MiRNA-disease association prediction method based on enhanced hypergraph convolution auto-encoding algorithm

Abstract

The application provides a miRNA-disease correlation prediction method based on an enhanced hypergraph convolution self-encoding algorithm, and comprises the following steps: S1: obtaining an adjacency matrix A of miRNA-disease, and describing the correlation relationship between the two through the miRNA-disease adjacency matrix; S2: calculating miRNA Gaussian similarity MG, miRNA cosine similarity MC, disease Gaussian similarity DG and disease cosine similarity DC through the adjacency matrix A, and taking them as the features of miRNA and the features of diseases respectively; S3: using weighted K-nearest neighbors to obtain graph hidden correlation information according to the Gaussian similarity, so as to obtain an enhanced hypergraph correlation matrix; S4: mapping the features of miRNA and the features of diseases to the same domain space through a full connection layer, and training to obtain embedding features in a hypergraph convolution mode; and S5: decoding the embedding features using a bilinear decoder, and calculating the correlation score, so that the application can more accurately predict the correlation information of miRNA-disease.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of deep learning combined with biological information, and more particularly relates to a miRNA-disease association prediction method based on an enhanced hypergraph convolution auto-encoding algorithm. BACKGROUND

[0002] miRNA is a kind of single-stranded non-coding RNA with a length of about 22 nucleotides. A large number of studies have shown that it can regulate genes to play biological functions after transcription, for example, in cell development and immune response, and its disorder can lead to cell disorder. In recent years, scientists have found that the disorder of miRNA is closely related to various human complex diseases, such as lung cancer, breast cancer, liver cancer and other diseases. Therefore, identifying the miRNA associated with diseases is of great significance for people to find disease biomarkers and diagnose human complex diseases. In the past research, experimenters basically predicted new association relationships through personal experimental experience and biological targeting characteristics, which was time-consuming and laborious and had a great relationship with the personal level of experimenters. Therefore, it is urgently hoped that a new method can quickly predict new association relationships, and calculating the original association through a computing method to infer the potential association between miRNA and diseases to a certain extent makes up for the vacancy, and in the previous research, the computing model plays an effective auxiliary role in miRNA-disease association prediction.

[0003] At present, the identification of disease-related miRNA can promote the pathological study of diseases and the detection of disease biomarkers. The discovery of potential miRNA-disease association is of great help to understand the pathogenesis of diseases and develop treatment methods for human diseases. For example, previous research results show that miR-204 can play the function of tumor suppressor in non-small cell lung cancer (NSCLC) by targeting JAK2, and miR-204 can be used as a biomarker for the diagnosis and treatment of NSCLC in medical treatment. The prognosis of hepatocellular carcinoma (HCC) patients is poor, and phosphoprotein (VASP) as a regulator of actin cytoskeleton and cell migration is overexpressed in HCC, which is related to malignant features and poor prognosis. The function of miRNA-204 down-regulated by hypoxia is to regulate the overexpression of VASP at the post-transcriptional level, which indirectly promotes the up-regulation of VASP at the transcriptional and post-transcriptional levels, and to a certain extent, it can be used as a prognostic biomarker.

[0004] To further understand the relationship between miRNAs and diseases, researchers have proposed several new hypotheses, one key one being that phenotypic diseases are often associated with miRNAs with similar functions. Based on this key hypothesis, researchers have proposed many methods based on similarity metrics and machine learning. However, while these methods can learn and capture important relationships between data and have achieved significant research results, most neglect the global network structure. Furthermore, these methods cannot uncover the complex connections deep within the network, making it extremely difficult to capture nonlinear relationships within the network. Deep learning methods, on the other hand, can capture nonlinear relationships in networks by constructing complex connections, and have achieved significant success in supervised learning tasks in multiple fields such as computer vision, natural language processing, and speech processing.

[0005] Deep learning, due to its ability to handle unstructured data, is also widely used in bioinformatics, and researchers have proposed a series of methods based on graph neural networks (GNNs). GNNs are derived from convolutional neural networks (CNNs) and graph embedding concepts, exhibiting superior performance in aggregating graph structural information. Currently, GNNs have demonstrated advanced performance in node classification tasks and have wide applications in life sciences and physical sciences. For example, Han et al. proposed a novel disease gene association task framework, GCN-MF, by capturing nonlinear interactions through GCN and matrix factorization. Tang et al. developed a multi-view, multi-channel attention graph convolutional network (MMGCN) to predict potential associations in MDA. Zhou et al. proposed a new method called NIMCGCN to incorporate GNNs with neural inductive matrix completion models. Recently, attention mechanisms have also been introduced into GNNs, enabling the network to focus on task-relevant input parts to identify hidden relationships between miRNAs and diseases. For example, Luo et al. created a hierarchical attention mechanism that learns node representations at the node level and uses graph-level attention to adjust contributions from different input graphs.

[0006] However, during the training process, much of the training is based on the original association graph. Since negative samples are difficult to obtain, a large number of unknown samples are assumed to be used as negative samples for training. Using unknown samples as negative samples for training may reduce the flow of data among these negative samples and affect the prediction accuracy of the method. Summary of the Invention

[0007] This invention provides a miRNA-disease association prediction method based on an enhanced hypergraph convolutional autoencoder algorithm, which can more accurately predict miRNA-disease associations.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0009] A miRNA-disease association prediction method based on an enhanced hypergraph convolutional autoencoder algorithm, which includes the following steps:

[0010] S1: Obtain the miRNA-disease adjacency matrix A, and describe the relationship between the two through the miRNA-disease adjacency matrix;

[0011] S2: Calculate the miRNA Gaussian similarity MG, miRNA cosine similarity MC, disease Gaussian similarity DG, and disease cosine similarity DC using the adjacency matrix A, and use them as features of miRNA and disease, respectively.

[0012] S3: Based on Gaussian similarity, the weighted k-nearest neighbor profile (WKNNP) is used to obtain the hidden association information of the graph, thereby obtaining the enhanced hypergraph association matrix;

[0013] S4: Map the features of miRNA and the features of the disease to the same domain space through a fully connected layer, and train the embedded features using hypergraph convolution.

[0014] S5: Decode the embedded features using a bilinear decoder and calculate their association score.

[0015] Preferably, the miRNA-disease adjacency matrix A mentioned in step S1 is generated in the following way:

[0016] A miRNA-disease association dataset was obtained from the Human MicroRNA Disease Database (HMDD). HMDD v3.2 contained 1206 miRNAs and 893 diseases, totaling 35547 associations. After removing non-human data, 913 miRNAs and 554 diseases were retained. The miRNA-disease adjacency matrix can be represented as follows: Where n m and n d The miRNA-disease adjacency matrix A represents the number of miRNAs and the number of diseases, respectively. The miRNA-disease adjacency matrix A is defined as follows:

[0017]

[0018] Preferably, the Gaussian similarity calculation method mentioned in step S2 can calculate the Gaussian similarity MG of the miRNA, specifically as follows:

[0019] Calculate the bandwidth for Gaussian interaction kernel similarity:

[0020]

[0021] n m The vector A(i,:) represents the number of miRNAs; the vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, where 1 in each row indicates whether there is an association between the miRNA and the disease.

[0022] Calculate the association similarity between each miRNA:

[0023] MG(i,j)=exp(-r i ||A(i,:)-A(j,:)|| 2 )

[0024] MG(i,j) represents the similarity between the i-th miRNA and the j-th miRNA; exp represents a power function with basis e; vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, and vector A(j,:) represents the j-th row of the miRNA-disease adjacency matrix A.

[0025] Preferably, the method for calculating the Gaussian similarity DG of the disease is the same as the method for calculating the Gaussian similarity MG of the miRNA, thereby calculating the Gaussian similarity DG of the disease, wherein the matrix used is the transpose of the miRNA-disease adjacency matrix A.

[0026] Preferably, the method for calculating cosine similarity mentioned in step S2 can calculate the cosine similarity of miRNA as MC, specifically as follows:

[0027] The similarity between two vectors is measured by the cosine angle between them; the smaller the angle, the more similar they are. For miRNA-miRNA associations, the similarity is calculated as follows:

[0028]

[0029] MC(i,j) represents the cosine similarity between the i-th miRNA and the j-th miRNA; vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, vector A(j,:) represents the j-th row of the miRNA-disease adjacency matrix A, and the symbol · represents the inner product between the two vectors.

[0030] Preferably, the method for calculating the cosine similarity DC of the disease is the same as the method for calculating the cosine similarity MC of the miRNA, wherein the matrix used is the transpose of the miRNA-disease adjacency matrix A.

[0031] Preferably, step S3 uses the weighted k-nearest neighbor profile (WKNNP) to obtain graph hidden association information, thereby obtaining an enhanced hypergraph association matrix. This uses Gaussian similarity (MG) for miRNAs and Gaussian similarity (DG) for diseases. The specific definition of the WKNNP method for constructing graph hidden links is as follows:

[0032] To improve message passing, the original adjacency matrix A is preprocessed to obtain more hidden relationships. Vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, indicating the association relationship of miRNAs, and vector A(:,j) represents the j-th column of the miRNA-disease adjacency matrix A, indicating the association relationship of diseases.

[0033] For each miRNA, obtain its k most relevant similar miRNAs from its Gaussian similarity, and let m j Let j represent the index of the j-th similar miRNA. The calculated new association vector is shown below:

[0034]

[0035] Where A m (i,:) represents the new association relationship calculated by miRNA in the i-th row, Q m =∑ 1≤j≤K MG(i,m j ) is a normalization term; Where α is the decay term, and its value is α∈[0,1];

[0036] Similarly, a new association vector can be calculated for each disease, using the following formula:

[0037]

[0038] Where A d (:,i) represents the calculated new association in the i-th column, let d j Q represents the index of the j-th similar disease. d =∑ 1≤j≤K DG(i,d j ) is a normalization term; Where α is the decay term, and its value is α∈[0,1];

[0039] Finally, the hidden association information of the graph is obtained, and the enhanced hypergraph association matrix is ​​obtained through the following formula:

[0040]

[0041] Preferably, in step S4, miRNA and disease features are mapped to the same domain space, specifically as follows:

[0042] The feature matrix of miRNA can be represented as [MG,Z] d ,MC,Z d Similarly, the characteristics of a disease can be represented as [Z]. m DG,Z m [,DC], where Z d Z m Let E represent zero matrices with column numbers equal to the disease length and miRNA length, respectively. The features of both miRNA and disease are passed through a linear layer for linear dimensionality reduction, and finally through a ReLU activation function. This yields the miRNA embedding E. m Embedded E of Disease d .

[0043] Preferably, step S4 employs a hypergraph convolution method, specifically defined as:

[0044] Based on the miRNA-disease association A, a miRNA hyperedge relationship is constructed, assuming the initial miRNA node characteristics are as follows:

[0045]

[0046] The hyperedge features of miRNAs are associated with all nodes within the hyperedge. Therefore, the hyperedge features of miRNAs can be derived as follows:

[0047]

[0048] in B represents the node features of the miRNA hypergraph in layer l. -1 For A T The normalized matrix;

[0049] The following hypergraph convolution formula is defined based on the graph convolution formula:

[0050]

[0051] Expanding the formula yields the following formula:

[0052]

[0053] This allows the features of nodes to flow globally, enabling the training of a data transfer matrix W.

[0054] Similarly, for diseases, we can obtain the features of a disease node.

[0055] Preferably, in step S5, the embedded features are decoded using a bilinear decoder to calculate their association score. The specific implementation of the bilinear decoder is as follows:

[0056]

[0057] The encoded miRNA and disease embeddings are used as input to the decoder. A bilinear decoder treats different interaction types as different categories. The miRNA m... i and disease d j Correlation score:

[0058]

[0059] Among them W r R is a trainable matrix of parameters whose parameter dimension is the dimension of the input embedding, where r∈R=[0,1].

[0060] Preferably, to optimize the results, cross-iterative entropy is used as the loss function to calculate the loss between the predicted result and the true value. Adam is used as the gradient optimizer to optimize the global network parameters.

[0061] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0062] This invention enhances the original graph by preprocessing the original association matrix using weighted K-nearest neighbor computation, reducing the impact of negative random sampling and enriching the information of the initial graph. This process reveals hidden association information within the graph, ultimately yielding an enhanced hypergraph association matrix. Subsequently, a new hypergraph network is constructed, incorporating the information from the enhanced hypergraph association matrix. This allows graph node features to flow across the hypergraph, collecting information from the entire domain. High-order sampling is then performed on the graph to extract nonlinear information from the graph nodes. Finally, a bilinear decoder is employed to calculate miRNA-disease association scores by decoding node embeddings, achieving more accurate miRNA-disease association prediction. Attached Figure Description

[0063] Figure 1 This is a flowchart of the method of the present invention;

[0064] Figure 2 In this embodiment, the five-fold cross-validation method utilizes SHGAEMDA, GAEMDA, and

[0065] A schematic diagram of AUCs implemented by DFELMDA and TDRC. Detailed Implementation

[0066] This embodiment provides a miRNA-disease association prediction method based on an enhanced hypergraph convolutional autoencoder algorithm, such as...Figure 1 As shown, it needs to include the following steps:

[0067] S1: Obtain the miRNA-disease adjacency matrix A, and describe the relationship between the two through the miRNA-disease adjacency matrix;

[0068] S2: Calculate the miRNA Gaussian similarity MG, miRNA cosine similarity MC, disease Gaussian similarity DG, and disease cosine similarity DC using the adjacency matrix A, and use them as features of miRNA and disease, respectively.

[0069] S3: Based on Gaussian similarity, the weighted k-nearest neighbor profile (WKNNP) is used to obtain the hidden association information of the graph, thereby obtaining the enhanced hypergraph association matrix;

[0070] S4: Map the features of miRNA and the features of the disease to the same domain space through a fully connected layer, and train the embedded features using hypergraph convolution.

[0071] S5: Decode the embedded features using a bilinear decoder and calculate their association score.

[0072] The miRNA-disease adjacency matrix A mentioned in step S1 is generated as follows:

[0073] A miRNA-disease association dataset was obtained from the Human MicroRNA Disease Database (HMDD). HMDD v3.2 contained 1206 miRNAs and 893 diseases, totaling 35547 associations. After removing non-human data, 913 miRNAs and 554 diseases were retained. The miRNA-disease adjacency matrix can be represented as follows: Where n m and n d These represent the number of miRNAs and the number of diseases, respectively. The miRNA-disease adjacency matrix A is defined as follows:

[0074]

[0075] The Gaussian similarity calculation method mentioned in step S2 can calculate the Gaussian similarity MG of miRNA, specifically as follows:

[0076] Calculate the bandwidth for Gaussian interaction kernel similarity:

[0077]

[0078] nm The vector A(i,:) represents the number of miRNAs; the vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, where 1 in each row indicates whether there is an association between the miRNA and the disease.

[0079] Calculate the association similarity between each miRNA:

[0080] MG(i,j)=exp(-γ i ||A(i,:)-A(j,:)|| 2 )

[0081] MG(i,j) represents the similarity between the i-th miRNA and the j-th miRNA; exp represents a power function with basis e; vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, and vector A(j,:) represents the j-th row of the miRNA-disease association matrix A.

[0082] The method for calculating the Gaussian similarity DG of a disease is the same as that for calculating the Gaussian similarity MG of miRNAs. In this way, the Gaussian similarity DG of a disease can be calculated, and the matrix used is the transpose of the miRNA-disease adjacency matrix A.

[0083] The method for calculating cosine similarity mentioned in step S2 can calculate the cosine similarity (MC) of miRNA, specifically as follows:

[0084] Based on the cosine angle between the two vectors,

[0085] This can measure the similarity between two vectors; the smaller the angle between them, the more similar they are. For miRNA-miRNA associations, the similarity is calculated as follows:

[0086]

[0087] MC(i,j) represents the cosine similarity between the i-th miRNA and the j-th miRNA; vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, vector A(j,:) represents the j-th row of the miRNA-disease adjacency matrix A, and the symbol · represents the inner product between the two vectors.

[0088] The method for calculating the cosine similarity DC of diseases is the same as the method for calculating the cosine similarity MC of miRNAs, where the matrix used is the transpose of the miRNA-disease adjacency matrix A.

[0089] Step S3 uses the weighted k-nearest neighbor profile (WKNNP) to obtain graph hidden association information, thereby obtaining an enhanced hypergraph association matrix. It uses Gaussian similarity (MG) for miRNAs and Gaussian similarity (DG) for diseases. The specific definition of the WKNNP method for constructing graph hidden links is as follows:

[0090] To improve message passing, the original adjacency matrix A is preprocessed to obtain more hidden relationships. Vector A(i,:) represents the i-th row of the miRNA-disease adjacency matrix A, indicating the association relationship of miRNAs, and vector A(:,j) represents the j-th column of the miRNA-disease adjacency matrix A, indicating the association relationship of diseases.

[0091] For each miRNA, its k most relevant similar miRNAs can be obtained from its Gaussian similarity, let m j Let j represent the index of the j-th similar miRNA. The calculated new association vector is shown below:

[0092]

[0093] Where A m (i,:) represents the new association relationship calculated by miRNA in the i-th row, Q m =∑ 1≤j≤K MG(i,m j ) is a normalization term; Where α is the decay term, and its value is α∈[0,1];

[0094] Similarly, a new association vector can be calculated for each disease, using the following formula:

[0095]

[0096] Where A d (:,i) represents the calculated new association in the i-th column, let d j Q represents the index of the j-th similar disease. d =∑ 1≤j≤K DG(i,d j ) is a normalization term; In this context, α is the decay term, and its value is α∈[0,1].

[0097] Finally, the hidden association information of the graph is obtained, and the enhanced hypergraph association matrix is ​​obtained through the following formula:

[0098]

[0099] In step S4, miRNA and disease features are mapped to the same domain space, specifically as follows:

[0100] The feature matrix of miRNA can be represented as [MG,Z] d ,MC,Z d Similarly, the characteristics of a disease can be represented as [Z]. m DG,Z m [,DC], where Z d Z m Let E represent zero matrices with column numbers equal to the disease length and miRNA length, respectively. The features of both miRNA and disease are passed through a linear layer for linear dimensionality reduction, and finally through a ReLU activation function. This yields the miRNA embedding E. m Embedded E of Disease d .

[0101] Step S4 employs a hypergraph convolution method, specifically defined as:

[0102] Based on the miRNA-disease association A, a miRNA hyperedge relationship can be constructed, assuming the initial miRNA node characteristics are as follows:

[0103]

[0104] The hyperedge features of miRNAs are associated with all nodes within the hyperedge. Therefore, the hyperedge features of miRNAs can be derived as follows:

[0105]

[0106] in B represents the node features of the miRNA hypergraph in layer l. -1 For A T The normalized matrix;

[0107] Based on the formula for graph convolution, the following hypergraph convolution formula can be defined:

[0108]

[0109] Expanding the formula yields the following formula:

[0110]

[0111] This allows the features of nodes to flow globally, enabling the training of a data transfer matrix W.

[0112] Similarly, for diseases, we can obtain the features of a disease node.

[0113] In step S5, the embedded features are decoded using a bilinear decoder to calculate their association score. The specific implementation of the bilinear decoder is as follows:

[0114]

[0115] The encoded miRNA and disease embeddings are used as input to the decoder. A bilinear decoder treats different interaction types as different categories. The miRNA m... i and disease d j Correlation score:

[0116]

[0117] Among them W r R is a trainable matrix of parameters whose parameter dimension is the dimension of the input embedding, where r∈R=[0,1].

[0118] To optimize the results, cross-iterative entropy was used as the loss function to calculate the loss between the predicted and actual values. Adam was used as the gradient optimizer to optimize the global network parameters.

[0119] To verify the prediction accuracy of the method (SHGAEMDA) in this embodiment, an experiment was conducted using five-fold cross-validation. First, an equal number of samples with validated miRNA-disease relationships were sampled as negative samples. These samples were then divided into five equal parts, with one part used as the test set each time. A new association matrix was constructed from the original four parts, and its Gaussian and cosine similarities were calculated. The miRNA-disease association score was calculated using the method described in this embodiment, and the area under the mean curve (AUC) was calculated to be 0.9367 ± 0.0011. Within the framework of five-fold cross-validation, compared with other algorithms (GAEMDA 0.9295 ± 0.0028, DFELMDA 0.9139 ± 0.0018, TDRC 0.8884 ± 0.0038), this method significantly outperforms the other comparative methods and is more effective in predicting miRNA-disease correlations. The results are shown in the comparison figure below. Figure 2 As shown.

[0120] To verify the applicability of the method, the SHGAEMDA method was used to predict unknown miRNA-disease associations based on known miRNA-disease associations. In predicting new miRNA-disease associations, known miRNA-disease association information was used as the training dataset for SHGAEMDA. Then, the prediction scores for each unknown miRNA-disease pair were calculated and ranked. Lung cancer, breast cancer, and other diseases were selected as examples.

[0121] As a case study, colorectal cancer was used to validate the top 20 cancer-associated miRNAs in a third-party database (dbDEMC). The results are shown in Tables 1, 2, and 3, with 100%, 85%, and 95% of the predicted miRNAs being associated with cancer, respectively.

[0122] In addition, SHGAEMDA predicted some unproven miRNA-related diseases, including hsa-mir-133 with breast cancer and hsa-mir-29 with colon cancer. These predicted associations have not been reported in the current literature, but there is a high probability of such associations.

[0123] Table 1. Top 20 potential miRNAs associated with lung cancer predicted by SHGAEMDA

[0124]

[0125] Table 2. Top 20 potential miRNAs associated with breast cancer predicted by SHGAEMDA

[0126]

[0127] Table 3. Top 20 potential miRNAs associated with colorectal cancer predicted by SHGAEMDA

[0128]

[0129] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0130] Obviously, the above embodiments of the present invention are merely illustrative examples to clearly illustrate the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A miRNA-disease association prediction method based on an enhanced hypergraph convolutional autoencoder algorithm, which needs to include the following steps: S1: Obtain the miRNA-disease adjacency matrix The association between the two is described using a miRNA-disease adjacency matrix; S2: Through the adjacency matrix The miRNA Gaussian similarity (MG), miRNA cosine similarity (MC), disease Gaussian similarity (DG), and disease cosine similarity (DC) were calculated and used as features of miRNA and disease, respectively. S3: Based on Gaussian similarity, weighted K-nearest neighbors are used to obtain hidden association information in the graph, thereby obtaining an enhanced hypergraph association matrix. This matrix uses miRNA Gaussian similarity (MG) and disease Gaussian similarity (DG), specifically defined as follows: For the original adjacency matrix Perform preprocessing to obtain hidden relationships; define vectors. Represents the miRNA-disease neighbor matrix The Rows represent miRNA associations, vectors represent... Represents the miRNA-disease neighbor matrix The Columns indicate the relationships between diseases; For each miRNA, obtain its k most relevant similar miRNAs from its Gaussian similarity, and let... The first one indicates that it is similar to the second one. The new association vector is calculated based on the sequence number of each miRNA, as shown below: in This indicates the number of miRNAs calculated. Establish new relationships. This is the normalization term for miRNA association; in This is the decay term, and its value is... Similarly, a new association vector can be calculated for each disease, using the following formula: in Represents the calculated first List new relationships, The first one indicates that it is similar to the second one. The serial number of the disease. This is a normalized term representing disease association; in This is the decay term, and its value is... Finally, the hidden association information of the graph is obtained, and the enhanced hypergraph association matrix is ​​obtained through the following formula: S4: Map the features of miRNA and the features of the disease to the same domain space through a fully connected layer, and train the embedded features using hypergraph convolution. S5: Decode the embedded features using a bilinear decoder and calculate their association score.

2. The miRNA-disease association prediction method according to claim 1, characterized in that: The miRNA-disease neighbor matrix in step S1 The specific generation method is as follows: A miRNA-disease association dataset was obtained from the human microRNA disease database, which contained 1206 miRNAs and 893 diseases, totaling 35547 associations. After removing non-human data, 913 miRNAs and 554 diseases were retained. The miRNA-disease adjacency matrix can be represented as follows: ,in and The miRNA-disease adjacency matrix represents the number of miRNAs and the number of diseases, respectively. The definition is as follows: .

3. The miRNA-disease association prediction method according to claim 1, characterized in that: The Gaussian similarity calculation method in step S2 can calculate the Gaussian similarity MG of miRNA, specifically as follows: Calculate the bandwidth for Gaussian interaction kernel similarity: Represents the number of miRNAs; vector Represents the miRNA-disease neighbor matrix The Each row contains a 1, indicating an association between the miRNA and the disease. Calculate the association similarity between each miRNA: Indicates the first The miRNA and the The similarity between miRNAs; exp represents a power function with basis e; vector Represents the miRNA-disease neighbor matrix The Rows, vectors miRNA-disease neighbor matrix The OK.

4. The miRNA-disease association prediction method according to claim 3, characterized in that: The method for calculating the Gaussian similarity (DG) of diseases is the same as that for calculating the Gaussian similarity (MG) of miRNAs, where the matrix used is the miRNA-disease adjacency matrix. The transpose of .

5. The miRNA-disease association prediction method according to claim 1, characterized in that: The cosine similarity of miRNA, MC, is calculated using the cosine similarity calculation method in step S2, specifically as follows: The similarity between two vectors is measured by the cosine angle between them; the smaller the angle, the more similar they are. For miRNA-miRNA associations, the similarity is calculated as follows: Indicates the first The miRNA and the Cosine similarity between miRNAs; vector Represents the miRNA-disease neighbor matrix The Rows, vectors Represents the miRNA-disease neighbor matrix The line, symbol This represents the inner product between two vectors.

6. The miRNA-disease association prediction method according to claim 5, characterized in that: The method for calculating the cosine similarity (DC) of diseases is the same as that for calculating the cosine similarity (MC) of miRNAs, where the matrix used is the miRNA-disease adjacency matrix. The transpose of .

7. The miRNA-disease association prediction method according to claim 1, characterized in that: In step S4, miRNA and disease features are mapped to the same domain space. Specifically, the feature matrix of miRNA features can be represented as follows: Similarly, the characteristics of a disease can be represented as ,in Let each matrix represent a 0 matrix with columns equal to the disease length and the miRNA length, respectively. The features of both the miRNA and the disease are passed through a linear layer for linear dimensionality reduction, and finally through a ReLU activation function to obtain the miRNA embedding. and the embedding of disease .

8. The miRNA-disease association prediction method according to claim 7, characterized in that: The hypergraph convolution method used in step S4 is specifically defined as follows: Based on the association between miRNA and disease Construct a miRNA hyperedge relationship, assuming the initial miRNA node characteristics are as follows: The hyperedge features of miRNAs are associated with all nodes within the hyperedge; therefore, the hyperedge features of miRNAs are: in Indicates the first Node features of the miRNA hypergraph of the layer, for The normalized matrix; The following hypergraph convolution formula is defined based on the graph convolution formula: Expanding the formula yields the following formula: This allows the features of nodes to flow globally, enabling the training of a data transition matrix. Similarly, to obtain the characteristics of a disease node. .

9. The miRNA-disease association prediction method according to claim 1, characterized in that: In step S5, the embedded features are decoded using a bilinear decoder to calculate their association score. The specific implementation of the bilinear decoder is as follows: The encoded miRNA and disease embeddings are used as input to the decoder, and a bilinear decoder is used to treat different interaction types as different categories. Calculate and disease Correlation score: in It is a trainable matrix of parameters, whose parameter dimension is the dimension of the input embedding, where .

10. The miRNA-disease association prediction method according to claim 9, characterized in that: To optimize the results, cross-iterative entropy was used as the loss function to calculate the loss between the predicted results and the true values; Adam was used as the gradient optimizer to optimize the global network parameters.

Images