graph attention mechanism and the second graph convolution layer (i.e., the second layer GAT+GCN of the encoder) to perform a second encoding on the heterogeneous correlation matrix, the first spliced ​​feature matrix Y1, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity to obtain the second miRNA output feature matrix and the second disease output feature matrix. The specific calculation formula is:

[0195]

[0196]

[0197] in, represents the second miRNA output feature matrix, represents the second disease output feature matrix, W3 and W4 are weight matrices (parameters to be learned), represents the miRNA node feature matrix obtained by the second graph attention mechanism (the specific acquisition process is similar to formulas (22) to (28)), Represents the disease node feature matrix obtained by the second graph attention mechanism (the specific acquisition process is similar to formulas (22) to (28)).

[0198] Step 5.4: Use the second splicing layer to splice the second miRNA output feature matrix and the second disease output feature matrix to obtain the second splicing feature matrix Y2. The specific formula is:

[0199]

[0200] Step 5.5: Use the third graph attention mechanism and the third graph convolution layer (i.e., the first layer GAT+GCN of the decoder) to perform the first decoding on the heterogeneous correlation matrix, the second splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity to obtain the third miRNA output feature matrix and the third disease output feature matrix. The specific calculation formula is:

[0201]

[0202]

[0203] in, represents the third miRNA output feature matrix, Represents the output feature matrix of the third disease, W5 and W6 are both weight matrices (parameters to be learned), represents the miRNA node feature matrix obtained by the third graph attention mechanism (the specific acquisition process is similar to formulas (22) to (28)), Represents the disease node feature matrix obtained by the third graph attention mechanism (the specific acquisition process is similar to formulas (22) to (28)).

[0204] Step 5.6: Use the third splicing layer to splice the third miRNA output feature matrix and the third disease output feature matrix to obtain the third splicing feature matrix Y3. The specific formula is:

[0205]

[0206] Step 5.7: Use the fourth graph attention mechanism and the fourth graph convolution layer (i.e., the second layer GAT+GCN of the decoder) to perform a second decoding on the heterogeneous correlation matrix, the third splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity to obtain the fourth miRNA output feature matrix and the fourth disease output feature matrix. The fourth miRNA output feature matrix and the fourth disease output feature matrix are the output matrices of the graph autoencoder. The specific calculation formula is:

[0207]

[0208]

[0209] in, represents the fourth miRNA output feature matrix, Represents the fourth disease output feature matrix, W7 and W8 are weight matrices (parameters to be learned), represents the miRNA node feature matrix obtained by the attention mechanism in the fourth figure (the specific acquisition process is similar to formulas (22) to (28)), Represents the disease node feature matrix obtained by the fourth graph attention mechanism (the specific acquisition process is similar to formulas (22) to (28)), and the output matrix includes and

[0210] For ease of description, and Expressed as:

[0211]

[0212]

[0213] Formulas (38) and (39) are the outputs of the decoder.

[0214] Step 6: Calculate the loss value based on the output matrix and the miRNA-disease association matrix.

[0215] In the process of graph autoencoder training, in order to reconstruct an adjacency matrix similar to the output matrix, a loss function is needed to minimize the error as much as possible, which is generally calculated using binary cross entropy. When training graph autoencoders for miRNA and disease separately, the loss function Γ m and Γ d The calculation formula is as follows:

[0216]

[0217]

[0218] Among them, h ij represents the element in row i and column j of the miRNA-disease association matrix, r m ij Represents the matrix R m The element in row i and column j in r d ij Represents the matrix R d T The element in row i and column j in R d T R d Therefore, the final loss function Γ can be calculated, and the specific formula is as follows:

[0219]

[0220] Step 7: Adjust the training parameters of the graph autoencoder according to the loss value to complete the training of the graph autoencoder.

[0221] Step 8: When the loss value is stable, the miRNA-disease association prediction score matrix is ​​calculated based on the output matrix of the graph autoencoder.

[0222] When the loss value tends to be stable and is less than or equal to the set threshold, the association prediction score matrix between miRNA and disease is calculated to achieve the association prediction between miRNA and disease. In this embodiment, the specific calculation formula of the association prediction score matrix is:

[0223]

[0224] Among them, R represents the association prediction score matrix, In this embodiment, the threshold is set to 0.001.

[0225] The higher the AUC value, the better the performance of the association prediction. In order to obtain association prediction results with higher AUC values, during the training process of the graph autoencoder, the AUC value is calculated at the same time as the loss value. When the loss value tends to be stable and is less than or equal to the set threshold, and the AUC value is greater than or equal to the set value (for example, 0.9), the association prediction score matrix is ​​calculated according to formula (45), which can obtain higher performance association prediction results.

[0226] The performance of the method of the present invention and related miRNA and disease prediction methods were evaluated and compared on a 5-fold cross validation basis:

[0227] For the method of the present invention, association prediction belongs to a two-classification problem. The two-classification problem is generally a combination of the true category of the data and the model prediction category, including true positive examples (TP), true negative examples (TN), false positive examples (FP), and false negative examples (FN). The sum of these four categories is the total number of data samples. Performance evaluation indicators generally include accuracy (Acc), recall (Recall, also known as sensitivity Sen), precision (Pre), F1 score (F1_score) and AUC (Area Under theCurve, referring to the area under the ROC curve surrounded by TPR and FPR) and AUPR value (referring to the area under the PR curve surrounded by precision and recall), and the specific calculation formula is:

[0228]

[0229]

[0230]

[0231]

[0232] The performance evaluation index values ​​of different methods are shown in Table 1:

[0233] Table 1 Performance evaluation index values ​​of different methods

[0234] Method Name AUC AUPR Pre Sen / Recall F1_score References[1] 0.7763 0.0182 0.01212 0.78073 0.02387 References[2] 0.8463 0.0626 0.01839 0.84929 0.03601 References[3] 0.8988 0.0835 0.0211 0.90022 0.04128 References[4] 0.8207 0.2229 0.2738 0.3389 0.3029 Method of the present invention 0.9069 0.3506 0.09444 0.89506 0.17085

[0235] In Table 1, the literature [1]: Chen X, Liu MX, Yan GY. RWRMDA: predicting novel human microRNA–disease associations [J]. Molecular BioSystems, 2012, 8(10): 2792-8., the literature [1] discloses a global network similarity measurement, which is different from the traditional local network similarity measurement. It takes into account the structure and information of the entire network and uses the restarted random walk algorithm to obtain the potential association between miRNA and disease from the known association matrix between miRNA and disease, the miRNA similarity matrix and the disease similarity matrix.

[0236] Reference [2]: Qu Y, Zhang H, Liang C, et al. KATZMDA: prediction of miRNA-disease associations based on KATZ model [J]. Ieee Access, 2017, 6: 3943-50. Reference [2] integrates miRNA-disease similarity information to establish a heterogeneous network and uses the KATZ algorithm based on the network structure to predict miRNA-disease associations, thereby improving the accuracy of prediction.

[0237] Reference [3]: Rao Y, Xie M, Wang H. Predict potential miRNA-disease associations based on bounded nuclear norm regularization [J]. Frontiers in Genetics, 2022, 13. Reference [3] constructed a heterogeneous network of miRNAs and diseases by integrating the known associations between miRNAs and diseases, the comprehensive similarity of miRNAs, and the comprehensive similarity of diseases, and predicted the potential associations between miRNAs and diseases based on the bounded nuclear norm regularization method.

[0238] Reference [4]: ​​Bian C, Lei XJ, Wu FX. GATCDA: predicting circRNA-disease associations based on graph attention network [J]. Cancers, 2021, 13(11): 2595. Reference [4] starts from disease similarity, circRNA similarity and the known association between circRNA and disease, uses the attention mechanism to learn the representation of nodes on the graph, assigns different weights to different nodes in the neighborhood, and finally predicts the association between circRNA and disease.

[0239] As can be seen from Table 1, compared with the related miRNA and disease association prediction methods, the method of the present invention has better performance, which shows that the method of the present invention has good reliability in predicting the potential association between miRNA and disease.

[0240] In the case analysis results of the method of the present invention on predicting miRNAs associated with a complex disease, the top 20 candidate miRNAs in the three diseases studied, namely breast tumors, non-small cell lung cancer, and colon tumors, were selected, and the association between the disease and miRNA was verified through references in the HMDDv3.2, dbDEMC, and PubMed databases. For details, see Tables 2, 3, and 4.

[0241] Table 2 Top 20 miRNAs predicted to be associated with breast tumors

[0242]

[0243] Table 3 Top 20 miRNAs predicted to be associated with non-small cell lung cancer

[0244]

[0245] Table 4 Top 20 miRNAs predicted to be associated with colon tumors

[0246]

[0247]

[0248] As can be seen from Tables 2, 3, and 4, the present method has been validated in real-world biomedical wet lab experiments using candidate miRNAs associated with three complex diseases, demonstrating its reliability. While these unverified predictions can assist bioinformatics experts in identifying candidate targets, significantly improving research efficiency and facilitating the rapid identification of disease causes, enabling timely and targeted treatment.

[0249] Based on the same concept, an embodiment of the present invention further provides an electronic device, comprising: a processor and a memory storing a computer program, wherein the processor is configured to implement the above-mentioned miRNA-disease association prediction method when executing the computer program.

[0250] Although not shown, the electronic device includes a processor that can perform various appropriate operations and processes based on programs and / or data stored in a read-only memory (ROM) or programs and / or data loaded from a storage portion into a random access memory (RAM). The processor can be a multi-core processor or can include multiple processors. In some embodiments, the processor can include a general-purpose main processor and one or more special coprocessors, such as a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), etc. Various programs and data required for the operation of the electronic device are also stored in the RAM. The processor, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0251] The processor and memory are used together to execute the program stored in the memory. When the program is executed by the computer, the methods, steps or functions described in the above embodiments can be implemented.

[0252] Although not shown, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the miRNA-disease association prediction method described above is implemented.

[0253] Storage media in embodiments of the present invention include permanent and non-permanent, removable and non-removable items that can be used to store information using any method or technology. Examples of storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0254] The above disclosure is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or modifications within the technical scope disclosed in the present invention, and they should all be covered by the scope of protection of the present invention.

Claims

1. A method for predicting the association between miRNA and disease, characterized in that: The following steps are involved: Acquire miRNA and disease data, obtain a miRNA and disease association matrix based on the miRNA and disease data, calculate a miRNA similarity matrix based on the miRNA data, and calculate a disease similarity matrix based on the disease data; Constructing a heterogeneous association matrix based on the miRNA-disease association matrix, the miRNA similarity matrix, and the disease similarity matrix, and defining a feature space matrix using the miRNA similarity matrix and the disease similarity matrix; A heterogeneous network diagram of the association between miRNA and disease was constructed based on the heterogeneous association matrix and feature space matrix; Calculating a Laplacian graph regularization matrix of miRNA similarity based on the miRNA similarity matrix, and calculating a Laplacian graph regularization matrix of disease similarity based on the disease similarity matrix; Constructing a graph autoencoder, the graph autoencoder comprising an encoder and a decoder, the encoder comprising a first graph attention mechanism, a first graph convolution layer, a first splicing layer, a second graph attention mechanism, a second graph convolution layer, and a second splicing layer connected in sequence, and the decoder comprising a third graph attention mechanism, a third graph convolution layer, a third splicing layer, a fourth graph attention mechanism, and a fourth graph convolution layer connected in sequence; Using the heterogeneous association matrix, the feature space matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity as inputs of the graph autoencoder to obtain an output matrix of the graph autoencoder; Calculating a loss value based on the output matrix and the miRNA-disease association matrix; Adjusting the training parameters of the graph autoencoder according to the loss value, and calculating the association prediction score matrix between miRNA and disease according to the output matrix of the graph autoencoder when the loss value is stable; The specific process of obtaining the output matrix is ​​as follows: Performing a first encoding on the heterogeneous correlation matrix, the feature space matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity using the first graph attention mechanism and the first graph convolution layer to obtain a first miRNA output feature matrix and a first disease output feature matrix; Using the first splicing layer to splice the first miRNA output feature matrix and the first disease output feature matrix to obtain a first spliced ​​feature matrix; Using the second graph attention mechanism and the second graph convolution layer, the heterogeneous correlation matrix, the first splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity are encoded for a second time to obtain a second miRNA output feature matrix and a second disease output feature matrix; Using the second splicing layer to splice the second miRNA output feature matrix and the second disease output feature matrix to obtain a second splicing feature matrix; Performing a first decoding on the heterogeneous association matrix, the second splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity using the third graph attention mechanism and the third graph convolutional layer to obtain a third miRNA output feature matrix and a third disease output feature matrix; Using the third splicing layer to splice the third miRNA output feature matrix and the third disease output feature matrix to obtain a third splicing feature matrix; The heterogeneous association matrix, the third splicing feature matrix, the Laplacian graph regularization matrix of miRNA similarity, and the Laplacian graph regularization matrix of disease similarity are decoded for a second time using the fourth graph attention mechanism and the fourth graph convolution layer to obtain a fourth miRNA output feature matrix and a fourth disease output feature matrix, which are the output matrices of the graph autoencoder.

2. The miRNA-disease association prediction method according to claim 1, wherein The specific calculation process of the miRNA similarity matrix is: Taking each miRNA as a string, the sequence similarity between different miRNAs is calculated using the Levenshtein distance. The specific calculation formula is: Among them, m i indicates miRNAi, m j miRNAj, Seq(m i ,m j ) represents m i With m j The sequence similarity between s(m i ) represents the sequence information of miRNAi, s(m j ) represents the sequence information of miRNAj, dis represents the sequence information of m i Convert to m j Minimum editing costs required; Calculate the functional similarity between different miRNAs. The specific calculation formula is: Among them, MF(m i ,m j ) represents m i With m j The functional similarity between i Indicates that m i Related disease collection, D j Indicates that m j Related disease collection, d ir Denotes the disease set D i Diseases in jr Denotes the disease set D j Diseases in Sim(d ir ,D j ) indicates disease d ir and disease set D j The similarity of Sim(d jr ,D i ) indicates disease d jr and disease set D i The similarity of |D i | represents the disease set D i The number of diseases in |D j | represents the disease set D j The number of diseases in Sim(d,d r ) represents the relationship between disease d and a disease d in the disease set D r The similarity between them, n represents the number of diseases in the disease set D; Calculate the Gaussian kernel similarity between different miRNAs. The specific calculation formula is: GM(m i ,m j )=exp(-υ m ||IP(m i )-IP(m j )|| 2 ) Among them, GM(m i ,m j ) represents m i With m j Gaussian kernel similarity between them, IP(m i ) represents m i The interaction spectrum, IP(m j ) represents m j The interaction spectrum, υ m represents the variable that controls the bandwidth of the miRNA kernel, υ m ' represents the normalized miRNA kernel bandwidth variable, Num m represents the number of miRNAs; The sequence similarity, functional similarity and Gaussian kernel similarity between different miRNAs are integrated to obtain miRNA similarity, and the miRNA similarity matrix is ​​constructed by the similarities of all miRNAs. The specific calculation formula of the miRNA similarity is: Among them, MM (m i ,m j ) represents m i With m j The similarity between IM(m i ,m j ) represents the average of sequence similarity and functional similarity.

3. The method for predicting the association between miRNA and disease according to claim 1, wherein The specific calculation process of the disease similarity matrix is: Construct disease d i Directed acyclic graph of The directed acyclic graph Including diseases i The node and the set of all its ancestor disease nodes and the set of all direct edges from the parent disease node to the child disease nodes Calculate the two disease nodes d i d j The semantic similarity DL (d i ,d j ), the specific calculation formula is: in, Indicates disease i The semantic value of the disease d, Indicates disease j The semantic value of the disease d, DValue(d i ) indicates disease d i The semantic value of DValue(d j ) indicates disease d j The semantic value of Indicates disease d versus disease d i The semantic contribution of Indicates disease d versus disease d j The semantic contribution of disease d, η represents the semantic contribution factor, and d' represents the sub-disease of disease d; Calculate the Gaussian kernel similarity between different diseases. The specific calculation formula is: GD(d i ,d j )=exp(-υ d ||IP(d i )-IP(d j )|| 2 ) Among them, GD(d i ,d j ) indicates disease d i and disease j Gaussian kernel similarity between IP(d i ) indicates disease d i The interaction spectrum, IP(d j ) indicates disease d j The interaction spectrum, υ d represents the variable that controls the bandwidth of the disease kernel, υ d ' represents the normalized disease kernel bandwidth variable, Num d Indicates the number of diseases; The semantic similarity and Gaussian kernel similarity between different diseases are integrated to obtain disease similarity, and a disease similarity matrix is ​​constructed by all disease similarities. The specific calculation formula of the disease similarity is: Among them, DD(d i ,d j ) indicates disease d i and disease j The similarities between them.

4. The method for predicting the association between miRNA and disease according to claim 1, wherein The specific implementation process of constructing a heterogeneous association matrix based on the miRNA-disease association matrix, miRNA similarity matrix and disease similarity matrix is ​​as follows: Construct the first zero matrix O m and the second zero matrix O d , the first zero matrix O m The miRNA similarity matrix is ​​a homotype matrix, and the second zero matrix O d The disease similarity matrix is ​​a homotypic matrix; The miRNA and disease association matrix, the first zero matrix O m and the second zero matrix O d The heterogeneous correlation matrix is ​​formed, and the specific expression is: Among them, HN is the heterogeneous association matrix, and H is the miRNA-disease association matrix.

5. The method for predicting the association between miRNA and disease according to claim 1, wherein The specific expression of the feature space matrix is: Among them, Y0 is the feature space matrix, Num m Indicates the number of miRNA, Num d Indicates the number of diseases; W mm and W dd represents the transformation matrix; MM represents the miRNA similarity matrix, DD represents the disease similarity matrix; Y mm Represents the characteristics of miRNA nodes, Y dd represents the characteristics of the disease node; cat(·) represents the splicing function.

6. The method for predicting the association between miRNA and disease according to claim 1, wherein The specific calculation formula of the Laplacian graph regularization matrix of the miRNA similarity is: Among them, M m represents the Laplacian graph regularization matrix of miRNA similarity, MM represents the miRNA similarity matrix, and the diagonal matrix B m represents the degree matrix of MM; The specific calculation formula of the Laplacian graph regularization matrix of the disease similarity is: Among them, M d represents the Laplacian graph regularization matrix of disease similarity, DD represents the disease similarity matrix, and the diagonal matrix B d represents the degree matrix of DD.

7. The method for predicting the association between miRNA and disease according to any one of claims 1 to 6, characterized in that: The specific calculation formulas for the first miRNA output feature matrix and the first disease output feature matrix are: in, represents the first miRNA output feature matrix, Represents the first disease output feature matrix, M m Laplacian graph regularization matrix representing miRNA similarity, M d The Laplace graph regularization matrix representing disease similarity, HN is the heterogeneous correlation matrix, Y0 is the feature space matrix, W1 and W2 are both weight matrices, represents the miRNA node feature matrix obtained by the first graph attention mechanism, represents the disease node feature matrix obtained by the first graph attention mechanism; The specific calculation formulas for the second miRNA output feature matrix and the second disease output feature matrix are: in, represents the second miRNA output feature matrix, represents the second disease output feature matrix, Y1 is the first splicing feature matrix, W3 and W4 are both weight matrices, represents the miRNA node feature matrix obtained by the second graph attention mechanism, represents the disease node feature matrix obtained by the second graph attention mechanism; The specific calculation formulas for the third miRNA output feature matrix and the third disease output feature matrix are: in, represents the third miRNA output feature matrix, represents the third disease output feature matrix, Y2 is the second splicing feature matrix, W5 and W6 are both weight matrices, represents the miRNA node feature matrix obtained by the third graph attention mechanism, represents the disease node feature matrix obtained by the third graph attention mechanism; The specific calculation formula of the output matrix of the graph autoencoder is: in, represents the fourth miRNA output feature matrix, represents the fourth disease output feature matrix, Y3 is the third splicing feature matrix, W7 and W8 are both weight matrices, represents the miRNA node feature matrix obtained by the attention mechanism in the fourth figure, Represents the disease node feature matrix obtained by the fourth graph attention mechanism.

8. The method for predicting the association between miRNA and disease according to claim 7, wherein: The miRNA node feature matrix is ​​obtained by the first graph attention mechanism and disease node feature matrix The specific implementation process is: The calculation coefficient of the attention mechanism is calculated according to the feature space matrix. The specific calculation formula is: in, Represents the parameterized attention mechanism; Represents the calculation coefficient of attention; W represents the weight matrix; Represents the i-th miRNA node m in the heterogeneous network graph i The eigenvector of Represents the jth disease node d in the heterogeneous network graph j The eigenvector of Represents the rth disease node d in the heterogeneous network graph r The eigenvector of r Represents node m i Neighbor nodes; || represents the concatenation operation of two matrices; Represents miRNA node m in the heterogeneous network graph i The set of all neighbor nodes of Multi-head attention is used to aggregate the features of the neighboring nodes of the disease node to obtain a new feature space matrix GAT(HN,Y0). The specific aggregation formula is: Among them, ζ represents the nonlinear LeakyReLU activation function, K represents the number of attention heads, represents the disease node d under the k-th attention head j Neighbor aggregation information, W k represents the weight matrix under the k-th attention head, Represents the calculation coefficient under the k-th attention; Determine the miRNA node feature matrix according to the new feature space matrix GAT(HN,Y0) and disease node feature matrix The specific formula is: Among them, GAT(HN,Y0)[1:Num m ,:] represents the first to Num of the new feature space matrix GAT(HN,Y0) m Row, GAT(HN,Y0)[(Num m +1):(Num m +Num d ),:] represents the Num-th matrix of the new feature space matrix GAT(HN,Y0) m +1~Num m +Num d OK.

9. The method for predicting the association between miRNA and disease according to claim 1, wherein The specific calculation formula of the loss value is: Among them, Γ represents the loss value, h ij represents the element in row i and column j of the miRNA-disease association matrix H, represents the fourth miRNA output feature matrix, Represents the fourth disease output feature matrix, r m ij Represents the matrix R m The element in row i and column j in r d ij Represents the matrix R d T The element at row i and column j in .

10. The method for predicting the association between miRNA and disease according to claim 1, wherein: The specific calculation formula of the association prediction score matrix is: Among them, R represents the association prediction score matrix, represents the fourth miRNA output feature matrix, represents the fourth disease output feature matrix, represents the equilibrium parameter.

11. An electronic device, comprising: A processor and a memory storing a computer program, wherein the processor is configured to implement the miRNA and disease association prediction method according to any one of claims 1 to 10 when executing the computer program.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting the association between miRNA and disease according to any one of claims 1 to 10 is implemented.

Images