MiRNA-disease relationship prediction method, system and model based on hybrid expert model and storage medium

Abstract

The application provides a miRNA-disease relationship prediction method, system and model based on a hybrid expert model and a storage medium. The method comprises the following steps: obtaining a miRNA-disease correlation matrix of multi-omics data of a miRNA-disease to be predicted, a miRNA similarity matrix and a disease similarity matrix; constructing a miRNA-disease heterogeneous graph, a miRNA homogeneous subgraph and a disease homogeneous subgraph, and fusing them into a multi-modal biological graph network; inputting the fusion network into a pre-trained gated hybrid multi-expert network model to output a correlation probability of a miRNA-disease pair to be predicted; and the gated hybrid multi-expert network model comprises multiple expert networks, a gating network and an output layer. The application adopts a gating mechanism to adjust the weights of each expert for miRNA-disease prediction, maintains high prediction accuracy on a small-scale data set, and has the advantages of portability and light weight.

Description

Technical Field

[0001] This application relates to the field of biological information technology, and in particular to a method, system, model, and computer program product for predicting miRNA-disease relationships based on a hybrid expert model. Background Technology

[0002] MicroRNAs (miRNAs), a class of non-coding RNA molecules approximately 22 nucleotides in length, play a crucial role in cell proliferation, differentiation, and disease development (such as cancer, cardiovascular disease, and neurodegenerative diseases) by regulating gene expression through targeting mRNA. Revealing the associations between miRNAs and diseases (miRNA-disease associations, MDAs) not only helps elucidate the molecular mechanisms of diseases but also provides new targets for drug development and precision medicine targeting miRNAs. In recent years, with the development of high-throughput sequencing technology, multiple databases (such as miR2Disease and HMDD) have systematically integrated experimentally validated MDA data, providing abundant resources for computational model training. However, experimental validation (such as luciferase reporter gene assays and microarray sequencing) has limitations such as high cost, long cycle time, and low throughput, making it difficult to meet the needs of large-scale MDA prediction. Therefore, developing efficient computational models has become a core challenge in this field.

[0003] Existing MDA prediction methods are mainly divided into two categories: methods based on traditional machine learning and methods based on deep learning. Traditional methods (such as the RWRMDA model based on random walk proposed by Chen et al., which uses miRNA functional similarity and disease semantic similarity to construct heterogeneous graphs and captures potential associations through local walks) and matrix factorization-based methods (such as the MCMDA model of Jiang et al., which integrates multi-source similarity matrices and predicts associations through co-factor matrix decomposition) can capture some features, but they have the following limitations: they rely on manually designed features, making it difficult to capture high-dimensional nonlinear relationships; and they cannot dynamically integrate different modalities of data (such as miRNA sequence features, disease phenotypic features, and gene expression data).

[0004] In recent years, graph neural networks (GNNs) have been widely used because they can directly model biological entities (miRNAs, diseases, genes, etc.) and their relationships (such as regulation and association). For example, the GCNMDA model proposed by Li et al. constructs heterogeneous graphs by using miRNAs and diseases as nodes, and predicts associations after extracting node embeddings through GCN; the GATMDA model by Wang et al. introduces an attention mechanism to assign different weights to miRNAs' neighboring diseases to capture key local associations. Although these methods have made progress in MDA prediction, they still have the following shortcomings: a single GNN model (such as GCN or GAT) is only good at capturing specific types of graph features (such as GCN being good at global smooth features, and GAT being good at local attention features), and cannot fully cover the multi-dimensional characteristics of heterogeneous graphs; the importance of different subgraphs (such as miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heterogeneous graphs) changes dynamically in the prediction task, but existing models use fixed weights to fuse features, resulting in the weakening of key information. Summary of the Invention

[0005] The purpose of this application is to provide a method, system, computer-readable storage medium, and computer program product for predicting miRNA-disease relationships based on a hybrid expert model, thereby solving the aforementioned problems existing in the prior art.

[0006] To achieve the above objectives, the technical solution adopted in this application is as follows:

[0007] This application provides a method for predicting miRNA-disease relationships based on a hybrid expert model, the method comprising:

[0008] Based on the multi-omics data of miRNA-disease to be predicted, obtain the miRNA-disease association matrix, miRNA similarity matrix and disease similarity matrix. The miRNA-disease association matrix includes known miRNA-disease associations and unknown miRNA-disease associations.

[0009] Based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, miRNA-disease heteromorphism map, miRNA isomorphism sub-map, and disease isomorphism sub-map are constructed respectively, and the three maps are fused into a multimodal biological atlas network;

[0010] The multimodal biomap network is input into a pre-trained gated hybrid multi-expert network model, which outputs the association probability of the miRNA-disease pair to be predicted. The gated hybrid multi-expert network model includes: multiple structurally distinct expert networks used to extract features from the multimodal biomap network from different feature perspectives and generate node embeddings; a gated network used to dynamically calculate the weights of each expert network based on the embedding features and select the Top-M expert networks based on the weights; and an output layer used to fuse the node embeddings generated by the selected expert networks and then process them to output the prediction result.

[0011] Optionally, the gated hybrid multi-expert network model includes: at least two expert networks of the following network types; a multi-hop graph convolutional network, including: a three-layer graph convolutional GCN that aggregates hop neighbor information layer by layer, with each layer followed by Dropout and residual connections; a graph sampling aggregation network, including: three layers of GraphSAGEConv convolutional layers, each layer using a mean aggregator, followed by ELU activation and Dropout; an autoencoder network, including: an encoder and a decoder, each being a three-layer fully connected network, the encoder extracting low-dimensional latent representations, and the decoder reconstructing input features; and a multilayer perceptron, including: a three-layer fully connected network, with each layer followed by ELU activation and Dropout, used for nonlinear feature transformation, independent of graph structure information.

[0012] Optionally, the gating network dynamically calculates the weights of each expert network based on the input features, including:

[0013] The initial weights of each expert network are calculated using a multilayer perceptron based on the node embedding features learned from the multimodal biological map network.

[0014] By integrating the dynamic performance evaluation metrics of each expert network with the initial weights, a comprehensive weight is generated; wherein, the dynamic performance evaluation metrics are based at least on the statistical characteristics of the current output and historical performance of each expert network.

[0015] Based on the comprehensive weights, the expert networks corresponding to the Top-K weights with the largest weight values ​​are selected as the currently selected expert networks, where K is a positive integer greater than or equal to 1;

[0016] The node embeddings output by the selected expert network are weighted and fused to obtain the final node embedding.

[0017] Optionally, the step of obtaining the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix based on the multi-omics data of the miRNA-disease to be predicted includes:

[0018] Based on the multi-omics data of miRNA-disease to be predicted, obtain the miRNA-disease association matrix, which includes known associations and unknown associations;

[0019] Based on the miRNA-disease association matrix, the Gaussian similarity matrix between miRNA and disease is calculated. Based on the functional information of miRNA and the semantic information of disease in the multi-omics data of miRNA-disease to be predicted, the functional similarity of miRNA and the semantic similarity of disease are calculated respectively.

[0020] The miRNA Gaussian similarity matrix and the functional similarity matrix were weighted and fused separately to form the miRNA similarity matrix, and the disease Gaussian similarity matrix and the semantic similarity matrix were weighted and fused separately to form the disease similarity matrix.

[0021] Optionally, the construction of miRNA-disease isomorphism maps, miRNA isomorphism sub-maps, and disease isomorphism sub-maps based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix respectively includes:

[0022] Using the miRNA-disease association matrix as the adjacency matrix, and miRNA and disease as two types of nodes, a miRNA-disease heterogeneous graph is constructed. The edges between miRNA nodes and disease nodes are provided by the known miRNA-disease association matrix. In the matrix, 1 indicates that there is an association and 0 indicates that there is no association.

[0023] Using the miRNA similarity matrix as the adjacency matrix and miRNAs as nodes, a miRNA isomorphic subgraph is constructed, wherein the similarity between miRNAs is used as the weight of the edges in the miRNA isomorphic subgraph.

[0024] Using the disease similarity matrix as the adjacency matrix and diseases as nodes, a disease isomorphic subgraph is constructed, where the similarity between diseases is used as the weight of the edges in the disease isomorphic subgraph.

[0025] Optionally, the training process of the gated hybrid multi-expert network model includes:

[0026] Based on the miRNA-disease association dataset, construct the miRNA-disease association matrix;

[0027] Calculate the Gaussian similarity matrix between miRNA and disease based on the association matrix. Calculate the functional similarity of miRNA and the semantic similarity of disease based on the functional information of miRNA and the semantic information of disease in the miRNA-disease association dataset.

[0028] The miRNA Gaussian similarity matrix and the miRNA functional similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the miRNA. The disease Gaussian similarity matrix and the disease semantic similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the disease.

[0029] The miRNA-disease association matrix of miRNA and disease is used as positive samples. Negative samples are extracted from miRNA-disease with unknown associations through hard negative sampling to form a labeled training sample set. The negative samples are those that are more similar to the positive samples in terms of features or topology.

[0030] Using known associated positive samples as edges, miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heteromorphic graphs are constructed by combining the similarity matrix after miRNA integration, the similarity matrix after disease integration, and the known association between miRNA and disease, and then fused into a multimodal biological atlas network.

[0031] The fused multimodal biological map network and the initial features of miRNA and disease nodes are input into the gated hybrid multi-expert network model to be trained, and the gated hybrid multi-expert network model is obtained.

[0032] Optionally, the hard negative sample sampling process includes:

[0033] Traverse the miRNA-disease association matrix to separate all positive samples and all possible negative samples; where miRNA-disease pairs with known associations are positive samples, and miRNA-disease pairs without associations are possible negative samples.

[0034] The similarity between the current miRNA in the negative sample and all positive miRNAs of the target disease is summed, as well as the similarity between the current disease in the negative sample and all positive diseases of the target miRNA, to calculate the association probability score between each negative sample and the positive samples. The association probability score represents the overall similarity between the negative sample and all related positive samples.

[0035] Sort all negative samples in descending order of score, and select the negative sample with the highest score as the hard negative sample.

[0036] This application also provides a miRNA-disease relationship prediction system based on a hybrid expert model, the system comprising:

[0037] The matrix construction module is used to obtain the miRNA-disease association matrix, miRNA similarity matrix and disease similarity matrix based on the multi-omics data of the miRNA-disease to be predicted. The miRNA-disease association matrix includes known miRNA-disease associations and unknown miRNA-disease associations.

[0038] The graph network construction module is used to construct miRNA-disease heteromorphism graphs, miRNA isomorphism subgraphs, and disease isomorphism subgraphs based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, respectively, and combine the three graphs into a multimodal biological graph network.

[0039] The prediction module is used to input multi-source heterogeneous features into a pre-trained gated hybrid multi-expert network model and output the association probability of the miRNA-disease pair to be predicted. The gated hybrid multi-expert network model includes: multiple structurally distinct expert networks for generating node embeddings from different feature perspectives; a gated network for dynamically calculating the weights of each expert network based on the input features and selecting the Top-M expert networks based on the weights; and an output layer for fusing and processing the node embeddings generated by the selected expert networks to output the prediction result.

[0040] This application also provides a gated hybrid multi-expert network model, which is the gated hybrid multi-expert network model in the method described above, and the gated hybrid multi-expert network model is used for miRNA and disease association prediction.

[0041] In another aspect, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0042] The beneficial effects of this application are:

[0043] The prediction method in this application adopts a Mixture-of-Experts architecture. Based on the differentiated advantages of different expert models in handling specific feature combinations, the optimal K experts are dynamically selected to participate in inference through a gating network. This avoids interference from redundant experts, improves the modeling ability and generalization ability of the model when facing the diverse semantic relationships between miRNA and disease, and enhances the prediction results. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating a miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application.

[0045] Figure 2 This is the training process of a gated hybrid multi-expert network model in the miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application.

[0046] Figure 3 This is a diagram of the multi-hop graph convolutional network structure of the gated hybrid multi-expert network model in the miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application;

[0047] Figure 4 This is a graph sampling aggregation network structure diagram of the gated hybrid multi-expert network model in the miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application;

[0048] Figure 5 This is a diagram of the autoencoder network structure of the gated hybrid multi-expert network model in the miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application;

[0049] Figure 6 This is a diagram of the multilayer perceptron structure of the gated hybrid multi-expert network model in the miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application;

[0050] Figure 7 This is a line graph showing the prediction performance of the miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application on two datasets, HMDDv3.2 and HMDDv2.0.

[0051] Figure 8 This is a schematic diagram of a five-fold cross-validation experiment demonstrating a miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application;

[0052] Figure 9 This is a heatmap of expert weights on different subgraphs after the gating network selects TOPT-2 experts in a miRNA-disease relationship prediction method based on a hybrid expert model proposed in this application. Among them, (a) is a miRNA isomorphic subgraph, (b) is a disease isomorphic subgraph, and (c) is a miRNA-disease heteromorphic subgraph.

[0053] Figure 10 This is a schematic diagram of the miRNA-disease relationship prediction system based on a hybrid expert model proposed in this application. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0055] like Figure 1 As shown, this application proposes a miRNA-disease relationship prediction method based on a hybrid expert model, the method comprising:

[0056] Step S100: Based on the multi-omics data of the miRNA-disease to be predicted, obtain the miRNA-disease association matrix, the miRNA similarity matrix, and the disease similarity matrix. The miRNA-disease association matrix includes known miRNA-disease associations and unknown miRNA-disease associations.

[0057] Before step S100 of this application, multi-omics data of miRNA-disease to be predicted are first obtained. Based on the multi-omics data of miRNA-disease to be predicted, a miRNA-disease association matrix, a miRNA similarity matrix, and a disease similarity matrix are constructed. Corresponding subgraphs are constructed based on the three matrices respectively.

[0058] The miRNA similarity matrix includes a weighted combination of the miRNA Gaussian kernel similarity matrix and the miRNA functional similarity matrix; the disease similarity matrix includes a weighted combination of the disease Gaussian kernel similarity matrix and the disease semantic similarity matrix.

[0059] As an example, the process of obtaining the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix based on the multi-omics data of the miRNA-disease to be predicted includes:

[0060] Step S110: Based on the multi-omics data of the miRNA-disease to be predicted, obtain the miRNA-disease association matrix, which includes known associations and unknown associations.

[0061] This application first obtains multi-omics data on miRNA-disease associations to be predicted. The multi-omics data on miRNA-disease associations to be predicted includes miRNA-disease association matrices with known associations and unknown associations. MiRNA-disease associations with unknown associations need to be predicted by the model.

[0062] The miRNA-disease association matrix refers to the association between miRNAs and diseases, which can be used to... A express, A ij Represents the first in the matrix i line, number j The elements of the column can be either 0 or 1. A ij =1 indicates the first i miRNA m i and the j a disease j There is a known association. A ij =0 indicates the first i miRNA mi and the j a disease j The relationship is unknown, which means the association needs to be predicted. A ij This represents the specific value in the i-th row and j-th column of matrix A. Each row corresponds to one miRNA, and each column corresponds to one disease. The values ​​in the matrix represent the association between miRNA i and disease j.

[0063] Step S120: Calculate the miRNA Gaussian interaction kernel similarity matrix and the disease Gaussian interaction kernel similarity matrix based on the miRNA-disease association matrix. Calculate the miRNA functional similarity and disease semantic similarity based on the functional information of the miRNA and the semantic information of the disease in the multi-omics data of the miRNA-disease to be predicted.

[0064] The calculation of the miRNA Gaussian similarity matrix and the disease Gaussian similarity matrix is ​​based on the interaction spectrum formed by all elements in the miRNA-disease association matrix A. This interaction spectrum is a binary vector, where '1' represents the existence of a known association and '0' represents the absence of a known association (i.e., the association state is unknown or has been verified to be non-existent). This calculation reflects the topological similarity of nodes in the entire association network (including known associations and unknown states).

[0065] Based on the miRNA-disease association matrix, calculate the miRNA Gaussian interaction similarity matrix. :

[0066] ,in, σ For kernel bandwidth parameters, f i M and f j M Representing the first i The and the first j The interaction spectrum of the miRNAs. That is, the interaction spectrum of the miRNAs in correlation matrix A. i row and number j The row vector is a binary vector whose elements are composed of the values ​​(0 or 1) at corresponding positions in matrix A. It fully describes the known associations between a miRNA and all diseases (1 for known associations, 0 for unknown or no associations). 2 This represents the calculation of Euclidean distance. This matrix quantifies the topological similarity of any two miRNAs in a known network of interconnected networks. The higher the kernel similarity, the more similar the two miRNAs are to each other in their row vector patterns.

[0067] Based on the miRNA-disease association matrix, calculate the Gaussian kernel similarity for the disease. :

[0068] ,in, σ For kernel bandwidth parameters, f i D and f j D Representing the first i The and the first j The interaction spectrum of the nth element, i.e., the nth element of the correlation matrix A. i Column and number j A column vector that describes the known association status between the disease and all miRNAs, ||·|| 2 This represents the calculation of Euclidean distance. This matrix quantifies the topological similarity of any two diseases in a known network of connections.

[0069] The miRNA functional similarity matrix is ​​calculated based on functional information: if two miRNAs participate in similar functions (e.g., they target genes or participate in pathways that highly overlap), then their functional similarity is high. The calculation formula is as follows:

[0070] ,in, The miRNA functional similarity matrix represents the first... i The and the first j The similarity of miRNAs is calculated by dividing the numerator by the dot product of the vectors and the denominator by the product of their magnitudes. and These are the feature vectors based on the functional information of miRNAi and j, respectively. The superscript F indicates that the similarity comes from functional feature information. The similarity value ranges from [0,1], with a larger value indicating a higher degree of functional similarity between the two miRNAs.

[0071] Calculate the disease semantic similarity matrix based on semantic information: for any two diseases d i With d j The semantic similarity can also be represented by cosine similarity. , in, and These are the semantic information feature vectors of diseases i and j (constructed from the hierarchical paths of diseases in the Medical Subject Headings (MeSH) or the Disease Ontology (DO), for example, using the Wang method to represent each disease as a set of weighted path vectors), ||·||2 represents the Euclidean norm of the vector, and the superscript S indicates that the similarity comes from semantic information.

[0072] The feature vectors of these two matrices are derived from the functional information of the corresponding miRNAs and the semantic information of the disease extracted from the original data, respectively. The feature vector of the miRNA functional information is calculated using standard functional similarity methods (Wang et al. or the Relevance method based on GO annotation) based on data such as miRNA target genes (fixed-length numerical vectors). The feature vector of the disease semantic information is calculated using the semantic similarity method proposed by Wang et al. based on the disease MeSH descriptor (fixed-length numerical vectors). The association matrix represents the numerical representation of the association information. DS indicates that f comes from semantic information, and MF indicates the functional similarity information of miRNAs.

[0073] Step S130: The miRNA Gaussian similarity matrix and the functional similarity matrix are weighted and fused into a miRNA similarity matrix, and the disease Gaussian similarity matrix and the semantic similarity matrix are weighted and fused into a disease similarity matrix.

[0074] The formula for integrating the miRNA Gaussian similarity matrix and functional similarity matrix is: F M = α +(1— α )

[0075] ,in, F M This represents the integrated miRNA similarity matrix. Represents the miRNA Gaussian kernel similarity matrix (based on eigenvectors) f i M calculate), Represents the miRNA functional similarity matrix (based on functional feature vectors) calculate), α ∈[0,1] is a weight parameter used to balance the contributions of the two similarities.

[0076] The Gaussian similarity matrix and semantic similarity matrix of the disease are integrated as follows: F D = β +(1— β ) ,in, F D This represents the integrated disease similarity matrix. Represents the Gaussian kernel similarity matrix of diseases (based on eigenvectors) f i D calculate), Represents the semantic similarity matrix of diseases (based on semantic feature vectors) (Calculate), β∈[0,1] is the weight parameter.

[0077] Step S200: Construct miRNA-disease isomorphism map, miRNA isomorphism sub-map, and disease isomorphism sub-map based on miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, respectively, and fuse the three maps into a multimodal biological atlas network.

[0078] Specifically, a miRNA-disease heterogeneous graph is constructed based on the miRNA-disease association matrix, where known association pairs constitute the edges in the miRNA-disease heterogeneous graph; a miRNA isomorphic subgraph is constructed based on the miRNA similarity matrix, where the similarity between miRNAs serves as the weight of the edges in the miRNA isomorphic subgraph; and a disease isomorphic subgraph is constructed based on the disease similarity matrix, where the similarity between diseases serves as the weight of the edges in the disease isomorphic subgraph. The miRNA and disease isomorphic subgraphs are constructed from miRNA-miRNA similarity matrices and disease-disease similarity matrices, respectively. The miRNA-disease heterogeneous graph is constructed from known associations between miRNAs and diseases (e.g., miRNA-disease pairings verified experimentally or recorded in databases). The miRNA-disease heterogeneous graph is built upon a matrix containing known associations; it includes both known edges and unknown, i.e., node pairs to be predicted. Finally, these three types of graphs are fused into a multimodal biological atlas network.

[0079] As an example, the process of constructing a multimodal biological atlas network includes:

[0080] Step S210: Construct an adjacency matrix of the miRNA isomorphic subgraph based on the miRNA similarity matrix, construct an adjacency matrix of the disease isomorphic subgraph based on the disease similarity matrix, and construct an adjacency matrix of the miRNA-disease heteromorphic graph based on the miRNA-disease association matrix.

[0081] In step S210, the miRNA isomorphic subgraph uses miRNAs as nodes, and the edge weights are derived from the fusion of the miRNA Gaussian similarity matrix and the miRNA functional similarity matrix. The row vectors of the fusion matrix serve as the initial feature vectors of the miRNAs, and the similarity obtained after fusion is the edge weight.

[0082] Disease isomorphic subgraph: Using diseases as nodes, the edge weights are derived from the fusion of the disease Gaussian similarity matrix and the disease semantic similarity matrix. The row vectors of the fusion matrix serve as the initial feature vectors of the diseases, and the similarity obtained after fusion is the edge weight.

[0083] miRNA-disease heterogeneity graph: The edges between miRNA nodes and disease nodes are provided by the known miRNA-disease association matrix, where 1 indicates that there is an association and 0 indicates that there is no association.

[0084] The similarity matrix between the matrix and the subgraph (miRNA–miRNA, disease–disease) directly corresponds to the adjacency matrix of the isomorphic subgraph.

[0085] miRNA-disease association matrices directly correspond to adjacency matrices of heterogeneous graphs; these matrices are numerical representations of graph structures used as inputs to subsequent expert network models, GNNs, or other graph learning models.

[0086] Step S220: The miRNA-disease isomorphism map, miRNA isomorphism map, and disease isomorphism map are fused into a multimodal biological atlas network.

[0087] This is achieved through node alignment and edge relationship fusion. The multimodal biological atlas network comprehensively reflects the relationship between miRNAs and diseases by fusing multi-source information (functional similarity, semantic similarity, and known associations). By aggregating information from edges from different modalities (similar edges and associated edges), a feature representation that integrates multi-source information is generated for each node.

[0088] Step S300: Input the multimodal biological atlas network into the pre-trained gated hybrid multi-expert network model (MOE), and output the association probability of the miRNA-disease pair to be predicted. The gated hybrid multi-expert network model includes: multiple expert networks with different structures, used to perform nonlinear transformations on the node features of the input graph from different feature perspectives to generate diverse node embedding representations; a gating network, used to dynamically calculate the weights of each expert network according to the input target features, and select the Top-M expert networks to participate in the current node embedding generation based on the weights; a fusion module, which performs weighted fusion of the node embeddings generated by the selected Top-M expert networks to obtain the final comprehensive node embedding; and an MLP output layer, used to process the fused comprehensive node embedding and output the association probability of the pair.

[0089] The gated hybrid multi-expert network model proposed in this application is a graph neural network model based on the Mixture of Experts (MOE) architecture, specifically designed for the identification and prediction of miRNA-disease associations. The method first constructs a multimodal biological graph network including miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heteromorphic graphs. The multimodal biological graph network (including node features, adjacency matrices, etc.) is input to the pre-trained gated hybrid multi-expert network model (MOE). Graph structure information includes the adjacency matrices of the miRNA-disease heteromorphic graph, miRNA isomorphic subgraphs, and disease isomorphic subgraphs, which are derived from the miRNA / disease similarity matrix and known association matrices. Node feature information includes the initial feature vector for each miRNA and disease node (this vector is formed in step S130 by weighted fusion of Gaussian interaction kernel similarity and functional / semantic similarity). The expert networks in the gated hybrid multi-expert network model collectively constitute a dynamic embedding generation system. Taking a multimodal biological graph network as input, they extract and generate higher-level, more task-relevant node embedding representations from the original input features through multi-level, multi-perspective nonlinear transformations (such as convolution, neighbor aggregation, and autoencoder reconstruction) of node features and graph structure. These embeddings implicitly contain rich information about topological relationships.

[0090] The multimodal biomap network is used as input to a gated hybrid multi-expert network model to predict unknown miRNA-disease associations.

[0091] During the usage phase, the model receives an integrated multimodal biological map as input (including graph structure and initial node features). The gating network dynamically selects the most relevant experts based on the input. The selected expert network generates node embeddings, which are then fused and the MLP output layer calculates and outputs the association probability of the miRNA-disease pair to be predicted.

[0092] The gated hybrid multi-expert network model includes at least two expert networks: MultiHopGCN (Multi-hop Graph Convolutional Network), GraphSAGE (Graph Sampling Aggregation Network), AutoEncoder (AutoEncoder Network), and Multilayer Perceptron (MLP). The MultiHopGCN in this application consists of three layers of graph convolutional GCNs, each followed by Dropout and residual connections. The objective is to capture the topological patterns of locally dense connections through layer-by-layer aggregation of 1→2→3-hop neighbor information, while preserving original features to prevent over-smoothing.

[0093] The GraphSAGE network consists of three layers: GraphSAGEConv, each using a MeanAggregator, followed by ELU activation and Dropout. Its purpose is to sample and aggregate features from neighboring nodes, enabling more robust extraction of global statistics in sparse graphs and alleviating the dependence of GCNs on dense adjacency matrices.

[0094] The AutoEncoder network consists of two three-layer fully connected networks: an encoder and a decoder. The encoder extracts low-dimensional latent representations, and the decoder reconstructs the input features. Its purpose is to learn latent structural features through reconstruction tasks, exhibiting strong robustness to missing or sparse data.

[0095] The Multilayer Perceptron (MLP) consists of three fully connected layers, each followed by ELU activation and Dropout, used for nonlinear feature transformations. It operates without graph structure information and involves graph-free convolutional operations. Its purpose is to utilize explicit features for nonlinear transformations to capture non-graph structure information.

[0096] Preferably, the gated hybrid multi-expert network model includes: a parallel multi-hop graph convolutional network MultiHopGCN, a graph sampling aggregation network GraphSAGE, an autoencoder network AutoEncoder, and a multilayer perceptron MLP.

[0097] The gated hybrid multi-expert network model is a classification (parallel) hybrid expert model. Four expert networks—MultiHopGCN, GraphSAGE, AutoEncoder, and MLP—exist concurrently, and the input data is processed by these expert networks simultaneously. Then, the gating network dynamically assigns weights based on the input features and selects the most relevant Top-M expert outputs for fusion, where M can be 2.

[0098] The multimodal biological graph network inputs various expert networks and gating networks. MultiHopGCN performs graph convolution to capture topological information of multi-hop neighbors. GraphSAGE performs neighbor sampling and aggregation to capture sparse graph features. AutoEncoder encodes and decodes the input features to learn latent representations. MLP experts only perform nonlinear transformations on the features of the nodes themselves, completely ignoring graph structural relationships.

[0099] MultiHopGCN generates embeddings based on multi-hop neighbor topology. GraphSAGE generates embeddings based on sampled and aggregated neighbor features. AutoEncoder generates embeddings based on feature compression and reconstruction. MLP generates embeddings based on node attributes.

[0100] The gating network uses a small neural network (such as a two-layer MLP and Softmax) to calculate an initial weight vector based on the feature representation of the target node extracted from the multimodal biological graph network, i.e., the node embedding vector, which corresponds to four experts.

[0101] The gating network dynamically calculates the weights of each expert network based on the input features, including:

[0102] Based on the node embedding features learned from the multimodal biological map network, the initial weights of each expert network are calculated using a multilayer perceptron; wherein, the embedding features include the initial feature vectors of miRNA and disease nodes, as well as the structural embedding features extracted from the multimodal biological map network.

[0103] Node embedding features encapsulate the information of a node within the entire multimodal biological map network.

[0104] The dynamic performance evaluation metrics of each expert network are combined with the initial weights to generate a comprehensive weight; wherein the dynamic performance evaluation metrics are based at least on the statistical characteristics of the current output and historical performance of each expert network.

[0105] Specifically, the initial weights are comprehensively adjusted based on the learnable historical performance parameters of each expert network and the quality scores calculated based on the variance and entropy of the current output features to obtain the comprehensive weights.

[0106] Historical performance includes historical performance parameters, which are learnable parameters acquired through backpropagation during training. Statistical properties include the variance and entropy of the current output features, or skewness and kurtosis, etc.

[0107] The quality score is calculated based on the variance and entropy of the output features of each expert network. The quality score (quality_scores) is calculated as follows: i =variance i entropy i Among them, variance i Let be the variance of the output of the i-th expert network, and entropy. i Let be the entropy output by the i-th expert network.

[0108] The final weights, combined_weights:

[0109] combined_weights = base_weights ⊙ expert_performance ⊙ quality_scores, where expert_performance represents the parameters that can be learned during backpropagation when training a gated hybrid multi-expert network model, and base_weights represents the initial weights.

[0110] The expert networks corresponding to the top-K largest weights from the comprehensive weights are selected as the currently selected expert networks, where K is a positive integer greater than or equal to 1; K can be 2.

[0111] The node embeddings output by the selected expert network are weighted and fused to obtain the final node embedding.

[0112] Dynamic weight calculation: The gating network receives features from the input nodes and calculates a set of initial weights (base_weights) corresponding to all expert networks through a multi-layer neural network (e.g., two fully connected layers + ELU activation). Performance-aware mechanism: The gating network dynamically adjusts the initial weights not only based on the input but also considering the historical performance (expert_performance) of each expert network and the "quality_scores" of the current output (calculated from the variance and entropy of the output features) (e.g., through element-wise multiplication). Top-K sparsity: From the adjusted weights, the Top-K (e.g., K=2) expert networks with the largest weights are selected, and the weights of the remaining expert networks are reset to zero. Only the outputs of the selected experts are used for weighted fusion to improve efficiency and focus on key features.

[0113] During the feature learning phase, i.e., when training the model, the gated hybrid multi-expert network model introduces four types of expert network modules with different structures: MultiHopGCN (Multi-hop Graph Convolutional Network), GraphSAGE (Graph Sampling Aggregation Network), AutoEncoder (AutoEncoder Network), and Multilayer Perceptron (MLP). These modules are adept at extracting global topological dependencies across multiple hop counts, sparse graph aggregation features, latent data structure features, and non-graph node attribute features, respectively. Each type of expert network models nodes from a different perspective, ensuring the diversity and complementarity of feature representations.

[0114] To fully leverage the strengths of each expert model and achieve adaptive feature fusion, this method designs an independent gating network module. Based on the structural characteristics of miRNA isomorphism maps, disease isomorphism maps, and heteromorphism maps, the participation weights of each expert are dynamically calculated. The gating mechanism can flexibly adjust the importance of each expert according to the contextual information of the specific input sample, effectively avoiding information redundancy and key feature loss problems caused by fixed fusion strategies.

[0115] Through a structural design of "multi-expert parallel modeling and dynamic gating fusion," this model achieves deep complementarity at the feature layer, structural layer, and semantic layer, comprehensively improving the accuracy, generalization ability, and biological interpretability of miRNA-disease association prediction. Figure 7 As shown, experimental results demonstrate that the MOE model proposed in this application significantly outperforms existing mainstream models on multiple public datasets (such as HMDDv3.2), achieving leading performance in key evaluation metrics such as AUC, AUPR, and F1. This provides an efficient and robust computational method for large-scale miRNA function research and related disease target identification.

[0116] like Figure 2 As shown, as a specific example, the training process of the gated hybrid multi-expert network model includes:

[0117] Step S310: Construct a miRNA-disease association matrix based on the miRNA-disease association dataset.

[0118] It should be noted that the miRNA-disease association dataset in this application is derived from publicly available data in the Human MicroRNA Disease Database (HMDD), and includes experimentally validated miRNA-disease associations. For example, the HMDDv3.2 dataset contains 853 miRNAs, 591 diseases, and 12,405 known associations (after removing duplicates); the HMDDv2.0 dataset contains 495 miRNAs, 383 diseases, and 5,430 known associations.

[0119] Read the miRNA-disease association dataset, construct an association matrix of miRNA-disease relationships for all verified miRNA-disease associations and miRNA-disease pairs not appearing in the known association list, where all verified miRNA-disease associations form the known association list.

[0120] Retriev all validated miRNA-disease associations from public databases (such as HMDD), and set the corresponding position in the matrix to 1 according to the miRNA and disease indices to form a known association list. For miRNA-disease pairs not appearing in the known association list, set their matrix elements to 0, constructing an association matrix of miRNA-disease relationships. A .

[0121] It should be noted that the association between miRNAs and diseases is established using an association matrix. A Indicated, A ij Represents the first in the matrix i line, number j The elements of the column can take the value 0 or 1. For example, A ij=1 indicates the first i miRNA m i and the j a disease j There is a known correlation (positive sample). A ij Samples with a value of 1 are positive samples. A ij =0 indicates the first i miRNA m i and the j a disease j Relationship unknown (possible negative samples), negative samples from A ij =0 is used to obtain samples. Positive samples are known association pairs, and negative samples are selected using a hard negative sampling strategy (the number of negative samples is the same as that of positive samples). A ij It is a matrix A The Middle i Line number j The columns correspond to specific values, with each row representing a single miRNA and each column representing a single disease. The matrix values ​​represent the association between miRNA i and disease j.

[0122] M is the set of miRNA nodes, D is the set of disease nodes, and |M| is a matrix. A The number of rows, i.e., the number of miRNAs, where |D| is the matrix. A The number of columns, i.e., the number of diseases. These parameters are used to determine the correlation matrix. A The size of the matrix A The size is |M| |D|.

[0123] Figure 2 In the process, preprocessing and feature construction include steps S310 to S350. Preprocessing refers to the process of obtaining the original, experimentally validated list of miRNA-disease association pairs from public databases (such as HMDD, miR2Disease), including basic cleaning work such as removing duplicates and handling naming inconsistencies (e.g., unifying "Breast cancer" and "Breast Neoplasms").

[0124] Step S320: Calculate the Gaussian similarity matrix between miRNA and disease based on the miRNA-disease association matrix. Calculate the functional similarity of miRNA and the semantic similarity of disease based on the functional information of miRNA and the semantic information of disease in the miRNA-disease association dataset.

[0125] Based on the correlation matrix A, calculate the kernel similarity matrix of the miRNA Gaussian interaction spectrum and the kernel similarity matrix of the disease Gaussian interaction spectrum, respectively. : ,in, f i M and f j M The interaction spectra of miRNAi and miRNAj, represented by the vectors in the i-th and j-th rows of matrix A, respectively, characterize the known associations between miRNAi and miRNAj and all diseases. 2 Represents the calculation of Euclidean distance.

[0126] For the Gaussian interaction spectrum kernel similarity of the disease :

[0127] ,in, σ For kernel bandwidth parameters, f i D and f j D The interaction spectra of diseases i and j, respectively, are represented by the i-th and j-th column vectors of matrix A, characterizing the known associations between diseases i and j and all miRNAs. 2 Represents the calculation of Euclidean distance.

[0128] Calculate the miRNA functional similarity matrix Disease semantic similarity matrix Cosine similarity was used to calculate the cosine similarity between any two miRNA functional information feature vectors and the cosine similarity between any two disease semantic information feature vectors, thus obtaining the miRNA functional similarity matrix and the disease semantic similarity matrix.

[0129] Cosine similarity is defined as ,in, f i and f j Feature vectors representing different diseases or miRNAs are randomly initialized from the correlation matrix. f i · f j It is the vector dot product, || f i ||2 represents f i The L2 norm (Euclidean norm), || fj ||2 represents f j The L2 norm.

[0130] The miRNA functional information feature vector is a fixed-length numerical vector calculated using standard functional similarity methods (Wang et al. or the Relevance method based on GO annotation) based on miRNA functional information extracted from miRNA-disease association datasets, i.e., miRNA target genes and other data. The disease semantic information feature vector is a fixed-length numerical vector calculated using semantic similarity methods proposed by Wang et al. based on disease semantic information extracted from miRNA-disease association datasets, i.e., disease MeSH descriptors.

[0131] Step S330: The miRNA Gaussian similarity matrix and the miRNA functional similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the miRNA. The disease Gaussian similarity matrix and the disease semantic similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the disease.

[0132] Similarity matrices from different sources are integrated to generate a rich feature vector for each node. The initial feature vector for each miRNA and disease node is output.

[0133] The formula for integrating the miRNA Gaussian similarity matrix and functional similarity matrix is: F M = α +(1— α )

[0134] ,in, F M This represents the integrated miRNA similarity matrix. Represents the miRNA Gaussian kernel similarity matrix (based on eigenvectors) f i M calculate), Represents the miRNA functional similarity matrix (based on functional feature vectors) calculate), α ∈[0,1] is the weight parameter, used to balance the contributions of the two similarities. The integrated matrix... F M The i-th row vector is used as the initial feature vector of miRNAmi.

[0135] The Gaussian similarity matrix and semantic similarity matrix of the disease are integrated as follows: F D =β +(1— β ) ,in, F D This represents the integrated disease similarity matrix. Represents the Gaussian kernel similarity matrix of diseases (based on eigenvectors) f i D calculate), Represents the semantic similarity matrix of diseases (based on semantic feature vectors) (Calculated), β∈[0,1] is the weight parameter. The integrated matrix F D Its j-th row vector is used as the disease d j The initial eigenvectors.

[0136] Each row of the integrated matrix serves as a feature vector for the corresponding miRNA (or disease), which is then used to construct miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heteromorphic subgraphs. These are then used as input features to train the graph neural network model. The goal is to fuse multiple similarity information sets, allowing the model to utilize richer prior knowledge in graph structure learning.

[0137] Step S340: Take the miRNA-disease with known association in the miRNA-disease association matrix as positive samples, and extract negative samples from miRNA-disease with unknown association through hard negative sampling to form a labeled training sample set. The negative samples are those that are more similar to the positive samples in terms of features or topological structure. That is, samples that are difficult to distinguish from the unknown association are selected as negative samples. The label of the positive samples is 1, and the label of the negative samples is 0.

[0138] The selection of negative samples directly affects the training effect and final performance of the model. Therefore, this application adopts a hard negative sampling strategy for selecting negative samples. The core idea of ​​the hard negative sampling strategy is to select negative samples that are more similar to positive samples in terms of features or topological structure. These samples are more difficult to distinguish than random negative samples, thereby increasing the learning difficulty and generalization ability of the model. In essence, it forces the model to learn more refined feature differences through "hard samples", rather than relying solely on simple discriminative features.

[0139] The hard negative sampling strategy includes the following four steps: First, traverse the miRNA-disease association matrix to separate all positive samples and all possible negative samples. Among them, miRNA-disease pairs that are known to be associated are positive samples, and miRNA-disease pairs that are not associated are possible negative samples, i.e., candidate negative samples.

[0140] First, positive and negative samples are separated by traversing the miRNA-disease association matrix to separate all positive samples (pairs with known associations) and all possible negative samples (pairs without associations). P ={( i , j ) A ij =1}, N ={( i , j ) A ij =0}, where, P Represents the set of positive samples. N This represents the possible set of negative samples, i.e., the set of non-positive samples.

[0141] First, the sample refers to miRNA-disease pairs, such as miRNA i With disease j In this pair, the sample is a relation instance.

[0142] Secondly, the similarity between the current miRNA in the negative sample and all positive sample miRNAs of the target disease, as well as the similarity between the current disease in the negative sample and all positive sample diseases of the target miRNA, are summed to calculate the association probability score between each negative sample and the positive sample. Here, the score represents the comprehensive similarity between the candidate negative sample and all related positive samples.

[0143] The association probability score is used to quantify the overall similarity between an unknown association miRNA-disease pair and a known positive sample; the higher the score, the closer the negative sample is to the positive sample in the feature space or topology, and the more difficult it is for the model to distinguish it.

[0144] Next, the negative samples are scored: For each sample in the candidate negative sample set N (i.e., miRNA-disease pair (m, d)), its association probability score with the positive sample S(m, d) is calculated according to the following rules.

[0145] For a candidate negative sample pair (m, d), the sum of similarities between its miRNA member (m) and all positive sample miRNAs of the target disease (d) is evaluated, as is the sum of similarities between its disease member (d) and all positive sample diseases of the target miRNA (m). These two scores are then summed to obtain a comprehensive similarity score. This score effectively reveals the biological semantic and network topological proximity of the negative sample pair to the known positive sample population. For each negative sample, its similarity to positive samples is calculated. The calculation rule is: under the same disease d, the sum of similarities between the current miRNA m and other miRNAs positively associated with d; under the same miRNA m, the sum of similarities between the current disease d and other diseases positively associated with miRNA m. This allows for the scoring and ranking of negative samples.

[0146] Where m represents the currently considered miRNA node; m´ represents other miRNA nodes positively correlated with the current disease d; d represents the currently considered disease node; and d´ represents other disease nodes positively correlated with the current miRNA m. miRNA_sim(m,m´) represents the similarity between miRNA m and miRNA m´, derived from the miRNA m functional similarity matrix calculated in step S320 or the integrated miRNA feature matrix. disease_sim(d,d´) represents the similarity between disease d and disease d´, derived from the disease semantic similarity matrix or the integrated disease feature matrix. The similarity can be derived from the functional similarity matrix, semantic similarity matrix, or Gaussian similarity matrix.

[0147] The third step is to sort all negative samples by score. S ( m , d Sort the samples in descending order and select the negative sample with the highest score as the hard negative sample.

[0148] Finally, select the hard negative samples and sort all negative samples by score. S ( m , d The negative samples are sorted in descending order, and the negative samples with the highest scores are selected as hard negative samples (i.e., the negative samples most similar to the positive samples). This negative sampling strategy is better than random sampling because random negative sampling selects "simple negative samples" that are significantly different from the positive samples. The model can easily distinguish between positive and negative samples through surface features, but it cannot learn more fundamental patterns. Hard negative sampling can guide the model to focus on negative samples that are more difficult to distinguish, thereby effectively narrowing the embedding distance boundary between positive and negative samples and improving the model's robustness and recognition ability under near-neighbor interference.

[0149] Step S350: Using known associated positive samples as edges, and combining the similarity matrix after miRNA integration, the similarity matrix after disease integration, and the known association between miRNA and disease, construct miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heteromorphic graphs respectively, and fuse them into a multimodal biological atlas network.

[0150] Step S350 of this application merges the subgraphs, including the miRNA isomorphic subgraph, the disease isomorphic subgraph, and the miRNA-disease isomorphic subgraph, which describe the same group of nodes from different perspectives, into a complex multimodal biological map network by using node alignment.

[0151] miRNA isomorphism subgraph and disease isomorphism subgraph: constructed from the similarity matrices between miRNAs and between diseases.

[0152] miRNA isomorphic subgraph: Using miRNAs as nodes, the edge weights are derived from the fusion of the miRNA Gaussian similarity matrix and the miRNA functional similarity matrix. The resulting similarity is the edge weight.

[0153] Disease isomorphic subgraph: Using diseases as nodes, the edge weights are derived from the fusion of the disease Gaussian similarity matrix and the disease semantic similarity matrix. The similarity obtained after fusion is the edge weight.

[0154] miRNA-disease heterogeneity graph: Constructed from known associations between miRNAs and diseases (such as experimentally validated or database-recorded miRNA-disease pairings). The edges between miRNA nodes and disease nodes are provided by a known miRNA-disease association matrix, where 1 indicates an association and 0 indicates no association.

[0155] The similarity matrix between the matrix and the subgraph (miRNA–miRNA, disease–disease) directly corresponds to the adjacency matrix of the isomorphic subgraph.

[0156] miRNA-disease association matrices directly correspond to the adjacency matrices of heterogeneous graphs; these matrices are numerical representations of graph structures used as inputs to subsequent graph neural networks (GNNs) or other graph learning models.

[0157] Multimodal biological mapping networks comprehensively reflect the relationship between miRNAs and diseases through the fusion of multi-source information (functional similarity, semantic similarity, and known associations).

[0158] Step S360: Input the fused multimodal biological map network and the initial feature vectors of miRNA and disease nodes into the gated hybrid multi-expert network model to be trained, and obtain the gated hybrid multi-expert network model through optimization algorithm training.

[0159] This method constructs a Mixture-of-Experts (MOE) framework comprising multiple expert models and, for the first time, introduces a gate network in the miRNA-disease association task to achieve a Top-K dynamic expert selection mechanism. This structure adaptively selects the most relevant sub-models to participate in prediction based on input features, thereby improving the model's ability to extract features from complex data networks. The technique uses various heterogeneous data (such as miRNA functional similarity, disease semantic information, and known association networks) as input and designs a unified embedding generation module to map features from different sources to the same representation space, serving as the input basis for the expert models, thus enhancing the model's adaptability and generalization ability.

[0160] Embedding learning is the process of learning embedding representations. The process of "inputting the graph structure and node features into a gated hybrid multi-expert network model to be trained and training it to obtain parameters" is essentially embedding learning. The model training process involves learning a function that generates a low-dimensional, dense, task-relevant vector representation (i.e., embedding) for each node based on its own attributes and its topological position in the network.

[0161] The location where embedded learning is manifested:

[0162] (1) At S330, the initial feature vectors of miRNA and disease are obtained by integrating similarity matrices (Gaussian kernel, functional similarity, semantic similarity, etc.), which are static initial representations.

[0163] (2) At S360, these initial feature vectors will be used as inputs and fed into the introductory hybrid multi-expert MOE model.

[0164] (3) During the training process, the four types of experts in the MOE model (MultiHopExpert, GraphSAGEExpert, AutoEncoderExpert, MLPExpert) will perform multi-layer nonlinear transformations, graph convolution, encoding compression, etc. on the input features. These processes will continuously update and optimize the representation of nodes, which is the process of embedding learning.

[0165] The goal of embedding learning is to map the original features into a low-dimensional, information-condensed vector space, making similar nodes closer together. During MOE model training, the weights of each expert network are updated through backpropagation to ensure that the representations of miRNAs and diseases in this space are most favorable for the prediction task (e.g., miRNA-disease association prediction). Therefore, after model training, what is obtained is not only the predicted probability, but also an optimized miRNA and disease embedding vector (embedding vector = embedding representation = feature vector).

[0166] The MOE model proposed in this application consists of four expert networks: MultiHopExpert, GraphSAGEExpert, AutoEncoderExpert, and MLPExpert. Each expert network is responsible for learning feature representations at different levels. A gating mechanism is used to selectively activate different expert networks.

[0167] like Figure 3 As shown, MultiHopExpert mainly consists of three stacked graph convolutional (GCN) modules, each aggregating neighbor information with different hop counts:

[0168] The first layer (1-hop aggregation) input is the original node features. H (0), using the adjacency matrix The normalized information is then aggregated with the features of each node and its direct neighbors. The output is an aggregated feature representation of the 1-hop neighbors.

[0169] The second layer (2-hop aggregation) takes the output of the first layer as its input. H (1) Based on the 1-hop features of the previous layer, information from 2-hop neighbors is further aggregated, enabling nodes to perceive a wider range of graph structures. The output is a feature representation containing both 1-hop and 2-hop information.

[0170] The input to the third layer (3-hop aggregation) is the output of the second layer. H (2) By stacking and aggregating multi-hop neighbors, semantic information of nodes at greater distances (3-hop) can be obtained, while the original features are preserved through residual connections to prevent over-smoothing.

[0171] Output: The final node embedding that incorporates information from 1 to 3 hop neighbors.

[0172] Improvements compared to existing technologies: (1) Explicit multi-hop design: Traditional GCNs can only indirectly obtain multi-hop information through the number of layers. This scheme is explicitly divided into three layers: 1-hop, 2-hop, and 3-hop, and the aggregation range of each layer is fixed, avoiding the oversmoothing problem caused by excessively deep GCNs. (2) Symmetric normalization of the adjacency matrix:

[0173] Using adjacency matrix symmetric normalization instead of simple normalization reduces the impact of degree imbalance on feature aggregation, making it more suitable for handling sparse or heterogeneous graphs. (3) Residual connection: While aggregating multi-hop information, residual connection is introduced to combine the original node features with multi-hop features, improve gradient fluidity and maintain the individual characteristics of nodes, and prevent the features of deep networks from becoming too smooth.

[0174] MultiHopExpert, a multi-hop graph convolution expert, aggregates 1-hop → 2-hop → 3-hop neighbor information layer by layer through three-layer graph convolution (GCN), and captures topological patterns of locally dense connections layer by layer through symmetric normalization of the adjacency matrix. , where H (l) Representing the features of the Lth layer, This represents the adjacency matrix, specifically the adjacency matrix with self-loops, ensuring that nodes retain their own characteristics when aggregating neighbor information. The degree matrix represents the node degree matrix, used for symmetric normalization to prevent highly connected nodes from dominating features during aggregation. W (l) Indicates the first l The layer has a trainable weight matrix that learns a linear transformation of features. σ represents a non-linear activation function (ReLU, LeakyReLU, etc.) to introduce non-linear expressive power. Each layer is followed by Dropout regularization, and residual connections are used to preserve the original features and prevent over-smoothing.

[0175] like Figure 4As shown, GraphSAGEExpert, a graph sampling and aggregation algorithm, consists of three GraphSAGEConv convolutional layers, each using a MeanAggregator, followed by ELU activation and Dropout. The input to GraphSAGEExpert is the center node and its neighboring nodes, updating the node representation layer by layer through "graph sampling and aggregation." The specific process includes: First, Neighbor Sampling selects a fixed number of neighboring nodes for each center node, ensuring controllable computational scale while enhancing the model's generalization ability; then, Mean Aggregator performs mean calculation on the sampled neighbor features to obtain the global statistical features of the neighbors; next, Concatenation concatenates the center node features with the aggregated neighbor features, preserving self-information; subsequently, linear transformation and ELU activation use a weight matrix W(l) for mapping and introduce non-linear expressive power; finally, Dropout regularization reduces the risk of overfitting.

[0176] Compared with existing technologies, GraphSAGEExpert does not rely on strict normalization of the adjacency matrix. It is more robust to sparse or noisy graph structures through sampling and mean aggregation, while supporting large-scale graph computation. It avoids information loss by splicing center and neighbor features and improves the discriminativeness of representation. The three-layer stack captures multi-hop neighbor information, taking into account both local topology and long-range dependencies. It is especially suitable for non-uniform graph structures such as disease co-occurrence networks and drug-target networks.

[0177] AGEExpert learns the global statistical features of neighbors through the MeanAggregator of GraphSAGE, emphasizing the average contribution of neighbor information. Unlike multi-hop graph convolution, mean aggregation avoids the strict normalization dependency of GCN on the adjacency matrix, making it more robust to sparse graphs (such as sparse scenarios for disease semantic similarity). SAGE aggregation uses three SAGEConv convolutional layers, each using mean aggregation, followed by ELU activation and Dropout. Where || represents the concatenation operation. W l Indicates the first l The trainable weight matrix of the layer is used to map the concatenated feature vector (center node features and neighbor aggregated features) to a new feature space, thereby learning the feature transformation of the layer. h i (l-1) Represents a node i In the l The feature representation vector of layer -1, that is, the initial features of the output or input of the previous layer, contains the attribute information of the node itself. hj (l-1) node j (i.e., the neighboring nodes of i) at the th l The feature representations at layer -1 represent the attribute information of neighboring nodes. These are aggregated through the MEAN(·) operation to obtain the node... i The neighborhood statistical characteristics. Let represent the set of neighbors of node i, and MEAN represents the mean aggregation operation. This expert is suitable for capturing global distribution characteristics.

[0178] like Figure 5 As shown, the AutoEncoderExpert architecture consists of two parts: an encoder and a decoder. The input is the original feature vector of a node or sample. First, the high-dimensional input is compressed into a low-dimensional latent representation through a multi-layer fully connected network of the encoder. Then, it is restored to a reconstructed feature with the same dimension as the input space through a multi-layer fully connected network of the decoder. During the training phase, the expert learns feature representations that preserve key information from the input through a self-supervised reconstruction task (input → encoding → decoding → reconstruction). Its optimization objective is to minimize the mean squared error (MSE) loss.

[0179]

[0180] Where N represents the total number of samples, Indicates the first i The original input feature vector of each sample, i.e. the feature data before entering the autoencoder. Indicates the first i Each sample is a feature vector reconstructed by an autoencoder, which is the result of the input being compressed into a low-dimensional latent space by the encoder and then restored to the original space by the decoder. This represents the square of the Euclidean distance.

[0181] AutoEncoderExpert learns compressed feature representations through a self-supervised reconstruction task (input features → encoding → decoding → reconstructing the input), using a 3-layer fully connected encoder and a symmetric 3-layer fully connected decoder. The reconstruction capability is optimized using MSE loss. The reconstruction loss is: ,in, This represents the input of the i-th sample. and reconstructed output The mean squared error is obtained by summing the squared Euclidean distances between the samples and dividing the sum by the number of samples N. The encoded features contain key information about the input. It is the input to the model. The output that the model attempts to reconstruct is the result of the input. The difference between the two is measured by mean squared error (MSE), which is used to optimize the reconstruction capability of the autoencoder so that the encoded latent representation retains as much key information as possible from the input.

[0182] like Figure 6 As shown, the core structure of MLPExpert consists of three fully connected layers stacked sequentially, each followed by an ELU activation function and a Dropout layer. It does not use graph convolution operations, therefore relying entirely on the explicit feature vectors of the input nodes for representation learning, without utilizing graph structure information such as adjacency matrices. Its computation process is as follows:

[0183]

[0184] Here, x represents the original feature vector of the input node; W1, W2, and W3 represent trainable weight matrices, corresponding to the parameters of the three fully connected network layers, used to map features layer by layer to the new representation space; the ELU activation function introduces non-linear modeling capabilities, avoiding gradient vanishing and improving the diversity of feature representations; Dropout randomly discards some neuron connections during training to prevent overfitting and improve generalization ability. Finally, the output layer, after non-linear transformation, yields the hidden representation of the node, which can be used for subsequent tasks.

[0185] The improvement over existing technologies lies in the removal of graph structure dependence: Compared to traditional graph neural networks (such as GCN and GraphSAGE) that rely on adjacency matrices for feature propagation, MLPExpert completely abandons graph convolution, avoiding the misleading effects of sparse graphs or noisy edges.

[0186] MLPExpert learns features through a fully connected network, relying on explicit information from the input features rather than graph structure information. Its core structure consists of a 3-layer fully connected network followed by ELU activation and Dropout, without graph convolution operations. .

[0187] Structural Relationship between Gating and Expert Networks: In this application, the gate network and the multi-expert networks form a typical dynamic mixed-experts (MOE) structure. The multimodal biomap network is first simultaneously fed into multiple expert networks with different structures (such as GraphSAGEExpert, MultiHopExpert, AutoEncoderExpert, MLPExpert) and the gate network. The gate network consists of multiple fully connected layers, which encode the input and output a weight vector equal to the number of experts. The multiple fully connected layers include: two fully connected layers (MLP) and a softmax activation function. The weights are normalized by softmax to ensure that their sum is 1. Each expert network undertakes a different information extraction task: GraphSAGEExpert captures local graph structure information through neighbor sampling and aggregation; MultiHopExpert models long-range dependencies using multi-hop convolutions; AutoEncoderExpert learns the latent compressed representation of the input through a self-supervised reconstruction task; and MLPExpert mines explicit global relationships of input features through a fully connected structure without relying on the graph structure. Finally, the output representations of each expert are weighted and fused according to gating weights. The gating network also includes: a dynamic adjustment module, a fusion function, a Top-K selector, and a weighted fusion unit.

[0188] During model training, after each expert updates the features of the input data, a gating network is used to implement the MOE (Mean of Effect) of experts. This is the core module of the MOEMDA model for dynamically selecting experts. The dynamic adjustment module is responsible for dynamically allocating the weights of each expert based on the input features and the "quality" of the expert output (i.e., measuring the effectiveness of each expert in learning the features), ensuring that the model is applicable to different subgraphs. Dynamic adaptability, through quality scores and expert performance parameters, allows the gating network to automatically select the most relevant expert based on the characteristics of the current input data (e.g., sparse / dense graphs) (e.g., MultiHopExpert for dense graphs, SAGEExpert for sparse graphs). First, the initial weights of each expert are calculated: ,in, These are the parameters of the first fully connected layer. The parameters of the second fully connected layer are as follows: the input feature x first passes through the first fully connected layer W1 to obtain a 128-dimensional hidden representation; the hidden representation is then activated by the ELU activation function to introduce non-linearity; then, the second fully connected layer W2 maps the dimensions to a vector of size "number of experts", and is activated by ELU again to ensure the non-linear expressiveness of the feature; finally, Softmax is used to normalize the vector into a probability distribution, ensuring that the sum of the initial weights of all experts is 1. Secondly, the dynamic adjustment module of the gating network calculates the "quality" of each expert's output to adjust the weights. The "quality" is determined by the product of the variance and entropy of each expert's network output. ,in, variance i Indicates the first i The variance of the output of an expert network. Indicates the first i The output of an expert network for the j-th sample μ i Indicates the first i The variance is calculated using the mean of the outputs of each expert network, where N represents the total number of samples. Expert networks with higher variance may be more sensitive to changes in the input data, while those with lower variance are more stable. In the variance formula, N represents the number of samples a single expert participates in the calculation for the current batch.

[0189]

[0190] Among them, entropy i Let i represent the entropy of the i-th expert. Indicates the first i The k-th dimension feature output by each expert is used to balance dispersion and information content to obtain a "quality" score, quality_scores, by multiplying variance and entropy. The final weights are:

[0191] combined_weights=base_weights⊙expert_performance⊙quality_scores

[0192] Here, `expert_performance` represents the learnable parameters during backpropagation when training a gated hybrid multi-expert network model. The improvement in `combined_weights` lies in simultaneously introducing variance and entropy to construct a quality score, which is then combined with the initial weights `base_weights` and the historical performance parameter `expert_performance`. ⊙ represents the Hadamard product, also known as element-wise multiplication. This triple-weighting method adjusts the final expert weights, achieving a balance between diversity and stability. The Top-K selector in the gated network selects the top-K experts to avoid dilution of expert contributions due to weight dispersion, allowing the model to focus on the two experts performing best under the current input, enhancing the ability to capture key features. Simultaneously, by dynamically selecting experts to adapt to data heterogeneity, a balance is ultimately achieved between efficiency and accuracy. Furthermore, retaining only the two experts with the highest weights significantly reduces computational complexity and improves forward propagation speed. Experiments have verified that K=2 is ultimately chosen in this application. The weighted fusion unit weights and fuses the outputs of the selected K experts to generate the final node embedding.

[0193] For each sample (m) in the training sample set S i ,d i ,y i The model learns and outputs miRNAm based on a complete graph structure. i and disease d i The node embedding is performed. The association probability of the node pair is calculated through a decoder (such as the output layer of an MLP or a dot product operation). .

[0194] The final cross-entropy loss is: ,in, It's a real label. This represents the probability predicted by the model, and N is the batch size. In the cross-entropy formula, N is the total number of samples in the current batch for the entire model. The parameters of the gating network and expert networks are iteratively optimized using the error gradient of the entire training set through backpropagation until the model converges. Different hyperparameters affect the model's predictive performance. To achieve optimal performance, this application experimented with a set of different hyperparameters, ultimately selecting the set that best demonstrated the model's predictive performance: dropout = 0.2, epoch = 200, learning rate = 0.001, and neighbor = 20.

[0195] This application aims to address the core problems of existing GNN models in MDA prediction: traditional single GNN models cannot efficiently capture the multi-dimensional features of heterogeneous graphs simultaneously, and fixed-weight fusion strategies result in insufficient utilization of key information. Through the dynamic weight allocation of the MOE architecture, the model can automatically select the optimal feature combination based on data characteristics, improving the accuracy and robustness of MDA prediction.

[0196] like Figure 7 As shown, the horizontal axis represents six commonly used evaluation metrics for classification tasks, including: Area Under the Curve (AUC): measures the model's ability to distinguish between positive and negative samples; Area Under the Precision-Recall Curve (AUPR): focuses more on the prediction quality of positive samples; Accuracy: the proportion of correctly predicted samples out of the total number of samples; F1 score: the harmonic mean of precision and recall; Recall: the proportion of positive samples that are correctly identified; Precision: the proportion of truly positive samples out of those predicted as positive. The vertical axis represents the numerical values ​​of each metric (ranging from 0 to 1), with higher values ​​indicating better model performance. Figure 7 The markers in the graph are: ○ circle: HMDD v2.0 dataset results; △ triangle: HMDD v3.2 dataset results; data labels: indicating the specific values ​​of each metric on the corresponding dataset. Curve interpretation: Overall, HMDD v3.2 slightly outperforms HMDD v2.0 on almost all metrics, especially showing a significant advantage in AUC, AUPR, and Recall; Accuracy, F1 score, and Precision are also slightly higher, but the differences are relatively small. This indicates that the model's overall prediction performance on the v3.2 dataset is better than that on the v2.0 dataset, but the trends of both are basically consistent, indicating high accuracy. Experimental results show that the prediction AUC of the HMDD v3.2 dataset exceeds 0.97, and the prediction AUC of the HMDD v2.0 dataset exceeds 0.96, both higher than the metrics of all current methods. There are many existing methods, which will not be shown here. MOEMDA has improved the AUC and AUPR metrics on public datasets (such as HMDDv3.2) compared with existing models and methods, which verifies the effectiveness of dynamic multi-expert fusion strategy in complex biological heterogeneous graph tasks and provides a more reliable computational tool for miRNA-driven disease mechanism research and targeted therapy.

[0197] like Figure 8As shown, this graph illustrates the AUC (Area Under Curve) distribution of the model in 5-Fold Cross Validation, used to evaluate the performance stability of the model during training at different folds. First, all samples are divided into 5 subsets that are as equal as possible. Each subset can be considered a "test set" for one validation cycle, and the remaining 4 subsets are used as the "training set." For example, in fold 1: the first subset is used as the test set, and the other 4 subsets are used as the training set. In fold 2: the second subset is used as the test set, and the other 4 subsets are used as the training set. This continues until all 5 subsets have been used as a test set in turn. The horizontal axis, Train, represents each fold in 5-Fold Cross Validation (5fCV_1 ~ 5fCV_5). The vertical axis, AUC, represents the AUC value of the model on the training set in each fold, ranging from approximately 0.92 to 0.98. The gray box represents the distribution range of the compromise AUC value (box plot), with the thick horizontal line in the middle representing the median, and the upper and lower boundaries of the box representing the upper and lower quartiles (Q1, Q3). The whiskers represent the range of non-outliers, with values ​​outside the range considered outliers. The gray scatter plot represents the specific AUC value obtained for each training iteration. The AUC value is a primary indicator of model accuracy; a more concentrated and larger value is generally better.

[0198] Figure 8 The results are explained as follows: (1) The median AUC of each fold is above 0.96, indicating that the model has high classification performance under different data partitions. (2) The box height (interquartile range) of each fold is relatively small, indicating that the performance fluctuation is small and the model is stable. (3) A small number of low outliers (below 0.94) may correspond to unstable performance in some training batches, but the overall impact is not significant. (4) The AUC of the 4th fold (5fCV_4) is slightly higher and more concentrated, indicating that the training effect of this fold is the best.

[0199] like Figure 9 As shown, the heterogeneous graph represents information about miRNAs and diseases. Unlike the miRNA and disease graphs, which learn miRNA and disease features separately, the heterogeneous graph learns features of both miRNAs and diseases simultaneously. This means that associated miRNA-disease pairs are initially input into the expert for learning, while those without a confirmed association learn miRNA and disease features separately. (The miRNA-disease heterogeneous graph mentioned in step S350) This is a heatmap; the horizontal axis represents different samples, and the vertical axis represents different experts. The color intensity represents the weight of that expert on the corresponding sample. Lighter colors indicate higher expert weights, and darker colors (close to black) indicate lower weights.

[0200] like Figure 10 As shown, this application also proposes a miRNA-disease relationship prediction system based on a hybrid expert model, the system comprising:

[0201] The matrix construction module 100 is used to obtain a miRNA-disease association matrix, a miRNA similarity matrix, and a disease similarity matrix based on the multi-omics data of the miRNA-disease to be predicted. The miRNA-disease association matrix includes known miRNA-disease associations and unknown miRNA-disease associations.

[0202] The graph network construction module 200 is used to construct miRNA-disease heteromorphism graphs, miRNA isomorphism subgraphs, and disease isomorphism subgraphs based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, respectively, and to fuse the three graphs into a multimodal biological graph network.

[0203] The prediction module 300 is used to input the multimodal biological atlas network into a pre-trained gated hybrid multi-expert network model and output the association probability of the miRNA-disease pair to be predicted. The gated hybrid multi-expert network model includes: multiple expert networks with different structures, used to generate node embeddings from different feature perspectives; a gated network, used to dynamically calculate the weights of each expert network according to the embedding features, and select the Top-M expert networks according to the weights; and an output layer, used to fuse the node embeddings generated by the selected expert networks and then process them to output the prediction result.

[0204] The system also includes a training module for training a gated hybrid multi-expert network model, thereby implementing the training process of the gated hybrid multi-expert network model in steps S310 to S360 above.

[0205] This application also provides a gated hybrid multi-expert network model, which is the gated hybrid multi-expert network model in the method described above, and the gated hybrid multi-expert network model is used for miRNA and disease association prediction.

[0206] In another aspect, this application proposes a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0207] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for predicting miRNA-disease relationships based on a hybrid expert model, characterized in that, The method includes: Based on the multi-omics data of miRNA-disease to be predicted, obtain the miRNA-disease association matrix, miRNA similarity matrix and disease similarity matrix. The miRNA-disease association matrix includes known miRNA-disease associations and unknown miRNA-disease associations. Based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, a miRNA-disease isomorphism map, a miRNA isomorphism sub-map, and a disease isomorphism sub-map are constructed respectively, and the three maps are fused into a multimodal biological atlas network. A multimodal biomap network is input into a pre-trained gated hybrid multi-expert network model, which outputs the association probability of the miRNA-disease pair to be predicted. The gated hybrid multi-expert network model includes: multiple structurally distinct expert networks used to extract features from the multimodal biomap network from different feature perspectives and generate node embeddings; a gated network used to dynamically calculate the weights of each expert network based on the embedding features and select the Top-M expert networks based on the weights; and an output layer used to fuse and process the node embeddings generated by the selected expert networks to output the prediction result. The training process of the gated hybrid multi-expert network model includes: Based on the miRNA-disease association dataset, construct the miRNA-disease association matrix; Calculate the Gaussian similarity matrix between miRNA and disease based on the association matrix. Calculate the functional similarity of miRNA and the semantic similarity of disease based on the functional information of miRNA and the semantic information of disease in the miRNA-disease association dataset. The miRNA Gaussian similarity matrix and the miRNA functional similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the miRNA. The disease Gaussian similarity matrix and the disease semantic similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the disease. The miRNA-disease association matrix of miRNA and disease is used as positive samples. Negative samples are extracted from miRNA-disease with unknown associations through hard negative sampling to form a labeled training sample set. The negative samples are those that are more similar to the positive samples in terms of features or topology. Using known associated positive samples as edges, miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heteromorphic graphs are constructed by combining the similarity matrix after miRNA integration, the similarity matrix after disease integration, and the known association between miRNA and disease, and then fused into a multimodal biological atlas network. The initial features of the fused multimodal biological map network, miRNA, and disease nodes are input into the gated hybrid multi-expert network model to be trained, thus obtaining the gated hybrid multi-expert network model.

2. The method according to claim 1, characterized in that, The gated hybrid multi-expert network model includes at least two expert networks of the following types; Multi-hop graph convolutional networks include: a three-layer graph convolutional network (GCN) that aggregates hop neighbor information layer by layer, with each layer followed by Dropout and residual connections; The graph sampling aggregation network consists of three GraphSAGEConv convolutional layers, each using a mean aggregator followed by ELU activation and Dropout. An autoencoder network consists of two fully connected networks, one encoder and one decoder. The encoder extracts low-dimensional latent representations, and the decoder reconstructs input features. The multilayer perceptron consists of three fully connected layers, each followed by ELU activation and Dropout, for nonlinear feature transformation, independent of graph structure information.

3. The method according to claim 1, characterized in that, The gating network dynamically calculates the weights of each expert network based on the input features, including: The initial weights of each expert network are calculated using a multilayer perceptron based on the node embedding features learned from the multimodal biological map network. By integrating the dynamic performance evaluation metrics of each expert network with the initial weights, a comprehensive weight is generated; wherein, the dynamic performance evaluation metrics are based at least on the statistical characteristics of the current output and historical performance of each expert network. Based on the comprehensive weights, the expert networks corresponding to the Top-K weights with the largest weight values ​​are selected as the currently selected expert networks, where K is a positive integer greater than or equal to 1; The node embeddings output by the selected expert network are weighted and fused to obtain the final node embedding.

4. The method according to any one of claims 1 to 3, characterized in that, The process involves obtaining a miRNA-disease association matrix, a miRNA similarity matrix, and a disease similarity matrix based on the multi-omics data of the miRNA-disease to be predicted, including: Based on the multi-omics data of miRNA-disease to be predicted, obtain the miRNA-disease association matrix, which includes known associations and unknown associations; Based on the miRNA-disease association matrix, the Gaussian similarity matrix between miRNA and disease is calculated. Based on the functional information of miRNA and the semantic information of disease in the multi-omics data of miRNA-disease to be predicted, the functional similarity of miRNA and the semantic similarity of disease are calculated respectively. The miRNA Gaussian similarity matrix and the functional similarity matrix were weighted and fused separately to form the miRNA similarity matrix, and the disease Gaussian similarity matrix and the semantic similarity matrix were weighted and fused separately to form the disease similarity matrix.

5. The method according to claim 1, characterized in that, The construction of miRNA-disease isomorphism maps, miRNA isomorphism sub-maps, and disease isomorphism sub-maps based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, respectively, includes: Using the miRNA-disease association matrix as the adjacency matrix, and miRNA and disease as two types of nodes, a miRNA-disease heterogeneous graph is constructed. The edges between miRNA nodes and disease nodes are provided by the known miRNA-disease association matrix. In the matrix, 1 indicates that there is an association and 0 indicates that there is no association. Using the miRNA similarity matrix as the adjacency matrix and miRNAs as nodes, a miRNA isomorphic subgraph is constructed, wherein the similarity between miRNAs is used as the weight of the edges in the miRNA isomorphic subgraph. Using the disease similarity matrix as the adjacency matrix and diseases as nodes, a disease isomorphic subgraph is constructed, where the similarity between diseases is used as the weight of the edges in the disease isomorphic subgraph.

6. The method according to claim 5, characterized in that, The hard negative sample sampling process includes: Traverse the miRNA-disease association matrix to separate all positive samples and all possible negative samples; where miRNA-disease pairs with known associations are positive samples, and miRNA-disease pairs without associations are possible negative samples. The similarity between the current miRNA in the negative sample and all positive miRNAs of the target disease is summed, as well as the similarity between the current disease in the negative sample and all positive diseases of the target miRNA, to calculate the association probability score between each negative sample and the positive samples. The association probability score represents the overall similarity between the negative sample and all related positive samples. Sort all negative samples in descending order of score, and select the negative sample with the highest score as the hard negative sample.

7. A miRNA-disease relationship prediction system based on a hybrid expert model, characterized in that, The system includes: The matrix construction module is used to obtain the miRNA-disease association matrix, miRNA similarity matrix and disease similarity matrix based on the multi-omics data of the miRNA-disease to be predicted. The miRNA-disease association matrix includes known miRNA-disease associations and unknown miRNA-disease associations. The graph network construction module is used to construct miRNA-disease heteromorphism graphs, miRNA isomorphism subgraphs, and disease isomorphism subgraphs based on the miRNA-disease association matrix, miRNA similarity matrix, and disease similarity matrix, respectively, and then merge the three graphs into a multimodal biological graph network. The prediction module is used to input multi-source heterogeneous features into a pre-trained gated hybrid multi-expert network model and output the association probability of the miRNA-disease pair to be predicted. The gated hybrid multi-expert network model includes: multiple expert networks with different structures, used to generate node embeddings from different feature perspectives; a gated network, used to dynamically calculate the weights of each expert network according to the embedding features, and select the Top-M expert networks according to the weights; and an output layer, used to fuse the node embeddings generated by the selected expert networks and then process them to output the prediction result. A gated hybrid multi-expert network model training module is used to construct a miRNA-disease association matrix based on the miRNA-disease association dataset. Calculate the Gaussian similarity matrix between miRNA and disease based on the association matrix. Calculate the functional similarity of miRNA and the semantic similarity of disease based on the functional information of miRNA and the semantic information of disease in the miRNA-disease association dataset. The miRNA Gaussian similarity matrix and the miRNA functional similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the miRNA. The disease Gaussian similarity matrix and the disease semantic similarity matrix are weighted and integrated, and the row vectors of the integrated matrix are used as the initial feature vectors of the disease. The miRNA-disease association matrix of miRNA and disease is used as positive samples. Negative samples are extracted from miRNA-disease with unknown associations through hard negative sampling to form a labeled training sample set. The negative samples are those that are more similar to the positive samples in terms of features or topology. Using known associated positive samples as edges, miRNA isomorphic subgraphs, disease isomorphic subgraphs, and miRNA-disease heteromorphic graphs are constructed by combining the similarity matrix after miRNA integration, the similarity matrix after disease integration, and the known association between miRNA and disease, and then fused into a multimodal biological atlas network. The initial features of the fused multimodal biological map network, miRNA, and disease nodes are input into the gated hybrid multi-expert network model to be trained, thus obtaining the gated hybrid multi-expert network model.

8. A gated hybrid multi-expert network model, characterized in that, The model is a gated hybrid multi-expert network model as described in any one of claims 1 to 6, and the gated hybrid multi-expert network model is used for miRNA and disease association prediction.

9. A computer program product comprising a computer program stored on a readable storage medium, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 6.

Images