Bi-directional hypergraph attention method for multi-type association prediction of miRNA-disease

Abstract

The application discloses a bidirectional hypergraph attention method for miRNA-disease multi-type correlation prediction, and belongs to the field of bioinformatics; the known miRNA-disease correlation is divided into different functional types, and a heterogeneous correlation network with a weight attribute is constructed accordingly; multi-order neighbor subgraphs are extracted and corresponding hyperedge representations are constructed; a node-aware attention mechanism is adopted to aggregate local neighborhood information, and a bidirectional hypergraph attention network is innovatively introduced to realize bidirectional information propagation from node to hyperedge embedding and then to node, effectively capturing high-order semantic structures in the network; a gating convolution strategy is used to adaptively fuse and filter node features, enhance the transmission of key features and suppress noise interference; a multi-task joint loss function is used to jointly optimize model representation and prediction performance; the BGMMDA model has reliability and effectiveness in predicting miRNA-disease multi-type correlation.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, and in particular relates to a bidirectional hypergraph attention method for predicting multi-type associations of miRNA diseases. Background Technology

[0002] MicroRNAs (miRNAs) are a class of small, single-stranded RNA molecules that participate in regulating the expression of many genes preserved during evolution. Studies have shown that abnormally high or low expression of miRNAs is closely related to the progression and development of various complex diseases, including cancer, cardiovascular disease, and neurodegenerative diseases. For example, miR-155 is significantly overexpressed in triple-negative breast cancer, promoting tumor cell proliferation and metastasis by inhibiting the tumor suppressor gene SOCS1, while miR-210 induces epithelial-mesenchymal transition in a hypoxic microenvironment, enhancing the invasiveness of breast cancer. Both miRNAs together constitute important biomarkers for breast cancer diagnosis and molecular subtyping. Furthermore, serum miR-155-5p has been found to be a potential biomarker for lumbar disc degeneration in patients with low back pain. Given their important regulatory functions and disease relevance, miRNAs are not only considered a promising class of disease diagnostic biomarkers but also hold promise for becoming a new breakthrough in the development of targeted therapies, making them a highly anticipated frontier area in interdisciplinary research.

[0003] In recent years, researchers have been working on the systematic classification of miRNA-disease associations. Based on this, a series of computational models that integrate multi-type association data have been proposed to more accurately predict potential miRNA-disease associations.

[0004] Existing computational methods can generally be divided into two categories, one of which is based on matrix factorization. Huang et al. proposed a tensor factorization model with association constraints, modeling multi-type miRNA-disease associations as a tensor. This model uses CP decomposition to extract the data matrix from the sparse tensor, while introducing the similarity matrices of miRNAs and diseases as constraints to standardize their respective feature representations, thereby enabling the prediction of miRNA-disease interactions and their specific association types. Ouyang et al. designed the WeightTDAIGN model, which integrates graph Laplacian regularization to preserve local structural information in biological similarity networks, and further applies L2,1-norm constraints to the projection matrix to reduce the influence of redundant information, while fusing multiple similarity networks to improve prediction performance.

[0005] Another category is based on graph neural networks. Yang et al.'s PDMDA model uses a fully connected network to extract miRNA feature representations and a graph neural network to extract disease feature representations, then designs a multi-layer classifier to predict the type of evidence for miRNA-disease associations. Wang et al. developed the NMCMDA framework, which runs an encoder on a heterogeneous miRNA-disease network, uses a graph neural network to learn the latent spatial embeddings of miRNAs and diseases respectively, and uses a decoder to generate prediction scores for miRNA-disease pairs under different evidence types. Although matrix factorization-based methods have shown some effectiveness, their dependence on sparse data and inherent linearity limit their ability to capture complex nonlinear relationships in miRNA-disease associations.

[0006] While existing studies have successfully integrated biological data from multiple sources and improved the predictive performance of miRNA-disease associations, most of these methods neglect a crucial functional dimension: the direction of miRNA expression changes in disease states—upregulation, downregulation, or other. This directional information is essential for understanding the specific mechanisms of miRNA action in pathological processes. For example, downregulation of miR-137 affects lung tumors, while upregulation of let-7a helps to weaken insulin signaling. Therefore, simply exploring the existence of miRNA disease associations while ignoring the specific direction of their expression changes will make it difficult to fully reveal the regulatory mechanisms of miRNAs in disease and will also limit their potential application in targeted therapy development. Furthermore, existing predictive models based on multi-type data generally still frame the problem as a binary classification task, failing to further distinguish the specific types of miRNA roles in disease, which has significant limitations. Summary of the Invention

[0007] The purpose of this invention is to provide a bidirectional hypergraph attention method (BGMMDA) for predicting miRNA-disease multi-type associations. This method addresses the problem in existing technologies that merely explore the existence of miRNA-disease associations while ignoring the specific direction of their expression changes. This approach makes it difficult to fully reveal the regulatory mechanisms of miRNAs in diseases and also limits their application in targeted therapy development.

[0008] To achieve the above objectives, this invention provides a bidirectional hypergraph attention method for predicting miRNA-disease multi-type associations, comprising the following steps: S1. Based on the channel attention mechanism, a weighted fusion of miRNA functional similarity, disease semantic similarity, miRNA sequence similarity, target-based disease similarity, and Gaussian interaction spectrum kernel similarity between miRNA and disease is performed to obtain the integrated similarity between miRNA and disease. S2. Obtain biologically validated multi-type miRNA-disease association data, integrate miRNA ensemble similarity and disease ensemble similarity, and construct a heterogeneous association network between miRNAs and diseases with weighted attributes; extract multi-hop neighbor subgraphs using a hierarchical subgraph sampler. S3. Introduce a node-aware attention mechanism to aggregate local neighborhood information, design a bidirectional hypergraph attention network, construct a bidirectional closed-loop information propagation path of "node-hyperedge-node", and learn discriminative embedding representations through the reverse attention mapping of hyperedges to central nodes; S4. Design a gated convolution module to filter and enhance features, suppress noise interference and highlight useful feature transmission, and perform unified modeling of local feature dependencies and global context information. S5. Reconstruct multi-type relationships of miRNA diseases using MLP; optimize node representation and association prediction by fusing supervised prediction loss function, contrastive learning loss function and regularization loss function based on multi-task joint loss function.

[0009] Preferably, the specific content of S1 is as follows: S101. Treat each similarity measure—miRNA functional similarity, disease semantic similarity, miRNA sequence similarity, target-based disease similarity, and Gaussian interaction spectrum kernel similarity between miRNA and disease—as an independent information channel. Apply a channel attention mechanism to assign different weights to miRNA and disease in multiple similarity views, and output the final weighted similarity by calculating the attention weights. The similarity matrix input is pooled and its dimensionality reduced based on global average pooling to form the information for each channel; Calculate the first Channel information for each sample The expression is as follows: ; In the formula, For the first The similarity matrix of the samples, The dimension of the similarity matrix; The first element in the similarity matrix row and number Column elements; S102. Calculate the attention weight for each entity. The expression is as follows: ; In the formula, and It is a learnable parameter matrix; It is the ReLU activation function; This is the SoftMax activation function; The final attention score expression for each channel is as follows: ; S103. Sum the weighted multi-view similarity matrices to calculate the aggregated miRNA similarity matrix. Disease similarity matrix The expression is as follows: .

[0010] Preferably, the biologically validated multi-type miRNA-disease association data in S2 are obtained from the HMDD v3.2 database. Based on the direction of miRNA expression changes in disease states, the association between miRNA and disease is divided into three types: upregulation, downregulation, and others.

[0011] Preferably, the specific content of S2 is as follows: S201. The aggregated miRNA similarity matrix Disease similarity matrix They are divided into three categories: high similarity, medium similarity, and low similarity. S202. Define the miRNA-disease association weighted heterogeneous network by using relation type encoding. The relational adjacency matrix expression is as follows: ; In the formula, For miRNAs and disease node sets; Let be the set of edges of type, where This is a miRNA-disease association adjacency matrix. This represents miRNA similarity. This indicates a disease similarity relationship. S203, the heterogeneous network with association weighting obtained in step S202, is processed by a hierarchical subgraph sampler respectively. and The expression for extracting multi-hop neighbor subgraphs is as follows: ; In the formula, and It is a node , The set of nodes consisting of itself and its neighbors; and For a given set of associated edges and their corresponding edges... and The set of edges whose attribute information is between nodes.

[0012] Preferably, the specific content of S3 is as follows: S301, For the central node and its neighbor node set Node-aware attention aggregation is used to compute nodes. Its first Jump Neighbor Attention weights between The expression is as follows: ; In the formula, For the modified linear unit activation function with leakage; The embeddings are for the central node and the neighboring nodes, respectively. for and Relational embedding; The weight matrix is ​​a learnable matrix; S302. SoftMax normalize the attention coefficients of all neighbors calculated in S301 to obtain the attention weights. The expression is as follows: ; In the formula, It is an exponential function; For traversing the set Each neighbor node in the array; For nodes The set consisting of all neighboring nodes; Node The self-representation is fused with the aggregated neighbor information to obtain the updated representation. The expression is as follows: ; In the formula, The weight matrix is ​​a learnable matrix; It is a learnable bias vector; S303, the central node and The initial hyperedge embedding is generated by splicing and performing a linear transformation. The expression is as follows: ; In the formula, To modify the activation function of the linear unit; The weight matrix is ​​a learnable matrix; ,in For nodes The initial node embedding representation, For nodes The initial node embedding representation; S304. Design a bidirectional hyperedge attention network. Query information for all first-order neighbors, weight the neighbors, and construct a hyperedge. The final expression The expression is as follows: ; ; In the formula, This is the attention vector; For super-edge The set of first-order neighbors, including the central node , And all first-order neighbor nodes and their correspondences; It is the hyperbolic tangent activation function; For neighboring nodes Embedded representation; For neighboring nodes Relationship type encoding vector; S305, the central node and The embedding is used as a query, and the hyperedge The final representation serves as both key and value. Within each attention head, a similarity score between the query and the key is calculated, and the values ​​are weighted and fused. The outputs of multiple attention heads are concatenated and linearly transformed to integrate semantic information from different representation subspaces. The aggregated result is then added to the original node representation via a residual connection to obtain the enhanced node representation. The expression is as follows: ; In the formula, This is a multi-head attention mechanism.

[0013] Preferably, S4 specifically includes: S401. Extract miRNA-disease pairs from the global association matrix. The corresponding row vector; S402. Design a gated convolution strategy to extract primary feature maps. ; S403, Primary feature map extracted from S402 Pairwise attribute matrices are generated through convolution operations. ,right Perform another convolution operation to generate a gated feature matrix. ; S404, Mapping the primary features The signals are fed into Conv2 and Pool2 to generate the output of the main characteristic path, which is modulated by a gating signal and then... and Forming the enhanced final feature representation The expression is as follows: ; In the formula, For Hadamard product operations; This is a residual connection.

[0014] Preferably, the specific content of S401 is as follows: A global association matrix is ​​constructed by adding the similarity between miRNAs and diseases to the adjacency matrix of the graph. The expression is as follows: ; Extract miRNA-disease pairs from the global correlation matrix The corresponding row vector expression is as follows: ; In the formula, and They are nodes and The strength of association with all other nodes in the network; By superimposing these two attribute features, we obtain the attribute feature matrix. The expression is as follows: .

[0015] Preferably, the specific content of S402 is as follows: Design a gated convolution strategy, including two convolutional pooling layers, Conv1 and Pool1; Extracting primary feature maps The expression is as follows: ; In the formula, Convolution is the operation; MaxPooling is the maximum pooling operation. For convolution kernel; This is the deviation vector.

[0016] Preferably, the specific content of S403 is as follows: Primary feature mapping based on S402 The data is fed into two independent convolutional pooling layers, denoted as Conv2 and Pool2, where convolution operations are used to increase the dimensionality and generate pairwise attribute matrices. ,right Perform another convolution operation to generate a gated feature matrix with the activation function tanh. The expression is as follows: ; ; In the formula, Learnable convolutional kernels are used for... Perform a linear transformation; It is a learnable bias vector used to adjust the output of the linear transformation; Learnable convolutional kernels are used to generate gated feature matrices. ; It is a learnable bias vector used to adjust the output of the gated features.

[0017] Preferably, the specific content of S5 is as follows: S501. Predict the probability that the current hyperedge has a correlation using a multilayer perceptron classifier. The expression is as follows: ; In the formula, To splice according to feature dimensions; is a trainable weight matrix; It is the bias vector; S502, Constructing a multi-class cross-entropy loss function The expression is as follows: ; In the formula, For the first The true label of each sample; For the model to predict the first The sample belongs to the first The probability value of the class; S503. Construct the contrastive loss function The expression is as follows: ; In the formula, and Samples and The predicted vector; This is the margin hyperparameter; and Samples and The corresponding category label; The number of sample pairs sampled; S504, Embedded Vector Regularization Loss Function The expression is as follows: ; In the formula, This is the embedding parameter matrix for the miRNA node; This is the embedding parameter matrix for disease nodes; S505. Construct a multi-task joint loss function by weighted summing of the multi-class cross-entropy loss function, the contrastive loss function, and the embedding vector regularization loss function, as shown in the following expression: ; In the formula, and These are the weight coefficients for the contrastive loss function and the regularized loss function, respectively, used to balance the contributions of each part in model training.

[0018] Therefore, the bidirectional hypergraph attention method for multi-type association prediction of miRNA diseases adopted in this invention has the following beneficial effects: (1) Based on the channel attention mechanism, this invention performs weighted fusion of miRNA functional / sequence similarity, disease semantic / target-based similarity and Gaussian interaction spectrum kernel similarity, assigns differentiated attention weights to different similarity views, effectively integrates multi-source biological feature information, avoids the limitations of a single similarity measure, greatly improves the ability to characterize miRNA and disease features, and makes the integrated similarity more in line with the actual biological association law; (2) This invention first aggregates local neighborhood information through node-aware attention, and then realizes bidirectional information transmission through the reverse attention mapping of hyperedge to the central node. At the same time, it combines hierarchical subgraph sampling to extract multi-hop neighbor subgraphs, which breaks through the limitation of traditional graph models that can only capture low-order neighborhood relationships, effectively mines the high-order semantic structure in miRNA-disease heterogeneous association network, and learns more discriminative node and association embedding representations. (3) The gated convolution module designed in this invention extracts features through multiple rounds of convolutional pooling and modulates the main feature pathway using the gated feature matrix. It combines the Hadamard product with residual connection to achieve adaptive enhancement of useful features and effective suppression of noisy features. At the same time, it extracts miRNA-disease pair features from the global correlation matrix and completes the unified modeling of local features and global context information, thereby improving the effectiveness and robustness of the features. (4) This invention subdivides miRNA-disease association into three categories: upregulation, downregulation, and other, according to the direction of expression change in disease state. This makes up for the limitation of existing models that simplify association prediction to a binary classification task, and is more in line with the actual regulatory mechanism of miRNA in disease. At the same time, it integrates multi-class cross-entropy loss, contrastive learning loss and regularization loss to construct a multi-task joint loss function. It ensures the accuracy of prediction through supervised loss, achieves "compact intra-class and separated inter-class" feature space through contrastive loss, and prevents model overfitting by using regularization loss. It optimizes node representation and association prediction effect in multiple dimensions.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a flowchart of the bidirectional hypergraph attention method for predicting multiple types of miRNA disease associations in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0022] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0023] The following is in conjunction with the appendix Figure 1 The embodiments of the present invention will be described in detail below.

[0024] Example 1 A bidirectional hypergraph attention method for predicting multi-type associations of miRNAs with diseases includes the following steps: S1. Based on the channel attention mechanism, a weighted fusion of miRNA functional similarity, disease semantic similarity, miRNA sequence similarity, target-based disease similarity, and Gaussian interaction spectrum kernel similarity between miRNA and disease is performed to obtain the integrated similarity between miRNA and disease. S101. Treat each similarity measure—miRNA functional similarity, disease semantic similarity, miRNA sequence similarity, target-based disease similarity, and Gaussian interaction spectrum kernel similarity between miRNA and disease—as an independent information channel. Apply a channel attention mechanism to assign different weights to miRNA and disease in multiple similarity views, and output the final weighted similarity by calculating the attention weights. The similarity matrix input is pooled and its dimensionality reduced based on global average pooling to form the information for each channel; Calculate the first Channel information for each sample The expression is as follows: ; In the formula, For the first The similarity matrix of the samples, The dimension of the similarity matrix; The first element in the similarity matrix row and number Column elements; S102. Calculate the attention weight for each entity. The expression is as follows: ; In the formula, and It is a learnable parameter matrix; It is the ReLU activation function; This is the SoftMax activation function; The final attention score expression for each channel is as follows: ; S103. Sum the weighted multi-view similarity matrices to calculate the aggregated miRNA similarity matrix. Disease similarity matrix The expression is as follows: .

[0025] Preferably, the biologically validated multi-type miRNA-disease association data in S2 are obtained from the HMDD v3.2 database. Based on the direction of miRNA expression changes in disease states, the association between miRNA and disease is divided into three types: upregulation, downregulation, and others.

[0026] S2. Obtain biologically validated multi-type miRNA-disease association data, integrate miRNA ensemble similarity and disease ensemble similarity, and construct a heterogeneous association network between miRNAs and diseases with weighted attributes; extract multi-hop neighbor subgraphs using a hierarchical subgraph sampler. S201. The aggregated miRNA similarity matrix Disease similarity matrix They are divided into three categories: high similarity, medium similarity, and low similarity. S202. Define the miRNA-disease association weighted heterogeneous network by using relation type encoding. The relational adjacency matrix expression is as follows: ; In the formula, For miRNAs and disease node sets; Let be the set of edges of type, where This is a miRNA-disease association adjacency matrix. This represents miRNA similarity. This indicates a disease similarity relationship. S203, the heterogeneous network with association weighting obtained in step S202, is processed by a hierarchical subgraph sampler respectively. and The expression for extracting multi-hop neighbor subgraphs is as follows: ; In the formula, and It is a node , The set of nodes consisting of itself and its neighbors; and For a given set of associated edges and their corresponding edges... and The set of edges whose attribute information is between nodes.

[0027] S3. Introduce a node-aware attention mechanism to aggregate local neighborhood information, design a bidirectional hypergraph attention network, construct a bidirectional closed-loop information propagation path of "node-hyperedge-node", and learn discriminative embedding representations through the reverse attention mapping of hyperedges to central nodes; S301, For the central node and its neighbor node set Node-aware attention aggregation is used to compute nodes. Its first Jump Neighbor Attention weights between The expression is as follows: ; In the formula, For the modified linear unit activation function with leakage; The embeddings are for the central node and the neighboring nodes, respectively. for and Relational embedding; The weight matrix is ​​a learnable matrix; S302. SoftMax normalize the attention coefficients of all neighbors calculated in S301 to obtain the attention weights. The expression is as follows: ; In the formula, It is an exponential function; For traversing the set Each neighbor node in the array; For nodes The set consisting of all neighboring nodes; Node The self-representation is fused with the aggregated neighbor information to obtain the updated representation. The expression is as follows: ; In the formula, The weight matrix is ​​a learnable matrix; It is a learnable bias vector; S303, the central node and The initial hyperedge embedding is generated by splicing and performing a linear transformation. The expression is as follows: ; In the formula, To modify the activation function of the linear unit; The weight matrix is ​​a learnable matrix; ,in For nodes The initial node embedding representation, For nodes The initial node embedding representation; S304. Design a bidirectional hyperedge attention network. Query information for all first-order neighbors, weight the neighbors, and construct a hyperedge. The final expression The expression is as follows: ; ; In the formula, This is the attention vector; For super-edge The set of first-order neighbors, including the central node , And all first-order neighbor nodes and their correspondences; It is the hyperbolic tangent activation function; For neighboring nodes Embedded representation; For neighboring nodes Relationship type encoding vector; S305, the central node and The embedding is used as a query, and the hyperedge The final representation serves as both key and value. Within each attention head, a similarity score between the query and the key is calculated, and the values ​​are weighted and fused. The outputs of multiple attention heads are concatenated and linearly transformed to integrate semantic information from different representation subspaces. The aggregated result is then added to the original node representation via a residual connection to obtain the enhanced node representation. The expression is as follows: ; In the formula, This is a multi-head attention mechanism.

[0028] S4. Design a gated convolution module to filter and enhance features, suppress noise interference and highlight useful feature transmission, and perform unified modeling of local feature dependencies and global context information. S401. Extract miRNA-disease pairs from the global association matrix. The corresponding row vector; A global association matrix is ​​constructed by adding the similarity between miRNAs and diseases to the adjacency matrix of the graph. The expression is as follows: ; Extract miRNA-disease pairs from the global correlation matrix The corresponding row vector expression is as follows: ; In the formula, and They are nodes and The strength of association with all other nodes in the network; By superimposing these two attribute features, we obtain the attribute feature matrix. The expression is as follows: .

[0029] S402. Design a gated convolution strategy to extract primary feature maps. ; Design a gated convolution strategy, including two convolutional pooling layers, Conv1 and Pool1; Extracting primary feature maps The expression is as follows: ; In the formula, Convolution is the operation; MaxPooling is the maximum pooling operation. For convolution kernel; This is the deviation vector.

[0030] S403, Primary feature map extracted from S402 Pairwise attribute matrices are generated through convolution operations. ,right Perform another convolution operation to generate a gated feature matrix. ; Primary feature mapping based on S402 The data is fed into two independent convolutional pooling layers, denoted as Conv2 and Pool2, where convolution operations are used to increase the dimensionality and generate pairwise attribute matrices. ,right Perform another convolution operation to generate a gated feature matrix with the activation function tanh. The expression is as follows: ; ; In the formula, Learnable convolutional kernels are used for... Perform a linear transformation; It is a learnable bias vector used to adjust the output of the linear transformation; Learnable convolutional kernels are used to generate gated feature matrices. ; It is a learnable bias vector used to adjust the output of the gated features.

[0031] S404, Mapping the primary features The signals are fed into Conv2 and Pool2 to generate the output of the main characteristic path, which is modulated by a gating signal and then... and Forming the enhanced final feature representation The expression is as follows: ; In the formula, For Hadamard product operations; This is a residual connection.

[0032] S5. Reconstruct multi-type relationships of miRNA diseases using MLP; optimize node representation and association prediction by fusing supervised prediction loss function, contrastive learning loss function and regularization loss function based on multi-task joint loss function.

[0033] S501. Predict the probability that the current hyperedge has a correlation using a multilayer perceptron classifier. The expression is as follows: ; In the formula, To splice according to feature dimensions; is a trainable weight matrix; It is the bias vector; S502, Constructing a multi-class cross-entropy loss function The expression is as follows: ; In the formula, For the first The true label of each sample; For the model to predict the first The sample belongs to the first The probability value of the class; S503. Construct the contrastive loss function The expression is as follows: ; In the formula, and Samples and The predicted vector; This is the margin hyperparameter; and Samples and The corresponding category label; The number of sample pairs sampled; S504, Embedded Vector Regularization Loss Function The expression is as follows: ; In the formula, This is the embedding parameter matrix for the miRNA node; This is the embedding parameter matrix for disease nodes; S505. Construct a multi-task joint loss function by weighted summing of the multi-class cross-entropy loss function, the contrastive loss function, and the embedding vector regularization loss function, as shown in the following expression: ; In the formula, and These are the weight coefficients for the contrastive loss function and the regularized loss function, respectively, used to balance the contributions of each part in model training.

[0034] Example 2 To verify the effectiveness of the present invention, this embodiment compares the association matrix predicted by the prediction model of the present invention with the actual association labels, and evaluates it using two types of five-fold cross-validation: CVtriplet and CVtype. Under the CVtriplet experiment, AUC, AUPR and F1-score values ​​are calculated, and under the CVtype experiment, Top-1 precision, Top-1 recall and Top-1 F1-score values ​​are calculated and compared with existing methods.

[0035] Table 1 shows the comparative experimental results of the bidirectional hypergraph attention method for miRNA-disease multi-type association prediction of this invention with five other excellent miRNA-disease multi-type association prediction models in CVtriplet and CVtype experiments, including TDRC, SPLDHyperAWNTF, MRFGMDA, HHAWMD, and MSHGANMDA. The prediction results show that, under the CVtype and CVtriplet experimental settings, this invention consistently outperforms all other comparative models on all six evaluation metrics, effectively improving the performance of miRNA-disease multi-type association prediction and demonstrating high practicality.

[0036] Table 1. Comparison of BGMMDA with five other prediction models in experimental results.

[0037] To evaluate the ability of BGMMDA to predict novel miRNA-disease multitype associations, all validated associations between miRNAs and diseases were used to train the model. Predictions were made for all unknown association pairs in the dataset. All miRNAs predicted to be disease-associated were sorted in descending order of prediction score, and the top 15 candidate miRNAs were selected for analysis. Table 2 shows the results of a case study experiment on lung cancer using the bidirectional hypergraph attention method for predicting miRNA-disease multitype associations of this invention. Of the top 15 lung cancer-associated miRNAs predicted by BGMMDA, 13 could be verified by relevant literature. The results of this case study can provide biologists with candidate miRNAs related to diseases, facilitating wet experiments to some extent.

[0038] Table 2. Case study results of BGMMDA in lung cancer.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A bidirectional hypergraph attention method for predicting miRNA-disease multi-type associations, characterized in that, Includes the following steps: S1. Based on the channel attention mechanism, a weighted fusion of miRNA functional similarity, disease semantic similarity, miRNA sequence similarity, target-based disease similarity, and Gaussian interaction spectrum kernel similarity between miRNA and disease is performed to obtain the integrated similarity between miRNA and disease. S2. Obtain biologically validated multi-type miRNA-disease association data, integrate miRNA ensemble similarity and disease ensemble similarity, and construct a heterogeneous association network between miRNAs and diseases with weighted attributes; extract multi-hop neighbor subgraphs using a hierarchical subgraph sampler. S3. Introduce a node-aware attention mechanism to aggregate local neighborhood information, design a bidirectional hypergraph attention network, construct a bidirectional closed-loop information propagation path of "node-hyperedge-node", and learn discriminative embedding representations through the reverse attention mapping of hyperedges to central nodes; S4. Design a gated convolution module to filter and enhance features, suppress noise interference and highlight useful feature transmission, and perform unified modeling of local feature dependencies and global context information. S5. Reconstruct multi-type relationships of miRNA diseases using MLP; optimize node representation and association prediction by fusing supervised prediction loss function, contrastive learning loss function and regularization loss function based on multi-task joint loss function.

2. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 1, characterized in that, The specific content of S1 is as follows: S101. Treat each similarity measure—miRNA functional similarity, disease semantic similarity, miRNA sequence similarity, target-based disease similarity, and Gaussian interaction spectrum kernel similarity between miRNA and disease—as an independent information channel. Apply a channel attention mechanism to assign different weights to miRNA and disease in multiple similarity views, and output the final weighted similarity by calculating the attention weights. The similarity matrix input is pooled and its dimensionality reduced based on global average pooling to form the information for each channel; Calculate the first Channel information for each sample The expression is as follows: ； In the formula, For the first The similarity matrix of the samples, The dimension of the similarity matrix; The first element in the similarity matrix row and number Column elements; S102. Calculate the attention weight for each entity. The expression is as follows: ； In the formula, and It is a learnable parameter matrix; It is the ReLU activation function; This is the SoftMax activation function; The final attention score expression for each channel is as follows: ； S103. Sum the weighted multi-view similarity matrices to calculate the aggregated miRNA similarity matrix. Disease similarity matrix The expression is as follows: 。 3. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 1, characterized in that, The biologically validated multi-type miRNA-disease association data in S2 were obtained from the HMDD v3.2 database. Based on the direction of miRNA expression changes in disease states, the association between miRNA and disease was divided into three types: upregulation, downregulation, and others.

4. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, The specific content of S2 is as follows: S201. The aggregated miRNA similarity matrix Disease similarity matrix They are divided into three categories: high similarity, medium similarity, and low similarity. S202. Define the miRNA-disease association weighted heterogeneous network by using relation type encoding. The relational adjacency matrix expression is as follows: ； In the formula, For miRNAs and disease node sets; Let be the set of edges of type, where This is a miRNA-disease association adjacency matrix. This represents miRNA similarity. This indicates a disease similarity relationship. S203, the heterogeneous network with association weighting obtained in step S202, is processed by a hierarchical subgraph sampler respectively. and The expression for extracting multi-hop neighbor subgraphs is as follows: ； In the formula, and It is a node , The set of nodes consisting of itself and its neighbors; and For a given set of associated edges and their corresponding edges... and The set of edges whose attribute information is between nodes.

5. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, The specific details of S3 are as follows: S301, For the central node and its neighbor node set Node-aware attention aggregation is used to compute nodes. Its first Jump Neighbor Attention weights between The expression is as follows: ； In the formula, For the modified linear unit activation function with leakage; The embeddings are for the central node and the neighboring nodes, respectively. for and Relational embedding; The weight matrix is ​​a learnable matrix; S302. SoftMax normalize the attention coefficients of all neighbors calculated in S301 to obtain the attention weights. The expression is as follows: ； In the formula, It is an exponential function; For traversing the set Each neighbor node in the array; For nodes The set consisting of all neighboring nodes; Node The self-representation is fused with the aggregated neighbor information to obtain the updated representation. The expression is as follows: ； In the formula, The weight matrix is ​​a learnable matrix; It is a learnable bias vector; S303, the central node and The initial hyperedge embedding is generated by splicing and performing a linear transformation. The expression is as follows: ； In the formula, To modify the activation function of the linear unit; The weight matrix is ​​a learnable matrix; ,in For nodes The initial node embedding representation, For nodes The initial node embedding representation; S304. Design a bidirectional hyperedge attention network. Query information for all first-order neighbors, weight the neighbors, and construct a hyperedge. The final expression The expression is as follows: ； ； In the formula, This is the attention vector; For super-edge The set of first-order neighbors, including the central node , And all first-order neighbor nodes and their correspondences; It is the hyperbolic tangent activation function; For neighboring nodes Embedded representation; For neighboring nodes Relationship type encoding vector; S305, the central node and The embedding is used as a query, and the hyperedge The final representation serves as both key and value. Within each attention head, a similarity score between the query and the key is calculated, and the values ​​are weighted and fused. The outputs of multiple attention heads are concatenated and linearly transformed to integrate semantic information from different representation subspaces. The aggregated result is then added to the original node representation via a residual connection to obtain the enhanced node representation. The expression is as follows: ； In the formula, This is a multi-head attention mechanism.

6. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, S4 specifically includes: S401. Extract miRNA-disease pairs from the global association matrix. The corresponding row vector; S402. Design a gated convolution strategy to extract primary feature maps. ; S403, Primary feature map extracted from S402 Pairwise attribute matrices are generated through convolution operations. ,right Perform another convolution operation to generate a gated feature matrix. ; S404, Mapping the primary features The signals are fed into Conv2 and Pool2 to generate the output of the main characteristic path, which is modulated by a gating signal and then... and Forming the enhanced final feature representation The expression is as follows: ； In the formula, For Hadamard product operations; This is a residual connection.

7. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, The specific details of S401 are as follows: A global association matrix is ​​constructed by adding the similarity between miRNAs and diseases to the adjacency matrix of the graph. The expression is as follows: ； Extract miRNA-disease pairs from the global correlation matrix The corresponding row vector expression is as follows: ； In the formula, and They are nodes and The strength of association with all other nodes in the network; By superimposing these two attribute features, we obtain the attribute feature matrix. The expression is as follows: 。 8. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, The specific details of S402 are as follows: Design a gated convolution strategy, including two convolutional pooling layers, Conv1 and Pool1; Extracting primary feature maps The expression is as follows: ； In the formula, Convolution is the operation; MaxPooling is the maximum pooling operation. For convolution kernel; This is the deviation vector.

9. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, The specific details of S403 are as follows: Primary feature mapping based on S402 The data is fed into two independent convolutional pooling layers, denoted as Conv2 and Pool2, where convolution operations are used to increase the dimensionality and generate pairwise attribute matrices. ,right Perform another convolution operation to generate a gated feature matrix with the activation function tanh. The expression is as follows: ； ； In the formula, Learnable convolutional kernels are used for... Perform a linear transformation; It is a learnable bias vector used to adjust the output of the linear transformation; Learnable convolutional kernels are used to generate gated feature matrices. ; It is a learnable bias vector used to adjust the output of the gated features.

10. The bidirectional hypergraph attention method for miRNA-disease multi-type association prediction according to claim 2, characterized in that, The specific details of S5 are as follows: S501. Predict the probability that the current hyperedge has a correlation using a multilayer perceptron classifier. The expression is as follows: ； In the formula, To splice according to feature dimensions; is a trainable weight matrix; It is the bias vector; S502, Constructing a multi-class cross-entropy loss function The expression is as follows: ； In the formula, For the first The true label of each sample; For the model to predict the first The sample belongs to the first The probability value of the class; S503. Construct the contrastive loss function The expression is as follows: ； In the formula, and Samples and The predicted vector; This is the margin hyperparameter; and Samples and The corresponding category label; The number of sample pairs sampled; S504, Embedded Vector Regularization Loss Function The expression is as follows: ； In the formula, This is the embedding parameter matrix for the miRNA node; This is the embedding parameter matrix for disease nodes; S505. Construct a multi-task joint loss function by weighted summing of the multi-class cross-entropy loss function, the contrastive loss function, and the embedding vector regularization loss function, as shown in the following expression: ； In the formula, and These are the weight coefficients for the contrastive loss function and the regularized loss function, respectively, used to balance the contributions of each part in model training.

Images