Graph convolution spatial domain identification method of fusion gene interaction network and medium

Abstract

The invention discloses a graph convolution spatial domain identification method of a fusion gene interaction network, and relates to the technical field of biological information, and the method comprises the steps: preprocessing spatial transcriptome data; constructing a spatial diagram according to the spatial position information of the spot and the gene expression matrix; constructing a conditional gene interaction network based on at least a part of key genes in the spatial transcriptome data, constructing a conditional gene interaction network matrix, combining the conditional gene interaction network matrix and the gene expression matrix to obtain an expression degree-interaction matrix, and constructing a feature map; performing graph neural network convolution on the space graph and the feature graph, and introducing an attention mechanism to obtain final embedding; and clustering and finally embedding, and identifying a spatial domain with a similar gene expression mode. According to the method, the spatial domain is identified through the fusion gene interaction network, the problem that the spatial domain is discontinuous is effectively solved, the robustness and accuracy of the spatial domain identification method are improved, and high performance is shown in complex and highly heterogeneous tissue samples.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics technology, and more specifically, to the field of spatial domain identification technology of spatial transcriptome. Background Art

[0002] In recent years, spatial transcriptomics, as an emerging branch of bioinformatics, has gradually attracted the attention of researchers. Spatial transcriptomics combines gene expression data with the exact spatial location of cells in tissues, allowing researchers to not only detect which genes are active, but also to determine the specific location of these gene activities in the tissue. In this technology, "spot" usually refers to a small area of ​​detection, each of which can capture and analyze RNA from a specific location in a tissue section, thereby forming a gene expression matrix with spatial coordinates. The data form of spatial transcriptomes usually includes a gene expression matrix, the spatial coordinates of each spot, and related image data (such as H&E staining images, immunofluorescence images, DAPI staining images, and antibody labeling images, etc.), which together form the basis for analyzing and interpreting complex biological processes within tissues. This technology can measure the spatial distribution of gene expression in tissue sections at single-cell resolution or near single-cell resolution, thus providing a new perspective for studying the behavior and interactions of cells in their microenvironment. Spatial transcriptome data are characterized by high dimensionality and high complexity, and effective processing and analysis of these data have become a hot topic and difficulty in current research.

[0003] Due to technical and biological limitations, spatial transcriptomic data are usually highly sparse, with many genes having expression levels close to zero in most spots. This sparsity poses a major challenge to data analysis because it can lead to increased noise and loss of signal. Sparse data processing is one of the key steps in spatial transcriptomic analysis. To address these issues, researchers have developed a variety of sparse data processing methods. For example, ALRA (Adaptively-thresholded Low Rank Approximation) uses a technique based on low-rank matrix decomposition to fill in missing data and reduces the impact of noise on analysis results through regularization. In addition, models using graph convolutional networks and attention mechanisms can also enhance the expression of sparse data by aggregating information from neighboring spots, thereby improving the accuracy and robustness of data analysis. Through these methods, sparse spatial transcriptomic data can be better processed and analyzed, providing a reliable basis for subsequent spatial domain identification.

[0004] Spatial domain identification is one of the important tasks in spatial transcriptomics research. Its purpose is to divide cells (or spots) or regions in tissue sections into spatial domains with similar gene spatial expression patterns based on gene expression data and spatial information. Although existing spatial domain identification methods have made some progress, they still have limitations. Existing spatial domain identification methods are mainly divided into non-spatial and spatial clustering methods. Non-spatial clustering methods such as Seurat and Louvain only rely on gene expression data, ignore spatial neighbor relationships, and often lead to discontinuous spatial domains. Spatial clustering methods such as SpatialDE and BayesSpace take spatial information into account during clustering, but they still face challenges in dealing with the inherent noise and data sparsity in spatial transcriptome data. In addition, these methods are usually unable to accurately capture the complex nonlinear characteristics of gene expression and spatial dependence, especially in highly heterogeneous and complex tissues. The clustering performance is poor. Summary of the invention

[0005] The purpose of the present invention is to overcome the spatial domain discontinuity existing in the prior art, face challenges in dealing with the inherent noise and data sparsity in spatial transcriptome data, and are usually unable to accurately capture the complex nonlinear characteristics of gene expression and spatial dependence, especially the problem of poor clustering performance in highly heterogeneous and complex tissues, and provide a graph convolution spatial domain recognition method that integrates gene interaction networks.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] Obtaining public spatial transcriptome data and preprocessing the spatial transcriptome data, wherein the spatial transcriptome data includes gene expression data of each spot and spatial position information of each spot, and the gene expression data includes a gene expression matrix;

[0008] Constructing a spatial map according to the spatial position information of the spot and the gene expression matrix;

[0009] Constructing a conditional gene interaction network based on at least a portion of key genes among all genes in the spatial transcriptome data, constructing a conditional gene interaction network matrix based on the conditional gene interaction network, combining the conditional gene interaction network matrix with the gene expression matrix to obtain an expression-interaction matrix, and constructing a feature map based on the expression-interaction matrix;

[0010] Using a multi-view graph convolutional network encoder to convolve the spatial graph and the feature graph to obtain a convolution result;

[0011] Introducing an attention mechanism into the result of the convolution to obtain the final embedding;

[0012] The final embeddings are clustered to identify spatial domains with similar gene expression patterns.

[0013] Furthermore, the preprocessing of the spatial transcriptome data includes: cleaning, filtering and denoising the spatial transcriptome data, eliminating problematic genes that appear in a few samples due to technical errors or other non-biological factors, and then performing normalization standard processing to adjust the gene expression levels of all samples in the spatial transcriptome data to the same level to eliminate technical deviations between different samples.

[0014] Furthermore, constructing a spatial map according to the spatial position information of the spot and the gene expression matrix comprises the following steps:

[0015] Initialize an adjacency matrix with the same number of elements as the number of spots, and the initial element values ​​of the adjacency matrix are all set to zero;

[0016] Use a radius-based nearest neighbor method to determine the neighbor spots of each of the spots, predefine a radius r, calculate the Euclidean distance between each of the spots based on the spatial position information of each spot, if the Euclidean distance between two spots is less than the predefined radius r, then the two spots are considered to be neighbors, update the adjacency matrix according to this neighbor relationship, set the matrix elements corresponding to these neighbor spots to 1, and construct an adjacency matrix reflecting the spatial relationship, in which each element represents the neighbor relationship between the corresponding two spots;

[0017] A spatial graph reflecting the spatial relationship between the spots is constructed by using the adjacency matrix and the gene expression matrix:

[0018] G s =(A s , X)

[0019] in, is the spatial adjacency matrix of N spots, is the normalized gene expression matrix, and M is the number of genes after preprocessing.

[0020] Furthermore, the construction of a conditional gene interaction network based on at least a portion of key genes in all genes of the spatial transcriptome data is specifically as follows:

[0021] The spatial transcriptome data also includes m genes and n spots. The element x in the gene expression matrix ij represents the expression level of gene i in cell j. For each pair of genes x and y, the following hypothesis test is performed:

[0022]

[0023] By calculating the standardized statistics of the number of spots for genes x, y, and z And with the significance level N α That is, compared with the α quantile of the standard normal distribution, determine the dependency of genes x and y under the condition of conditional gene z. If Then gene x and gene y are dependent on conditional gene z, and their edge weight is assigned to be 1, otherwise it is 0;

[0024] Finally, a conditional gene interaction network is constructed according to the independence, conditional independence and dependency of at least a part of the key genes x, V, z in all the genes in the spatial transcriptome data.

[0025] Furthermore, at least a portion of the key genes among all the genes in the spatial transcriptome data include: highly expressed genes, genes with highly variable expression, key transcription factors, and genes with high network degree in the initial network.

[0026] Furthermore, the conditional gene interaction network matrix is ​​constructed based on the conditional gene interaction network, and the centrality and importance of each gene in the conditional gene interaction network are quantified by counting the connectivity of each gene in the conditional gene interaction network.

[0027] Furthermore, the conditional gene interaction network matrix and the gene expression matrix are combined to obtain an expression-interaction matrix, and constructing a feature map according to the expression-interaction matrix includes:

[0028] The characteristic adjacency matrix A is calculated based on the expression-interaction matrix EDIM f , and then according to the characteristic adjacency matrix A f The feature map was constructed by combining the expression-interaction matrix EDIM:

[0029] G f =(A f , EDIM).

[0030] Furthermore, the convolution of the spatial graph and the feature graph using a multi-view graph convolutional network encoder to obtain a convolution result comprises the following steps:

[0031] Performing spatial convolution on the spatial graph includes the following steps: adding the identity matrix to the spatial adjacency matrix A of the spatial graph s In, get calculate The diagonal matrix Apply the following layered propagation rule to convolve the spatial graph and use the activation function ReLU to perform nonlinear processing on the convolution result to obtain a new spatial embedding Its expression is:

[0032]

[0033] in is the weight matrix of the lth layer, the initial X is the normalized gene expression matrix, ReLU is the activation function, yes The diagonal matrix of

[0034] Performing feature convolution on the feature map includes the following steps: adding a unit matrix to the feature adjacency matrix A of the feature map f In, get calculate The diagonal matrix Apply the following hierarchical propagation rule to convolve the feature map and use the activation function to perform nonlinear processing on the convolution result to obtain a new feature embedding Its expression is:

[0035]

[0036] in is the weight matrix of the lth layer, the initial X is the normalized gene expression matrix, ReLU is the activation function, yes The diagonal matrix of

[0037] The spatial map and feature map are co-convolved using the following layer-wise propagation rule:

[0038]

[0039] in is the weight matrix of the lth layer of co-convolution, and The first, layer embedding extracted from the spatial map and feature map, respectively, and the initial

[0040] Through co-convolution, we get the co-embedding H c , whose expression is:

[0041]

[0042] In order to enhance H sc and H fc The commonality between them introduces a consistency constraint, which is defined as follows:

[0043]

[0044] Furthermore, the introduction of the attention mechanism into the result of the convolution to obtain the final embedding matrix is ​​specifically as follows:

[0045] Introducing an attention mechanism to adaptively learn the importance of latent embeddings for the joint embedding H c , perform a nonlinear transformation on it, and use a shared attention vector W cv To obtain the attention coefficient a c , the formula is as follows:

[0046] a c =svftmax(w c0 ·σ(WH c +b))

[0047] Where σ is the tanh activation function, W is the trainable weight matrix, and b is the bias vector;

[0048] The attention coefficient a is obtained by using c The same method is used to calculate the spatial embedding H s and feature embedding H f The attention coefficient a s and a f , and finally, the final embedding matrix H is obtained by combining these embeddings:

[0049] H=F(a s ·H s +a f ·H f +a c ·H c )

[0050] Here, F is a single linear layer used to learn the latent representation of highly variable features.

[0051] On the other hand, the present invention also provides a computer-readable storage medium, which includes a program for a graph convolution spatial domain identification method for a fusion gene interaction network. When the program for the graph convolution spatial domain identification method for a fusion gene interaction network is executed by a processor, the steps of the graph convolution spatial domain identification method for a fusion gene interaction network as described in the above technical solution are implemented.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The present invention recognizes the spatial domain by fusing graph convolutional networks and gene interaction networks, and combines multi-perspective information, attention mechanisms, sparse data processing and other technologies to effectively overcome the problem of spatial domain discontinuity and enhance the adaptability in dealing with the noise and data sparsity inherent in spatial transcriptome data. It also significantly improves the robustness and accuracy of spatial domain recognition by converting unstable gene expression into a more stable gene interaction network, and exhibits higher performance in complex and highly heterogeneous tissue samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flowchart of a graph convolution spatial domain recognition method for a fusion gene interaction network of the present application;

[0055] Figure 2 This is a schematic diagram of the adjusted Rand Index ARI and effect comparison between the inventive method EDIM-MGCN (our method) of this application and other public methods in Example 2. DETAILED DESCRIPTION

[0056] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0057] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0058] Example 1

[0059] See also Figure 1 , Example 1 of the present application provides a graph convolution spatial domain identification method for a fusion gene interaction network, comprising the following steps:

[0060] S1: Data collection and preprocessing: Obtain public spatial transcriptome data and preprocess the spatial transcriptome data. The spatial transcriptome data includes the gene expression data of each spot and the spatial location information of each spot. The gene expression data includes the gene expression matrix.

[0061] Preprocessing includes: first, cleaning, filtering and denoising the data to remove problematic genes that appear in a few samples, which may be caused by technical errors or other non-biological factors. Next, normalization and standardization are performed to adjust the gene expression levels of all samples to the same level to eliminate technical deviations between different samples. In order to improve the sensitivity and specificity of subsequent analysis, highly variable genes (HVGs) are selected for key analysis. These genes show significant expression differences in different cells or conditions and are more likely to be involved in key biological processes. Through these steps, data noise can be significantly reduced and the accuracy and robustness of data processing can be improved.

[0062] S2: Constructing a spatial map: constructing a spatial map according to the spatial position information of the spot and the gene expression matrix.

[0063] First, the spatial coordinates of each spot are extracted from the spatial transcriptome data and stored in a matrix. Subsequently, an adjacency matrix with the same number of spots is initialized, and the initial element values ​​of the matrix are all set to zero. Then, a radius-based nearest neighbor method is used to determine the neighbor spots of each spot, and the neighbors of each spot are determined by calculating the Euclidean distance between spots. If the distance between two spots is less than the predefined radius r, they are considered neighbors. Based on this neighbor relationship, the adjacency matrix is ​​updated and the matrix elements corresponding to these neighbor spots are set to 1. In this way, an adjacency matrix reflecting spatial relationships is constructed, in which each element represents the neighbor relationship between the corresponding two spots.

[0064] Through the above steps, a spatial map reflecting the spatial relationship between spots in the organization is successfully constructed:

[0065] G s =(A s , X)

[0066] in is the spatial adjacency matrix of N spots, is the normalized gene expression matrix, and M is the number of genes after screening.

[0067] S3: Construct feature map: construct a conditional gene interaction network based on at least a part of the key genes in all genes of the spatial transcriptome data, construct a conditional gene interaction network matrix based on the conditional gene interaction network, combine the conditional gene interaction network matrix with the gene expression matrix to obtain the expression-interaction matrix, and construct a feature map based on the expression-interaction matrix.

[0068] The construction of the feature map in step S3 includes the following steps:

[0069] Constructing conditional gene interaction network:

[0070] In the process of constructing the expression-interaction matrix (EDIM), we first start from the conditional gene interaction network (CGIN), which is a network constructed based on the conditional dependencies between genes. This network is established through advanced statistical methods and bioinformatics analysis, with the goal of identifying and verifying the direct relationship between genes under specific biological conditions. This network not only captures the physical connection between genes, but also integrates the functional connection of genes in a specific environment, thereby significantly improving the biological relevance and prediction accuracy of the network.

[0071] The spatial transcriptome data also includes m genes and n spots. The element x in the gene expression matrix ij represents the expression level of gene i in cell j. For each pair of genes x and y, the following hypothesis test is performed:

[0072]

[0073] By calculating the standardized statistics of the number of spots for genes x, y, and z And with the significance level N α That is, compared with the α quantile of the standard normal distribution, determine the dependency of genes x and y under the condition of conditional gene z. If Then gene x and gene y are dependent on conditional gene z, and their edge weight is assigned to be 1, otherwise it is 0;

[0074] Finally, a conditional gene interaction network (CGIN) is constructed based on the probability estimates of the independence, conditional independence and dependency of at least some of the key genes x, y, z among all genes in the spatial transcriptome data.

[0075] In theory, all remaining m-2 genes should be tested for conditional independence, but this is computationally very time-consuming. Therefore, in practice, only a small number of key genes are selected as conditional genes for testing, which significantly reduces the computational cost. Selecting a small number of key genes as conditional genes can be achieved in a variety of ways, including highly expressed genes, genes with highly variable expression, key transcription factors, and genes with high network degree in the initial network.

[0076] Construct a conditional gene interaction network matrix:

[0077] After constructing CGIN, the next step is to generate the Conditional Gene Interaction Network Matrix (CGINM). This matrix is ​​obtained by counting the connectivity of each gene in the conditional gene interaction network to quantify its centrality or importance in the conditional gene interaction network. The purpose of this step is to further strengthen the role of genes in specific biological environments and reveal which genes play key roles in regulatory networks. CGINM provides a quantitative perspective to observe the activity of genes in the network, which is crucial to understanding how genes affect cell function and disease states.

[0078] Construct expression-interaction matrix:

[0079] The conditional gene interaction network matrix is ​​combined with the gene expression matrix to form the expression-interaction matrix (EDIM). Through this combination, EDIM not only contains the expression level of the gene, but also incorporates the strength of each gene in the biological network. This dual information integration greatly improves the interpretation of the data. In this way, EDIM enhances the understanding of the function of genes in cells, especially those genes that are at the core of the network, which may play a vital role in regulating cell behavior and pathological processes.

[0080] By constructing EDIM, we can more accurately identify and analyze the functions and interactions of genes in specific cell states, thus providing a powerful tool that combines expression information and network dynamics for subsequent spatial domain identification. The innovation of this method is that it not only improves the accuracy of analysis, but also enhances the understanding of complex biological systems, enabling the model to better process and analyze high-dimensional and complex spatial transcriptome data.

[0081] The characteristic adjacency matrix A is constructed by the expression-interaction matrix EDIM. f , and then according to the characteristic adjacency matrix A f The feature graph G is constructed by combining the expression-interaction matrix EDIM to capture the similarity structure in gene expression data. f :

[0082] G f =(A f , EDIM)

[0083] Through the feature map G f ,The model can capture the potential similarities and relationships represented in the EDIM matrix, which helps to reveal the regulatory networks and biological functions among genes.

[0084] S4: Use graph convolutional network to convolve spatial graph and feature map: The multi-view GCN encoder is a powerful graph neural network that can directly process graph data and utilize graph structure information. In order to extract the most relevant information from gene expression and spatial structure, the multi-view GCN encoder is used to perform convolution operations on feature maps and spatial maps.

[0085] In step S4, the convolution operation includes the following steps S41-S43:

[0086] S41: Spatial Convolution:

[0087] In order to combine gene expression information with spatial information, we first construct the spatial adjacency matrix A s The convolution operation is performed on the spatial graph to aggregate the spatial information of the neighbors in the Fourier domain. The following layered propagation rule is applied to convolve the spatial graph and the activation function ReLU is used to perform nonlinear processing on the convolution result to obtain a new spatial embedding Its expression is:

[0088]

[0089] Among them, W s (l) is the weight matrix of the lth layer, the initial X is the normalized gene expression matrix, ReLU is the activation function, yes The diagonal matrix of .

[0090] Specifically, in the first step of this process, we add the identity matrix to the spatial adjacency matrix A s In, forming This ensures that the self-connection is included. Then, by calculating The angle matrix Next, the activation function ReLU is used to perform nonlinear processing on the convolution result to obtain a new spatial embedding The weight matrix is a parameter that needs to be learned, which can capture the association between gene expression information and spatial neighbors.

[0091] S42: Feature convolution:

[0092] In order to learn more comprehensive gene expression information and fill in the gene expression of spots in the feature map, in the feature adjacency matrix A f The feature convolution operation is performed on the gene expression matrix X, applying the following layer-wise propagation rule:

[0093]

[0094] in is the weight matrix of the lth layer, the initial X is the normalized gene expression matrix, ReLU is the activation function, yes The diagonal matrix of .

[0095] Feature convolution is similar to spatial convolution, adding the identity matrix to the feature adjacency matrix A of the feature map f In, get calculate The angle matrix Apply the following hierarchical propagation rule to convolve the feature map and use the activation function to perform nonlinear processing on the convolution result to obtain a new feature embedding

[0096] S43: Co-convolution:

[0097] Since there is a certain correlation between gene expression and spatial distribution, it is necessary to consider the specific information and extract the commonality between them. To this end, a model using a parameter sharing strategy is designed to extract the co-embedding of gene expression and spatial distribution. Its propagation rule is defined as follows:

[0098]

[0099] in is the weight matrix of the lth layer of co-convolution, and The first, layer embedding extracted from the spatial map and feature map, respectively, and the initial

[0100] Through co-convolution, we can get the co-embedding H c , co-embedding is defined as:

[0101]

[0102] To enhance the spatial embedding H sc and feature embedding H fc The commonality between them introduces a consistency constraint, which is defined as follows:

[0103]

[0104] S5: Introducing attention mechanism: Introduce the attention mechanism into the convolution result obtained in step S4 to obtain the final embedding.

[0105] In the spatial domain recognition task, gene expression, spatial information and their joint representation should have different contributions to the results. Therefore, an attention mechanism is introduced to adaptively learn the importance of latent embeddings. Specifically, for the joint embedding Hc , firstly transform it nonlinearly and apply a shared attention vector W co To obtain the attention coefficient a c , the formula is as follows:

[0106] a c =softmax(w c0 ·σ(WH c +b))

[0107] Among them, σ is the tanh activation function, W is the trainable weight matrix, and b is the bias vector. Similarly, the attention coefficient a can be obtained by c The same method is used to calculate the embedding H s and H f The attention coefficient a s and a f Finally, by combining these embeddings, we get the final embedding H:

[0108] H=F(a s ·H s +a f ·H f +a c ·H c )

[0109] Here, F is a single linear layer used to learn the latent representation of highly variable features.

[0110] This attention mechanism ensures that the contributions of different types of information can be effectively balanced and optimized.

[0111] S6: Spatial domain identification: The final embedding obtained in step S5 is processed using clustering techniques to accurately identify spatial domains with similar gene expression patterns.

[0112] Example 2

[0113] Based on Example 1, this Example 2 uses a graph convolution spatial domain recognition method that integrates a gene interaction network in Example 1 to test on a public dataset to evaluate the performance of the method in the spatial domain recognition clustering task.

[0114] See also Figure 2 ,In order to better demonstrate the effectiveness of the method of this application, this method was tested on the DLPFC dataset. Taking slices #151670 and #151671 in the DLPFC dataset as examples, the analysis results of these two slices not only verified the application potential of ,the method in high variability areas, but also demonstrated its advantages in capturing subtle ,spatial expression patterns.

[0115] Specifically, in slices #151670 and #151671 of the DLPFC dataset, this method shows significant advantages. In slice #151670, the clustering is more continuous and closely follows the anatomical structure of the DLPFC. The boundaries between the cortical layers (layers3, layers4, layers5, layers6, and WM) are clearly defined, showing precise divisions consistent with known anatomical landmarks. In slice #151671, despite the complex anatomical structure, this method can still effectively distinguish different cortical layers, showing better spatial resolution and clustering effects than other methods. These results further confirm that our method can maintain the integrity of the underlying biological characteristics while ensuring the accuracy of spatial clustering.

[0116] The other steps of this embodiment 2 are the same as those of embodiment 1 and will not be repeated here.

[0117] Example 3

[0118] This embodiment 3 further provides a computer-readable storage medium based on embodiment 1, wherein the computer-readable storage medium includes a program for a graph convolution spatial domain identification method for a fusion gene interaction network. When the program is executed by a processor, the steps of the graph convolution spatial domain identification method for a fusion gene interaction network as described in embodiment 1 are implemented.

[0119] The storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required by at least one function; the data storage area may store used or received data, etc. In addition, the storage medium may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage devices.

[0120] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A graph convolution spatial domain recognition method integrating gene interaction networks, characterized in that: The following steps are involved: Obtaining public spatial transcriptome data and preprocessing the spatial transcriptome data, wherein the spatial transcriptome data includes gene expression data of each spot and spatial location information of each spot, and the gene expression data includes a gene expression matrix; Constructing a spatial map according to the spatial position information of the spot and the gene expression matrix; constructing a conditional gene interaction network based on at least a portion of key genes among all genes in the spatial transcriptome data, constructing a conditional gene interaction network matrix based on the conditional gene interaction network, combining the conditional gene interaction network matrix with the gene expression matrix to obtain an expression-interaction matrix, and constructing a feature map based on the expression-interaction matrix; Convolving the spatial map and the feature map using a multi-view graph convolutional network encoder to obtain a convolution result; Introducing an attention mechanism into the result of the convolution to obtain the final embedding; The final embeddings are clustered to identify spatial domains with similar gene expression patterns.

2. The method for identifying a fusion gene interaction network using graph convolutional spatial domain according to claim 1, characterized in that: The preprocessing of the spatial transcriptome data includes: cleaning, filtering and denoising the spatial transcriptome data, eliminating problem genes that appear in a small number of samples due to technical errors or other non-biological factors, and then performing normalization standard processing to adjust the gene expression levels of all samples in the spatial transcriptome data to the same level to eliminate technical deviations between different samples.

3. The method for identifying a fusion gene interaction network using graph convolution in spatial domain according to claim 1, wherein: The step of constructing a spatial map based on the spatial position information of the spot and the gene expression matrix comprises the following steps: Initialize an adjacency matrix with the same number of elements as the number of spots, and the initial element values ​​of the adjacency matrix are all set to zero; Using a radius-based nearest neighbor method to determine the neighbor spots of each of the spots, a predefined radius r is used, and the Euclidean distance between each of the spots is calculated based on the spatial position information of each spot. If the Euclidean distance between two spots is less than the predefined radius r, the two spots are considered to be neighbors. The adjacency matrix is ​​updated according to this neighbor relationship, and the matrix elements corresponding to these neighbor spots are set to 1 to construct an adjacency matrix reflecting the spatial relationship, in which each element represents the neighbor relationship between the corresponding two spots; A spatial graph reflecting the spatial relationship between the spots is constructed by using the adjacency matrix and the gene expression matrix: G s =(A s ,X) in, is the spatial adjacency matrix of N spots, is the normalized gene expression matrix, and M is the number of genes after preprocessing.

4. The method for identifying a fusion gene interaction network using graph convolution in spatial domain according to claim 1, wherein: The construction of a conditional gene interaction network based on at least a portion of key genes in all genes of the spatial transcriptome data is specifically as follows: The spatial transcriptome data also includes m genes and n spots. The element x in the gene expression matrix ij represents the expression level of gene i in cell j. For each pair of genes x and y, the following hypothesis test is performed: By calculating the standardized statistics of the number of spots for genes x, y, and z And with the significance level N α That is, compared with the α quantile of the standard normal distribution, determine the dependence of genes x and y under the condition of conditional gene z. If Then gene x and gene y are dependent on conditional gene z, and their edge weight is assigned to 1, otherwise it is 0; Finally, a conditional gene interaction network is constructed based on the independence, conditional independence and dependency relationships of at least a portion of the key genes x, y, and z among all the genes in the spatial transcriptome data.

5. The method for identifying a fusion gene interaction network using graph convolution in spatial domain according to claim 4, wherein: At least a portion of the key genes among all the genes in the spatial transcriptome data include: highly expressed genes, genes with highly variable expression, key transcription factors, and genes with high network degree in the initial network.

6. The method for identifying a fusion gene interaction network using graph convolution in spatial domain according to claim 4, wherein: The conditional gene interaction network matrix is ​​obtained by counting the connectivity of each gene in the conditional gene interaction network, and the conditional gene interaction network matrix quantifies the centrality or importance of each gene in the conditional gene interaction network.

7. The method for identifying a fusion gene interaction network using graph convolution spatial domain according to claim 1, characterized in that: Combining the conditional gene interaction network matrix and the gene expression matrix to obtain an expression-interaction matrix, and constructing a feature map according to the expression-interaction matrix includes: The characteristic adjacency matrix A is calculated based on the expression-interaction matrix EDIM f , and then according to the characteristic adjacency matrix A f The characteristic graph is constructed by combining the expression-interaction matrix EDIM: G f =(A f ,EDIM)。 8. The graph convolution spatial domain recognition method for a fusion gene interaction network according to any one of claims 3 or 7, characterized in that: The convolution operation of the spatial graph and the feature graph using a multi-view graph convolutional network encoder to obtain a convolution result comprises the following steps: Performing spatial convolution on the spatial graph includes the following steps: adding the identity matrix I to the spatial adjacency matrix A of the spatial graph s In, get calculate The diagonal matrix Apply the following layered propagation rule to convolve the spatial graph and use the activation function ReLU to perform nonlinear processing on the convolution result to obtain a new spatial embedding Its expression is: in, is the weight matrix of the lth layer, the initial X is the normalized gene expression matrix, ReLU is the activation function, yes The diagonal matrix of Performing feature convolution on the feature map includes the following steps: adding the identity matrix I to the feature adjacency matrix A of the feature map f In, get calculate The diagonal matrix Apply the following layered propagation rule to convolve the feature map and use the activation function to perform nonlinear processing on the convolution result to obtain a new feature embedding Its expression is: in is the weight matrix of the lth layer, the initial X is the normalized gene expression matrix, ReLU is the activation function, yes The diagonal matrix of The spatial map and feature map are co-convolved using the following layer-wise propagation rule: in is the weight matrix of the lth layer of co-convolution, and The l-th layer embedding extracted from the spatial map and feature map, respectively, the initial Through co-convolution, we get the co-embedding H c , whose expression is: In order to enhance H sc and H fc The commonality between them introduces a consistency constraint, which is defined as follows:

9. The method for identifying a fusion gene interaction network using graph convolution in spatial domain according to claim 1, wherein: The introduction of the attention mechanism into the result of the convolution to obtain the final embedding matrix is ​​specifically as follows: The attention mechanism is introduced to adaptively learn the importance of latent embeddings for the joint embedding H c , perform a nonlinear transformation on it, and use a shared attention vector W co To obtain the attention coefficient a c , the formula is as follows: a c =softmax(w c0 ·σ(WH c +b)) Where σ is the tanh activation function, W is the trainable weight matrix, and b is the bias vector; The attention coefficient a is obtained by using c The same method is used to calculate the spatial embedding H s and feature embedding H f The attention coefficient a s and a f , and finally, the final embedding matrix H is obtained by combining these embeddings: H=F(a s ·H s +a f ·H f +a c ·H c ) where F is a single linear layer used to learn the latent representation of highly variable features.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a program for a graph convolution spatial domain identification method for a fusion gene interaction network. When the program for a graph convolution spatial domain identification method for a fusion gene interaction network is executed by a processor, the steps of a graph convolution spatial domain identification method for a fusion gene interaction network as described in any one of claims 1 to 9 are implemented.

Images