Gene network regulation relation reconstruction method based on deep directed graph convolution

Abstract

The invention discloses a gene network regulation relation reconstruction method based on deep directed graph convolution. The method comprises the following steps: 1, acquiring topological information of a saccharomyces cerevisiae gene regulation network by using a Node2Vec method; 2, for the saccharomyces cerevisiae, a graph data enhancement means is used due to the regulation and control network, so that more efficient information expression is obtained; 3, for the regulation direction problem of the saccharomyces cerevisiae gene regulation network, predicting regulation relation direction information by using a directed graph neural network; 4, constructing a deep directed graph convolution model, and obtaining high-order neighbor information of the gene; and 5, splicing expression characteristics of saccharomyces cerevisiae genes and the characteristics obtained in the steps 1 and 2, inputting the spliced expression characteristics into the model to obtain a saccharomyces cerevisiae regulatory network prediction score matrix, and judging whether a regulatory relationship exists among the genes or not according to a relationship between the prediction score matrix and a threshold value. The method provided by the invention can accurately predict the regulatory relationship between the saccharomyces cerevisiae genes, and is beneficial to more efficient research on the essence of biochemical reaction of organisms.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent information processing and bioinformatics, and in particular to a method for reconstructing a gene regulatory network of Saccharomyces cerevisiae based on a deep directed graph convolutional network. Background Art

[0002] Genes are common attributes of all life. The activity of genes is related to the expression level of genes, and there are complex relationships between the expression products of genes (such as proteins, transcription factors) and their gene targets. Gene regulation networks (GRNs) explain the complex life phenomena of organisms at the genome level. They are important mechanisms for maintaining life processes, controlling biochemical reactions, and regulating compound levels. For cells, the dynamic control of biochemical levels by gene expression products allows cells to make timely adjustments to the stimulation of changes in the external environment. Therefore, GRNs play an important role in predicting gene functions, identifying cancer biomarkers, assisting drug development, and discovering potential drug targets. Reconstructing the network of Saccharomyces cerevisiae genes and understanding the regulatory relationship of Saccharomyces cerevisiae genes in environmental changes can better carry out industrial production.

[0003] There are some methods for predicting gene regulatory networks to study the regulatory mechanism between genes. For example, the method based on correlation theory: for a fully connected network composed of genes, the correlation between the edges between genes is directly calculated as a reference for importance, and the results of the correlation are sorted, and some set thresholds are used to select and construct the gene regulatory network. Based on feature selection technology: It is a kind of unsupervised learning method. The scheme based on feature selection often converts the p gene network construction problem into p independent feature selection problems, and uses feature selection technology to obtain a subset of regulatory factors with strong correlation with the target factor. The GENIE3 method is based on regression trees, which can be learned using random forest methods and (EXTREA-TREE) extreme tree methods, Jump3, which is a hybrid method for GRN reasoning. It is based on the formal dynamic model of each gene expression of GRN, and uses greedy, non-parametric methods to reconstruct the topological structure of GRN. In recent years, with the development of deep learning and neural networks, more and more research methods based on neural networks have emerged. They are based on supervised learning schemes: CNNC, DGRNS, GRGNN, TDL, DeepDRIM, etc.

[0004] The above methods are based on traditional methods such as information theory or traditional machine learning, or use basic convolutional neural networks or combine with recurrent neural networks. For natural directed graph data such as gene regulatory networks, the use of directed graph neural networks can be more efficient and convenient, but because the process of constructing data is cumbersome and graph data information is difficult to obtain, few people use directed graph neural networks. Secondly, the final prediction effect of the method that only uses gene expression data will be poor because it cannot mine deeper information, and the natural topological structure information of gene regulatory networks is not utilized. At the same time, the large number of low-degree gene nodes in the gene regulatory network will cause the previous model to be unable to be fully trained, resulting in a loss of model performance. In addition, the gene regulatory network has problems such as network sparseness and long regulatory chains, which make the above model unable to fully capture information, resulting in poor model effect. Summary of the invention

[0005] The present invention aims to solve the deficiencies of the above-mentioned prior art and proposes a method for reconstructing the regulatory relationship of gene networks based on deep directed graph convolution, in order to more efficiently and accurately predict the regulatory relationship between genes, thereby helping to more effectively study the regulatory relationship between genes and facilitate industrial production.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The invention provides a method for reconstructing gene network regulatory relationships based on deep directed graph convolution, which comprises the following steps:

[0008] Step 1: Obtain the gene network of Saccharomyces cerevisiae , where V is the set of gene nodes of Saccharomyces cerevisiae, and , represents the nth Saccharomyces cerevisiae gene node; represents the total number of Saccharomyces cerevisiae gene nodes, E represents the directed edge set formed by the regulatory relationship between each Saccharomyces cerevisiae gene node, and ; Indicates A directed edge, Represents the total number of directed edges, and the adjacency matrix with dimension N×N is obtained from E ;

[0009] Get The M-dimensional gene expression profile of N Saccharomyces cerevisiae gene nodes constitutes an expression profile matrix with a dimension of N×M. ;

[0010] Step 2: Build a neighborhood enhancement network and use it to and A to obtain enhanced features ;

[0011] Step 3: Input into the Node2Vec model and generated by the neighborhood sampling strategy The network neighborhood of Traverse the network neighborhood of The topological information embedding of N Saccharomyces cerevisiae gene nodes is obtained and an N×T-dimensional network topological information embedding matrix F is formed; where T represents the final representation dimension of each gene;

[0012] Step 4: and After concatenation, the dimension is The characteristic matrix , and They are input together into the feature fusion module for processing, and then the fusion feature matrix B is obtained using formula (1):

[0013] (1)

[0014] In formula (1), Indicates the form of message dissemination. represents regularization, Indicates splicing, Represents the information learning weight matrix;

[0015] Step 5: Build a directed graph neural network and train A, B and Processing is performed to obtain the regulatory relationship prediction score matrix ;

[0016] Step 6: Use the stochastic gradient descent method to train the directed graph convolutional neural network, and calculate the cross entropy loss function to update the network parameters until the cross entropy loss function converges, thereby obtaining a trained brewer's yeast gene regulatory network inference model, which is used to output the prediction score of the regulatory relationship between genes, and finally determine whether there is a directed regulatory relationship through the set threshold.

[0017] The method for reconstructing gene network regulatory relationships based on deep directed graph convolution of the present invention is also characterized in that the neighborhood enhancement network in step 2 includes: an encoder network and decoder network :

[0018] Step 2.1: We sample the neighborhood of The H neighboring nodes of The M-dimensional gene expression profiles of the H neighboring nodes form a neighborhood node feature matrix with dimensions of H×M After that, it is input into the encoder network Encode in to obtain a neighborhood node feature encoding vector with a dimension of H×1;

[0019] Step 2.2: Input the neighborhood node feature encoding matrix of dimension H×1 into the decoder network Processed in to generate a 1×M The neighborhood node feature decoding vector ;

[0020] Step 2.3: The neighborhood node feature decoding vectors of N Saccharomyces cerevisiae gene nodes are combined into an enhanced feature matrix .

[0021] Further, the step 5 comprises:

[0022] Step 5.1: Use formula (2) to get and the mth Saccharomyces cerevisiae gene node Adjacency information between , thus the adjacency matrix is ​​formed by the adjacency information between N Saccharomyces cerevisiae gene nodes ;

[0023] (2)

[0024] In formula (2), express and The regulatory relationship between express and The regulatory relationship between

[0025] Step 5.2: Use formula (3) to get The degree matrix , thus obtaining the degree matrix of N Saccharomyces cerevisiae gene nodes ;

[0026] (3)

[0027] Step 5.3: Use formula (4) to construct the and The phase difference between , thus obtaining the phase matrix under the condition of parameter q ;

[0028] (4)

[0029] In formula (4), is pi, q is a constant;

[0030] Using equations (5) and (6), we construct the complex Hermitian matrix under the condition of parameter q: :

[0031] (5)

[0032] (6)

[0033] In formula (5) and formula (6), represents component-wise multiplication, represents the imaginary unit, and equation (6) is the expanded form of Euler's equation;

[0034] Step 5.4: Use equation (7) to construct the Laplace operator under the condition of parameter q :

[0035] (7)

[0036] Step 5.5: Use formula (8) to get the final direction information transmission paradigm :

[0037] (8)

[0038] In formula (8), It means equivalent to; express Plus the identity matrix form, and , is the identity matrix, express Plus the form of the identity matrix;

[0039] Step 5.6: Use formula (9) to get Feature information matrix after layer learning :

[0040] (9)

[0041] In formula (9), represents the weight, Represents the feature information matrix after learning the l-1th layer, when =1, let =B;

[0042] Step 5.7: Use formula (10) to get and The regulatory relationship prediction score , thus obtaining the regulatory relationship prediction score matrix of N Saccharomyces cerevisiae gene nodes :

[0043] (10)

[0044] In formula (10), ( ) is the classification function, ( ) is the concatenation function, It is the feature information after the last layer of learning Middle Row, No. Row feature information vector.

[0045] An electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the reconstruction method, and the processor is configured to execute the program stored in the memory.

[0046] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the reconstruction method when the computer program is executed by a processor.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The gene regulatory network is a natural network data with network topology information. The present invention uses a graph representation learning method to obtain the graph topology information, which solves the problem that the neural network model only uses the expression value of the cell gene as the feature matrix of the neural network model input and cannot mine more useful information from the gene expression data, providing new features and ideas for the prediction of the gene regulatory relationship of brewer's yeast.

[0049] 2. The present invention uses conditional variational autoencoder CVAE to perform local enhancement of graph neural network, which solves the problem that there are a large number of low-degree nodes in GRN, making it difficult for directed graph neural network to effectively obtain surrounding neighbor information, resulting in the model being unable to efficiently predict the regulatory relationship between genes. The present invention enhances the local features of nodes in graph neural network and obtains efficient feature representation to enhance the expression ability of graph neural network, thereby solving the problem of difficulty in mining the association information of low-degree gene nodes in gene regulatory network;

[0050] 3. The present invention uses a deep graph neural network to learn the knowledge carried by higher-order neighbors by traversing deeper neighbor nodes. Therefore, the feature values ​​of genes can be extracted from the global perspective of the graph, and efficient feature representation can be learned, thereby improving the prediction performance of the gene regulation model of Saccharomyces cerevisiae;

[0051] 4. The present invention uses a directed graph convolutional neural network to solve the problem that traditional convolutional neural networks or those used in combination with recurrent neural networks do not focus on the directionality of regulatory relationships during message transmission. Therefore, the present invention can more accurately predict gene regulatory relationships while efficiently processing directed graph data. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of the method of the present invention;

[0053] Figure 2 A structural diagram of a neighborhood enhancement network of the present invention;

[0054] Figure 3 Schematic diagram of the Node2Vec model of the present invention;

[0055] Figure 4 is a schematic diagram of a feature fusion module of the present invention;

[0056] Figure 5 The area under the receiver operating characteristic curve and the area under the precision-recall curve of each method of the present invention are compared with other existing methods under ten five-fold cross validation runs in five different networks in DREAM4 simulated data;

[0057] Figure 6 The area under the receiver operating characteristic curve and the area under the precision-recall curve of each method of the present invention under ten five-fold cross validation runs in three different networks in real Escherichia coli data are compared with other existing methods;

[0058] Figure 7 The figure shows the comparison results of the area under the working characteristic curve and the area under the precision-recall curve of the present invention under the gene expression profile of Saccharomyces cerevisiae with other existing methods. DETAILED DESCRIPTION

[0059] In this embodiment, a gene regulation relationship prediction method based on multi-layer directed graph convolution is provided. Figure 1 As shown, the following steps are followed:

[0060] Step 1: Obtain the gene network of Saccharomyces cerevisiae , where V is the set of gene nodes of Saccharomyces cerevisiae, and , represents the nth Saccharomyces cerevisiae gene node; represents the total number of Saccharomyces cerevisiae gene nodes, E represents the directed edge set formed by the regulatory relationship between each Saccharomyces cerevisiae gene node, and ; Indicates A directed edge, Represents the total number of directed edges, and the adjacency matrix with dimension N×N is obtained from E ;

[0061] Get The M-dimensional gene expression profile of N Saccharomyces cerevisiae gene nodes constitutes an expression profile matrix with a dimension of N×M. ;

[0062] Step 2: Construct a neighborhood enhancement network including: encoder network and decoder network , and used for and A to obtain enhanced features ,like Figure 2 As shown;

[0063] Step 2.1: We sample the neighborhood of The H neighboring nodes of The M-dimensional gene expression profiles of the H neighboring nodes form a neighborhood node feature matrix with dimensions of H×M After that, it is input into the encoder network Encode in to obtain a neighborhood node feature encoding vector with a dimension of H×1;

[0064] Step 2.2: Input the neighborhood node feature encoding matrix of dimension H×1 into the decoder network Processed in to generate a 1×M The neighborhood node feature decoding vector , as shown in formula (11):

[0065] (11)

[0066] In formula (11), is the decoder network The activation function, is the encoder network The activation function, and Denote the decoder network and encoder network The weight matrix of is the decoder network Medium bias term;

[0067] The encoder and decoder structures of the neighborhood enhancement network are both composed of two layers of multi-layer perceptrons. is the input condition information, and the gene node The characteristics of the composition of the neighbor gene nodes (Data information input by the encoder, which is gene node-generated by the encoder network The mean and standard deviation of the distribution of are calculated. Then, a neighborhood node feature encoding matrix with a dimension of H×1 is sampled from the distribution obtained by the encoder and used as the input data of the decoder, finally obtaining a 1×M The neighborhood node feature decoding vector .

[0068] Step 2.3: The neighborhood node feature decoding vectors of N Saccharomyces cerevisiae gene nodes are combined into an enhanced feature matrix ;

[0069] Step 3: Input into the Node2Vec model, such as Figure 3 As shown, it is generated by the neighborhood sampling strategy The network neighborhood of traverse the network neighborhood to obtain a 1×T dimensional The topological information of N Saccharomyces cerevisiae gene nodes is embedded, and the network topological information embedding matrix F with N×T dimensions is formed; where T represents the final representation dimension that each gene may obtain. During the experiment, 4 is taken as the final output dimension for this T.

[0070] Step 4: and After concatenation, the dimension is The characteristic matrix , and Input together into the feature fusion module for processing as follows Figure 4 As shown, the fusion feature matrix B is obtained using formula (1):

[0071] (1)

[0072] In formula (1), represents regularization, Indicates splicing, represents the information learning weight matrix, Indicates the message propagation form. In this embodiment, Is a message aggregation function used to first The network neighborhood is screened using Calculate the embedding of all nodes, and then aggregate the embeddedness of the selected neighboring nodes to get the node The final embedding of This process is performed on N gene nodes to obtain the final embedding of N gene nodes.

[0073] Step 5: Build a directed graph neural network and train A, B and Processing is performed to obtain the regulatory relationship prediction score matrix ;

[0074] Step 5.1: Use (2) to get and the mth Saccharomyces cerevisiae gene node Adjacency information between , thus the adjacency matrix is ​​formed by the adjacency information between N Saccharomyces cerevisiae gene nodes ;

[0075] (2)

[0076] In formula (2), express and The regulatory relationship between express and The regulatory relationship between

[0077] Step 5.2: Use formula (4) to get The degree matrix , thus obtaining the degree matrix of N Saccharomyces cerevisiae gene nodes ;

[0078] (3)

[0079] In formula (3), Indicates Saccharomyces cerevisiae gene nodes exist The sum of all out-degrees in .

[0080] Step 5.3: Use formula (4) to construct the and The phase difference between , thus obtaining the phase matrix under the condition of parameter q ;

[0081] (4)

[0082] In formula (4), is pi, q is a constant;

[0083] Using equations (5) and (6), we construct the complex Hermitian matrix under the condition of parameter q: :

[0084] (5)

[0085] (6)

[0086] In formula (5) and formula (6), represents component-wise multiplication, represents the imaginary unit, and formula (7) is the expanded form of the Euler equation. When q is not equal to 0, Can be used to encode the direction of an edge.

[0087] Step 5.4: Use equation (7) to construct the Laplace operator under the condition of parameter q :

[0088] (7)

[0089] Step 5.5: Use formula (8) to get the final direction information transmission paradigm :

[0090] (8)

[0091] In formula (8), It means equivalent to; express Plus the identity matrix form, and , is the identity matrix, express Plus the identity matrix form, express The out-degree information of , express and The adjacency information between them plus the identity matrix.

[0092] Step 5.6: Use formula (9) to get Feature information matrix after layer learning :

[0093] (9)

[0094] In formula (9), represents the weight, Represents the feature information matrix after learning the l-1th layer, when =1, let =B;

[0095] Step 5.7: Figure 1 As shown, using formula (10) we get and The regulatory relationship prediction score , thus obtaining the regulatory relationship prediction score matrix of N Saccharomyces cerevisiae gene nodes :

[0096] (10)

[0097] In formula (10), ( ) is the classification function, ( ) is the concatenation function, It is the feature information after the last layer of learning Middle Row, No. Row feature information vector;

[0098] Prediction score matrix Indicates the prediction score of the regulatory relationship between any two gene nodes. If the prediction score is greater than the set threshold, there is a regulatory relationship between the corresponding two genes, and the regulatory direction is from the regulatory factor to the target gene. Otherwise, it means that there is no regulatory relationship between the two genes.

[0099] Step 6: Use the stochastic gradient descent method to train the directed graph convolutional neural network, and use formula (12) to calculate the cross entropy loss function The network parameters are updated until the cross entropy loss function converges, thereby obtaining a trained Saccharomyces cerevisiae gene regulatory network inference model, which is used to output the prediction score of the regulatory relationship between genes and finally determine whether there is a directed regulatory relationship through the set threshold.

[0100] (12)

[0101] In formula (12), Indicates A True labels of regulatory relationships with other genes, Indicates R Prediction of regulatory relationships with other genes.

[0102] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0103] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

[0104] In order to evaluate the performance of the model, the results were compared with a single classifier and other methods. The evaluation index was the area under the receiver operating characteristic curve (AUROC). AUROC is the area under the receiver operating characteristic curve. The value of this area is between 0 and 1. It can intuitively evaluate the quality of the classifier. The larger the AUC value, the better the classifier effect. The experiment used ten five-fold cross validations and calculated the average value as the model result.

[0105] Figure 5 The results of the model of the method of the present invention in the DREAM4 simulation data set are compared with other prediction models Jump3, dynamical GEne Network Inference with Ensemble of trees (dynGENIE3), Bidirectional eXtreme Gradient Boosting (BiXGBoost) and Multi-source Multi-model Fusion method (MMFGRN). The experimental results show that the results of the model of the present invention in the five network data of DREAM4 from network 1 to network 5 are better than those of other prediction models, verifying the effectiveness and superiority of the model.

[0106] Figure 6 The results of the model of the method of the present invention in the real data set of E. coli were compared with other prediction models Jump3, dynGENIE3 and BiXGBoost. In the three network data of cold pressure (Cold), heat pressure (Heat) and oxygen stress pressure (Oxidative), the prediction results of the model of the present invention on the real data set are also better than those of other prediction models, which shows that the model has superiority and also indicates that the model can more accurately predict the regulatory relationship between genes.

[0107] Figure 7The results of implementing the model of this method in the data of Saccharomyces cerevisiae (S. cerevisiae) are shown, and CNNGRN (A Convolutional Neural Network-Based Method for Gene Regulatory Network), XGBnet (XGBoost-Based Gene Network), iRafNet (Integrative Random Forestfor Gene Regulatory Network), and PLSNET (Gene regulatory network inference using PLS-based methods) methods are compared. The experimental results show that the prediction results of the model of the present invention in the actual case of Saccharomyces cerevisiae data are better than those of other prediction models, verifying the feasibility and effectiveness of the model of the present invention in practical applications.

Claims

1. A method for reconstructing gene network regulatory relationships based on deep directed graph convolution, characterized in that: The steps include: Step 1: Obtain the gene network of Saccharomyces cerevisiae , where V is the set of gene nodes of Saccharomyces cerevisiae, and , represents the nth Saccharomyces cerevisiae gene node; represents the total number of Saccharomyces cerevisiae gene nodes, E represents the directed edge set formed by the regulatory relationship between each Saccharomyces cerevisiae gene node, and ; Indicates A directed edge, Represents the total number of directed edges, and the adjacency matrix with dimension N×N is obtained from E ; Get The M-dimensional gene expression profile of N Saccharomyces cerevisiae gene nodes constitutes an expression profile matrix with a dimension of N×M. ; Step 2: Build a neighborhood enhancement network and use it to and A to obtain enhanced features ; Step 3: Input into the Node2Vec model and generated by the neighborhood sampling strategy The network neighborhood of Traverse the network neighborhood of The topological information embedding of N Saccharomyces cerevisiae gene nodes is obtained and an N×T-dimensional network topological information embedding matrix F is formed; where T represents the final representation dimension of each gene; Step 4: and After concatenation, the dimension is The characteristic matrix , and with They are input together into the feature fusion module for processing, and then the fusion feature matrix B is obtained using formula (1): (1) In formula (1), Indicates the form of message dissemination. represents regularization, Indicates splicing, Represents the information learning weight matrix; Step 5: Build a directed graph neural network and train A, B and Processing is performed to obtain the regulatory relationship prediction score matrix ; Step 6: Use the stochastic gradient descent method to train the directed graph convolutional neural network, and calculate the cross entropy loss function to update the network parameters until the cross entropy loss function converges, thereby obtaining a trained brewer's yeast gene regulatory network inference model, which is used to output the prediction score of the regulatory relationship between genes, and finally determine whether there is a directed regulatory relationship through the set threshold.

2. According to claim 1, a method for reconstructing gene network regulatory relationships based on deep directed graph convolution, characterized in that: The neighborhood enhancement network in step 2 includes: an encoder network and decoder network : Step 2.1: We sample the neighborhood of The H neighboring nodes of The M-dimensional gene expression profiles of the H neighboring nodes form a neighborhood node feature matrix with dimensions of H×M After that, it is input into the encoder network Encode in to obtain a neighborhood node feature encoding vector with a dimension of H×1; Step 2.2: Input the neighborhood node feature encoding matrix of dimension H×1 into the decoder network Processed in to generate a 1×M The neighborhood node feature decoding vector ; Step 2.3: The neighborhood node feature decoding vectors of N Saccharomyces cerevisiae gene nodes are combined into an enhanced feature matrix .

3. The method for reconstructing gene network regulatory relationships based on deep directed graph convolution according to claim 2, characterized in that: The step 5 comprises: Step 5.1: Use formula (2) to get and the mth Saccharomyces cerevisiae gene node Adjacency information between , thus the adjacency matrix is ​​formed by the adjacency information between N Saccharomyces cerevisiae gene nodes ; (2) In formula (2), express and The regulatory relationship between express and The regulatory relationship between Step 5.2: Use formula (3) to get The degree matrix , thus obtaining the degree matrix of N Saccharomyces cerevisiae gene nodes ; (3) Step 5.3: Use formula (4) to construct the and The phase difference between , thus obtaining the phase matrix under the condition of parameter q ; (4) In formula (4), is pi, q is a constant; Using equations (5) and (6), we construct the complex Hermitian matrix under the condition of parameter q: : (5) (6) In formula (5) and formula (6), represents component-wise multiplication, represents the imaginary unit, and equation (6) is the expanded form of Euler's equation; Step 5.4: Use equation (7) to construct the Laplace operator under the condition of parameter q : (7) Step 5.5: Use formula (8) to get the final direction information transmission paradigm : (8) In formula (8), It means equivalent to; express Plus the identity matrix form, and , is the identity matrix, express Plus the identity matrix form; Step 5.6: Use formula (9) to get Feature information matrix after layer learning : (9) In formula (9), represents the weight, Represents the feature information matrix after learning the l-1th layer, when =1, let =B; Step 5.7: Use formula (10) to get and The regulatory relationship prediction score , thus obtaining the regulatory relationship prediction score matrix of N Saccharomyces cerevisiae gene nodes : (10) In formula (10), ( ) is the classification function, ( ) is the concatenation function, It is the feature information after the last layer of learning Middle Row, No. Row feature information vector.

4. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports a processor to execute the reconstruction method described in any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the reconstruction method according to any one of claims 1 to 3 are executed.

Images