Gene identification method based on mask pattern auto-encoder

Abstract

The invention discloses a gene identification method based on a mask graph auto-encoder. The method comprises the following steps: 1, acquiring gene interaction data and omics characteristic data and preprocessing the data; 2, masking the processed data, specifically comprising two branches, a node masking module and an edge masking module; 3, inputting the masked network into a graph automatic encoder for training, wherein the graph automatic encoder learns embedded representation of the network by reconstructing nodes and edges of the network; and 4, obtaining low-dimensional embedding of the features through the trained encoder, and finally performing gene classification by using a logic regression classifier. According to the method, the node information and the structure information of the graph are concerned at the same time by respectively masking the nodes and the edges in the network, and the dependence of the feature training model on the label information is reduced in a self-supervised learning mode, so that the genes can be accurately classified.

Description

Technical Field

[0001] The invention belongs to the field of gene recognition, in particular to a gene recognition method based on a mask graph autoencoder. Background Art

[0002] With the rapid development of high-throughput sequencing technology, a huge amount of data has been accumulated in the field of genomics. These data include genome sequences, transcriptome data, proteome data, etc., which are highly complex and multidimensional. The surge in the amount of data provides rich resources for gene identification research, but also poses severe challenges to traditional data analysis methods.

[0003] The rapid development of artificial intelligence technology has provided new ideas and methods for gene identification research. Artificial intelligence algorithms have powerful data processing and pattern recognition capabilities, can automatically extract features from large-scale data sets, and achieve accurate identification of genes. Accurate identification of genes can provide strong support for early diagnosis of diseases, personalized treatment, and drug development. This not only helps to improve treatment outcomes and the quality of life of patients, but also helps to promote innovation and development in the medical industry. At present, algorithms for gene identification are mainly based on traditional machine learning and deep learning.

[0004] Traditional machine learning-based gene identification methods usually train a classifier, such as random forest and support vector machine, to identify genes by extracting different features from different types of omics data. However, most existing machine learning-based methods only extract network-based features by using network analysis, and they cannot effectively combine the network topology features of genes with multi-omics features.

[0005] In recent years, deep learning models have also achieved remarkable results in the field of gene identification. Some researchers have applied graph convolutional networks to gene identification tasks, such as the Explainable Multi-Omics Graph Integration (EMOGI) method, which uses multiple types of omics features as gene features and uses graph convolutional networks to naturally combine gene networks with gene features to learn the embedding representation of nodes. However, the model performance of such methods mainly depends on a limited amount of labeled data, which results in the feature information of a large number of unlabeled nodes not being fully utilized, making it difficult to accurately identify genes. Summary of the invention

[0006] In order to solve the shortcomings of the above-mentioned prior art, the present invention proposes a gene identification method based on mask graph autoencoder, so as to effectively combine the network characteristics and multi-omics characteristics of genes in the gene identification task, and make full use of the information of a large number of unlabeled nodes, thereby reducing the model's dependence on label data, and providing a new solution for gene identification.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0008] The gene identification method based on the mask image autoencoder of the present invention is characterized in that it is performed according to the following steps:

[0009] Step 1: Collect multi-omics data of normal genes and abnormal genes, and establish a gene network G={V, A, X}, where V represents the set of gene nodes and A represents the adjacency matrix of gene nodes. represents the feature set of gene nodes, N is the number of gene nodes, and M represents the number of omics types of each gene node feature; let any i-th gene node be denoted as v i , let x i Represents the i-th gene node v i characteristics, and ; Let e i,j Indicates v i With the jth gene v j If the interaction between i,j =1, indicating v i With v j There is an interaction relationship between them. If e i,j =0, indicating v i With v j There is no interaction between them; let the i-th gene node v i The true label is y i ,y i ∈{0,1}, 0 represents v i is a normal gene, 1 means v i For abnormal genes;

[0010] Step 2: Construct a graph representation learning network based on mask graph autoencoder, including: graph mask module, encoder, and decoder;

[0011] Step 2.1, the graph mask module is divided into a node mask unit and an edge mask unit, wherein the node mask unit performs node masking on V to obtain a node mask graph G1, and the edge mask unit performs edge masking on A to obtain an edge mask graph G2;

[0012] Step 2.2: The encoder module consists of L GCN layers, and processes G1 and G2 respectively to obtain node feature representations accordingly. and ;

[0013] Step 2.3, the decoder module includes: a node decoder unit and an edge decoder unit, and respectively and Processing is performed and the reconstructed feature vector z is obtained accordingly. i and G2i and v j The probability P(v i ,v j );

[0014] Step 2.4: Use formula (5) to construct the total loss function Loss;

[0015] (5)

[0016] In formula (5), loss1 is the node reconstruction loss, loss2 is the edge reconstruction loss, α represents the hyperparameter used to balance the node reconstruction and edge reconstruction loss functions, and 0<α<1;

[0017] Step 3, using the Adam optimizer to train the weight parameters of the graph representation learning network, and calculating the total loss function Loss to update the network parameters, and stopping the training when the number of iterations reaches the maximum number of iterations or the Loss reaches the minimum, thereby obtaining the optimal graph autoencoder model after training;

[0018] Step 4: Input G into the encoder of the trained optimal graph autoencoder model and generate a low-dimensional embedding feature S; let s i Represents the i-th gene node v in S i Low-dimensional embedding feature representation of ;

[0019] Step 5: Construct a logistic regression classifier by a fully connected layer, and use formula (10) to transform s i Project from feature space to probability space and output the i-th gene node v i Probability of abnormal gene ;

[0020] (10)

[0021] In formula (10), ω is the weight parameter to be learned in the fully connected layer, b is the bias term to be learned, and T represents the transpose;

[0022] Step 6: Use formula (11) to construct the cross entropy loss function L BCE , and use the gradient descent algorithm and Adam optimizer to train the logistic regression classifier so that the cross entropy loss function L BCE Convergence is achieved, thus obtaining a trained normal and abnormal gene classification model, which is used to classify normal and abnormal genes:

[0023] (11)

[0024] In formula (11), n ​​represents the number of gene nodes with labels. Represents the i-th gene node vi The real label.

[0025] The gene identification method based on mask image autoencoder described in the present invention is also characterized in that step 2.1 is performed as follows:

[0026] Step 2.1.1: The node mask unit samples V according to the mask ratio p and Bernoulli distribution to obtain the gene node subset V to be masked. mask , where 0 <p<1;

[0027] If v i The subset of gene nodes to be masked V mask , then v i The eigenvector x i Assign an all-zero vector, otherwise, v i Features of x i Remain unchanged, thus obtaining the feature set X of the gene node after the mask operation mask , and then get the node mask graph G1={V, A, X mask};

[0028] Step 2.1.2: The edge mask unit samples V according to the sampling ratio q and Bernoulli distribution to obtain the root node set V root , where 0 <q<1;

[0029] V root Take any node in as the starting point and use the random walk algorithm to find V root Sampling is performed, and a set of edges generated by random walk is used as the edge set to be masked, denoted as E mask ;

[0030] If e i,j The edge set E that belongs to the mask mask , then set the element value of the i-th row and j-th column in the adjacency matrix A to "0", otherwise, set the element value of the i-th row and j-th column in the adjacency matrix A to "1", so as to obtain the adjacency matrix A after the mask operation mask ; Then we get the edge mask graph G2={V, A mask , X}.

[0031] Further, the step 2.2 is performed as follows:

[0032] Step 2.2.1: transform the node mask graph G1={V, A, X mask} is input into the encoder module, and the embedded feature representation of the node mask graph G1 output by the lth GCN layer is obtained using formula (1): :

[0033] (1)

[0034] In formula (1), represents the weight matrix to be learned of the lth GCN layer, σ(·) represents the nonlinear ReLU activation function; D represents the angle matrix, A represents the adjacency matrix of the gene node, I represents the identity matrix, Represents the node feature representation of the node mask graph G1 output by the lth GCN layer. When l=1, let = ;

[0035] Step 2.2.2: transform the edge mask graph G2={V, A mask , X} is input into the encoder module, and the edge mask graph G2 is obtained by using formula (2) after the embedding feature representation output by the lth GCN layer :

[0036] (2)

[0037] In formula (2), represents the embedded feature representation of the edge mask graph G2 output by the lth GCN layer. When l=1, let = ;

[0038] Step 2.2.3: According to the process of step 2.2.1 and step 2.2.2, the node mask graph G1 is obtained through the node feature representation output by the L-layer GCN layer. The embedded feature representation of the edge mask graph G2 output by the L-layer GCN layer .

[0039] Further, step 2.3 is performed as follows:

[0040] Step 2.3.1. Input into the node decoder for processing, and use formula (3) to obtain the i-th gene node v in the node mask graph G1 i The reconstructed eigenvector z of i ;

[0041] (3)

[0042] In formula (3), MLP represents multi-layer perceptron, represents the i-th gene v in G1 i The embedded feature representation, h i ∈ ;

[0043] Step 2.3.2: Input to the edge decoder for processing, and use formula (4) to obtain the node mask graph G2 v i and v j The probability P(v i ,v j );

[0044] (4)

[0045] In formula (4), represents the i-th gene v in G2 i The embedded feature representation of represents the jth gene v in G2 j Embedded feature representation of , ∈ , ⊙ represents the multiplication of corresponding elements, and sigmoid is a nonlinear activation function.

[0046] Further, step 2.4 is performed as follows:

[0047] Step 2.4.1, use formula (6) to construct the loss function loss1 of node reconstruction;

[0048] (6)

[0049] In formula (5), γ is the scaling factor, and γ ≥ 1;

[0050] Step 2.4.2, use equations (7) to (9) to construct the edge reconstruction loss function loss2;

[0051] (7)

[0052] (8)

[0053] (9)

[0054] In formula (7) to formula (9), is a randomly generated set of edges that do not exist in graph G; Indicates e i,j = 1, the difference between the predicted probability distribution and the probability distribution of the true label, Indicates e i,j = 0, the difference between the predicted probability distribution and the probability distribution of the true label.

[0055] The 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 gene identification method, and the processor is configured to execute the program stored in the memory.

[0056] 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 gene identification method when the computer program is executed by a processor.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. The present invention designs a network architecture of a graph autoencoder, which effectively combines the network characteristics of genes with multi-omics characteristics. Compared with existing gene identification methods, this method can better integrate gene features from different omics, thereby being more conducive to improving the accuracy of gene identification.

[0059] 2. The present invention designs an encoder module composed of a graph convolutional network (GCN). During the encoding process, the graph convolutional network is used to perform graph representation learning on the gene network. Compared with the existing gene recognition methods, the representation ability of embedded features is enhanced, thereby improving the accuracy of gene recognition.

[0060] 3. The method proposed in the present invention divides the mask module into node mask units and edge mask units through a dual-mask architecture, and simultaneously masks and reconstructs gene node features and topological connections, which enables the model to simultaneously take into account node information and structural information in the gene network, thereby improving the accuracy of gene identification.

[0061] 4. The method proposed in the present invention performs feature learning in a self-supervised learning manner, ensuring that the model can fully utilize the information of a large number of unlabeled nodes in the gene network. Compared with the existing gene identification methods, it avoids the model's dependence on label information during the feature learning process. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 is a structural diagram of the network in the present invention;

[0063] Figure 2 A diagram of the data processing process used in the present invention. DETAILED DESCRIPTION

[0064] In this embodiment, a gene identification method based on a masked graph autoencoder mainly uses a graph autoencoder that masks both nodes and edges to mask and reconstruct nodes and edges in a gene interaction network, and then obtains an embedded representation through a trained GNN encoder, and then feeds the embedded representation into a logistic regression classifier to complete gene identification. Figure 1 As shown, the specific steps of this method are as follows:

[0065] Step 1: Collect multi-omics data of normal genes and abnormal genes, and establish a gene network G = {V, A, X}, where V represents the set of gene nodes, A represents the adjacency matrix of gene nodes, X represents the set of gene node features, N is the number of gene nodes, and M represents the number of omics types of each gene node feature; let any i-th gene node be denoted as v i i, let x i i represents the feature of the i-th gene node v i i, and x i,j i ∈ X; let e i ij represents the interaction between v j i and the j-th gene v i,j j, if e i ij = 1, it means there is an interaction relationship between v j i and v i,j j, if e i ij = 0, it means there is no interaction relationship between v j i and v i j; let the true label of the i-th gene node v i i be y i i, y i i ∈ {0, 1}, 0 means v i i is a normal gene, and 1 means v Figure 2 i is an abnormal gene; in this embodiment, the gene interaction network data used is the high-confidence interaction network in the Consensus Path DB database, including 14,092 nodes and 344,862 edges; the gene features are 4 types of omics feature data from 3,318 samples, including copy number variation features, gene mutation features, DNA methylation features, and gene expression features. Therefore, N = 14,092 and M = 4. As

[0066] shown, the 4 types of feature data from different omics are spliced to form a feature matrix, and then the feature matrix is combined with the gene interaction network to form a gene network with multi-omics features.

[0067] Step 2: Construct a graph representation learning network based on a masked graph autoencoder, including: a graph masking module, an encoder, and a decoder;

[0068] Step 2.1: The graph masking module is divided into a node masking unit and an edge masking unit. Among them, the node masking unit performs node masking on V to obtain a node-masked graph G1, and the edge masking unit performs edge masking on A to obtain an edge-masked graph G2; mask Step 2.1.1: The node masking unit samples V according to the masking ratio p and follows the Bernoulli distribution to obtain a subset V i m of gene nodes to be masked, where 0 < p < 1; in this embodiment, the masking ratio p = 0.5;

[0069] If v i belongs to the subset V of gene nodes to be masked mask , then assign the feature vector x i of v i to the all-zero vector. Otherwise, the feature x i of v i remains unchanged, so as to obtain the feature set X of the gene nodes after the masking operation mask , and further obtain the node masked graph G1 = {V, A, X mask};

[0070] Step 2.1.2: The edge masking unit samples V according to the sampling ratio q and follows the Bernoulli distribution to obtain the root node set V root , where 0 < q < 1; in this embodiment, the sampling ratio q = 0.6;

[0071] Taking any root node in V root as the starting point, use the random walk algorithm to sample V root , and take a group of edges generated by the random walk as the set of edges to be masked, denoted as E mask ;

[0072] If e i,j belongs to the set of edges E to be masked mask , then set the element value of the i-th row and j-th column in the adjacency matrix A to "0". Otherwise, set the element value of the i-th row and j-th column in the adjacency matrix A to "1", so as to obtain the adjacency matrix A mask after the masking operation; and further obtain the edge masked graph G2 = {V, A mask , X}.

[0073] Step 2.2: The encoder module consists of L GCN layers; in this embodiment, the encoder uses 2 GNN layers, that is, L = 2, and each GNN layer is processed using the batch normalization function;

[0074] Step 2.2.1: Input the node masked graph G1 = {V, A, X mask} into the encoder module, and use Equation (1) to obtain the embedded feature representation of the node masked graph G1 output by the l-th GCN layer :

[0075] (1)

[0076] In Equation (1), represents the weight matrix to be learned in the l-th GCN layer, σ(·) represents the non-linear activation function; D represents the diagonal matrix, A represents the adjacency matrix of gene nodes, and I represents the identity matrix, Represents the node feature representation of the node mask graph G1 output by the lth GCN layer. When l=1, let = ; In this embodiment, the activation function σ(·) uses the ELU activation function;

[0077] Step 2.2.2: transform the edge mask graph G2={V, A mask , X} is input into the encoder module, and the edge mask graph G2 is obtained by using formula (2) after the embedding feature representation output by the lth GCN layer :

[0078] (2)

[0079] In formula (2), represents the embedded feature representation of the edge mask graph G2 output by the lth GCN layer. When l=1, let = ;

[0080] Step 2.2.3: According to the process of step 2.2.1 and step 2.2.2, the node mask graph G1 is obtained through the node feature representation output by the L-layer GCN layer. The embedded feature representation of the edge mask graph G2 output by the L-layer GCN layer .

[0081] Step 2.3, the decoder module includes: a node decoder unit and an edge decoder unit;

[0082] Step 2.3.1. Input into the node decoder for processing, and use formula (3) to obtain the i-th gene node v in the node mask graph G1 i The reconstructed eigenvector z of i ;

[0083] (3)

[0084] In formula (3), MLP represents multi-layer perceptron, represents the i-th gene v in G1 i The embedded feature representation, h i ∈ ; In this embodiment, the multilayer perceptron MLP includes two layers of fully connected neural networks;

[0085] Step 2.3.2: Input to the edge decoder for processing, and use formula (4) to obtain the node mask graph G2 v i and v j The probability P(v i ,vj );

[0086] (4)

[0087] In formula (4), represents the i-th gene v in G2 i The embedded feature representation of represents the jth gene v in G2 j Embedded feature representation of , ∈ , ⊙ represents the multiplication of corresponding elements, and sigmoid is a nonlinear activation function.

[0088] Step 2.4: Use formula (5) to construct the total loss function Loss;

[0089] (5)

[0090] In formula (5), loss1 is the node reconstruction loss, loss2 is the edge reconstruction loss, α represents the hyperparameter used to balance the node reconstruction and edge reconstruction loss functions, and 0<α<1; in this embodiment, the hyperparameter α is 0.4;

[0091] Step 2.4.1, use formula (6) to construct the loss function loss1 of node reconstruction;

[0092] (6)

[0093] In formula (5), γ is a scaling factor, and γ≥1; in this embodiment, the scaling factor γ is 3;

[0094] Step 2.4.2, use equations (7) to (9) to construct the edge reconstruction loss function loss2;

[0095] (7)

[0096] (8)

[0097] (9)

[0098] In formula (7) to formula (9), is a randomly generated set of edges that do not exist in graph G; Indicates when e i,j =1 is the difference between the predicted probability distribution and the probability distribution of the true label, Indicates when e i,j =0 is the difference between the predicted probability distribution and the probability distribution of the true label.

[0099] Step 3. Use the Adam optimizer to train the weight parameters of the graph representation learning network, calculate the total loss function Loss to update the network parameters, and stop training when the number of iterations reaches the maximum number of iterations or the Loss reaches the minimum, so as to obtain the optimal graph autoencoder model after training; in this embodiment, the node reconstruction task and the edge reconstruction task are trained simultaneously, and the training is terminated after 50 training iterations. The initial learning rate and weight decay are set to 0.01 and 0.0005, respectively.

[0100] Step 4: Input G into the encoder of the trained optimal graph autoencoder model and generate a low-dimensional embedding feature S; let s i Represents the i-th gene node v in S i Low-dimensional embedding feature representation of ;

[0101] Step 5: Construct a logistic regression classifier by a fully connected layer, and use formula (10) to transform s i Project from feature space to probability space and output the i-th gene node v i Probability of abnormal gene ;

[0102] (10)

[0103] In formula (10), ω is the weight parameter to be learned in the fully connected layer, b is the bias term to be learned, and T represents the transpose.

[0104] Step 6: Use formula (11) to construct the cross entropy loss function L BCE , and use the gradient descent algorithm and Adam optimizer to train the logistic regression classifier so that the cross entropy loss function L BCE Convergence is reached, thereby obtaining a trained normal and abnormal gene classification model for classifying normal and abnormal genes; in this embodiment, 20% of all labeled genes are used as a test set, and 80% are used as a training set. The logistic regression classifier is trained using 200 epochs, and the learning rate is set to 0.001;

[0105] (11)

[0106] In formula (11), n ​​represents the number of gene nodes with labels. Represents the i-th gene node v i The real label.

[0107] 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.

[0108] 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.

Claims

1. A gene recognition method based on mask graph autoencoder, characterized in that: The steps are as follows: Step 1: Collect multi-omics data of normal genes and abnormal genes, and establish a gene network G={V, A, X}, where V represents the set of gene nodes and A represents the adjacency matrix of gene nodes. represents the feature set of gene nodes, N is the number of gene nodes, and M represents the number of omics types of each gene node feature; let any i-th gene node be denoted as v i , let x i Represents the i-th gene node v i characteristics, and ; Let e i,j Indicates v i With the jth gene v j If the interaction between i,j =1, indicating v i With v j There is an interaction relationship between them. If e i,j =0, indicating v i With v j There is no interaction between them; let the i-th gene node v i The true label is y i ,y i ∈{0,1}, 0 represents v i is a normal gene, 1 means v i For abnormal genes; Step 2: Construct a graph representation learning network based on mask graph autoencoder, including: graph mask module, encoder, and decoder; Step 2.1, the graph mask module is divided into a node mask unit and an edge mask unit, wherein the node mask unit performs node masking on V to obtain a node mask graph G1, and the edge mask unit performs edge masking on A to obtain an edge mask graph G2; Step 2.2: The encoder module consists of L GCN layers, and processes G1 and G2 respectively to obtain node feature representations accordingly. and ; Step 2.3, the decoder module includes: a node decoder unit and an edge decoder unit, and respectively and Processing is performed and the reconstructed feature vector z is obtained accordingly. i and G2 i and v j The probability P(v i ,v j ); Step 2.4: Use formula (5) to construct the total loss function Loss; (5) In formula (5), loss1 is the node reconstruction loss, loss2 is the edge reconstruction loss, α represents the hyperparameter used to balance the node reconstruction and edge reconstruction loss functions, and 0<α<1; Step 3, using the Adam optimizer to train the weight parameters of the graph representation learning network, and calculating the total loss function Loss to update the network parameters, and stopping the training when the number of iterations reaches the maximum number of iterations or the Loss reaches the minimum, thereby obtaining the optimal graph autoencoder model after training; Step 4: Input G into the encoder of the trained optimal graph autoencoder model and generate a low-dimensional embedding feature S; let s i Represents the i-th gene node v in S i Low-dimensional embedding feature representation of ; Step 5: Construct a logistic regression classifier by a fully connected layer, and use formula (10) to transform s i Project from feature space to probability space and output the i-th gene node v i Probability of abnormal gene ; (10) In formula (10), ω is the weight parameter to be learned in the fully connected layer, b is the bias term to be learned, and T represents the transpose; Step 6: Use formula (11) to construct the cross entropy loss function L BCE , and use the gradient descent algorithm and Adam optimizer to train the logistic regression classifier so that the cross entropy loss function L BCE Convergence is achieved, thus obtaining a trained normal and abnormal gene classification model, which is used to classify normal and abnormal genes: (11) In formula (11), n ​​represents the number of gene nodes with labels. Represents the i-th gene node v i The real label.

2. A gene identification method based on mask image autoencoder according to claim 1, characterized in that: The step 2.1 is carried out as follows: Step 2.1.1: The node mask unit samples V according to the mask ratio p and Bernoulli distribution to obtain the gene node subset V to be masked. mask , where 0 <p<1; If v i The subset of gene nodes to be masked V mask , then v i The eigenvector x i Assign an all-zero vector, otherwise, v i Features of x i Remain unchanged, thus obtaining the feature set X of the gene node after the mask operation mask , and then get the node mask graph G1={V, A, X mask }; Step 2.1.2: The edge mask unit samples V according to the sampling ratio q and Bernoulli distribution to obtain the root node set V root , where 0 <q<1; V root Take any node in as the starting point and use the random walk algorithm to find V root Sampling is performed, and a set of edges generated by random walk is used as the edge set to be masked, denoted as E mask ; If e i,j The edge set E that belongs to the mask mask , then set the element value of the i-th row and j-th column in the adjacency matrix A to "0", otherwise, set the element value of the i-th row and j-th column in the adjacency matrix A to "1", so as to obtain the adjacency matrix A after the mask operation mask ; Then we get the edge mask graph G2={V, A mask , X}.

3. A gene identification method based on mask image autoencoder according to claim 2, characterized in that: The step 2.2 is performed as follows: Step 2.2.1: transform the node mask graph G1={V, A, X mask } is input into the encoder module, and the embedded feature representation of the node mask graph G1 output by the lth GCN layer is obtained using formula (1): : (1) In formula (1), represents the weight matrix to be learned of the lth GCN layer, σ(·) represents the nonlinear ReLU activation function; D represents the angle matrix, A represents the adjacency matrix of the gene node, I represents the identity matrix, Represents the node feature representation of the node mask graph G1 output by the lth GCN layer. When l=1, let = ; Step 2.2.2: transform the edge mask graph G2={V, A mask , X} is input into the encoder module, and the edge mask graph G2 is obtained by using formula (2) after the embedding feature representation output by the lth GCN layer : (2) In formula (2), represents the embedded feature representation of the edge mask graph G2 output by the lth GCN layer. When l=1, let = ; Step 2.2.3: According to the process of step 2.2.1 and step 2.2.2, the node mask graph G1 is obtained through the node feature representation output by the L-layer GCN layer. The embedded feature representation of the edge mask graph G2 output by the L-layer GCN layer .

4. A gene identification method based on mask image autoencoder according to claim 3, characterized in that: Step 2.3 is performed as follows: Step 2.3.

1. Input into the node decoder for processing, and use formula (3) to obtain the i-th gene node v in the node mask graph G1 i The reconstructed eigenvector z of i ; (3) In formula (3), MLP represents multi-layer perceptron, represents the i-th gene v in G1 i The embedded feature representation, h i ∈ ; Step 2.3.2: Input to the edge decoder for processing, and use formula (4) to obtain the node mask graph G2 i and v j The probability P(v i ,v j ); (4) In formula (4), represents the i-th gene v in G2 i The embedded feature representation of represents the jth gene v in G2 j Embedded feature representation of , ∈ , ⊙ represents the multiplication of corresponding elements, and sigmoid is a nonlinear activation function.

5. A gene identification method based on mask image autoencoder according to claim 4, characterized in that: Step 2.4 is performed as follows: Step 2.4.1, use formula (6) to construct the loss function loss1 of node reconstruction; (6) In formula (5), γ is the scaling factor, and γ ≥ 1; Step 2.4.2, use equations (7) to (9) to construct the edge reconstruction loss function loss2; (7) (8) (9) In formula (7) to formula (9), is a randomly generated set of edges that do not exist in graph G; Indicates e i,j = 1, the difference between the predicted probability distribution and the probability distribution of the true label, Indicates e i,j = 0, the difference between the predicted probability distribution and the probability distribution of the true label.

6. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the gene identification method according to any one of claims 1 to 5, and the processor is configured to execute the program stored in the memory.

7. 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 gene identification method according to any one of claims 1 to 5 are executed.

Images