Similarity fusion and mixed gating-based circular RNA and disease association prediction method

Abstract

The invention discloses a circular RNA (Ribonucleic Acid) and disease association prediction method based on similarity fusion and hybrid gating, and belongs to the cross technical field of computers and bioinformatics. The objective of the invention is to solve the problem of poor prediction effect caused by fusion information loss and insufficient capability of capturing complex relationships in current circular RNA and disease association prediction based on deep learning. The method comprises the following steps: firstly, respectively constructing comprehensive similarity of circular RNA and comprehensive similarity of diseases based on circular RNA and disease data; performing inter-node feature aggregation on the similarity network by using an isomorphic network with a residual error; performing feature propagation between heterogeneous graph networks on the obtained features; and finally, performing association prediction by using a multi-layer perceptron.

Description

Technical Field

[0001] The present invention belongs to the interdisciplinary technical field of computer and bioinformatics, and relates to a method for predicting the association between circular RNA and diseases. Background Art

[0002] Circular RNAs (circRNAs) are a special class of non-coding RNAs that form closed circular structures through reverse splicing, offering unique advantages in stability and functional regulation. Although once considered abnormal byproducts of gene transcription, recent advances in high-throughput sequencing technologies have revealed that circular RNAs are widely present in eukaryotes and play a key role in disease progression through a variety of mechanisms. Circular RNAs can influence disease progression by acting as sponges for miRNAs, binding to regulatory elements such as transcription factors, and adsorbing RBPs to regulate splicing, playing a crucial role in cancer, cardiovascular disease, and neurodegenerative diseases. Given the crucial role of circular RNAs in regulating gene expression and disease development, predicting their function is crucial for uncovering the pathogenic mechanisms of disease. This provides crucial clues for discovering potential therapeutic targets and developing precise treatment strategies.

[0003] In recent years, as circular RNA has gained increasing attention, many studies have used wet lab methods to investigate the association between circular RNA and disease. However, wet lab methods require a lot of manpower and financial resources, so it is necessary to develop effective computational models to predict associations.

[0004] Existing methods for predicting the association between circular RNA and diseases are divided into three categories: network propagation-based, traditional machine learning-based, and deep learning-based. Among them, the network propagation-based method predicts new associations by constructing heterogeneous networks and calculating the similarity of known associations. Traditional machine learning models usually use manually extracted features to train classifiers to predict associations. Although many methods have made certain progress, and although these models have significantly improved the efficiency of association prediction, their performance relies heavily on manually extracted features. However, not all circular RNAs and diseases have appropriate similarity properties. For example, when calculating disease semantic similarity, some diseases are not included in certain databases, which will result in a similarity value of zero, which seriously affects the performance of the model.

[0005] To address these issues, deep learning-based methods have been used to construct deep learning networks that automatically learn features and predict associations. While many methods have made progress, enabling learning associations through different feature representations and providing different insights into circular RNAs and diseases, existing methods lack the ability to capture complex relationships during network propagation and are prone to information loss during data fusion. Consequently, existing deep learning-based prediction methods still suffer from suboptimal prediction accuracy. Summary of the Invention

[0006] The present invention aims to solve the problems of poor prediction results caused by loss of fusion information and insufficient ability to capture complex relationships in the current deep learning-based circular RNA and disease association prediction.

[0007] A circular RNA and disease association prediction method based on similarity fusion and hybrid gating, including:

[0008] S200, based on the identifier of each disease, calculate the disease and The semantic similarity between , using Gaussian interaction kernel similarity to calculate the similarity between diseases ; Calculate the ring Functional similarity , using Gaussian interaction kernel to calculate the similarity between circular RNAs ;

[0009] S300, the similarity matrix of circular RNA Normalize to get the normalized h-th similarity matrix , Represents the hth similarity matrix of circular RNA, h=1,2 represents the corresponding The similarity matrix and The similarity matrix is ​​constructed, and for each similarity matrix, the neighbor information matrix is ​​calculated ;Will As initial value , iteratively update the similarity matrix: , represents the similarity matrix after the tth iteration, yes The initial value is the standardized Similarity matrix; after the tth iteration, the comprehensive similarity of circular RNA ;

[0010] For diseases, the same method is used to obtain the comprehensive similarity matrix of the disease ;

[0011] Constructing a comprehensive heterogeneous network of circular RNAs and diseases , is the association matrix corresponding to the association data between circular RNA and diseases;

[0012] S400, and The features are respectively input into the homogeneous network with residual connection for feature aggregation, and the aggregation results are obtained respectively. Then, the initial feature matrix of the comprehensive heterogeneous network HA is obtained by splicing. ;

[0013] S500, Input to the heterogeneous graph propagation based on hybrid gating to obtain the feature fusion result , and based on the fusion results Obtaining the final characteristics of circular RNA and the final characteristics of the disease ;

[0014] S600, for each circular RNA-disease pair, the final circular RNA signature and the final characteristics of the disease Splicing to obtain circRNA and diseases The paired feature vector of , and then a multi-layer perceptron was used to predict the association between circular RNA and diseases.

[0015] Furthermore, the disease and The semantic similarity between as follows:

[0016]

[0017] in, and Represents diseases and ancestral disease collection; Indicates ancestral disease right Semantic contribution of:

[0018]

[0019] in, Represents the child nodes of t.

[0020] Furthermore, the similarity between diseases is calculated using Gaussian interaction kernel similarity The normalized bandwidth parameters used in the process are as follows:

[0021]

[0022] in, is the number of columns of the incidence matrix, for The norm of the column vector in the column.

[0023] Furthermore, the annular Functional similarity as follows:

[0024]

[0025] in, Indicates a ring The associated disease set, Indicates a ring The associated disease set, 、 express 、 capacity; disease With collection Semantic similarity scores of all diseases in .

[0026] Furthermore, the Gaussian interaction kernel was used to calculate the similarity between circular RNAs. The normalized bandwidth parameters used in the process are as follows:

[0027]

[0028] in, is the number of rows of the incidence matrix, for The norm of the row vector in the row.

[0029] Furthermore, the similarity matrix of circular RNA The standardization is done as follows:

[0030]

[0031] in, Represents the h-th similarity matrix of circular RNA.

[0032] Furthermore, for each similarity matrix, the neighbor information matrix is ​​calculated as follows:

[0033]

[0034] in, is the number of neighbors of node i and node j, represents the neighbors of node i, represents the neighbors of node j.

[0035] Further, Input to the heterogeneous graph propagation based on hybrid gating to obtain the feature fusion result The process includes:

[0036] Building a heterogeneous network Afterwards, the features Input into a graph attention layer to obtain the feature representation of the node; the obtained features are used to aggregate heterogeneous node features to obtain the aggregated features , and then a linear transformation is performed to obtain, and then a bilinear pooling layer is used to fuse the linearly transformed aggregate features from different types of nodes ;

[0037] Then, a gating mechanism is used to optimize the feature fusion process. The gating mechanism achieves feature fusion through the following formula:

[0038]

[0039] in, represents the sigmoid activation function, is the gating weight vector; is the activation function.

[0040] Furthermore, the features of heterogeneous nodes are aggregated to obtain the aggregated features as follows:

[0041]

[0042] in, Representation node Heterogeneous neighbor features, Representation node Neighbor set in heterogeneous graph HA; 、 Representation node and the degree of j, is the feature of node j obtained after the graph attention layer.

[0043] Furthermore, in the process of using bilinear pooling layers to fuse the linearly transformed aggregate features from different types of nodes, for nodes , its associated Heterogeneous nodes will participate in bilinear pooling, and then the results are averaged to obtain the bilinear pooling features of the node.

[0044] Beneficial effects:

[0045] Against this background, existing methods have insufficient ability to capture complex relationships during network propagation, and are prone to information loss during data fusion. The present invention not only can fully capture the characteristics of complex relationships through homogeneous networks and heterogeneous networks, but also can ensure that the extracted complex relationships are fully mined during later fusion; in addition, the present invention combines the similarity network fusion module and the hybrid gated heterogeneous network, making full use of their respective advantages for learning, which can not only reduce the information loss in the data fusion process, but also fully mine the complex relationship characteristics, thereby improving the prediction effect of circular RNA and disease association prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Flowchart of the circular RNA and disease association prediction method based on similarity fusion and hybrid gating;

[0047] Figure 2 Schematic diagram of the overall structure of the circular RNA and disease association prediction method based on similarity fusion and hybrid gating. DETAILED DESCRIPTION

[0048] Specific implementation method 1: Combination Figure 1 and Figure 2 To explain this embodiment,

[0049] The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating described in this embodiment includes the following steps:

[0050] Step 1: Collect and organize known association datasets between circular RNA and diseases from multiple public databases:

[0051] The association data between circular RNA and diseases were collected from the public circular RNA database, and the association matrix between circular RNA and diseases was constructed based on this data. The public databases used in the present invention are: CircR2Diseasev2.0, Circ2Disease, and circRNADisease. The association data were downloaded from the above databases, the human association data were screened out, and the data were deduplicated. Among them, the CircR2Diseasev2.0 database contains 725 associations between 661 circular RNAs and 100 diseases. Circ2Disease collected 273 associations, including 237 circular RNAs and 54 diseases. circRNADisease contains 354 associations between 330 circular RNAs and 48 diseases.

[0052] The processed association data is converted into an association matrix, represented by matrix A, where rows represent circular RNAs and columns represent diseases. Indicates the Circular RNA and There is an association with the disease, otherwise it is 0.

[0053] Step 2: Based on the original association dataset, construct the similarity network of circular RNA and the similarity network of disease:

[0054] Step 2.1. Retrieve the identifier of each disease from the Disease Ontology website and calculate the disease and The semantic similarity between :

[0055]

[0056] in, and Represents diseases and ancestral disease collection; Indicates ancestral disease right The semantic contribution of is calculated as follows:

[0057]

[0058] in, Represents the child nodes of t.

[0059] Step 2.2: Use Gaussian interaction kernel similarity to calculate the similarity between diseases. The formula is:

[0060]

[0061] in, and Diseases and Column vector in the circRNA-disease association matrix; To normalize the bandwidth parameter, the calculation formula is:

[0062]

[0063] in, is the number of columns of the incidence matrix, for The norm of the column vector in the column.

[0064] Step 2.3, use the following formula to calculate the ring Functional similarity:

[0065]

[0066] in, Indicates a ring The associated disease set, Indicates a ring The associated disease set, 、 express 、 capacity; disease With collection Semantic similarity scores of all diseases in .

[0067] Step 2.4: Use the Gaussian interaction kernel to calculate the similarity between circular RNAs. The formula is:

[0068]

[0069] In fact, Used to represent the corresponding vector in the association matrix, if it is a disease It represents the column vector corresponding to the disease. If it is a ring It represents the row vector corresponding to circular RNA;

[0070] To normalize the bandwidth parameter, the calculation formula is:

[0071]

[0072] in, is the number of rows of the incidence matrix, for The norm of the row vector in the row.

[0073] Step 3: In order to effectively fuse different similarity data, the present invention designs a similarity fusion module SNFM to alleviate the problem of information loss that is prone to occur during the fusion process of similarity data. The following takes the comprehensive similarity construction process of circular RNA as an example:

[0074] Step 3.1: First, the similarity matrix of circular RNA Standardize.

[0075]

[0076] in, represents the hth similarity matrix of circular RNA, h=1,2 represents two similarity matrices of circular RNA, respectively The similarity matrix and The similarity matrix formed; Represents the normalized h-th similarity matrix.

[0077] Step 3.2: After that, for each similarity matrix, calculate the neighbor information matrix.

[0078]

[0079] in, is the number of neighbors of node i and node j, represents the neighbors of node i, represents the neighbors of node j.

[0080] Step 3.3: Use the normalized similarity matrix and the neighbor information matrix to iteratively update the similarity matrix.

[0081]

[0082] in, represents the similarity matrix after the tth iteration; , yes The initial value of is the normalized similarity matrix.

[0083] Step 3.4: After the tth iteration, the final similarity matrix is

[0084]

[0085] Based on steps (2.1) and (2.2), the comprehensive similarity matrix of the disease is obtained in the same way .

[0086] In the circular RNA similarity network, the weight of the edge is the comprehensive similarity In the disease similarity network, the weight of the edge is the comprehensive similarity value.

[0087] Step 3.5: Based on this, the present invention combines the circular RNA-disease association matrix , circular RNA comprehensive similarity matrix , disease comprehensive similarity matrix Constructing a comprehensive heterogeneous network of circular RNAs and diseases:

[0088]

[0089] Comprehensive heterogeneous network HA actually represents the relationship between nodes;

[0090] Step 4: Similarity matrix between circular RNA and disease and Input them into the homogeneous network with residual connection respectively for feature aggregation:

[0091] Since the similarity matrices of circular RNA and disease obtained belong to different feature spaces, the two similarity matrices are first mapped to the same dimension, which is set to 128, as the input features of the similarity network.

[0092] In the circular RNA similarity network, the embedding representation of circular RNA is updated by aggregating the features of its neighbor nodes. The output features of the layer are calculated by the following formula:

[0093]

[0094] in, Indicates that from circular RNA Features obtained by aggregating neighbor nodes of Representing circular RNAs in the circular RNA similarity network neighbors; AGG is the aggregation function, and the aggregation function here uses the sum function; Represents the RNA similarity network The features obtained by layer node j.

[0095] Subsequently, a residual connection operation is performed on the aggregated features:

[0096]

[0097] Then the obtained features Perform splicing and nonlinear transformation, and finally serve as The output of the layer is:

[0098]

[0099] in, is the LeakyReLU nonlinear activation function. is a learnable vector parameter that represents the importance of the circular RNA node itself and its neighbors’ features; Indicates splicing processing.

[0100] Similar to circular RNA, the features of disease nodes are updated in the same way on the disease similarity network:

[0101]

[0102] The residual connection operation is used for the aggregated features:

[0103]

[0104] Perform splicing and nonlinear transformation to get the final output:

[0105]

[0106] in, Indicates the Layer disease The output features of Indicates disease Neighbor nodes in the disease similarity network, is a learnable vector parameter that expresses the disease The importance of its own features and the features of its neighboring nodes. Here, the number of layers of the propagation network is set to 3.

[0107] After obtaining the characteristics of circular RNA and disease respectively, splicing processing is performed to obtain the initial feature matrix of the comprehensive heterogeneous network HA.

[0108]

[0109] Step 5: The obtained features are further input into the hybrid gating-based heterogeneous graph propagation to capture the complex relationships between different nodes:

[0110] Step 5.1: Building a heterogeneous network Afterwards, the features The input is fed into a graph attention layer to capture important interactions between nodes and identify the nodes that contribute most to the feature representation of the target node. These nodes are then assigned higher weights. Specifically, the graph attention layer calculates the feature representation of node i using the following formula:

[0111]

[0112] in, Representation node The initial feature representation of Representation node The set of neighbor nodes of is a node With node The attention weights between is the linear transformation matrix.

[0113] Attention weight The calculation formula is as follows:

[0114]

[0115] in, is a learnable weight vector, and the symbol || represents a vector concatenation operation.

[0116] Step 5.2: Use the obtained features to aggregate the features of heterogeneous nodes, that is, aggregate features from different types of neighboring nodes. The feature aggregation formula for heterogeneous neighbors is as follows:

[0117]

[0118] in, Indicates the Layer Time Node Heterogeneous neighbor features, Representation node Neighbor set in heterogeneous graph HA; Representation node degree.

[0119] After obtaining the aggregated features, these features are further processed through linear transformation. The specific process is as follows:

[0120]

[0121] in, Linear transformation matrix representing heterogeneous features.

[0122] Step 5.3, then use the bilinear pooling layer to fuse the linearly transformed aggregate features from different types of nodes. The bilinear pooling layer achieves feature fusion by calculating the outer product of two feature vectors and applying linear transformation. , its associated Heterogeneous nodes will participate in bilinear pooling, and then the results are averaged to obtain the bilinear pooling features of the node. The specific process is as follows:

[0123]

[0124] in, represents bilinear processing; Represents the set of N neighbors in the heterogeneous graph HA.

[0125] Step 5.4: Next, use the gating mechanism to optimize the feature fusion process. The core of this mechanism is its ability to dynamically adjust the weights of various features during the feature fusion process. The gating mechanism achieves feature fusion through the following formula:

[0126]

[0127] in, represents the sigmoid activation function, is the gating weight vector, is the dimension of the output feature; is the activation function.

[0128] This gating mechanism achieves more flexible feature combination by dynamically adjusting the weights between fusion features and neighbor features. Determining circular RNA characteristics and disease characteristics In this embodiment, , ;use and Representing the final signatures of circular RNA and disease, respectively.

[0129] Step 6: Use a multi-layer perceptron to predict the results:

[0130] For each circRNA-disease pair, their feature vectors learned in the final GCN layer are concatenated. Subsequently, the method predicts the association probability between each circRNA and disease pair.

[0131] The formula for concatenating eigenvectors is as follows:

[0132] in, Indicates circRNA and diseases The paired feature vectors of .

[0133] After representation learning, the learned features of circular RNA and disease can be used to predict associations. The present invention uses a multi-layer perceptron for prediction.

[0134]

[0135] In this embodiment, the number of neurons in each layer of the multilayer perceptron is set to 128. This enables the association prediction of circular RNA and diseases based on similarity fusion and hybrid gating.

[0136] The present invention constructs the comprehensive similarity of circular RNA and the comprehensive similarity of diseases based on circular RNA and disease data, respectively; uses a homogeneous network with residuals to aggregate features between nodes in the similarity network; and propagates the obtained features between heterogeneous graph networks. Therefore, the present invention can overcome the limitation of previous learning methods that can only propagate in similarity networks, and adopts heterogeneous network propagation on the basis of homogeneous network propagation, thereby obtaining richer features and improving model performance.

[0137] The above examples are merely illustrative of the calculation model and process of the present invention and are not intended to limit the embodiments of the present invention. Persons skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. This list of embodiments is not exhaustive; however, any obvious variations or modifications derived from the technical solution of the present invention remain within the scope of protection of the present invention.

Claims

1. A method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating, characterized in that: include: S200, based on the identifier of each disease, calculate the disease and The semantic similarity between , using Gaussian interaction kernel similarity to calculate the similarity between diseases ; Calculate the ring Functional similarity , using Gaussian interaction kernel to calculate the similarity between circular RNAs ; S300, the similarity matrix of circular RNA Normalize to get the normalized h-th similarity matrix , Represents the hth similarity matrix of circular RNA, h=1,2 represents the corresponding The similarity matrix and The similarity matrix is ​​constructed, and for each similarity matrix, the neighbor information matrix is ​​calculated ;Will As initial value , iteratively update the similarity matrix: , represents the similarity matrix after the tth iteration, yes The initial value is the standardized Similarity matrix; after the tth iteration, the comprehensive similarity of circular RNA ; For diseases, the same method is used to obtain the comprehensive similarity matrix of the disease ; Constructing a comprehensive heterogeneous network of circular RNAs and diseases , is the association matrix corresponding to the association data between circular RNA and diseases; S400, and The features are respectively input into the homogeneous network with residual connection for feature aggregation, and the aggregation results are obtained respectively. Then, the initial feature matrix of the comprehensive heterogeneous network HA is obtained by splicing. ; S500, Input to the heterogeneous graph propagation based on hybrid gating to obtain the feature fusion result , and based on the fusion results Obtaining the final characteristics of circular RNA and the final characteristics of the disease ; S600, for each circular RNA-disease pair, the final circular RNA signature and the final characteristics of the disease Splicing to obtain circRNA and diseases The paired feature vector of , and then a multi-layer perceptron was used to predict the association between circular RNA and diseases.

2. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 1, characterized in that: disease and The semantic similarity between as follows: in, and Represents diseases and ancestral disease collection; Indicates ancestral disease right Semantic contribution of: in, Represents the child nodes of t.

3. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 1, characterized in that: Calculating the similarity between diseases using Gaussian interaction kernel similarity The normalized bandwidth parameters used in the process are as follows: in, is the number of columns of the incidence matrix, for The norm of the column vector in the column.

4. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 1, characterized in that: ring Functional similarity as follows: in, Indicates a ring The associated disease set, Indicates a ring The associated disease set, 、 express 、 capacity; disease With collection Semantic similarity scores of all diseases in .

5. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 1, characterized in that: Calculating similarity between circular RNAs using Gaussian interaction kernel The normalized bandwidth parameters used in the process are as follows: in, is the number of rows of the incidence matrix, for The norm of the row vector in the row.

6. A method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to any one of claims 1 to 5, characterized in that: The similarity matrix of circular RNA The standardization is done as follows: in, Represents the h-th similarity matrix of circular RNA.

7. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 6, characterized in that: For each similarity matrix, the neighbor information matrix is ​​calculated as follows: in, is the number of neighbors of node i and node j, represents the neighbors of node i, represents the neighbors of node j.

8. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 7, characterized in that: Will Input to the heterogeneous graph propagation based on hybrid gating to obtain the feature fusion result The process includes: Building a heterogeneous network Afterwards, the features Input into a graph attention layer to obtain the feature representation of the node; the obtained features are used to aggregate heterogeneous node features to obtain the aggregated features , and then a linear transformation is performed to obtain, and then a bilinear pooling layer is used to fuse the linearly transformed aggregate features from different types of nodes ; Then, a gating mechanism is used to optimize the feature fusion process. The gating mechanism achieves feature fusion through the following formula: in, represents the sigmoid activation function, is the gating weight vector; is the activation function.

9. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 8, characterized in that: Aggregate heterogeneous node features to obtain aggregated features as follows: in, Representation node Heterogeneous neighbor features, Representation node Neighbor set in heterogeneous graph HA; 、 Representation node and the degree of j, is the feature of node j obtained after the graph attention layer.

10. The method for predicting the association between circular RNA and disease based on similarity fusion and hybrid gating according to claim 8, characterized in that: In the process of using bilinear pooling layer to fuse the linearly transformed aggregate features from different types of nodes, for node , its associated Heterogeneous nodes will participate in bilinear pooling, and then the results are averaged to obtain the bilinear pooling features of the node.

Images