A spatial transcriptomic domain identification method, system, device and medium

Abstract

The application provides a spatial transcriptome domain identification method, system, device and medium, and belongs to the field of bioinformatics. The method comprises the following steps: constructing a multi-scale feature representation fusing molecular expression information and spatial context information, and constructing an initial spatial adjacency graph; constructing complementary double-masked views based on the multi-scale feature representation; learning a latent representation through a double-encoder structure, constructing a cross-view representation alignment loss for the latent embedding of the double-encoder, and jointly reconstructing gene expression features and spatial adjacency structures based on the latent representation obtained by the online encoder; constructing a local-global collaborative contrast learning strategy based on the latent representation, constructing positive and negative samples according to spatial neighborhood relationships and cross-region consistency, and sampling heterogeneous nodes; jointly optimizing a total loss function, and dividing tissue regions through a clustering method. The application can simultaneously retain molecular expression information and spatial structure information, and is used for spatial domain identification and downstream analysis of spatial transcriptome data.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, and in particular to a method, system, device, and medium for spatial transcriptome domain identification. Background Technology

[0002] Tissues in multicellular organisms are composed of populations of cells with specific molecular characteristics and functions. Through complex intercellular communication and microenvironment regulation, these cells form spatial structures with well-defined boundaries and functional zones, and their spatial distribution is closely related to the organism's physiological state, developmental processes, and disease progression. However, traditional transcriptome sequencing typically requires tissue dissociation, leading to the loss of cellular spatial location information. In recent years, the emergence of spatial transcriptomics has overcome this limitation. This technology can measure gene expression levels while preserving the original spatial structure of the tissue, thus achieving simultaneous acquisition of molecular and spatial location information. Currently, this technology is mainly divided into two categories: in situ hybridization-based methods (such as MERFISH, BaristaSeq, STARmap, etc.) and in situ capture sequencing-based methods (such as Visium, Slide-seq, Stereo-seq, etc.). With the continuous improvement of spatial resolution and detection coverage, spatial transcriptomics has become an important cutting-edge technology for elucidating complex tissue structures, cellular heterogeneity, and intercellular interactions.

[0003] However, accurately reconstructing the anatomical structure and functional regions of tissues from high-throughput, multi-dimensional spatial expression data has become one of the core challenges in spatial transcriptome data analysis. To address this issue, spatial domain identification has gradually become a key task in spatial transcriptome analysis. Its goal is to divide tissues into spatial regions with similar biological functions and molecular characteristics based on gene expression patterns and spatial neighborhood relationships. Accurate spatial domain identification not only helps in the reconstruction of tissue anatomy but also provides important support for intercellular communication analysis, tissue development research, and the exploration of disease mechanisms.

[0004] Early studies primarily employed non-spatial clustering methods, such as K-means, Louvain, and Seurat. However, these methods mainly relied on gene expression information, neglecting spatial continuity and neighborhood dependencies within tissues, thus making it difficult to accurately recover true tissue boundaries. Subsequently, researchers gradually incorporated spatial information into the recognition process. For example, stLearn used the histological image similarity between spatial points to smooth feature embeddings; Giotto introduced a Hidden Markov Random Field model to characterize spatial dependencies; and BayesSpace enhanced spatial consistency through a Bayesian framework. With the rapid development of Graph Neural Networks (GNNs), modeling spatial transcriptome data as a graph structure has gradually become the mainstream method for spatial domain recognition. In this framework, spatial locations are treated as graph nodes, and spatial neighborhood relationships are represented as graph edges, achieving joint modeling of gene expression features and spatial structural information through graph representation learning.

[0005] Typical methods include: SpaGCN, which uses graph convolutional networks to fuse gene expression, spatial location, and histological information; STAGATE, which adaptively adjusts neighborhood weights through graph attention mechanisms to improve tissue boundary resolution; and SEDR, which combines deep autoencoders and variational graph convolutional networks to achieve joint embedding of expression features and spatial structures. These studies fully demonstrate that representation learning frameworks based on graph neural networks can effectively fuse spatial information and molecular features. However, most existing graph models still mainly rely on traditional clustering or reconstruction objectives for optimization, making it difficult to fully explore the complex latent semantic structures in spatial transcriptome data. This limitation has further promoted the development of self-supervised learning frameworks, especially contrastive learning-based methods, in spatial transcriptome analysis.

[0006] Self-supervised learning offers advantages such as eliminating the need for manual annotation and adaptively mining intrinsic data features, effectively alleviating the problems of scarce annotations and high manual costs in spatial transcriptome data. Currently, the combination of graph neural networks and self-supervised learning has led to a series of efficient recognition methods that deeply integrate expressive features and spatial topological information. Among these, SpaceFlow utilizes graph convolutional networks to fuse spatial information and expressive features, and corrects feature biases through self-supervised optimization strategies; GraphST employs a data augmentation-driven self-supervised training approach to improve the robustness of feature representations to noise; STMGAC combines mask reconstruction and structural association learning to optimize spatial representation quality from two dimensions; ConSpaS, based on graph autoencoders and contrastive learning, achieves spatial domain recognition by integrating local and global similarity; SpaGRA uses graph augmentation and geometric contrastive learning strategies, treating spatial distance as prior knowledge and dynamically adjusting edge weights through a multi-head graph attention network to capture latent representations; and stHGC, as a self-supervised graph representation learning framework, achieves spatial domain recognition by constructing a hybrid neighborhood graph and introducing spatial regularization constraints. These methods have demonstrated superior recognition performance compared to traditional unsupervised methods on multiple spatial transcriptome datasets.

[0007] Although the above methods have achieved good results on multiple benchmark datasets, they still have the following limitations: First, most methods rely primarily on raw gene expression data, with limited utilization of spatial location information and local topological features, making it difficult to effectively characterize multi-scale spatial patterns in complex tissues, thus limiting the accuracy of spatial domain identification.

[0008] Second, existing autoencoder architectures typically focus on reconstructing gene expression features while neglecting the reconstruction and preservation of spatial topology. This results in the learned representations failing to fully retain the spatial structural information within the organization, affecting the accurate characterization of spatial neighborhood relationships.

[0009] Third, when constructing positive and negative samples for contrastive learning, most methods rely on fixed physical distances or simple clustering results for sample selection. They lack boundary resampling mechanisms designed for complex tissue boundary regions. Heterogeneous sites in transition regions are easily confused during the sampling process, thus limiting the model's ability to identify tissue boundaries.

[0010] In summary, how to construct a unified learning framework that can simultaneously achieve effective modeling of multi-scale spatial heterogeneity, preservation of spatial topology, and construction of reliable contrastive supervision mechanisms remains a significant challenge to be addressed in the field of spatial transcriptome analysis. Summary of the Invention

[0011] The purpose of this invention is to provide a method, system, device, and medium for spatial transcriptome domain identification. By constructing a unified framework of spatially guided multi-scale features, complementary double-mask views, double-mask collaborative reconstruction learning, and local and global collaborative contrastive representation learning, accurate identification of spatial domains in complex tissue structures can be achieved.

[0012] To achieve the above objectives, the present invention provides a method for spatial transcriptome domain identification, comprising the following steps: Step S1: Based on gene expression information and spatial location information in spatial transcriptome data, construct a multi-scale feature representation that integrates molecular expression information and spatial context information; Step S2: Construct an initial spatial adjacency graph based on spatial location information, construct a spatial topology graph using the spatial location information and the initial spatial adjacency graph, and construct a complementary double-mask view based on multi-scale feature representation, and enhance information interaction between nodes by combining graph structure propagation mechanism; Step S3: Based on the complementary dual-masked view, learn the latent representation through the dual encoder structure, construct the cross-view representation alignment loss for the latent embedding of the dual encoder, and jointly reconstruct the gene expression features and spatial adjacency structure in the complementary dual-masked view based on the latent representation obtained by the online encoder. Step S4: Based on the latent representation, construct a collaborative contrastive learning strategy, construct positive and negative samples according to spatial neighborhood relationships and cross-regional consistency, and construct an adaptive sampling mechanism to sample heterogeneous nodes in the transition region of the organizational structure and calculate the local-global collaborative contrastive loss. Step S5: Jointly optimize the total training loss function, divide the organization region through clustering methods, and carry out downstream tasks using latent representations.

[0013] The present invention also provides a spatial transcriptome domain identification system for implementing the spatial transcriptome domain identification method described above, comprising: The multi-scale feature construction module is used to construct multi-scale feature representations that integrate molecular expression information and spatial context information based on gene expression information and spatial location information in spatial transcriptome data. The spatial adjacency graph and masked view construction module is used to construct an initial spatial adjacency graph based on spatial location information, construct a spatial topology graph using spatial location information and the initial spatial adjacency graph, and construct a complementary dual-masked view based on multi-scale feature representation, and enhance the information interaction between nodes by combining graph structure propagation mechanism; The dual encoder and reconstruction module is used to learn latent representations based on complementary dual-masked views through the dual encoder structure, construct cross-view representation alignment loss on the latent embeddings of the dual encoder, and jointly reconstruct gene expression features and spatial adjacency structure in complementary dual-masked views based on the latent representations obtained by the online encoder. The collaborative contrastive learning module is used to construct a collaborative contrastive learning strategy based on latent representations. It constructs positive and negative samples according to spatial neighborhood relationships and cross-regional consistency, and constructs an adaptive sampling mechanism to sample heterogeneous nodes in the transition region of the organizational structure and calculate the local-global collaborative contrastive loss. The joint training and clustering module is used to jointly optimize the total training loss function, divide the organization region through clustering methods, and carry out downstream tasks using latent representations.

[0014] The present invention also provides a computer device, including: a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described spatial transcriptome domain identification method.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described spatial transcriptome domain identification method.

[0016] Therefore, the present invention employs the above-described method, system, device, and medium for spatial transcriptome domain identification, and the beneficial technical effects are as follows: (1) This invention uses a spatially guided multi-scale feature construction strategy to tokenize the original gene expression matrix, extract local co-expression relationships using one-dimensional convolution, construct multi-frequency spatial location codes by combining multi-frequency sine and cosine spatial location codes, and fuse gene token features, local convolution features and spatial location codes. After pooling and principal component analysis to reduce dimensionality, multi-scale feature representations are obtained, which can provide feature inputs that fuse molecular expression information and spatial context information for spatial domain recognition.

[0017] (2) This invention learns latent representation by using online encoder and target encoder through dual-mask collaborative reconstruction learning, and uses feature decoder to reconstruct gene expression features of mask nodes. At the same time, it uses structure decoder to predict the consistency between adjacency matrix and real spatial adjacency matrix to reconstruct spatial topology. By jointly optimizing dual reconstruction loss, cross-view representation alignment loss and local-global collaborative contrast loss, the latent representation can simultaneously retain molecular expression information and spatial structure information.

[0018] (3) This invention uses local and global collaborative comparative representation learning to select Top-K similar neighbors based on cross-view similarity matrix and jointly constrain them with the initial spatial adjacency matrix to construct local positive samples. It constructs a consensus matrix through multi-round clustering and constructs global positive samples by combining feature space neighborhood relations. At the same time, it calculates boundary scores based on consensus matrix and defines node sampling weights to adjust the sampling probability of nodes in boundary areas. It also obtains semi-difficult negative samples by filtering negative sample candidate sets through similarity threshold. It can construct positive and negative samples by utilizing spatial neighborhood relations and cross-regional consistency, and realize the sampling of heterogeneous nodes in the transition area of ​​organizational structure. Attached Figure Description

[0019] Figure 1 This is a model framework diagram of SpaMCAR; Figure 2 The results of SpaMCAR and contrastive methods in the spatial domain identification of a human breast cancer dataset are shown; among them, Figure 2 (A) in the table represents the actual label result; Figure 2 (B) in the image represents the SpaMCAR spatial domain identification result; Figure 2 (C) in the figure is a bar chart of the indicators of SpaMCAR and 8 comparison methods; Figure 2 (D) in the figure represents the spatial domain identification result of the comparison method; Figure 3 Heterogeneity analysis of SpaMCAR in the human breast cancer dataset; where, Figure 3 (A) in the figure is a visualization of the original gene expression and the spatial expression after SpaMCAR correction; Figure 3 (B) in the diagram represents communication networks in different spatial domains under the CXCL pathway; Figure 3 In the diagram, (C) represents the spatial expression distribution of representative genes in spatial domains 4 (healthy region), 7 (tumor border region), and 14 (tumor region); Figure 3 (D) in the figure represents the GO functional enrichment analysis bar chart between the healthy region and the tumor region; Figure 4 Spatial domain identification results of SpaMCAR and contrastive methods in a mouse anterior brain dataset; among which, Figure 4 (A) in the diagram represents the visualization of the real label space domain; Figure 4 (B) shows the enrichment analysis results of the main olfactory bulb region; (C) is a bar chart of the indexes of SpaMCAR and 7 comparison methods. Figure 4 In the diagram, (D) represents the spatial domain identification spatial mapping results of SpaMCAR and seven comparison methods; Figure 4 In the figure (E), the spatial expression distribution of representative marker genes (Myl4, Gpsm1, Gpr88, Enpp2, Ttr) is shown. Figure 5The results of SpaMCAR and contrastive methods in spatial domain identification of a mouse posterior brain dataset are shown; among them, Figure 5 (A) in the image represents the tissue H&E staining image and the SpaMCAR spatial domain recognition result; Figure 5 (B) in the figure is a bar chart of the silhouette coefficient (SC) and Davies-Bouldin index (DB) of SpaMCAR compared with 8 other comparison methods. Figure 5 (C) in the figure represents the spatial domain identification results of the eight comparison methods; Figure 5 (D) in the figure represents the spatial expression distribution of representative marker genes (Grm4, Ppp1r17, Kit), the violin diagram of expression levels of cluster 4 (synaptic region) and cluster 8 (synaptic edge region), and the comparison and verification with the results of the real in situ hybridization (ISH) experiment. Figure 6 The results of SpaMCAR and contrastive methods in spatial domain identification of mouse embryo datasets are shown; among them, Figure 6 (A) in the figure represents the spatial domain identification results of the embryo sample's true spatial domain label and SpaMCAR. Figure 6 (B) in the figure is a bar chart of quantitative indicators of ARI and NMI between SpaMCAR and 7 comparison methods; Figure 6 (C) in the diagram represents the spatial domain partitioning visualization results corresponding to the seven comparison methods; Figure 6 In the diagram, (D) represents the spatial distribution of each representative spatial cluster, as well as the spatial expression patterns of characteristic genes in the original expression and after SpaMCAR enhancement correction. Figure 6 (E) in the figure is a violin plot showing the cluster distribution of characteristic genes under original expression and SpaMCAR enhanced expression; Figure 7 The results of SpaMCAR and contrastive methods in the spatial domain identification of the melanoma dataset are shown; among them, Figure 7 Image (A) in the image is an H&E stained image of melanoma tissue; Figure 7 (B) in the figure represents the results of the identification of the melanoma real spatial domain label and the SpaMCAR spatial domain; Figure 7 (C) in the figure is a bar chart of quantitative indicators of ARI and NMI between SpaMCAR and 8 comparison methods; Figure 7 (D) in the figure represents the spatial domain partitioning visualization results of the eight comparison methods; Figure 7 In the diagram (E), there are bar charts showing GO functional enrichment and bubble charts showing KEGG pathway enrichment for the melanoma domain (Domain2) and the lymphoid domain (Domain4). Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0022] Example 1 like Figure 1 As shown, this embodiment provides a method for spatial transcriptome domain identification, including the following steps: Step S1: Based on gene expression information and spatial location information in spatial transcriptome data, construct a multi-scale feature representation that integrates molecular expression information and spatial context information.

[0023] Step S11: Tokenize the spatial transcriptome gene expression matrix.

[0024] Let the input original gene expression matrix be... ,in Indicates the number of spatial locations. Represents the dimension of gene characteristics. Representing the real number field, all genes in this matrix have been filtered in preprocessing to have non-zero expression values ​​in at least three spatial locations to reduce sequencing noise. First, a fixed token dimension is set. When the genetic feature dimension Unable to be During integer division, the gene dimension is expanded using zero-padding. Subsequently, the gene feature dimension is divided into... The original gene expression matrix is ​​transformed into a three-dimensional token representation using non-overlapping subspaces. ,in Gene token features This represents the number of gene tokens. Through the above operations, high-dimensional gene expression information is converted into a structured token sequence, enhancing the model's ability to learn high-order nonlinear relationships in gene expression.

[0025] Step S12: Extract local convolutional features based on 3D token representation.

[0026] Considering the potential co-expression relationships and coordinated regulatory patterns among genes, a one-dimensional convolution operation is used at the token sequence level to extract local dependencies: ; in, This represents a one-dimensional convolution operation with a kernel size of 3, used to aggregate local representation patterns between adjacent tokens. After convolution mapping, the resulting local convolutional features are... , This indicates the dimension of the convolutional output features. This process can effectively capture local gene association information and reduce the impact of high-frequency expression noise on feature learning.

[0027] Step S13: Construct a multi-frequency spatial location code based on spatial coordinate information.

[0028] To address the issue of explicit topological constraints in spatial locations within spatial transcriptome data, spatial coordinate information is utilized to enhance the model's ability to perceive tissue spatial structure. For the first... The original coordinates of each node First, standardization is performed to obtain the first... Standardized horizontal and vertical coordinates of each node , : ; in, and These represent the mean values ​​of the horizontal and vertical coordinates, respectively. and This represents the standard deviation in the corresponding direction.

[0029] Subsequently, multi-frequency sine and cosine functions are used to encode the spatial coordinates. For the spatial coordinate axes... Its position code is: ; ; in, Indicates the frequency dimension index. Indicates the first The node, the first The frequency, the first Cosine position encoding output corresponding to each channel dimension; Indicates the first The node, the first The frequency, the first Cosine position encoding output corresponding to each channel dimension; Indicates the first Spatial coordinate scalar of each node; frequency parameter .when hour, ;when hour, ,in The position encoding branch representing the X-space coordinate direction. The position encoding branch representing the Y-space coordinate direction. Indicates the first The normalized X-axis coordinates of each node Indicates the first The normalized Y-axis coordinates of each node.

[0030] By encoding the horizontal and vertical spatial directions separately and then concatenating the position embeddings from the two directions, a multi-frequency spatial position code is obtained. , of which The node, the first The one-dimensional vector corresponding to each token position is represented as follows: ,in This represents vector concatenation. , These represent the position encoding vectors in the horizontal and vertical directions, respectively.

[0031] Step S14: Jointly represent the gene token features, local convolution features, and multi-frequency spatial location encoding.

[0032] By concatenating along the feature dimension, we obtain Simultaneously, fused features containing gene expression patterns, local co-expression information, and spatial location information are obtained. Subsequently, max pooling is performed along the token dimension to obtain pooled features that aggregate key feature responses from different gene subspaces. ; in, The eigenmatrix after pooling is represented by the first... All elements of the row, This indicates taking the maximum value in the Token dimension. The first 3D fusion tensor after splicing represents the first 3D fusion tensor. The node, the first All the characteristics of a token.

[0033] Finally, principal component analysis was used to reduce the dimensionality of the fused features, resulting in a low-dimensional multi-scale feature representation. This is then used as input for the subsequent spatial representation learning module.

[0034] Step S2: Construct an initial spatial adjacency graph based on spatial location information, construct a spatial topology graph using the spatial location information and the initial spatial adjacency graph, and construct a complementary double-mask view based on multi-scale feature representation, and enhance information interaction between nodes by combining graph structure propagation mechanism.

[0035] Step S21: First, construct an initial spatial adjacency graph based on spatial coordinate information. Treat each spatial location in the spatial transcriptome data as a graph node, and use spatial distance relationships... The nearest neighbor method constructs the initial spatial adjacency graph. ,in Represents a set of spatial nodes. Indicates the first A spatial node, This represents the set of undirected edges between spatially neighboring nodes. An initial adjacency matrix is ​​constructed based on the spatial neighborhood relationships between nodes. : ; in, Indicates the first Nodes of The set of nearest neighbor nodes, This represents the index of any node belonging to this set. To improve the stability of the graph neural network propagation process and reduce the oversmoothing problem that may occur during multi-layer graph convolution, a self-loop structure is added to the initial adjacency matrix using an identity matrix, resulting in a self-loop adjacency matrix. : , Represents the initial adjacency matrix. This represents the identity matrix. Subsequently, the adjacency matrix after the self-loop is symmetrically normalized based on the node degree information to obtain the normalized adjacency matrix. : , express The corresponding degree matrix. The normalized spatial graph is used for subsequent spatial information propagation and feature updates.

[0036] Step S22: To enhance the model's adaptability to missing information and noise interference in spatial transcriptome data, a complementary mask input view is constructed. This method uses random sampling to sample the spatial node set. Divide the image into sections. Set the mask ratio. Randomly select a set of mask nodes Its quantity satisfies The remaining nodes then constitute the set of visible nodes. In this embodiment Set to 0.5. Both node sets must satisfy the following constraints: .

[0037] Step S23: For the online encoding branch, replace the masked nodes with learnable mask vectors while preserving the original features of the non-masked nodes. For nodes... Its input features are defined as: ; in, This indicates the input to the online encoder. The characteristics of each node, superscript Indicates online. This represents the learnable mask embedding in the online branch. The dimension is The real vector space, This represents the node features after dimensionality reduction.

[0038] Step S24: To form an information perspective that complements the online coding branch, an opposite masking strategy is adopted in the target encoder, that is, to recover the occluded node features in the online coding branch, while masking another part of the nodes: ; in, Indicates the first input to the target encoder The characteristics of each node, superscript Indicate the goal, This represents the independent learnable mask embeddings in the target branch. Through the above input mechanism, a complementary dual-mask view is formed.

[0039] The input features of all nodes in the online branch and the target branch are respectively arranged into matrices, denoted as the online input view. and target input view .

[0040] The online encoder and the target encoder learn spatial node representations from different information perspectives, enabling the model to infer the complete spatial semantic structure using partial observation information, thereby enhancing cross-view... Figure 1 Consistent learning ability improves the robustness of spatial representation.

[0041] Step S3: Based on the complementary dual-masked view, learn the latent representation through the dual encoder structure, construct the cross-view representation alignment loss for the latent embedding of the dual encoder, and jointly reconstruct the gene expression features and spatial adjacency structure in the complementary dual-masked view based on the latent representation obtained by the online encoder.

[0042] Step S31: For the online input view generated in step S2 and target input view The input features are extracted from both the online encoder and the target encoder. Both encoders use the same network structure, including a multilayer perceptron and a graph convolutional network. First, the input features are non-linearly mapped using the multilayer perceptron: ; ; Subsequently, the mapped features are input into a graph convolutional network, which utilizes the normalized adjacency matrix. To disseminate information within the neighborhood: ; ; in, , These represent the output representations of the in-circuit encoder and the target encoder, respectively. , , and Indicates learnable parameters, Indicates batch normalization. and This represents a learnable bias vector. and These represent the ELU and ReLU activation functions, respectively. This represents the graph projection output of the online encoder branch. The above process yields the output representation of the online encoder. (Latent representation); The target encoder uses the same coding structure for the complementary mask input. Feature propagation is performed to obtain the target encoder output. Target encoder function Indicates will Mapping to output representation The encoding process, specifically the forward propagation, is consistent with that of the online encoder.

[0043] To avoid model representation collapse, the target encoder parameters are not included in gradient updates, but are updated using an exponential moving average of the online encoder parameters. ; ; in, This represents the updated target encoder parameters. , These represent all parameters of the target encoder and the in-circuit encoder, respectively. This represents the momentum coefficient during the current training phase, with a value range of []. ,1], This represents the initial momentum coefficient, which defaults to 0.998. This indicates the current training iteration number, with a value range of [0, 1000]. This indicates the total number of training rounds, with a default value of 1000.

[0044] Step S32: To enhance the model's ability to recover latent biological semantic information from incomplete input, the online encoder output representation is... The image is mapped to the prediction space via an asymmetric projection network. First, a feature transformation is performed on the output representation of the online encoder, and the mask nodes of the online branches are then mapped. The corresponding representation is replaced with a learnable placeholder vector to obtain the intermediate representation. Subsequently, an online branch prediction representation is obtained by combining spatial topology information with a graph projection module. : ; in, This represents an asymmetric graph projection network. The target encoder output... Then it is directly used as the monitoring target: ,in This represents the projected representation of the target encoder. Also, it applies only to the mask node set. Calculate cross-view representation alignment loss : ; in, , Representing nodes respectively Online branch prediction representation and target branch supervision representation, This represents the L2 norm.

[0045] Step S33: Reconstruct gene expression features. Utilize online branch prediction representation. and the normalized adjacency matrix Gene expression characteristics of mask nodes are recovered using a graph decoder. : ; in, This is a learnable weight matrix. Considering the sparsity and dropout characteristics of spatial transcriptome data, weights are only applied to the mask node set. The above calculation expresses the reconstruction loss. : ; in, Cosine similarity is used to measure the original features. and reconstructed features In terms of similarity in direction, This represents the temperature coefficient scaling factor, with a default value of 2.

[0046] Step S34: Reconstruct the spatial topology. To maintain spatial neighborhood relationships, the output representation of the online encoder is... This is used for spatial graph structure reconstruction. First, the node representation is normalized, and then the connection probabilities between nodes are calculated. : ; in, Output for online encoder via projection head Transform and then proceed The normalized embedding matrix, denoted by norm Normalization function; and They are respectively The row and number Row, corresponding node and The normalized embedding vector; This indicates transposition; subsequently, binary cross-entropy loss is used. Constrain the consistency between the predicted adjacency matrix and the initial adjacency matrix: ; in, This represents the predicted adjacency matrix.

[0047] Step S35: By jointly reconstructing gene expression characteristics and spatial adjacency structures, a dual reconstruction loss is obtained. This allows the model to maintain both continuity of molecular expression and consistency of spatial structure. ; in, and These represent the weight coefficients for different reconstruction objectives, with default values ​​of 0.1 and 1 respectively.

[0048] Step S4: Based on the latent representation, construct a collaborative contrastive learning strategy, construct positive and negative samples according to spatial neighborhood relationships and cross-regional consistency, and construct an adaptive sampling mechanism to sample heterogeneous nodes in the transition region of the organizational structure and calculate the local-global collaborative contrastive loss.

[0049] Step S41: Latent representation based on the outputs of the online encoder and the target encoder in step S3. and First, Normalization is performed, and the cross-view cosine similarity matrix between nodes is calculated. : ; in, It is used to measure the potential semantic similarity between different spatial nodes, providing a basis for the subsequent construction of positive and negative samples.

[0050] Step S42: Considering that adjacent locations in spatial transcriptome data often have similar cell states and gene expression patterns, based on the cross-view cosine similarity matrix... Select Top- for each node Construct an adjacency matrix based on similarity representation of similar neighbors. : ; in, Represents the cross-view cosine similarity matrix The OK, The first two lines represent the highest similarity. A set of indexes for each node.

[0051] Subsequently, the spatial proximity relationship and the physical spatial adjacency relationship are jointly constrained to form a local positive sample set. ,in This represents element-wise matrix multiplication. Only node pairs that simultaneously satisfy spatial proximity and semantic similarity are retained as local positive samples to reduce erroneous associations caused by noise.

[0052] Step S43: Construct a set of positive samples. Since some spatial regions with the same biological function may be physically far apart, relying solely on local spatial relationships is insufficient to capture long-distance semantic consistency. Therefore, a target encoder representation is used... Perform multiple rounds of clustering and utilize cluster consistency to construct global semantic relationships.

[0053] Specifically, different random initialization methods are used for execution. Secondary clustering yields multiple sets of pseudo-labels. ,in Indicates the first The partitioning results table of the secondary clustering is used to construct a consensus matrix based on the consistency of node labels in different clustering results. : ; in, , They represent the first In round clustering, nodes , The cluster number to which it was assigned.

[0054] By combining the neighborhood relationships in the feature space, a global positive sample matrix is ​​constructed as follows: Finally, the local positive samples and global positive samples are fused together. For nodes Its positive sample set is defined as ,in Represents the first positive sample in the global positive sample matrix. Line 1 Column elements. By integrating local spatial continuity and global semantic consistency, the reliability of positive sample selection is improved.

[0055] Step S44: To address the issues of mixed cell states and blurred spatial transitions in the boundary regions of tissue structures, calculate the boundary score for each node: ; in, Represents the total number of neighbors. , Representing nodes respectively , Stable pseudo-labels, Represents a node Nearest neighbor set in feature space This indicates the indicator function. Furthermore, the node sampling weights are defined as follows: , This is a boundary weight adjustment parameter; the default value is 1. Represents a node The boundary score is calculated. During training, anchor point sampling is performed based on node weights, giving nodes in boundary regions a higher sampling probability, thereby enhancing the model's ability to learn complex spatial boundaries.

[0056] Step S45: To amplify the representational differences between different spatial domains, for each anchor node... Construct a negative sample candidate set as Based on the cross-view similarity matrix Select Top- from the negative sample candidate set Highly similar nodes: ; in, This indicates selecting Top- from the negative sample candidate set. Set of highly similar nodes Represents the first cross-view similarity matrix Line 1 Column elements.

[0057] Subsequently, a similarity threshold was applied. Filter out potential false negative samples and select samples that meet the constraints as semi-hard negative samples. , The default value is 0.95. For each positive sample... Training triples were constructed. .

[0058] Step S46: For the constructed training triples, in the online encoder representation space The distance between anchor nodes, positive samples in the positive sample set, and semi-difficult negative samples is calculated, and the interval constraint loss (local-global collaborative contrast loss) is applied. optimization: ; in, This represents the positive and negative sample interval parameter. , , These represent the anchor nodes output by the online encoder, the positive samples in the positive sample set, and the features of the semi-difficult negative samples, respectively. This represents the hinge loss function. Through the aforementioned local and global collaborative comparative learning mechanism, the model simultaneously maintains spatial neighborhood continuity and cross-regional semantic consistency, and combines boundary-aware sampling to enhance its ability to identify heterogeneous regions and ambiguous boundaries within an organization.

[0059] Step S5: Jointly optimize the total training loss function, divide the organization region through clustering methods, and carry out downstream tasks using latent representations.

[0060] Step S51: Construct a multi-objective joint optimization function. To simultaneously constrain gene expression recovery ability, spatial topology preservation ability, cross-view representation consistency, and spatial domain discrimination ability, the dual reconstruction loss, cross-view representation alignment loss, and local-global collaborative contrast loss are jointly optimized. The overall objective function is defined as follows: : in, This represents the dual reconstruction loss, used to constrain gene expression feature recovery and spatial topology preservation; This represents the alignment loss across views, used to enhance semantic consistency between the in-line encoder and the target encoder; This represents the local-global co-contrast loss, used to improve the representational discriminative ability between different spatial domains; and These represent the weighting coefficients of the corresponding loss terms, set to 2 and 0.1 respectively.

[0061] Step S52: Minimize the joint objective function using a gradient optimization method to learn a latent representation that integrates molecular expression information, spatial topological information, and biological semantic information. Based on the spatial representation obtained through training, spatial domain identification is achieved through clustering, and further downstream tasks such as gene denoising, functional enrichment analysis, and intercellular communication analysis are carried out.

[0062] To verify the effectiveness of this method, this embodiment systematically evaluates its performance in terms of spatial domain identification, gene expression denoising, and cancer heterogeneity analysis.

[0063] 1. Data set and evaluation metrics.

[0064] (1) Data set.

[0065] To comprehensively validate the applicability of this method across different tissue types, species, and spatial transcriptome sequencing platforms, multiple publicly available spatial transcriptome datasets were selected for experimental verification. The datasets used cover various spatial transcriptome technology platforms, including 10x GenomicsVisium and Stereo-seq, and encompass diverse tissue types such as human breast cancer, mouse brain tissue, mouse embryos, and human melanoma. These datasets exhibit significant differences in spatial resolution, gene capture efficiency, tissue heterogeneity, and sequencing technologies, and can be used to comprehensively evaluate the robustness and generalization ability of this method in resolving complex spatial structures.

[0066] (2) Evaluation indicators.

[0067] For datasets with manually labeled spatial regions, the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) are used to evaluate the consistency between the spatial domain recognition results and the actual organized regions. Higher ARI and NMI values ​​indicate stronger consistency between the predicted spatial domain and the manually labeled regions, suggesting higher spatial domain recognition accuracy. For datasets lacking manually labeled information, the Silhouette Coefficient (SC) and Davies-Bouldin Index (DB) are used to evaluate the quality of clustering results. Higher SC values ​​indicate more compact intra-cluster samples and better separation between different categories, while lower DB values ​​indicate higher discriminative power and a more stable clustering structure between different spatial domains.

[0068] 2. Comparative analysis of experimental results.

[0069] To verify the effectiveness of this method, this embodiment systematically evaluates its performance in terms of spatial domain identification, gene expression denoising, and cancer heterogeneity analysis.

[0070] (1) Spatial domain recognition experiment.

[0071] To verify the spatial domain recognition capability of this method in different tissue types and spatial transcriptome platforms, experiments were conducted on multiple spatial transcriptome datasets, including human breast cancer, mouse brain tissue, mouse embryos, mouse visual cortex, and human melanoma, and the results were compared with existing representative methods.

[0072] a. Human breast cancer dataset.

[0073] First, the spatial domain identification capability of this method was validated on the human breast cancer (HBC) spatial transcriptome dataset. For example... Figure 2As shown, the dataset contains histological images, manually annotated regions, and multiple spatial regions with different biological characteristics.

[0074] Experimental results show that this method can accurately identify the main spatial structures in tissues and obtain spatial partitioning results that are highly consistent with manually labeled areas. In comparison with several existing methods, this method demonstrates superior performance on both ARI and NMI evaluation metrics, indicating that it can fully utilize gene expression information and spatial topological information to improve the accuracy of identifying complex tissue structures.

[0075] From the spatial domain visualization results, this method can maintain the continuity of tissue regions and form clear boundaries between different spatial regions. Compared with other methods, this method shows better region segmentation ability in complex structures such as normal tissue regions, ductal carcinoma in situ regions, and invasive ductal carcinoma regions.

[0076] b. Mouse brain tissue dataset.

[0077] Furthermore, the method's ability to resolve complex hierarchical tissue structures was validated using anterior and posterior spatial transcriptome datasets of mouse brain tissue. Figure 4 As shown, this mouse anterior brain dataset contains artificially labeled brain region structures as true labels for spatial domain identification.

[0078] Experimental results show that this method can accurately recover multiple spatial regions in mouse brain tissue and achieve good recognition results on ARI and NMI evaluation indicators. Compared with other comparative methods, this method can more clearly distinguish different brain layer structures and reduce erroneous mixing between spatial regions.

[0079] Meanwhile, analysis of brain region-specific marker gene expression revealed that the spatial regions identified by this method highly corresponded to known brain tissue structures. Further functional enrichment analysis showed that different spatial regions possessed significantly different functional characteristics, validating the effectiveness of this method in the analysis of complex neural tissues.

[0080] To further validate the generalization ability of the method in different brain regions of mice, experiments were conducted on a mouse posterior spatial transcriptome dataset. Figure 5 As shown, this method can effectively identify the spatial structure of different brain regions, obtaining relatively continuous and clear spatial domain division results, and achieving good recognition performance on the SC and DB evaluation indicators. Compared with other methods, this method has better discriminative ability at tissue boundaries and can effectively reduce the mixing between adjacent regions, indicating that it has good spatial domain recognition ability and tissue structure resolution ability.

[0081] c. Mouse embryo dataset.

[0082] To verify the applicability of this method in developmental tissues, further experiments were conducted on the mouse embryo E9.5 spatial transcriptome dataset. Figure 6 As shown, this method can effectively identify multiple spatial regions in embryonic tissue and maintains a high degree of consistency with the results of manual annotation.

[0083] Compared with other methods, this method provides more continuous spatial region divisions and can more accurately recover tissue boundaries. Furthermore, combined with regional characteristic gene expression analysis, this method effectively captures molecular expression differences between different developmental regions, indicating its applicability to the analysis of developmental tissues with high spatial heterogeneity.

[0084] d. Mouse visual cortex and human melanoma datasets.

[0085] The generalization ability of this method was further validated on a human melanoma dataset. Figure 7 As shown, in the human melanoma dataset, this method can accurately distinguish different spatial structures such as tumor regions and lymph nodes. This result indicates that this method is not only suitable for the analysis of normal tissue structures, but can also be effectively applied to the identification of spatial regions in complex pathological tissues.

[0086] (2) Gene expression denoising analysis.

[0087] To further verify the ability of this method to handle spatial transcriptome noise, the original gene expression results and the expression results after denoising by this method were compared in multiple datasets.

[0088] Experimental results show that the spatial gene expression patterns processed by this method are more continuous, and it can effectively reduce the impact of random noise and low expression fluctuations on spatial structure recognition. Figure 3 and Figure 6 As shown, for representative spatial marker genes, the expression distribution after noise reduction is clearer and can more accurately reflect the expression differences between different tissue regions.

[0089] Furthermore, analysis of the expression changes of multiple characteristic genes in different spatial regions revealed that this method can maintain region-specific gene expression patterns while enhancing signal stability, indicating that the method can effectively achieve a balance between noise suppression and preservation of biological information.

[0090] (3) Cancer heterogeneity analysis experiment.

[0091] To further validate the capabilities of this method in resolving complex tumor microenvironments, cancer heterogeneity analysis was conducted on human breast cancer and human melanoma spatial transcriptome datasets. In the human breast cancer dataset, this method, building upon overall tissue structure identification, further elucidates fine-grained spatial heterogeneity within tumor regions. For example... Figure 3 As shown, this method can identify spatial differences between invasive ductal carcinoma (IDC), ductal carcinoma in situ (DCIS), and normal tissue regions, and further discover spatial subregions with different molecular characteristics within the same tumor region.

[0092] Characteristic gene expression analysis revealed significantly different gene expression patterns in different spatial regions. Further analysis, including differentially expressed gene analysis, violin plot visualization, and spatial expression distribution analysis, validated the method's ability to accurately capture molecular differences between different cellular states and functional regions within the tumor. Furthermore, GO functional enrichment analysis, KEGG pathway analysis, and intercellular communication analysis were conducted based on different spatial regions. The results showed significant differences in tumor-related signaling pathways, immune regulation, and cell interactions across different regions. For example, different spatial regions exhibited differential activation of communication pathways such as CXCL, further demonstrating the effectiveness of this method in revealing the spatial heterogeneity of the tumor microenvironment.

[0093] In human melanoma datasets, this method is also effective in distinguishing between tumor regions and lymph node regions. For example... Figure 7 As shown, GO functional enrichment analysis and KEGG pathway analysis revealed different functional enrichment patterns in different spatial regions, further validating the effectiveness of this method in resolving spatial heterogeneity in cancer tissues. In summary, the results of this embodiment demonstrate that this method can achieve accurate spatial domain identification in data from different species, tissue types, and spatial transcriptome sequencing platforms, while effectively reducing expression noise and providing in-depth analysis of spatial heterogeneity and potential biological functions in cancer tissues, exhibiting good robustness and generalization ability.

[0094] Example 2 A spatial transcriptome domain identification system, comprising: The multi-scale feature construction module is used to construct multi-scale feature representations that integrate molecular expression information and spatial context information based on gene expression information and spatial location information in spatial transcriptome data. The spatial adjacency graph and masked view construction module is used to construct an initial spatial adjacency graph based on spatial location information, construct a spatial topology graph using spatial location information and the initial spatial adjacency graph, and construct a complementary dual-masked view based on multi-scale feature representation, and enhance the information interaction between nodes by combining graph structure propagation mechanism; The dual encoder and reconstruction module is used to learn latent representations based on complementary dual-masked views through the dual encoder structure, construct cross-view representation alignment loss on the latent embeddings of the dual encoder, and jointly reconstruct gene expression features and spatial adjacency structure in complementary dual-masked views based on the latent representations obtained by the online encoder. The collaborative contrastive learning module is used to construct a collaborative contrastive learning strategy based on latent representations. It constructs positive and negative samples according to spatial neighborhood relationships and cross-regional consistency, and constructs an adaptive sampling mechanism to sample heterogeneous nodes in the transition region of the organizational structure and calculate the local-global collaborative contrastive loss. The joint training and clustering module is used to jointly optimize the total training loss function, divide the organization region through clustering methods, and carry out downstream tasks using latent representations.

[0095] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0096] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0097] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0098] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0099] Therefore, this invention employs the aforementioned spatial transcriptome domain identification method, system, device, and medium to obtain a feature representation that integrates molecular expression and spatial context information through multi-scale feature construction, and to simultaneously retain molecular expression information and spatial structure information through dual-mask collaborative reconstruction learning. Furthermore, it achieves sampling of heterogeneous nodes in the transition region of tissue structure through local and global collaborative contrastive learning.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for spatial transcriptome domain identification, characterized in that, Includes the following steps: Step S1: Based on gene expression information and spatial location information in spatial transcriptome data, construct a multi-scale feature representation that integrates molecular expression information and spatial context information; Step S2: Construct an initial spatial adjacency graph based on spatial location information, construct a spatial topology graph using the spatial location information and the initial spatial adjacency graph, and construct a complementary double-mask view based on multi-scale feature representation, and enhance information interaction between nodes by combining graph structure propagation mechanism; Step S3: Based on the complementary dual-masked view, learn the latent representation through the dual encoder structure, construct the cross-view representation alignment loss for the latent embedding of the dual encoder, and jointly reconstruct the gene expression features and spatial adjacency structure in the complementary dual-masked view based on the latent representation obtained by the online encoder. Step S4: Based on the latent representation, construct a collaborative contrastive learning strategy, construct positive and negative samples according to spatial neighborhood relationships and cross-regional consistency, and construct an adaptive sampling mechanism to sample heterogeneous nodes in the transition region of the organizational structure and calculate the local-global collaborative contrastive loss. Step S5: Jointly optimize the total training loss function, divide the organization region through clustering methods, and carry out downstream tasks using latent representations.

2. The spatial transcriptome domain identification method according to claim 1, characterized in that, Step S1 specifically includes: Step S11: Tokenize the spatial transcriptome gene expression matrix by dividing the gene dimension into several non-overlapping subspaces and converting them into three-dimensional token representations as gene token features. Step S12: For the obtained 3D Token representation, use a one-dimensional convolution operation on the Token sequence dimension to extract local dependencies and obtain local convolution features; Step S13: Construct a multi-frequency spatial location code based on spatial location information. Use multi-frequency sine and cosine functions to encode and splice the horizontal and vertical spatial directions respectively to obtain the multi-frequency spatial location code. Step S14: Concatenate the gene token features, local convolution features, and multi-frequency spatial location codes along the feature dimensions, and then reduce the dimensionality through pooling and principal component analysis to obtain a low-dimensional multi-scale feature representation.

3. The spatial transcriptome domain identification method according to claim 1, characterized in that, Step S2 specifically includes: Step S21: Based on spatial location distance, use The nearest neighbor method constructs an initial spatial adjacency graph, and after adding self-loops to the initial adjacency matrix, it performs symmetric normalization. Step S22: Divide the spatial node set based on random sampling, set the mask ratio, and randomly select the mask node set and the visible node set; Step S23: For the online encoding branch, replace the mask nodes with learnable mask vectors while preserving the original features of the non-mask nodes; Step S24: For the target encoder, a masking strategy opposite to that of the online coding branch is adopted, that is, the features of the occluded nodes in the online coding branch are restored, while the other part of the nodes are masked to form a complementary double mask view.

4. The spatial transcriptome domain identification method according to claim 3, characterized in that, The mask ratio is set to 0.4~0.

7.

5. The spatial transcriptome domain identification method according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Input the online view and the target view in the complementary dual-mask view into the online encoder and the target encoder respectively. Both the online encoder and the target encoder include a multilayer perceptron and a graph convolutional network. The parameters of the target encoder are updated by the exponential moving average of the parameters of the online encoder. Step S32: Map the latent representation obtained by the online encoder to the prediction space through an asymmetric projection network, replace the representation corresponding to the mask node with a learnable placeholder vector, and obtain the prediction representation by combining the graph projection module. With the output of the target encoder as the supervision target, calculate the cross-view representation alignment loss. Step S33: Using the online branch prediction representation and the normalized adjacency matrix, recover the gene expression features of the mask nodes through the graph decoder, and calculate the expression reconstruction loss; Step S34: Calculate the connection probability between nodes using the output of the online encoder, and predict the consistency between the adjacency matrix and the initial adjacency matrix by constraining the binary cross-entropy loss. Step S35: By jointly reconstructing gene expression characteristics and spatial adjacency structure, a dual reconstruction loss is obtained.

6. The spatial transcriptome domain identification method according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Normalize the latent representations output by the online encoder and the target encoder, and calculate the cross-view cosine similarity matrix; Step S42: Select Top-K similar neighbors for each node based on the cross-view similarity matrix, construct an adjacency matrix based on the representation of similarity, and jointly constrain it with the initial spatial adjacency matrix, retaining node pairs that simultaneously satisfy spatial proximity and semantic similarity as local positive samples; Step S43: Perform multi-round clustering based on the target encoder output representation, construct a consensus matrix based on the consistency of node labels in different clustering results, construct a global positive sample matrix by combining the neighborhood relationship in the feature space, and merge the local positive samples and global positive samples into a positive sample set. Step S44: Calculate the boundary score of each node based on the consensus matrix and define the node sampling weight; Step S45: Select anchor nodes according to node sampling weights. For each anchor node, construct a negative sample candidate set. Select highly similar nodes according to the cross-view similarity matrix and filter the samples in the negative sample candidate set through a similarity threshold to obtain semi-difficult negative samples. Step S46: For the training triples, calculate the distance between the anchor node, the positive sample in the positive sample set, and the semi-difficult negative sample in the online encoder representation space, and optimize using the margin constraint loss.

7. The spatial transcriptome domain identification method according to claim 1, characterized in that, Step S5 specifically includes: Step S51: Jointly optimize the cross-view representation alignment loss, dual reconstruction loss, and local-global collaborative contrast loss to construct a multi-objective joint optimization function; Step S52: Minimize the multi-objective joint optimization function using gradient optimization method, learn the latent representation, and realize spatial domain identification through clustering based on the latent representation, as well as carry out downstream tasks such as gene denoising, functional enrichment analysis and intercellular communication analysis.

8. A spatial transcriptome domain identification system, characterized in that, A method for implementing the spatial transcriptome domain identification method as described in any one of claims 1-7 includes: The multi-scale feature construction module is used to construct multi-scale feature representations that integrate molecular expression information and spatial context information based on gene expression information and spatial location information in spatial transcriptome data. The spatial adjacency graph and masked view construction module is used to construct an initial spatial adjacency graph based on spatial location information, construct a spatial topology graph using spatial location information and the initial spatial adjacency graph, and construct a complementary dual-masked view based on multi-scale feature representation, and enhance the information interaction between nodes by combining graph structure propagation mechanism; The dual encoder and reconstruction module is used to learn latent representations based on complementary dual-masked views through the dual encoder structure, construct cross-view representation alignment loss on the latent embeddings of the dual encoder, and jointly reconstruct gene expression features and spatial adjacency structure in complementary dual-masked views based on the latent representations obtained by the online encoder. The collaborative contrastive learning module is used to construct a collaborative contrastive learning strategy based on latent representations. It constructs positive and negative samples according to spatial neighborhood relationships and cross-regional consistency, and constructs an adaptive sampling mechanism to sample heterogeneous nodes in the transition region of the organizational structure and calculate the local-global collaborative contrastive loss. The joint training and clustering module is used to jointly optimize the total training loss function, divide the organization region through clustering methods, and carry out downstream tasks using latent representations.

9. A computer device, comprising: Memory and processor; The memory stores a computer program, characterized in that when the processor executes the computer program, it implements the steps of the spatial transcriptome domain identification method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When a computer program is executed by a processor, it implements the steps of the spatial transcriptome domain identification method according to any one of claims 1-7.

Images