A method for miRNA-disease association prediction based on global and local graph representation learning

Abstract

The application provides a miRNA disease correlation prediction method based on global and local graph representation learning, and belongs to the technical field of miRNA disease correlation prediction. The miRNA disease correlation prediction method based on global and local graph representation learning comprises the following steps: calculating a plurality of similarity matrices by using multi-source information of miRNAs and diseases and fusing the plurality of similarity matrices, constructing a miRNA-disease attribute bipartite graph according to known correlation information and fused similarity features; extracting global feature representation of nodes by using a graph Transformer neural network; extracting local feature representation of nodes by using an adaptive graph convolutional neural network; fusing the global and local feature representations, and predicting potential miRNA disease correlation by using a multilayer perceptron classifier; and the method can comprehensively and deeply depict miRNAs and diseases by fusing the global and local feature representations of nodes, high-precision prediction is realized, and credible candidate miRNAs are provided for researchers to determine correlations between miRNAs and diseases.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, and specifically relates to the application of computer technology to miRNA disease association prediction, specifically involving a miRNA disease association prediction method based on global and local graph representation learning. Background Technology

[0002] miRNAs are non-coding single-stranded RNAs that regulate gene expression at the post-transcriptional level by base pairing with messenger RNA, leading to mRNA degradation or inhibiting its translation. Dysregulation or abnormal expression of miRNAs is closely related to the occurrence and development of diseases. Understanding the association between miRNAs and diseases helps to comprehend the pathological mechanisms of diseases, aids in disease diagnosis and treatment, and promotes the development of new treatment strategies. Determining miRNA-disease associations using traditional biological experimental methods is time-consuming, labor-intensive, and costly, hindering widespread application. Therefore, using computational methods to predict potential miRNA-disease associations, thereby providing researchers with candidate miRNAs for disease management, is of great significance.

[0003] Current computational methods mainly include similarity-based methods and machine learning-based methods. Similarity-based methods calculate the similarity between miRNAs and diseases in various ways and rank the association probabilities of miRNA-disease pairs accordingly. However, due to limitations in the learning capabilities of the methods used, the predictive performance of similarity-based methods is not high. Machine learning-based methods use neural networks to extract feature representations of miRNAs and diseases for subsequent prediction. With the rise of graph representation learning methods, represented by Graph Convolutional Networks (GCNs), and the accumulation of large amounts of miRNA-disease association data, which makes it possible to model graph structure data, many GCN-based miRNA-disease association prediction methods have been developed.

[0004] In existing GCN-based prediction methods, each node undergoes the same number of message passing and feature updates—a feature smoothing process. However, this operation cannot meet the needs of all nodes in the graph, preventing the full extraction of local feature information. Furthermore, to aggregate feature information from distant nodes, multiple GCN layers need to be stacked, leading to over-smoothing of node features and impacting the prediction performance of downstream classifiers, thus hindering the extraction of global feature information. The inability to extract high-quality, comprehensive node feature representations limits the performance of existing prediction methods. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a miRNA disease association prediction method based on global and local graph representation learning. This method solves the problem that the inability to extract high-quality and comprehensive miRNA and disease node feature representations limits the effectiveness of miRNA disease association prediction.

[0006] The solution adopted in this invention is as follows:

[0007] Multiple similarity matrices between miRNAs and diseases are calculated using multi-source information. A comprehensive similarity feature matrix of nodes is constructed and fused using two autoencoders (AE) to obtain a fused similarity feature matrix between miRNAs and diseases.

[0008] Based on fusion similarity and known miRNA disease association information, a bipartite graph G of miRNA-disease attributes is constructed, with miRNA and disease as nodes, fusion similarity as the initial feature of the nodes, and known associations as the edges.

[0009] Calculate the degree of nodes in graph G, construct a degree fusion module and add it to the Transformer framework to construct a graph Transformer neural network; input the attribute bipartite graph G into the network to extract the global feature embedding representation of the nodes.

[0010] Based on graph G, an adaptive graph convolutional neural network is used to calculate the required number of feature smoothing operations for each node in the graph, so as to fully extract the local feature embedding representation of the node and avoid the problem of feature over-smoothing.

[0011] The global and local features of nodes are fused using a fully connected neural network; the features of miRNA and the features of the disease are concatenated to construct a miRNA-disease pair feature representation; and a multilayer perceptron classifier is used to predict potential associations.

[0012] The beneficial effects of this invention include:

[0013] 1. Calculate the required number of feature smoothing iterations for each node to adaptively aggregate local features of the node and avoid oversmoothing problems in the GCN feature extraction process;

[0014] 2. Use graph Transformer neural networks to fully extract global features of nodes, alleviating the problem that GCN cannot effectively capture long-distance dependencies in the graph;

[0015] 3. By fusing global and local features, high-quality and comprehensive miRNA and disease node feature representations are obtained, thereby improving the performance of prediction methods;

[0016] 4. Provide researchers with reliable disease candidate miRNAs, reducing experimental time and costs. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention;

[0018] Figure 2 This is a comparison chart of the AUC curves of the method of this invention and the neural network method in Figure 4;

[0019] Figure 3 This is a comparison chart of the AUPR curves of the method of this invention and the neural network method in Figure 4;

[0020] Figure 4 The figure shows the experimental results for verifying the effectiveness of each part of the method of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0022] Example 1

[0023] This invention provides a method for predicting miRNA disease associations based on global and local graph representation learning. For example... Figure 1 As shown, the specific implementation of this method is as follows:

[0024] S1. Utilize multi-source information to calculate multiple similarity matrices between miRNAs and diseases, construct a comprehensive similarity feature matrix for nodes, and fuse them using two autoencoders (AE) to obtain a fused similarity feature matrix between miRNAs and diseases:

[0025] In this embodiment, data was obtained from version 3.2 of the Human MicroRNA Disease Database (HMDD), and after processing, 14,550 association records between 792 diseases and 917 miRNAs were obtained. An adjacency matrix was used. Record associated information.

[0026] In this embodiment, the directed acyclic graph (DAG) structure of diseases is obtained from the medical subject thesaurus mesh and used to calculate the semantic similarity matrix (DSM) of diseases.

[0027] For target disease Any disease in its DAG right The semantic contribution value is:

[0028]

[0029] disease and The semantic similarity is:

[0030]

[0031] in and These are diseases and The set of nodes in a DAG.

[0032] The Gaussian interaction spectrum kernel similarity matrix (DGM) of the disease is calculated as follows:

[0033]

[0034]

[0035] in Adjacency matrix The List.

[0036] Gaussian interaction spectrum kernel similarity matrix of miRNA The calculation is as follows:

[0037]

[0038]

[0039] In this embodiment, miRNA sequence information is obtained from the miRBase database, and the Niedermann-Onsch algorithm is used to calculate the similarity between pairs of sequences to obtain the miRNA sequence similarity matrix (MSM). The functional similarity matrix (MFM) of the miRNA is obtained by analyzing the semantic similarity between disease sets associated with two miRNAs.

[0040] The formulas for the combined disease similarity matrix ID and miRNA similarity matrix IM are as follows:

[0041]

[0042]

[0043] Two AEs were used to perform dimensionality reduction and fusion on ID and IM respectively to obtain the fusion similarity matrix of the disease. Fusion similarity matrix with miRNA .

[0044] S2. Based on the fusion characteristics and known association information, construct a bipartite graph of miRNA-disease attributes:

[0045] In this embodiment, miRNA and disease are used as nodes, the similarity matrix is ​​fused as the initial feature of the nodes, and known associations are used as edges to construct a bipartite graph G of miRNA-disease attributes. The adjacency matrix of graph G is A, and the node feature matrix is ​​denoted as X.

[0046] S3. Based on the attribute bipartite graph, a graph Transformer neural network is used to extract the global feature embeddings of the nodes, as implemented below:

[0047] S301. Based on the degree information of the nodes in the attribute bipartite graph G, construct a degree fusion module to inject the inductive bias information of the graph.

[0048] S302. Build a node feature extraction framework based on Transformer, add the degree fusion module to the framework, and construct a graph Transformer neural network to extract global features of nodes.

[0049] The formula for feature extraction using a multi-head Transformer is as follows:

[0050]

[0051]

[0052] in The positional encoding is obtained based on the Laplace transform matrix of the adjacency matrix A, where W is the parameter matrix. A multi-head attention mechanism is used to learn node features at different levels, and the outputs of each head are concatenated and dimensionality reduced to obtain... .

[0053] In this embodiment, to utilize the structural information of the graph to construct the degree fusion module, the calculation formula is as follows:

[0054]

[0055] in For nodes The degree, and These are learnable parameters.

[0056] Feedforward neural networks and batch normalization layers are used to stabilize the model training. The final result is a global feature representation of the nodes. :

[0057]

[0058]

[0059] S4. Based on the attribute bipartite graph, an adaptive graph convolutional neural network is used to calculate the required number of feature smoothing iterations for each node in the graph, in order to fully extract the local feature embeddings of the nodes. The implementation is as follows:

[0060] S401, for each node in the attribute bipartite graph G, calculate the number of feature smoothing steps required;

[0061] S402. Based on the adaptive feature smoothing times, a simplified graph convolutional neural network is used to extract local features of nodes, avoiding the problem of over-smoothing features.

[0062] In this embodiment, the NDLS model is used to calculate the number of adaptive feature smoothing iterations required for each node in graph G, as shown in the following formula:

[0063]

[0064] in Let be any small constant greater than 0. Based on the number of feature smoothing iterations, the activation functions in the GCN are removed to extract local features of the nodes. :

[0065]

[0066] S5. By fusing global and local features, a miRNA-disease pair feature representation is constructed. A multilayer perceptron classifier is then used to predict potential associations. The implementation is as follows:

[0067] S501. Utilize a fully connected neural network to fuse the global and local features of a node to obtain the final feature representation of the node;

[0068] S502. The features of miRNA and the features of disease are spliced ​​together to obtain the feature representation of miRNA-disease pair;

[0069] S503. Using a multilayer perceptron classifier, predict potential associations based on the characteristics of miRNA-disease pairs.

[0070] In this embodiment, the formula for fusing features using a fully connected neural network is as follows:

[0071]

[0072] in Let b be the parameter matrix, and b be the deviation.

[0073] The features of miRNAs and disease are concatenated to obtain a feature representation of the miRNA-disease pair. This representation is then input into a multilayer perceptron (MLP) classifier for association prediction.

[0074]

[0075] In this embodiment, the cross-entropy loss function is used to optimize the parameters of the neural network.

[0076] In this embodiment, five-fold cross-validation is used to test the performance of the method provided by this invention, and it is compared with four other graph neural network-based models. Figure 2 and Figure 3 As shown, the AUC and AUPR values ​​of our method (our) are both higher than those of the four comparative models.

[0077] In this embodiment, the effectiveness of each part of the proposed method was verified through ablation experiments. A comparison was made between methods using only global features (our_global) and local features (our_local), and methods using Transformer to extract global features and GCN to extract local features. Figure 4 As shown, the global and local features extracted by the proposed method are both effective and necessary, and using both types of features simultaneously can achieve better prediction performance.

[0078] In this embodiment, to fully demonstrate the applicability of the proposed method in real-world scenarios, a case study experiment was conducted using breast cancer. The potential miRNA disease associations predicted by the method were validated by searching relevant literature evidence in the PubMed database. As shown in Table 1, the top 30 potentially associated miRNAs predicted by the method were all validated.

[0079] Table 1. Results of a case study of breast tumors

[0080] 1 hsa-mir-150 36340453 16 hsa-mir-181c 35611706 2 hsa-mir-142 36443711 17 hsa-mir-494 32372307 3 hsa-mir-15b 33519190 18 hsa-mir-99b 36381329 4 hsa-mir-106a 34837907 19 hsa-mir-30e 28288641 5 hsa-mir-192 31485620 20 hsa-mir-92b 29661250 6 hsa-mir-99a 33177704 21 hsa-mir-424 36305147 7 hsa-mir-378a 26255816 22 hsa-mir-491 33488122 8 hsa-mir-138 36340453 23 hsa-mir-184 34000513 9 hsa-mir-130a 35797350 24 hsa-mir-144 26124344 10 hsa-mir-212 33187481 25 hsa-mir-98 28232182 11 hsa-mir-196b 30416655 26 hsa-mir-483 34488224 12 hsa-mir-370 29698983 27 hsa-mir-542 34402683 13 hsa-mir-372 29456685 28 hsa-mir-130b 35191803 14 hsa-mir-208a 30911286 29 hsa-mir-134 34378285 15 hsa-mir-28 34593318 30 hsa-mir-181d 36340453

Claims

1. A miRNA disease association prediction method based on global and local graph representation learning, characterized in that, The miRNA disease association prediction method based on global and local graph representation learning includes the following steps: S1. Utilize multi-source information to calculate multiple similarity matrices between diseases and miRNAs, construct a comprehensive similarity feature matrix for nodes, and then fuse them: The comprehensive similarity feature matrix of nodes includes the comprehensive similarity matrix of diseases and the comprehensive similarity matrix of miRNAs, and the corresponding expressions are as follows: Comprehensive similarity matrix of diseases: Comprehensive similarity matrix of miRNAs: in This represents the semantic similarity matrix computed based on a directed acyclic graph of a disease. This represents the Gaussian interaction spectrum kernel similarity matrix calculated based on the disease-related information representation vectors. This represents the miRNA sequence similarity matrix calculated using the Niedermann-Onsch algorithm. This represents the miRNA functional similarity matrix calculated based on associated disease groups. The Gaussian interaction spectrum kernel similarity matrix is ​​calculated based on the miRNA-based association information representation vector. After obtaining the comprehensive similarity matrices of miRNA and disease respectively, two autoencoders are used to map the features of miRNA and disease into the same low-dimensional space, thereby achieving feature dimensionality reduction and fusion. S2. Based on the fusion characteristics and known miRNA disease associations, construct a bipartite miRNA-disease attribute graph: Using miRNAs and diseases as nodes, and features fused by an autoencoder as the initial features of the nodes, edges are set based on known miRNA-disease associations recorded in public databases to construct a bipartite miRNA-disease attribute graph. ;picture The adjacency matrix is ​​denoted as The node feature matrix is ​​denoted as ; S3. Based on the attribute bipartite graph, use the graph Transformer neural network to extract the global feature embeddings of the nodes; S4. Based on the attribute bipartite graph, use an adaptive graph convolutional neural network to calculate the number of feature smoothing steps required for each node in the graph, so as to fully extract the local feature embedding of the node and avoid the problem of feature over-smoothing. S5. Integrate global and local features to construct a miRNA-disease pair feature representation, and use a multilayer perceptron classifier to predict potential associations.

2. The miRNA disease association prediction method based on global and local graph representation learning according to claim 1, characterized in that, Step S3 includes the following steps: S301, Based on the attribute bipartite diagram The degree information of the nodes is used to construct a degree fusion module to inject the inductive bias information of the graph; S302. Build a node feature extraction framework based on Transformer, add the degree fusion module to the framework, and construct a graph Transformer neural network to extract global features of nodes.

3. The miRNA disease association prediction method based on global and local graph representation learning according to claim 1, characterized in that, Step S4 includes the following steps: S401, respectively, attribute bipartite diagram For each node in the array, calculate the number of feature smoothing iterations required. S402. Based on the adaptive feature smoothing times, use a graph convolutional neural network to extract local features of nodes to avoid over-smoothing of features.

4. The miRNA disease association prediction method based on global and local graph representation learning according to claim 1, characterized in that, Step S5 includes the following steps: S501. Utilize a fully connected neural network to fuse the global and local features of a node to obtain the final feature representation of the node; S502. The features of miRNA and the features of disease are spliced ​​together to obtain the feature representation of miRNA-disease pair; S503. Using a multilayer perceptron classifier, predict potential associations based on the characteristics of miRNA-disease pairs.

Images