A spatial domain recognition method based on artificial intelligence

Abstract

The application provides a space domain recognition method based on artificial intelligence, belongs to the field of space domain recognition, combines a traditional graph convolution network with multimodality, learns potential representation of different modal data, pre-processes and standardizes gene expression data and histological image data to obtain gene features and image features, generates a space neighborhood graph and a space adjacency matrix based on space coordinate information, constructs a cross-modal graph convolution network model, learns through multi-view graph convolution, a self-attention mechanism with pruning operation and a cross-modal joint embedding learning module, reconstructs a feature vector, performs clustering, generates a space clustering graph and evaluates clustering effect by using evaluation indexes. The method provided by the application can ensure that different representations are consistent in space and improve robustness of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of spatial domain identification, and in particular to a spatial domain identification method based on artificial intelligence. BACKGROUND

[0002] In bioinformatics, spatial domain identification techniques combine single-cell transcriptomics, tissue section analysis, and biological information to enable researchers to accurately locate and quantitatively analyze RNA expression in tissue sections. This technique reveals the diversity of cell types, the heterogeneity of gene expression patterns, and the interactions between cells.

[0003] Traditional molecular biology research often relies on complex and resource-intensive experiments to determine the interactions of cells within a specific spatial domain. Using data mining techniques to reconstruct the spatial domain of transcriptomics and determine the location of cells effectively compensates for the limitations of these methods, providing a new perspective for understanding the biological functions of complex tissues.

[0004] With the development of high-throughput sequencing technology, spatial transcriptomics (ST) has become an important tool for revealing cell-cell interactions and tissue structure. Based on spatial transcriptome data, researchers have proposed a variety of non-spatial clustering and spatial clustering methods to identify and analyze cell populations and their gene expression patterns in spatial domains.

[0005] For spatial domain identification methods based on spatial transcriptome data, some computational methods have been proposed.

[0006] Traditional techniques such as K-means, Louvain, and Leiden algorithms are commonly used in Scanpy or Seurat packages to form an integrated analysis workflow. In addition, people also use graph neural networks (GNN) to identify spatial domains. Although there are currently a large number of algorithms attempting to identify the spatial domain of spatial transcriptome data, the accuracy and robustness of these algorithms often need to be further improved.

[0007] Traditional non-spatial algorithms, such as Seurat, perform linear dimensionality reduction clustering based only on expression data. Although they can identify spatial domains, they often exhibit discontinuity because they ignore the relationship between adjacent spatial domains. Probabilistic non-spatial clustering algorithms, such as Giotto, use hidden Markov random fields to identify spatial domains by comparing the gene expression patterns of adjacent cells. Another method, BayesSpace, is based on Bayesian statistics and uses a pre-defined spatial prior to facilitate the grouping of adjacent cells into the same group. However, these methods fail to effectively utilize available spatial coordinate information. SUMMARY

[0008] The application aims to provide an artificial intelligence-based spatial domain identification method, which can enhance the flexibility and generalization ability of the model.

[0009] To achieve the above-mentioned purpose, the application is implemented through the following technical solutions.

[0010] An artificial intelligence-based spatial domain identification method, comprising the steps of:

[0011] Obtaining spatial transcriptome data, spatial information, gene expression, and histological images;

[0012] Pretreating and standardizing the gene expression data and histological image data to obtain gene features and image features;

[0013] Generating a spatial neighborhood graph based on spatial coordinate information and converting it into a spatial adjacency matrix;

[0014] Constructing a cross-modal graph convolution network model according to the gene features, image features, and spatial adjacency matrix, wherein the network comprises a multi-view graph convolution layer, a self-attention mechanism with pruning operation, and a cross-modal joint embedding learning module;

[0015] Using a decoder to reconstruct the feature vectors learned by the cross-modal joint embedding learning;

[0016] Clustering the reconstructed feature vectors, assigning cluster labels to each node, and generating a spatial clustering graph;

[0017] Using evaluation indexes to evaluate the clustering effect.

[0018] Further, the preprocessing of the original gene expression comprises the steps of:

[0019] Performing gene screening preprocessing on the original gene expression data to eliminate genes with an expression coverage lower than a preset cell number standard;

[0020] Using the Seurat V3 method to screen high-variable genes;

[0021] Normalizing each cell to eliminate sequencing depth differences;

[0022] Log-converting the gene data to obtain a converted gene feature matrix , wherein P is the number of high-variable genes, and N is the total number of points.

[0023] For the histological image data, using a pre-trained Vision Transformer model to extract image features corresponding to each pixel point to generate an image feature matrix , wherein M is the feature dimension of the model output.

[0024] The standardization processing of the original gene expression comprises the steps of:

[0025] The StandardScaler method is used to transform the features into a distribution with a mean of 0 and a standard deviation of 1, and the standardized features are obtained:

[0026] Let , ( ) represent the ith gene feature and image feature, respectively.

[0027] The standardized features are , where and are the mean and standard deviation of the feature, respectively, and the standardized gene expression data and histological image data are and , respectively.

[0028] Further, KNN and Radius are used to generate spatial neighborhood graphs and transform them into spatial adjacency matrices, and two different spatial neighborhood graphs are , where n is the graph number, n=1 represents the KNN method, n=2 represents the Radius method, V represents the set of points, R represents the set of real numbers, represents the set of connecting edges between points in the nth graph, and the adjacency matrix of graph is represented as .

[0029] Further, generating a spatial adjacency matrix using the KNN method includes the following steps:

[0030] Calculate the Euclidean distance between all nodes to generate a distance matrix:

[0031]

[0032] where , represent the ith and jth nodes in V, respectively.

[0033] Adjust the threshold t to find the appropriate number of neighbors, so that the number of neighbors of each node is close to the specified k value, if , then , otherwise 0.

[0034] Set the diagonal to 0 to remove the self-loop.

[0035] Determine the optimal k value based on the median number of neighbors , and adjust the adjacency matrix if the specified k is different from the optimal k value.

[0036] Construct positive and negative adjacency matrices , and generate a sparse adjacency matrix, where is the element in the i-th row and j-th column of matrix A;

[0037] Generating a spatial adjacency matrix using the Radius method includes the steps of:

[0038] Extracting spatial coordinates from spatial transcriptomic data and initializing an adjacency matrix:

[0039] ;

[0040] Using nearest neighbor search with different radius values, calculating the average number of neighbors for each cell to find the optimal radius value;

[0041] Based on the optimal radius, constructing an adjacency matrix representing the connections between cells and creating positive and negative graph representations, the positive graph representing the adjacency relationship between cells , and the negative graph representing the non-adjacency relationship and generating a sparse adjacency matrix, where is the element in the i-th row and j-th column of matrix A.

[0042] Further, the construction of the cross-modal graph convolutional network model includes the steps of:

[0043] Multi-view graph convolutional layer: combining two spatial neighborhood graphs with gene expression, combining PCA dimensionality reduction, generating spatial embedding, and using soft attention mechanism to feature fusion of generated embedding;

[0044] Self-attention mechanism with pruning operation: linear transformation of feature-fused embedding, calculation of attention weight, pruning of low-weight connections through pre-set threshold, and obtaining pruned features;

[0045] Cross-modal joint embedding learning module: constructing projection head through multilayer perception, linear transformation and GELU activation of pruned features, introducing residual connection, feature fusion of generated latent representation, and generating cross-modal joint embedding learning feature vector through fully connected layer.

[0046] Further, the multi-view graph convolutional layer step includes:

[0047] Graph convolution of two spatial neighborhood graphs and gene expression: ;

[0048] PCA dimensionality reduction operation: where is the principal component matrix calculated by PCA;

[0049] Two-layer graph convolution:

[0050]

[0051]

[0052] two spatial embeddings and , where is a corresponding diagonal matrix, , are weight matrices of two-layer graph convolution, , are bias terms of the n-th spatial adjacency matrix of two-layer graph convolution;

[0053] perform graph convolution on the spatial neighborhood graph and image features, i.e. , and then perform PCA operation: , where is a principal component matrix calculated by PCA, to obtain two spatial embeddings and ;

[0054] perform feature fusion on the generated embeddings using a soft attention mechanism: , , where and , and are trainable weight parameters;

[0055] input the obtained and into a self-attention mechanism.

[0056] Further, the self-attention mechanism with pruning operation includes the steps of:

[0057] perform linear transformation on the features: , , where and are learnable weight matrices;

[0058] calculate attention weights: , , where and are dimensions of each feature vector and ;

[0059] perform weighting to obtain , , and perform pruning according to a preset threshold, attention weights lower than the threshold will be set to 0;

[0060] input the obtained and input into the cross-modal joint embedding learning for joint learning.

[0061] Further, the cross-modal joint embedding learning module comprises steps of:

[0062] performing a first layer linear transformation on each modal data to calculate a projection head: , wherein and are learnable bias terms;

[0063] applying a GELU activation function to the projection result to introduce nonlinearity:

[0064]

[0065] ;

[0066] performing a second layer linear transformation on and : , wherein and are learnable parameters, and a residual connection is introduced to obtain: , ;

[0067] performing feature fusion on the generated latent representations, concatenating and to generate , and passing the generated through a fully connected layer: to map to an output dimension to generate a cross-modal joint embedding learning feature vector , which is input into a ZINB decoder for reconstruction.

[0068] Further, the ZINB decoder is used to reconstruct the feature vector :

[0069] Given the gene expression data x of the ST data, the gene expression data x follows a ZINB distribution: wherein are respectively a zero inflation parameter, a mean, a parameter matrix of a decoder output discretization and a bias vector, and a ZINB loss is minimized: .

[0070] Further, the KMeans algorithm is used to perform clustering analysis on the trained feature vector :

[0071] A KMeans clustering model is created, and the fit method is used to perform clustering on the input data The data is migrated out of the CPU, cleaned, divided into k clusters, and each point is assigned a cluster label, and a spatial clustering graph is generated;

[0072] The evaluation of the clustering effect uses the formula:

[0073]

[0074] Wherein represents the number of samples in N cell points whose clustering results are consistent with the true label, represents the number of samples in N cell points whose clustering results are inconsistent with the true label, is the expected random consistency number, is the accuracy of clustering, The higher the value of the, the more accurate the clustering.

[0075] The advantages of the present application are:

[0076] The cross-modal graph convolutional network is constructed, the data of multiple modes is fused in the graph structure, the node features of each mode are extracted by using the graph convolutional layer (GCN), and the information sharing and fusion between modes are realized through cross-modal information transmission. The graph convolutional layer of each mode updates the node features through aggregation operation of local neighborhood information, so as to capture the semantic relationship between different modes.

[0077] The most suitable neighbor number k can be adaptively found, and the overfitting or underfitting problem caused by fixed k value is avoided, so that the connectivity of the graph structure is neither too sparse nor too dense, thereby improving the flexibility and accuracy of the model.

[0078] By traversing different radius values, the number of neighbors of each node under the radius is calculated, and the most matched radius value is selected by comparing with the expected value of the radius value. In this way, the neighborhood structure of each node can be ensured to be most reasonable, and the sparsity or excessive density of the graph structure caused by too large or too small radius can be avoided.

[0079] The projection head maps the input of each mode to a unified feature space through linear transformation, so that the features of different modes can be understood by each other. Feature fusion integrates these features together, so that the model obtains more rich information as a whole.

[0080] The process of combining the projection head with the feature fusion aims to realize the spatial consistency of the latent representations between different modes. This method ensures that the features of different modal data are processed in a unified feature space, thereby reducing the difference between modes.

[0081] By adding spatial regularization constraints during training, it ensures that the nodes of the graph follow their relationships in the spatial neighborhood in the latent space. Specifically, spatially adjacent points should have a smaller distance in the latent space, while non-adjacent points should have a larger distance. This regularization mechanism helps the model better maintain the consistency of the spatial structure during learning, avoiding mismatch between the node distribution in the latent space and the actual spatial structure.

[0082] A self-attention mechanism with pruning operation is adopted, and dynamic pruning is introduced in attention calculation to adjust the attention weights between nodes. By pruning unimportant connections, the computational burden is reduced and efficiency is improved. During training, the pruning threshold can be dynamically adjusted, thereby reducing the model's dependence on noise or irrelevant connections, enhancing the model's generalization ability, and avoiding overfitting. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 A flowchart of the spatial domain recognition method based on artificial intelligence of the present application;

[0084] Figure 2 A flowchart of the preprocessing of gene expression data and histological image data of the present application;

[0085] Figure 3 A flowchart of the construction of the cross-modal graph convolutional network model of the present application. DETAILED DESCRIPTION

[0086] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0087] The present embodiment discloses a spatial domain recognition method based on artificial intelligence, which evaluates the performance on five ST data sets:(1) human dorsolateral prefrontal cortex cells (DLPFC);(2) human breast cancer cells (Human breast cancer);(3) lung cancer data set (Lung cancer);(4) human pancreatic ductal adenocarcinoma (PDAC);(5) mouse brain data set (Mouse brain). For specific process, please refer to Figure 1 , each step will be described in detail below.

[0088] S1, the spatial transcriptome data is taken as the input of the model, including gene expression, spatial information and histological image, and is preprocessed, please refer to the flowchart of Figure 2 .

[0089] Gene filtering preprocessing was performed on the original gene expression data, genes expressed in less than 3 cells were removed, 3000 high variable genes were screened out using Seurat V3 method, each cell was normalized to eliminate sequencing depth differences, and finally the data was log transformed to obtain the gene feature matrix , where P is the number of high variable genes, and N is the total number of points.

[0090] For histological image data, a pre-trained Vision Transformer model was used to extract image features corresponding to each pixel point to generate an image feature matrix , where M is the feature dimension of the model output.

[0091] In order to ensure that the features of different modalities are in the same scale, the StandardScaler method is used to transform the features into a distribution with mean 0 and standard deviation 1, and the standardized features are obtained:

[0092] Definition , ( ) represents the i-th gene feature and image feature, respectively.

[0093] The standardized features are , where and are the mean and standard deviation of the feature, respectively, and the standardized gene expression data and histological image data are and , respectively.

[0094] S2, for spatial information, we use different similarity measures (i.e. KNN and Radius) to generate two different spatial neighborhood graphs , n is the graph number, n=1 represents the KNN method, n=2 represents the Radius method, V represents the set of points, R represents the set of real numbers, represents the set of connecting edges between points in the nth graph, and the adjacency matrix of graph is represented as .

[0095] Generating a spatial adjacency matrix using the KNN method includes the following steps:

[0096] Calculate the Euclidean distance between all nodes to generate a distance matrix:

[0097]

[0098] where , represent the i-th and j-th nodes in V, respectively.

[0099] The appropriate number of neighbors is found by adjusting the threshold t so that the number of neighbors of each node is close to the specified k value, if then otherwise 0. Diagonal is set to 0 to remove self-loops; the best k value is determined based on the median of the number of neighbors and the adjacency matrix is adjusted if the specified k is different from the best k value. Positive and negative adjacency matrices are constructed , and a sparse adjacency matrix is generated for facilitating spatial regularization constraints, where is the element in the ith row and jth column of matrix A.

[0100] Generating spatial adjacency matrices with Radius method includes the steps of:

[0101] Extracting spatial coordinates from spatial transcriptomic data and initializing adjacency matrices:

[0102] .

[0103] Using nearest neighbor search with different radius values, the average number of neighbors for each cell is calculated to find the optimal radius value. Based on the optimal radius, adjacency matrices are constructed to represent the connections between cells and positive and negative graph representations are created, where the positive graph represents the adjacency relationship between cells and the negative graph represents the non-adjacency relationship and a sparse adjacency matrix is generated for facilitating spatial regularization constraints, where is the element in the ith row and jth column of matrix A.

[0104] S3, constructing a cross-modality graph convolutional network model, please refer to Figure 3 , which includes a multi-view graph convolutional network, cross-modality joint embedding learning, and a self-attention mechanism with pruning operation.

[0105] Multi-view graph convolutional layer: first, different spatial adjacency matrices are graph-convolved, which follows the following propagation rule:

[0106] where is the weight matrix of the lth layer, is the ReLU activation function.

[0107] Graph-convolving two types of spatial neighborhood graphs with gene expression: To reduce the complexity of the calculation and preserve important feature information, PCA operation is added during the convolution process: where is the principal component matrix calculated by PCA. Two layers of graph convolution are then performed:

[0108]

[0109]

[0110] get two spatial embeddings and where is the corresponding diagonal matrix, , are the weight matrices of two-layer graph convolution, , are the bias terms of the n-th spatial adjacency matrix of two-layer graph convolution.

[0111] Graph convolution is performed on the spatial neighborhood graph and image features, i.e. , and then PCA operation is performed: where is the principal component matrix calculated by PCA, and two spatial embeddings and are obtained in the same way. Feature fusion is performed on the generated embeddings using a soft attention mechanism: , where and , and are trainable weight parameters: , , , and are obtained in the same way. The obtained and are input into the self-attention mechanism.

[0112] Secondly, the self-attention mechanism with pruning operation is used to adaptively fuse the features of each point of each modal data.

[0113] Linear transformation is performed on the features: , where and are learnable weight matrices. The attention weight is calculated: , where and are the dimensions of each feature vector and . Finally, weighting is performed to obtain , In order to reduce the amount of calculation, pruning operation is used to remove low-weight connections, and attention weights below the threshold will be set to 0. The obtained and input into the cross-modal joint embedding learning for joint learning.

[0114] Finally, in order to learn the data of different modalities, feature learning is performed using a multi-layer perceptron (MLP), ensuring that the model captures similar structures when processing different modal data, thereby improving the generalization ability of the model.

[0115] A first layer linear transformation is performed on each modal data to calculate the projection head: , where and are learnable bias terms. A GELU activation function is applied to the projection result to introduce nonlinearity:

[0116]

[0117] ;

[0118] A second layer linear transformation is performed on and : , where and are learnable parameters. In order to facilitate the flow of information, a residual connection is introduced, resulting in: , The generated latent representation is fused by feature, and and are spliced together to generate , which is then passed through a fully connected layer to map to the output dimension to generate . The vector learned by the cross-modal joint embedding learning is input into the ZINB decoder for reconstruction.

[0119] S4, in order to solve the high sparsity and discreteness of ST data, the ZINB decoder is used to reconstruct the feature vector .

[0120] Given the gene expression data x of ST data, it is assumed to follow the ZINB distribution, and the ZINB is defined as follows: where are zero inflation parameters, mean, decoder output discrete and bias vector parameter matrix, respectively. By minimizing the ZINB loss, i.e. , the feature vector contains more biological information.

[0121] S5, the trained feature vector is clustered using the KMeans algorithm The clustering analysis is performed. A KMeans clustering model is created, k specifies the number of clusters of clustering, and random_state ensures the reproducibility of the clustering process. The input data is trained using the fit method, the data is migrated out of the CPU, the data is cleaned, the data is divided into k clusters, and each point is assigned a cluster label, and a spatial clustering graph is generated.

[0122] S6, evaluate the clustering effect using evaluation indexes:

[0123]

[0124] Among them represents the number of samples in N cell points whose clustering results are consistent with the true labels, represents the number of samples in N cell points whose clustering results are inconsistent with the true labels, is the expected random consistency number. is the accuracy of clustering, The higher the value of, the more accurate the clustering is, and the spatial clustering graph generated can reveal the complex biological characteristics of cells at important spatial positions.

[0125] Finally, it should be pointed out that: the above only for the preferred embodiments of the present application, and not for limiting the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.​

Claims

1. An artificial intelligence-based spatial domain recognition method, characterized by, The method comprises the steps of: obtaining spatial information, gene expression and histological images of spatial transcriptome data; preprocessing and standardizing gene expression data and histological image data to obtain gene features and image features; generating a spatial neighborhood graph based on spatial coordinate information and converting it into a spatial adjacency matrix; constructing a cross-modal graph convolution network model based on gene features, image features and the spatial adjacency matrix, wherein the network comprises a multi-view graph convolution layer, a self-attention mechanism with pruning operation and a cross-modal joint embedding learning module; reconstructing the feature vector learned by the cross-modal joint embedding learning using a decoder; clustering the reconstructed feature vector, assigning a cluster label to each node and generating a spatial clustering graph; evaluating the clustering effect using evaluation indexes; The construction of the cross-modal graph convolution network model comprises the steps of: the multi-view graph convolution layer: combining two spatial neighborhood graphs with gene expression, generating spatial embedding by PCA dimension reduction and performing feature fusion on the generated embedding using a soft attention mechanism; the self-attention mechanism with pruning operation: performing linear transformation on the embedding after feature fusion, calculating attention weights, pruning low-weight connections through a preset threshold to obtain pruned features; the cross-modal joint embedding learning module: constructing a projection head through a multilayer perception mechanism, performing linear transformation and GELU activation on the pruned features, introducing a residual connection, performing feature fusion on the generated latent representation and generating a feature vector learned by cross-modal joint embedding through a fully connected layer. 2.The space domain recognition method based on artificial intelligence according to claim 1, characterized in that, The preprocessing of the original gene expression comprises the steps of: performing gene screening preprocessing on the original gene expression data to remove genes with an expression coverage lower than a preset cell number standard; screening high-variable genes using the Seurat V3 method; normalizing each cell to eliminate sequencing depth differences; log transforming the gene data to obtain a transformed gene feature matrix where P is the number of high variance genes and N is the total number of points. For the histological image data, the image features corresponding to each pixel point are extracted using a pre-trained Vision Transformer model to generate an image feature matrix where M is the feature dimension of the model output; The standardization processing of the original gene expression comprises the steps of: using the StandardScaler method to transform the features into a distribution with a mean of 0 and a standard deviation of 1 to obtain standardized features: Definitions , ( ) respectively denote the i-th gene feature and image feature; The standardized feature is , where and are the mean and standard deviation of the feature, respectively, and the standardized gene expression data, histology image data are and , respectively. 3.The AI-based spatial domain recognition method of claim 2, wherein, The spatial neighborhood graphs are generated using KNN and Radius and transformed into spatial adjacency matrices, two different spatial neighborhood graphs are where n is the graph index, n = 1 denotes the KNN method, n = 2 denotes the Radius method, V denotes the set of points, R denotes the set of real numbers, denotes the set of connecting edges between points in the nth graph, graph The adjacency matrix of graph is denoted as 4.The AI-based spatial domain recognition method of claim 3, wherein, the steps of generating a spatial adjacency matrix using the KNN method comprise: calculating the Euclidean distance between all nodes to generate a distance matrix: wherein, , respectively represent the i-th, j-th node in V; The proper number of neighbors is found by adjusting the threshold t so that the number of neighbors for each node is close to the specified k value, if then else 0; setting the diagonal line to 0 and removing the self-loop; Determine the optimal k value based on the median number of neighbors Adjust the adjacency matrix if the specified k is different from the optimal k value. Constructing the positive and negative adjacency matrices , , and generating a sparse adjacency matrix where is the element in the ith row and jth column of matrix A; the steps of generating a spatial adjacency matrix using the Radius method comprise: extracting spatial coordinates from spatial transcriptome data and initializing an adjacency matrix: ； using nearest neighbor search with different radius values to calculate the average number of neighbors for each cell to find the optimal radius value; An adjacency matrix is constructed based on the optimal radius, representing the connection between cells, and a positive and negative graph representation is created, the positive graph representing the adjacency relationship between cells , and the negative graph representing the non-adjacency relationship , and a sparse adjacency matrix is generated, where is the element of the ith row and jth column in matrix A. 5.The AI-based spatial domain recognition method of claim 1, wherein, the multi-view graph convolution layer step comprises: convolving two spatial neighborhood graphs with gene expression: ; Performing PCA dimensionality reduction operation: where is the principal component matrix computed by PCA; performing two-layer graph convolution: two spatial embeddings are obtained and where is a corresponding diagonal matrix, , are weight matrices of two-layer graph convolution, respectively, , are bias terms of the n-th spatial adjacency matrix of two-layer graph convolution, respectively. The spatial neighborhood graph is graph-convolved with the image features, i.e. , followed by a PCA operation: , where is the principal component matrix computed by PCA, resulting in two spatial embeddings and ; Feature fusion is performed on the generated embedding using a soft attention mechanism: , wherein and , and are trainable weight parameters; The resulting and are input into the self-attention mechanism. 6.The space domain recognition method based on artificial intelligence according to claim 5, characterized in that, the self-attention mechanism with pruning operation comprises the steps of: Linear transformation of features: , where and are learnable weight matrices; Computing attention weights: , where and are the dimensions of each feature vector and , respectively. Weighting is performed to obtain , and pruning is performed according to a preset threshold, and attention weights below the threshold are set to 0; The resulting and are input into the cross-modal joint embedding learning for joint learning. 7.The space domain recognition method based on artificial intelligence according to claim 6, characterized in that, the cross-modal joint embedding learning module comprises the steps of: A first layer linear transformation is performed on each modality data to compute the projection heads: , where and are learnable bias terms; applying a GELU activation function to the projection result to introduce nonlinearity: ； To and perform the second layer linear transformation: , where and are learnable parameters, introduce a residual connection to get: , ; The generated potential representation is fused with features, and and are spliced together to generate , and then pass through a fully connected layer: , and are mapped to the output dimension to generate a cross-modal joint embedding learned feature vector , input to the ZINB decoder for reconstruction. 8.The space domain recognition method based on artificial intelligence according to claim 7, characterized in that, Decoding the feature vector using the ZINB decoder Reconstruction is performed: Given the gene expression data x of ST data follows ZINB distribution: where are the zero inflation parameter, mean, parameter matrix of the decoder output's discretization and bias vector, respectively, minimizing the ZINB loss: . 9.The space domain recognition method based on artificial intelligence according to claim 7, characterized in that, Using KMeans algorithm on the trained feature vectors Performing cluster analysis: Create a KMeans clustering model, train it on the input data using the fit method out of the CPU, clean the data, divide it into k clusters, and assign a cluster label to each point, generating a spatial clustering plot. The evaluation of the clustering effect uses the formula: ； wherein represents the number of samples in N cell points whose clustering results are consistent with the true labels, represents the number of samples in N cell points whose clustering results are inconsistent with the true labels, is the expected random consistency number, is the accuracy of clustering, The higher the value of is, the more accurate the clustering is.

Images