MiRNA-disease association prediction method based on adaptive content guided fusion mechanism and multi-scale gated convolution

Abstract

The invention designs a miRNA-disease association prediction method based on an adaptive content guided fusion mechanism and multi-scale gated convolution. Firstly, miRNA function similarity, disease semantic similarity and Gaussian interaction spectrum kernel similarity of the miRNA function similarity and the disease semantic similarity are calculated, the weight of each similarity matrix is adjusted through a self-adaptive content guide fusion mechanism, and similarity information contributing to a prediction task is automatically concerned. The method comprises the following steps: firstly, carrying out feature extraction through multi-scale gating convolution by combining Laplacian features and Weisfeier-Lehman absolute role coding features, and finally, connecting learned feature representations of miRNA and diseases, and sending the feature representations into an XGBoost classifier for training and predicting potential association of miRNA and diseases. The working mechanism of miRNA in diseases is deeply explored, practical application value is achieved for clinical diagnosis and prevention, and meanwhile discovery of a novel therapy is promoted.

Description

Technical Field:

[0001] The present invention designs a method for predicting the association between miRNA and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution, which relates to the field of bioinformatics disease prediction. Background Art:

[0002] Small interfering RNA, also known as microRNA, is a kind of RNA molecule with a length of about 21 to 23 nucleotides widely existing in eukaryotes, which can regulate the expression of other genes. miRNA binds to the target mRNA, thereby inhibiting the post-transcriptional gene expression, and plays an important role in regulating gene expression, cell cycle, organism development and disease occurrence. The earliest example of the discovery of the association between miRNA and cancer is the occurrence of miR-15 and miR-16 in chronic lymphocytic leukemia. Subsequently, the participation of miRNA has been found in various different types of cancers. Therefore, the specific expression of miRNA makes it of great value in the judgment of disease occurrence. The detection of the expression level of specific miRNA can help evaluate the risk of the development or metastasis of cancer patients. Although biological experimental methods provide direct evidence in the discovery of miRNA-disease associations, due to their long time period and high cost, it limits the possibility of their large-scale application. Therefore, the present invention attempts to solve these problems through a method based on an adaptive content-guided fusion mechanism and multi-scale gated convolution, which helps to study the mechanism of miRNA in diseases, and further provides assistance for clinical diagnosis and disease prevention, and explores new therapies for diseases. Summary of the Invention:

[0003] The purpose of the present invention is to solve the problem that it takes time and cost to discover disease-related miRNA through biological experiments, and to assist in studying the mechanism of miRNA in diseases.

[0004] In order to achieve the above purpose, the present invention provides the following technical solution: A method for predicting the association between miRNA and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution, comprising the following steps:

[0005] (1) First, construct a miRNA functional similarity network, a miRNA Gaussian interaction spectrum kernel similarity network, a disease semantic similarity network, and a disease Gaussian interaction spectrum kernel similarity network by using the multi-view data of miRNA and diseases;

[0006] (2) Secondly, an adaptive content-guided fusion mechanism is used to fuse the similarity networks of miRNAs and diseases respectively; the integrated miRNA similarity matrix SM, the integrated disease similarity matrix SD, and the association matrix A between miRNAs and diseases are used to construct a heterogeneous network H, and at the same time, a Laplacian feature matrix and a Weisfeeller-Lehman absolute role encoding are generated for feature enhancement;

[0007] (3) Multi-scale gated convolution operations are adopted to mine the hidden feature information of miRNAs and diseases;

[0008] (4) Finally, an XGBoost classifier is used for classification, and the association prediction is finally completed. The association prediction between miRNAs and diseases is regarded as a binary classification task, divided into two categories: associated and non-associated. We set the threshold to 0.5. If the miRNA-disease association score exceeds 0.5, it is considered associated; otherwise, if the score is lower than 0.5, it is considered non-associated;

[0009] (5) A confusion matrix is constructed, and evaluation is carried out using indicators such as AUC, AUPR, Accuracy, Recall, Precision, and F1-score. Description of the Drawings:

[0010] Figure 1 It is a schematic flowchart of the steps of a method for predicting the association between miRNAs and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution according to the present invention.

[0011] Figure 2 It is a schematic structural diagram of the present invention. Detailed Embodiment:

[0012] In order to clearly and elaborately explain the technical solutions of the embodiments of the present invention, we will further describe the present invention in detail in combination with the examples in the drawings.

[0013] As Figure 1 shown, an embodiment of the present invention provides a method for predicting the association between miRNAs and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution, including the following steps:

[0014] Step 1: Obtain miRNA and disease data and perform data preprocessing.

[0015] Specifically:

[0016] Step 1-1: Obtain miRNA-disease association data from the microRNA Disease Dataset (HMDD) v2.0

[0017] Step 1-2: Delete duplicate and missing data.

[0018] Step 2: Construct multiple similarity networks using multi-view data of miRNAs and diseases, including miRNA functional similarity network, miRNA Gaussian interaction spectrum kernel similarity network, disease semantic similarity network, and disease Gaussian interaction spectrum kernel similarity network.

[0019] Specifically:

[0020] Step 2-1 Calculate miRNA functional similarity. Based on the assumption that miRNAs with similar functions are often associated with diseases with similar phenotypes and vice versa, we calculate miRNA functional similarity, obtain miRNA functional correlations from the MISIM database 2.0, and construct the corresponding functional similarity matrix. miRNAm i and m j 's functional similarity is labeled as FM(m i ,m j ).

[0021] Step 2-2 Calculate the Gaussian interaction spectrum kernel similarity between miRNAs and diseases. The calculation formula for the Gaussian interaction spectrum kernel similarity between miRNAs and diseases is as follows:

[0022] GM(m i ,m j ) = exp(-γ m ||IP(m i ) - IP(m j )|| 2 )

[0023] GD(d i ,d j ) = exp(-γ d ||IP(d i ) - IP(d j )|| 2 )

[0024]

[0025] Among them, m i and m j represent different miRNAs, d i and d j represent different diseases respectively. GM and GD are the Gaussian interaction spectrum kernel similarities of miRNAs and diseases respectively. In addition, γ m and γ d are parameters for adjusting the kernel bandwidth. Based on previous studies, γ′ m , γ′ d are set to 1, n m , n drespectively represent the number of miRNAs and diseases, and P(m t ) represents the t-th row of the miRNA-disease association matrix, and P(d t ) represents the t-th column of the miRNA-disease association matrix.

[0026] Step 2-3 Calculate the semantic similarity of diseases. The Medical Subject Headings (MESH) provides a comprehensive disease classification system, and its descriptors can be downloaded from https: / / www.nlm.nih.gov / mesh / meshhhome. Based on this, a directed acyclic graph (DAG) of disease information is constructed. In this study, we adopted two different methods to calculate the semantic similarity of diseases. Among them, the semantic contribution degree DSC j of the first disease semantic similarity d 1 is calculated according to the topological structure of the directed acyclic graph (DAG) of disease d i , and the formula is as follows:

[0027]

[0028] Among them, θ represents the weight of semantic contribution. Based on previous studies, θ is set to 0.5. C(d j ) represents the set of child disease nodes of disease d j .

[0029] Therefore, the semantic value DSV i of disease d 1 includes:

[0030]

[0031] Among them, A(d i ) represents the set of nodes of disease d i , including disease d i itself and its ancestor nodes. According to previous research assumptions, the larger the intersection of the directed acyclic graphs (DAGs) of disease d j and disease d i , the higher the semantic similarity between disease d j and disease d i . The semantic similarity between disease d j and disease d 2 can be derived according to the following formula:

[0032]

[0033] Represents the disease semantic similarity matrix obtained by the first method. Different from the first method that emphasizes the semantic contribution degree in the disease DAG topology, the second method focuses on evaluating the frequency of occurrence of a specific disease in other disease DAGs. Therefore, the second method calculates the disease semantic contribution degree DSC 2 The formula for

[0034]

[0035] is as follows: j where N(d j ) represents the number of DAGs containing disease d d . n i represents the number of diseases. Therefore, the semantic value (DSV 2 ) of disease d

[0036]

[0037] is calculated as follows: i and disease d j The semantic similarity DSS 2 between them can also be determined by the common nodes shared in the two DAGs:

[0038]

[0039] Finally, we combine the disease semantic similarity matrices calculated by the two methods, that is, perform weighted averaging, to obtain the final disease semantic similarity matrix DS(d i and disease d j , d i , d j ), and the calculation formula is as follows:

[0040]

[0041] Step 3: Use the adaptive content-guided fusion mechanism to fuse the similarity networks of miRNAs and diseases respectively.

[0042] Specifically:

[0043] Step 3-1 Use graph convolution to aggregate neighbors of miRNAs and diseases. GCN has been widely used due to its excellent ability to capture complex structural information and implicit interaction patterns. In this study, GCN is used to aggregate potential similarity information in the network to effectively extract the discriminative features of nodes. Here, we take the miRNA similarity network as an example, and the same method is used for diseases. First, we normalize the adjacency matrix of the miRNA similarity network:

[0044]

[0045] Among them, represents the adjacency matrix plus the identity matrix Ι, is the matrix of the degree matrix. Then, according to the layer propagation rule of GCN, we calculate the node feature representation of miRNA, and the formula is as follows:

[0046]

[0047] Among them, represents the embedding representation of the node at layer l, where F m represents the feature dimension of each miRNA node. When l = 0, is the initial feature GM / FM. represents the learnable weight matrix, and σ(·) represents the non-linear activation function. Then the miRNA similarity matrix and the Gaussian spectral kernel similarity matrix after passing through GCN are H FM , H GM .

[0048] Step 3-2 generates the importance mapping matrix SIMs of a specific input through content-guided attention. Here, still taking the similarity fusion of miRNA as an example. The miRNA similarity matrices H FM , H GM after passing through GCN are used as inputs, represents their sum. We first calculate the channel attention weight matrix and the spatial attention weight matrix The formula is as follows:

[0049]

[0050] Among them, max(0, x) represents the ReLU activation function, C k×k (·) represents the convolutional layer with a kernel size of k×k, and [·] represents the channel connection operation. respectively represent global average pooling across the spatial dimension, global average pooling across the channel dimension, and max pooling across the channel dimension. Then, we fuse and through a simple addition operation with a broadcasting mechanism to obtain the rough SIMs The formula is as follows:

[0051]

[0052] Finally, through the guidance of the input content, the rough SIMs are adjusted for corresponding features to finally generate the fine SIMs Matrix. Specifically, the channels of W and X are rearranged alternately through a channel shuffle operation, and the formula is as follows: coa and X sm each channel, the formula is as follows:

[0053]

[0054] where σ represents the sigmoid activation function, CS(·) represents the shuffle operation, represents the grouped convolutional layer with kernel size K×K, and the final fused feature representation the formula is as follows:

[0055] SM = X sm + H GM × W sm + H FM × (1 - W sm )

[0056] Similarly, the fusion of the disease similarity matrix is carried out in the same way as above. The difference is that the dimension of the disease feature matrix will change accordingly, that is, the node numbers in the disease graph. The node features can be flexibly adjusted as needed. When there is no longer a large change, the learning stops and the final node representation is obtained.

[0057] Step 4: Construct a heterogeneous network H from the fused miRNA similarity matrix SM, disease similarity matrix SD, and miRNA-disease association matrix A, and at the same time generate a Laplacian feature matrix and Weisfeeller-Lehman absolute role coding for feature enhancement.

[0058] Specifically:

[0059] Step 4-1: Based on the miRNA-miRNA similarity matrix SM, disease-disease similarity matrix SD, and miRNA-disease association matrix A, we construct a heterogeneous network as the feature representation of miRNA-disease. The adjacency matrix H of the heterogeneous network is represented as follows:

[0060]

[0061] where A T is the transpose matrix of the association matrix A. If N m represents the number of miRNAs and N d represents the number of diseases, then the first N m rows in the matrix H represent the feature representation of miRNAs, and the last N d rows represent the feature representation of diseases.

[0062] Step 4-2 Calculate the Laplacian matrix on the graph structure, and then obtain the eigenvector representation of each node. The general definition and normalization of the eigenvector are as follows:

[0063] L = D - A

[0064]

[0065] where D represents the degree matrix and A represents the original miRNA-disease association matrix.

[0066] Step 4-3 The Weisfeeller-Lehman absolute role algorithm is implemented based on the assumption that nodes with similar roles are more likely to have the same feature representation. It means that a unique feature set will be obtained on the graph, which implies that each node on the graph has a unique role positioning. Specifically, first obtain the features of all neighbor nodes j of node v i Then aggregate the features of the surrounding nodes to the current node by using a hash injective function, that is Repeat this step until each node converges, and finally obtain the WL eigenvalue WL(v ) of the node. The formula is as follows: j )

[0067]

[0068] where represents the h-th iteration of node v i , represents the set of all neighbor nodes of node v i . After obtaining the WL eigenvalue of the node, calculate the node feature embedding through the position encoding formula as follows

[0069]

[0070] where i represents the index traversing from 0 to , and d h represents the dimension of the input data. This process can be regarded as a balance between absolute role embedding capturing global information and relative position embedding capturing local information.

[0071] Step 5: Adopt multi-scale gated convolution operations to mine hidden feature information.

[0072] Specifically:

[0073] Step 5-1: The features obtained in the previous step are divided into two parts before and after according to the number of channels after convolution. The first half is non-linearly processed by the Tanh function, and the second half controls the output information flow by the sigmoid function. Then, the two are combined by matrix multiplication. By stacking multiple gated convolution operations, the ability of long- and short-distance dependencies is improved. The specific implementation formula is as follows:

[0074] Z out = ReLU(Concat(Pooling(Γ1*Z),Pooling(Γ2*Z),Pooling(Γ3,Z))+Z)

[0075] Γ*Z = φ(E)⊙σ(F)

[0076] Among them, Γ1, Γ2, and Γ3 represent the sizes of the convolution kernels as 1×S1, 1×S2, and 1×S3 respectively. The Concat(·) operation connects the features obtained from three gated convolutions of different scales. Γ*Z represents one gated convolution operation. φ represents the Tanh function, σ represents the sigmoid function, and E and F represent the first half and the second half of the number of channels after the convolution operation respectively. The final output is obtained through the ReLU activation function and in cooperation with the residual connection.

[0077] Step 6: Finally, the XGBoost classifier is used to calculate the association score between miRNA and disease. The association prediction between miRNA and disease is a binary classification problem, including two cases: associated and unassociated. The threshold is set to 0.5. miRNA-disease association pairs with an association score greater than 0.5 are considered associated, and those less than 0.5 are considered unassociated.

[0078] The above describes the embodiments of the present invention in detail with reference to the accompanying drawings. It should be emphasized that the provided implementation methods are only used to help understand the method of the present invention. Experts in the technical field can adjust and modify it according to the principle of the present invention. All modifications should be within the scope protected by the claims.

Claims

1. A method for predicting the association between miRNA and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution, mainly including the following steps: Step 1: Retrieve and download miRNA and disease data from multiple online databases and perform data preprocessing; Step 2: Construct multiple similarity networks using the multi-view data of miRNA and diseases, including miRNA functional similarity network, miRNA Gaussian interaction spectrum kernel similarity network, disease semantic similarity network, and disease Gaussian interaction spectrum kernel similarity network; Step 3: Use the adaptive content-guided fusion mechanism to fuse the similarity networks of miRNA and diseases respectively; Step 4: Construct a heterogeneous network H from the fused miRNA comprehensive similarity matrix SM, disease comprehensive similarity matrix SD, and the association matrix A between miRNA and diseases, and generate Laplacian feature matrix and Weisfeeller-Lehman absolute role encoding for feature enhancement; Step 5: Adopt multi-scale gated convolution operations to mine the hidden feature information of miRNA and diseases; Step 6: Use the XGBoost classifier for classification to finally complete the association prediction.

2. The method for predicting the association between miRNA and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution according to claim 1, in the said Step 1, obtain miRNA and disease data and perform data preprocessing, the specific steps are: Step 1-1: Obtain the miRNA-disease association data from microRNA Disease Dataset (HMDD) v2.

0. Step 1-2: Delete the duplicate and missing data.

3. The method for predicting the association between miRNA and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution according to claim 1, in the said Step 2, construct miRNA functional similarity network, miRNA Gaussian interaction spectrum kernel similarity network, disease semantic similarity network, and disease Gaussian interaction spectrum kernel similarity network using the multi-view data of miRNA and diseases, the specific steps are: Step 2-1 Calculate miRNA functional similarity. Based on the assumption that miRNAs with similar functions are often associated with diseases with similar phenotypes and vice versa, we calculated the miRNA functional similarity, obtained the miRNA functional correlation from the MISIM database 2.0, and constructed the corresponding functional similarity matrix miRNA m i and m j 's functional similarity is labeled as FM(m i , m j ). Step 2-2: Calculate the Gaussian interaction spectrum kernel similarity between miRNA and diseases, and the calculation formula for the Gaussian interaction spectrum kernel similarity between miRNA and diseases is as follows: GM(m i ,m j ) = exp(-γ m ||IP(m i ) - IP(m j )|| 2 ) GD(d i ,d j ) = exp(-γ d ||IP(d i ) - IP(d j )|| 2 ) Among them, m i and m j represent different miRNAs, d i and d j represent different diseases respectively. GM and GD are the Gaussian interaction spectrum kernel similarities of miRNAs and diseases respectively. In addition, γ m and γ d are parameters for adjusting the kernel bandwidth. Based on previous studies, γ′ m , γ′ d are set to 1, n m , n d represent the numbers of miRNAs and diseases respectively. P(m t ) represents the t-th row of the miRNA-disease association matrix, and P(d t ) represents the t-th column of the miRNA-disease association matrix. Step 2-3 Calculate disease semantic similarity. The Medical Subject Headings (MESH) provides a comprehensive disease classification system, and its descriptors can be downloaded from https: / / www.nlm.nih.gov / mesh / meshhhome . Based on this, a directed acyclic graph (DAG) of disease information is constructed. In this study, we used two different methods to calculate the semantic similarity of diseases. Among them, the semantic contribution degree DSC j of the first disease semantic similarity d 1 is calculated according to the topological structure of the directed acyclic graph (DAG) of disease d i . The formula is as follows: Among them, θ represents the weight of semantic contribution. Based on previous studies, θ is set to 0.5, and C(d j ) represents the set of child disease nodes of disease d j . Therefore, the semantic value DSV i of disease d 1 includes: Among them, A(d i ) represents the set of nodes of the disease, including the disease d i itself and its ancestor nodes. According to previous research hypotheses, the disease d i and the disease d j the larger the intersection of the directed acyclic graph (DAG) of the disease d i and the disease d j the higher the semantic similarity between these two diseases. The semantic similarity between the disease d i and the disease d j can be derived according to the following formula: Among them, represents the disease semantic similarity matrix obtained by the first method. Different from the first method that emphasizes the semantic contribution degree in the disease DAG topology, the second method focuses on evaluating the frequency of occurrence of a specific disease in other disease DAGs. Therefore, the second method calculates the disease semantic contribution degree DSC 2 The formula for is: where N(d j ) represents the number of DAGs containing disease d j , n d represents the number of diseases. Therefore, the semantic value of disease d i (DSV 2 ) is calculated as follows: Therefore, disease d i and disease d j The semantic similarity DSS 2 between them can also be determined by the common nodes shared in the two DAGs: Finally, we combine the disease semantic similarity matrices calculated by the two methods, that is, perform weighted averaging, to obtain the final disease semantic similarity matrix between disease d i and disease d j , denoted as DS(d i , d j ), and the calculation formula is as follows:

4. The method for predicting the association between miRNA and diseases based on an adaptive content-guided fusion mechanism and multi-scale gated convolution according to claim 1, in the said Step 3, use the adaptive content-guided fusion mechanism to fuse the similarity networks of miRNA and diseases respectively, the specific steps are: Step 3-1 uses graph convolution to aggregate neighbors of miRNA and disease. GCN has been widely used due to its excellent ability to capture complex structural information and implicit interaction patterns. In this study, GCN is used to aggregate potential similarity information in the network, so that it can effectively extract the distinguishing features of nodes. Here we take the miRNA similarity network as an example, and the disease is also carried out in the same way. First, we normalize the adjacency matrix of the miRNA similarity network: Among them, The adjacency matrix representing the addition of the identity matrix Ι, is the matrix of the degree matrix. After that, according to the layer propagation rule of GCN, we calculate the node feature representation of miRNA, and the formula is as follows: Among them, represents the embedding representation of the node at layer l, where F m represents the feature dimension of each miRNA node. When l = 0, is the initial feature GM / FM, represents the learnable weight matrix, σ(·) represents the non-linear activation function. Then the miRNA similarity matrix and the Gaussian spectral kernel similarity matrix after passing through GCN are H FM , H GM . Step 3-2 generates the importance mapping matrix SIMs of specific inputs by guiding attention through content. Still taking the similarity fusion of miRNA as an example, the miRNA similarity matrix H after passing through GCN FM ,H GM is used as the input, representing their sum. We first calculate the channel attention weight matrix and the spatial attention weight matrix The formulas are as follows: Among them, max(0, x) represents the ReLU activation function, and C k×k (·) represents the convolutional layer with a kernel size of k×k, and [·] represents the channel connection operation. respectively represent global average pooling across the spatial dimension, global average pooling across the channel dimension, and max pooling across the channel dimension. Then, through a simple addition operation with a broadcasting mechanism, we and are fused to obtain the rough SIMs The formula is as follows: Finally, guided by the input content, the rough SIMs are adjusted for corresponding features and finally refined SIMs matrices are generated. Specifically, by performing a channel shuffle operation to rearrange each channel of W coa and X sm alternately, the formula is as follows: where σ represents the sigmoid activation function, CS(·) represents the shuffle operation, represents the grouped convolutional layer with a kernel size of K×K, and the final fused feature representation is given by the following formula: SM = X sm + H GM × W sm + H FM × (1 - W sm ) Similarly, the similarity matrix fusion of the disease is carried out in the same way as above, except that the dimension of the disease feature matrix will change accordingly, that is, the node number in the disease graph, and the node features can be flexibly adjusted as needed. When the node features no longer change significantly, the learning stops and the final node representation is obtained.

5. According to the method for predicting the association between miRNA and disease based on adaptive content-guided fusion mechanism and multi-scale gated convolution according to claim 1, in step 4, the fused miRNA similarity matrix SM, disease similarity matrix SD and miRNA-disease association matrix A are used to construct a heterogeneous network H, and Laplace feature matrix and Weisfeeller-Lehman absolute role coding are generated for feature enhancement, and the specific steps are as follows: Step 4-1 Based on the miRNA-miRNA similarity matrix SM, the disease-disease similarity matrix SD and the miRNA-disease association matrix A, we constructed a heterogeneous network as the feature representation of miRNA-disease. The adjacency matrix H of the heterogeneous network is expressed as follows: Among them, A T is the transpose matrix of the incidence matrix A. If N m represents the number N of miRNAs d represents the number of diseases, then in the matrix H, the first N m rows represent the feature representations of miRNAs, and the last N d rows represent the feature representations of diseases. Step 4-2 calculates the Laplacian matrix on the graph structure, and then obtains the eigenvector representation of each node. The general definition and normalization of the eigenvector are as follows: L=DA Among them, D represents the degree matrix and A represents the original association matrix. Step 4-3 The Weisfeeller-Lehman absolute role algorithm is implemented based on the assumption that nodes with similar roles are more likely to have the same feature representation. It means that a unique set of features will be obtained on the graph, which implies that each node on the graph has a unique role positioning. Specifically, first, obtain the features of all neighbor nodes j of node v i Then, aggregate the features of the surrounding nodes to the current node by using a hash injective function, that is Repeat this step until each node converges, and finally obtain the WL feature value WL(v ) of the node. The formula is as follows: j ) Among them represents the h-th iteration of node v i , and the represents the set of all neighbor nodes of node v i . After obtaining the WL eigenvalue of the node, the node feature embedding is calculated through the position encoding formula as follows: Among them, i represents an index that traverses from 0 to , d h represents the dimension of the input data, and this process can be regarded as a balance between the absolute role embedding capturing global information and the relative position embedding capturing local information.

6. According to the method for predicting the association between miRNA and disease based on adaptive content-guided fusion mechanism and multi-scale gated convolution according to claim 1, in step 5, a multi-scale gated convolution operation is used to mine hidden feature information, and the specific steps are: Step 5-1: After convolution, the features obtained in the previous step are divided into two parts according to the number of channels. The first half is processed nonlinearly by the Tanh function, and the second half is controlled by the sigmoid function to output information flow. The two are then combined by matrix multiplication. By superimposing multiple gated convolution operations, the ability of long- and short-distance dependencies is improved. The specific implementation formula is as follows: Z out = ReLU(Concat(Pooling(Γ1*Z), Pooling(Γ2*Z), Pooling(Γ3, Z)) + Z) Γ*Z=φ(E)⊙σ(F) Among them, Γ1, Γ2, Γ3 represent the sizes of the convolution kernels, which are 1×S1, 1×S2, and 1×S3 respectively. The Concat(·) operation connects the features obtained from the three gated convolutions of different scales. Γ*Z represents a gated convolution operation, φ represents the Tanh function, σ represents the sigmoid function, E and F represent the first and second half of the number of channels after the convolution operation, respectively. The final output is obtained through the ReLU activation function and the residual connection.

Images