Spatial transcriptomic domain identification method with spatial relationship perception and cross-view semantic collaboration

Abstract

The application provides a spatial transcriptome domain identification method with spatial relationship perception and cross-view semantic cooperation, and belongs to the technical field of bioinformatics. The method comprises the following steps: obtaining spatial transcriptome data, constructing a graph structure of a structure view and a semantic view and learning node representations of the views, and obtaining initial representations through reconstruction constraints; constructing spatial relationship perception-based local representation learning based on the node representations of the structure view, and enhancing local representation consistency within and between the structure views; constructing cross-view semantic cooperation optimization based on the node representations of the structure view and the semantic view, establishing global semantic consistency constraints, and optimizing the representations of different views; and jointly optimizing the targets to obtain final node representations. The application can effectively support spatial domain identification and downstream analysis tasks, and improve identification accuracy.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method for spatial transcriptome domain identification based on spatial relationship perception and cross-view semantic collaboration. Background Technology

[0002] Spatial transcriptomics technology can simultaneously acquire gene expression information and its spatial location information in tissues, providing an important technical means for analyzing tissue spatial structure, cellular heterogeneity, and intercellular interactions. Spatial domain identification, as a key task in spatial transcriptomics data analysis, aims to comprehensively utilize gene expression information and spatial location information to accurately delineate different functional regions in tissues, providing crucial data support for tissue development, disease mechanism research, and precision medicine.

[0003] In recent years, graph neural networks have been widely applied to spatial domain recognition tasks. Existing methods typically construct graph structures based on spatial adjacency relationships, utilize graph neural networks to learn low-dimensional representations of nodes, and combine self-supervised learning or contrastive learning strategies to enhance local spatial consistency. Some methods construct multiple complementary views, employ multi-view graph convolutional networks to learn node representations for different views, and utilize attention mechanisms to achieve multi-view feature fusion, thereby improving the collaborative modeling ability of graph structure information and node attribute information. Furthermore, some methods utilize graph attention networks to adaptively learn the importance of neighboring nodes, enhancing the representation ability of local spatial structures by dynamically aggregating spatial neighborhood information.

[0004] Although existing methods have made some progress in spatial domain identification of the spatial transcriptome, the following shortcomings still exist: First, the ability to perceive and model spatial relationship information is insufficient, making it difficult to fully utilize the spatial neighborhood relationship between nodes to enhance the consistency of local representations; Second, the collaborative utilization of complementary information between multiple views is insufficient, and there is a lack of a global semantic alignment mechanism between semantic and structural views, resulting in inconsistencies in the representations of different views in the semantic space, which affects the accuracy of spatial domain identification results. Summary of the Invention

[0005] The purpose of this invention is to provide a spatial transcriptome domain identification method that combines spatial relationship perception and cross-view semantic collaboration. By constructing a structural view and a semantic view and learning the node representations of each view, and combining local representation learning for spatial relationship perception with cross-view semantic collaboration optimization, this method solves the problems of insufficient spatial relationship perception capability and inadequate multi-view semantic collaboration in existing methods, thereby improving the accuracy and robustness of spatial domain identification.

[0006] To achieve the above objectives, this invention provides a spatial transcriptome domain identification method based on spatial relationship awareness and cross-view semantic coordination, comprising the following steps: Step S1: Obtain spatial transcriptome data, construct the graph structure of structural view and semantic view and learn the node representation of each view, and obtain the initial representation that integrates spatial structure and gene expression information by reconstructing constraints; Step S2: Based on the node representation of the structure view, construct a spatial relationship-aware local representation learning, and use spatial neighborhood relationships to enhance the consistency of local representations within and between the structure view to obtain the local representation learning target; Step S3: Based on the node representations of the structure view and semantic view, construct cross-view semantic collaborative optimization, establish global semantic consistency constraints, and collaboratively optimize the representations of the structure view and semantic view to obtain the cross-view semantic collaborative optimization objective. Step S4: Jointly optimize the reconstruction constraints, local representation learning objectives, and cross-view semantic collaborative optimization objectives to obtain the final node representation, which is used for spatial domain recognition and downstream analysis tasks.

[0007] Therefore, the spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration described above has the following beneficial technical effects: (1) This invention constructs a structural view by building a spatial structure graph and enhancing the spatial structure graph, introduces a local representation learning module for spatial relationship awareness, and constructs a multi-positive sample comparison learning mechanism based on the spatial neighborhood relationship between nodes, including cross-view self-consistency, intra-view neighborhood consistency and cross-view neighborhood consistency, so that the local spatial structure representation inside and between the structural views remains consistent, which can enhance the perception and utilization of the spatial neighborhood relationship between nodes and improve the ability of node representation to express spatial topological relationship.

[0008] (2) This invention establishes a global semantic consistency constraint between the semantic view and the structural view, combines variance constraint and covariance constraint to reduce the difference in the representation of different views in the semantic space, and introduces a graph structure smoothing regularization term on the semantic view to keep the semantic representation of adjacent nodes smooth. This can promote the collaboration between the structural view and the semantic view and improve the fusion effect of complementary information from multiple views.

[0009] (3) The present invention obtains a node low-dimensional representation that integrates spatial structure information and global semantic information by jointly optimizing spatial structure reconstruction constraints, local representation learning objectives and cross-view semantic collaborative optimization objectives. It can be applied to downstream tasks such as spatial domain recognition, trajectory inference and cancer heterogeneity analysis. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the STWLCG process. Figure 2 The results are for spatial domain identification using STWLCG and other methods on the human dorsolateral prefrontal cortex dataset; among them, Figure 2(A) shows the performance comparison results of STWLCG with eight other methods (SpaGCN, SEDR, STAGATE, ConSpaS, GraphST, STMGraph, GAAEST, SLGCA) on 12 slices of DLPFC for ARI and NMI. Figure 2 (B) in the figure shows the spatial domain recognition visualization results of each method on the DLPFC 151673 slice; Figure 3 The results show the spatial domain identification of STWLCG and other methods on the mouse E10.5 embryo dataset; among them, Figure 3 (A) in the figure shows the performance comparison results of STWLCG with other comparative methods on the mouse embryo E10.5 dataset for ARI and NMI; Figure 3 (B) in the figure shows the spatial domain recognition visualization results of each method on the mouse embryo E10.5 dataset; Figure 4 The results are for spatial domain identification using STWLCG and other methods on a human breast cancer dataset; among them, Figure 4 (A) in the figure represents STWLCG compared with other methods in human breast cancer 10 Performance comparison results of ARI and NMI on the Visium dataset; Figure 4 (B) in the figure shows the spatial domain recognition visualization results of each method on the human breast cancer dataset; Figure 5 Results of visualization and trajectory inference for STWLCG and other methods on the human dorsolateral prefrontal cortex dataset; Figure 6 The results of STWLCG's genetic heterogeneity analysis on the human breast cancer dataset; among them, Figure 6 (A) in the figure is a spatial distribution visualization of spatial domain 7 (invasive ductal carcinoma region) and spatial domain 9 (ductal carcinoma in situ / lobular carcinoma in situ region) identified by STWLCG on the human breast cancer dataset; Figure 6 (B) in the figure is the pathway enrichment analysis diagram of spatial domain 7 and spatial domain 9 by STWLCG. Detailed Implementation

[0011] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0012] 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.

[0013] Example 1 like Figure 1As shown, the spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration includes the following steps: Step S1: Obtain spatial transcriptome data, construct graph structures of structural and semantic views and learn the node representations of each view, and obtain an initial representation that integrates spatial structure and gene expression information by reconstructing constraints.

[0014] Step S11: Based on the spatial coordinate information in the spatial transcriptome data, calculate the spatial distance between any two nodes using Euclidean distance, and then utilize... - The nearest neighbor strategy constructs spatial adjacency relationships for each node. Select the closest Establish edges connecting each neighbor to obtain the initial adjacency matrix. : ; in, The initial adjacency matrix The Middle Line 1 The specific element values ​​of the column, For nodes of The set of nearest neighbors Indicates the first 1 node This represents an undirected spatial view (hereinafter referred to as a spatial view).

[0015] Then the initial adjacency matrix is... Symmetry transformation is performed to obtain an undirected spatial structure graph. ,in This is the gene expression matrix after selecting the first 3000 highly variable genes.

[0016] Step S12: Based on the undirected spatial structure graph of step S11, process the initial adjacency matrix. Perform a random edge perturbation operation, randomly deleting some edges according to a preset edge discard ratio to generate an enhanced adjacency matrix. : ; in, For edge perturbation operators, The preset edge discard ratio is set to 0.05. This represents an enhanced spatial view (hereinafter referred to as an enhanced view); an enhanced spatial structure diagram is constructed by randomly changing local connectivity relationships. .

[0017] Step S13: Based on the gene expression matrix of the selected top 3000 highly variable genes, perform dimensionality reduction on the node expression features, calculate the expression similarity between nodes, and combine it with... - The nearest neighbor strategy is used to select candidate neighbors. A similarity threshold is then introduced based on this. =0.6, only retain connections with similarity exceeding the similarity threshold to construct a semantic graph. : ; in, Adjacency matrix of semantic view The Middle Line 1 The specific element values ​​of the column, For nodes of The set of nearest neighbors For nodes With nodes Gene expression similarity between them Represents a semantic view.

[0018] Step S14: Use a multi-view graph encoding mechanism to perform differentiated modeling of the structural view and semantic view: input the undirected spatial structural graph and the enhanced spatial structural graph into the shared encoder respectively. The corresponding spatial structure graph node representation is generated. and enhanced spatial structure diagram node representation Input the semantic graph into a separate encoder. Generate semantic graph node representations : ; encoder It is implemented using a single-layer GCN (Graph Convolutional Network), and its propagation rule is defined as follows: ; in, This is the node representation matrix output by the graph convolutional network (GCN) encoder. To add the adjacency matrix after adding self-loops, The initial adjacency matrix, It is the identity matrix. For degree matrix, To represent the original input gene expression matrix, For a trainable parameter matrix, The ReLU activation function is used. Simultaneously, the feature dimensions represented by the output nodes of both encoders are constrained to remain consistent; a shared encoder is used. The network parameters are updated and optimized jointly by the spatial view and the augmented view branches; independent encoder Configure a set with Independent and non-shared trainable parameters enable the extraction of differentiated features from different views.

[0019] Step S15: Introduce reconstruction constraints into the branches of the undirected spatial structure graph. Based on the node representation of the spatial structure graph. Feature reconstruction and adjacency matrix reconstruction are performed using the following formulas: ; ; in, , These are the node feature matrix and the structural adjacency matrix reconstructed by the decoder, respectively. For decoder networks, It is the Sigmoid activation function. To transpose, the reconstruction result is mapped to Interval. Based on the above reconstruction results, a weighted reconstruction loss function is defined. : ; in, The total number of nodes. It is the Frobenius norm. The parameters contributing to the balance between feature reconstruction and structural reconstruction are set to fixed values. .

[0020] Step S2: Based on the node representation of the structure view, construct a spatial relationship-aware local representation learning, and use spatial neighborhood relationships to enhance the consistency of local representations within and between the structure view to obtain the local representation learning target.

[0021] Step S21: Node representation of the spatial structure diagram obtained in step S1 and enhanced spatial structure diagram node representation Based on this, a shared projection network is constructed to map the node representations of different views to a unified representation space, whose projection representations... , Calculate using the following formula: ; in, This is a shared projection head consisting of two fully connected network layers. After mapping, the projected representations corresponding to the spatial structure map and the enhanced spatial structure map are obtained respectively: ; Among them, symbols This indicates that the matrix representation is expanded into a list of nodes.

[0022] Step S22: Normalize the projected representation and calculate the similarity between nodes using temperature-scaled cosine similarity: ; in, For in view The next node With in view The next node Similarity score between them , They are nodes In view Projection representation and nodes In view The projection representation below, Represents cosine similarity. The temperature parameter is set to 0.5. The Euclidean distance between any two nodes is calculated based on their spatial coordinates. With nodes The spatial coordinates are respectively and Its Euclidean distance is And construct a continuous spatial relation weight function. : ; in, This is set to the average neighborhood distance to adapt to different data scales. Combined with the adjacency relationships in the spatial structure graph, a spatial relationship-weighted adjacency matrix is ​​obtained. This provides spatial relation constraints for local representation learning.

[0023] Step S23: Based on the spatial structure graph and the enhanced spatial structure graph, construct multiple positive sample relationships for each node, including: cross-view self-consistency. (i.e., the positive sample relationship formed between corresponding nodes in the spatial structure graph and the enhanced spatial structure graph); intra-view neighborhood consistency (i.e., the positive sample relationship formed between a node and its spatial neighboring nodes in the same spatial structure graph); cross-view neighborhood consistency (That is, the positive sample relationship formed between nodes in the spatial structure graph and their corresponding neighboring nodes in the enhanced spatial structure graph), where For matrix The Middle Line 1 The specific element values ​​of the column, For positive sample relationships within the view, The positive sample relationships between views are defined as follows: apart from the positive sample relationships mentioned above, all other nodes are treated as negative samples and participate in local representation learning.

[0024] Step S24: Based on the above spatial relationship constraints and multiple positive sample relationships, nodes... Learning objective function from local representation of the direction from spatial view to augmented view Calculate using the following formula: ; in, Represents the set of all node indices. For nodes and Similarity within the spatial view For nodes In spatial view and Similarity in enhanced view For nodes The similarity to itself in the spatial view. Considering the imbalance of neighborhood size among different nodes, a normalization factor is introduced to modify the local representation learning objective. The modified local representation learning objective... : ; Finally, symmetric optimization is performed on the objective functions for both the spatial structure graph and the enhanced spatial structure graph to obtain the final spatial relationship-aware local representation learning objective. : ; in, For nodes Learn the objective function from the local representation of the augmented view to the spatial view direction.

[0025] Step S3: Based on the node representations of the structure view and semantic view, construct cross-view semantic collaborative optimization, establish global semantic consistency constraints, and collaboratively optimize the representations of the structure view and semantic view to obtain the cross-view semantic collaborative optimization objective.

[0026] Step S31: Normalize the spatial structure graph node representation and semantic graph node representation obtained in step S1: , ,in , These are the normalized spatial structure graph node representation and the semantic graph node representation, respectively. The normalization function, through semantic consistency constraints, reduces the representational differences of the same node across different views, ensuring consistency between the spatial structure graph and the semantic graph within the shared semantic space. It is calculated using the following formula: ; in, For invariance loss, , They are nodes The normalized representation vector in the spatial view and semantic view.

[0027] Step S32: Introduce variance constraints to each feature dimension to prevent feature collapse in the node representation, ensuring sufficient distribution variance for each feature dimension and improving the discriminative ability of the node representation. Calculate using the following formula: ; in, For variance loss, Let be the vector dimension represented by the node. For view indexing, For the first The first view Standard deviation of dimensional features To achieve hinge loss, the standard deviation of each feature is forced to be at least 1 to ensure the discriminative power between different features.

[0028] Step S33: Represent the off-diagonal elements of the covariance matrix through constraint nodes to reduce the correlation between different feature dimensions and achieve feature decoupling. Calculate according to the following formula: ; in, For covariance loss, For the first The covariance matrix elements of the view embedding matrix.

[0029] Step S34: Combining the above constraints, construct a cross-view global semantic alignment objective function. Calculate using the following formula: ; in, , , To balance the hyperparameters of various losses, the default values ​​are 5, 5, and 1. A graph structure smoothing regularization term is introduced into the semantic view, based on the Laplacian matrix corresponding to the semantic graph. ,in The matrix is ​​a degree-diagonal matrix. Smoothing constraints are applied to the representations of adjacent nodes, calculated using the following formula: ; in, For Laplace smoothing loss, Let be the trace of the matrix.

[0030] Step S35: Jointly optimize the cross-view global semantic alignment objective function with the graph structure smoothing constraint to obtain the cross-view semantic collaborative optimization objective, calculated according to the following formula: ; in, The total loss function for the cross-view semantic collaborative optimization module. The hyperparameter controls the strength of semantic structure constraints and has a default value of 0.1.

[0031] Step S4: Jointly optimize the reconstruction constraints, local representation learning objectives, and cross-view semantic collaborative optimization objectives to obtain the final node representation, which is used for spatial domain recognition and downstream analysis tasks.

[0032] Step S41: Jointly optimize the reconstruction loss in step S1 The local representation learning objective in step S2 and the cross-view semantic collaborative optimization objective in step S3 Overall objective function Calculate using the following formula: ; in, , , To weigh the weighting factors of each loss term, they were set to 10, 0.5, and 0.5, respectively.

[0033] Step S42: The gradient optimization algorithm is used to iteratively optimize the overall objective function to obtain a low-dimensional node representation that integrates spatial structure information and global semantic information, and the low-dimensional node representation is applied to spatial domain identification and downstream analysis of spatial transcriptomics.

[0034] To verify the effectiveness of this method, the performance of this method was tested in three aspects: spatial domain identification, UMAP visualization, and cancer heterogeneity analysis.

[0035] 1. Data set and evaluation metrics.

[0036] (1) Data set.

[0037] Gene expression matrices, spatial location information, and histological images were obtained from three datasets from the 10X Visium and Stereo-seq platforms. The first dataset consisted of 12 slices from the dorsolateral prefrontal cortex (DLPFC) of three individuals acquired using the 10X Visium platform. Each slice contained 3498 to 4789 spots, capturing 33538 genes, with manual annotations covering 5 to 7 regions: the DLPFC layer and the white matter layer. The second dataset came from the Stereo-seq platform. The Stereo-seq technology's MOSTA (Mouse Organogenesis Spatiotemporal Transcriptome Atlas) aims to systematically map the spatiotemporal dynamics of gene expression throughout the development of a mouse embryo. Stereo-seq data from mouse embryos at E10.5 days were used here. This dataset contained 18408 spots and 25201 genes, with manual annotations covering 13 spatial domains. The third dataset is human breast cancer tissue from the 10x Visium platform, containing 3,798 points and 36,601 genes, manually annotated by SEDR into 20 regions based on H&E staining and pathological features.

[0038] (2) Evaluation indicators.

[0039] The most direct method for evaluating clustering performance is to quantify the consistency between different clustering results. The choice of specific evaluation metrics depends on whether the spatial dataset has known true labels. If the true labels of the dataset are known, the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) are used as metrics for clustering performance. ARI is calculated by comparing the consistency between true labels and predicted labels, taking into account the impact of random assignment. This index measures the similarity between two types of labels, with a value ranging from [-1, 1], where a higher value indicates higher clustering accuracy. NMI, based on information theory, measures the normalized mutual information between the clustering results and the true labels. As a normalized extension of mutual information (MI), NMI evaluates the amount of information shared between two variables and normalizes it to the range of [0, 1], facilitating interpretation and comparison between different datasets.

[0040] 2. Comparative analysis of experimental results.

[0041] (1) Spatial domain recognition experiment.

[0042] a. Human dorsolateral prefrontal cortex dataset.

[0043] This embodiment employs the proposed method STWLCG (Spatial relationship perception and Cross-view semantic collaboration for spatial transcriptomics domain recognition), which is first tested on 12 slices of DLPFC. The performance of STWLCG is compared with eight current state-of-the-art methods (SpaGCN, SEDR, STAGATE, ConSpaS, GraphST, STMGraph, GAAEST, and SLGCA). Figure 2 (A)). Experimental results show that STWLCG achieves the best performance in both ARI and NMI across 12 slices, with mean and median ARI values ​​of 0.604 and 0.594, respectively, and mean and median NMI values ​​of 0.691 and 0.688, respectively. Furthermore, STWLCG exhibits relatively small performance fluctuations across different slices, demonstrating good stability and cross-slice generalization ability.

[0044] To further explore the effectiveness of STWLCG in spatial domain recognition, a detailed analysis was conducted using slice 151673 of DLPFC as an example. The results are as follows: Figure 2 As shown in (B), STWLCG achieved the best spatial domain recognition results, with ARI and NMI values ​​of 0.660 and 0.742, respectively. In comparison, SLGCA (ARI=0.638, NMI=0.741) and GraphST (ARI=0.636, NMI=0.734) performed second best, while the remaining methods all exhibited varying degrees of interlayer aliasing, boundary blurring, or local misclustering, making it difficult to accurately recover the spatial hierarchy of the organization. The above results demonstrate that this method can effectively integrate spatial relationship information and global semantic information, improving the accuracy and stability of spatial domain recognition while maintaining the spatial topological continuity of the organization.

[0045] b. Mouse embryo E10.5 dataset.

[0046] To further validate the spatial domain recognition capability of this method in complex non-layered tissues, the spatial transcriptome dataset of mouse embryos E10.5 from the MOSTA platform was selected for experiments. Figure 3 STWLCG achieved the best recognition performance, with ARI and NMI reaching 0.399 and 0.552 respectively, both outperforming the second-best method STMGraph (ARI=0.351, NMI=0.550) and other comparative methods. Figure 3 (A)). The spatial domain identification results of different methods show that ( Figure 3In (B) of the study, SEDR and STAGATE exhibit significant boundary distortion and region misclassification in some tissue areas; GraphST and GAAEST generate numerous discrete noise points, affecting the continuity of the tissue structure; SLGCA can recover the main tissue contours, but still suffers from insufficient local continuity in areas such as the mesenchyme of the head; STMGraph can maintain the overall tissue structure well, but still exhibits boundary blurring in the junction area between the spinal cord and the surface ectoderm. In contrast, STWLCG can more accurately recover the spatial distribution characteristics of different tissues, maintaining high recognition accuracy and good spatial continuity even in complex tissue structures.

[0047] c. Human breast cancer dataset.

[0048] To further validate STWLCG's spatial domain recognition capability in complex tumor microenvironments, it was tested in human breast cancer 10 Experiments were conducted using the Visium spatial transcriptome dataset. STWLCG achieved the best recognition performance, with ARI and NMI reaching 0.616 and 0.702, respectively, both superior to the compared methods, indicating that this method can effectively improve the accuracy of spatial domain recognition in complex tumor tissues. Figure 4 (A) in the text. The spatial domain identification results show that ( Figure 4 In (B) of the results, STWLCG can accurately recover the complex spatial distribution of breast cancer tissue and clearly distinguish between ductal carcinoma in situ, invasive ductal carcinoma, normal tissue, and tumor margins. Its identification results show good consistency with the histological structure reflected in H&E stained images, with no obvious regional misclassification or noise interference. In contrast, STAGATE shows significant confusion at the tumor-normal tissue boundary; GAAEST produces many local misclassifications, resulting in poor structural continuity in some tumor regions; while SEDR, GraphST, STMGraph, and SLGCA can identify the main tissue regions, they still exhibit blurred boundaries or local misclassification at the tumor margin and at the boundary between ductal carcinoma in situ and invasive ductal carcinoma. These results further demonstrate that STWLCG can fully integrate spatial relationship information and global semantic information, maintaining high spatial domain recognition accuracy and good spatial structural continuity even in complex tumor microenvironments.

[0049] (2) Trajectory inference experiment.

[0050] To demonstrate that the learned features of STWLCG can be well generalized to other downstream tasks, a comparative trajectory inference experiment was further conducted on DLPFC. The clustering distribution was visualized using the Unified Manifold Approximation and Projection (UMAP), and the spatial hierarchy trajectory was reconstructed by calculating inter-domain distances using PAGA. Figure 5UMAP clustering methods such as SpaGCN, SEDR, ConSpaS, GraphST, and GAAEST exhibit varying degrees of inter-cluster overlap and boundary blurring. The PAGA topology generates numerous cross-regional connections, failing to accurately recover the hierarchical relationships of cortical tissue. STAGATE and STMGraph can recover some cortical hierarchical structures, but still suffer from missing local connections or cross-layer connections. While SLGCA demonstrates good clustering performance, its spatial domain boundary continuity still has room for improvement, and the recovery of some hierarchical relationships in the PAGA topology is not accurate enough. In contrast, our proposed method, STWLCG, forms a more compact spatial domain distribution with clear boundaries and good separation between spatial domains. The PAGA topology constructed based on the learned node representations can effectively recover the hierarchical organizational relationships that gradually transition from the superficial cortex to the deep white matter, showing high consistency with the actual anatomical structure of the tissue. Experimental results show that the low-dimensional node representations obtained by our method are not only suitable for spatial domain recognition but also effectively support downstream spatial transcriptome analysis tasks such as trajectory inference, demonstrating good representational and generalization capabilities.

[0051] (3) Cancer heterogeneity analysis experiment.

[0052] To further explore the application value of STWLCG in the analysis of functional heterogeneity of the tumor microenvironment, two representative spatial domains (spatial domain 7 and spatial domain 9) identified by STWLCG from the human breast cancer spatial transcriptome dataset were selected for analysis. Spatial domain 7 mainly corresponds to regions related to invasive ductal carcinoma, while spatial domain 9 mainly corresponds to regions related to ductal carcinoma in situ / lobular carcinoma in situ. Figure 6 Analysis of differentially expressed genes in the two spatial domains revealed distinct transcriptomic characteristics. Spatial domain 7 was primarily enriched in metabolism-related genes such as IER3, COX6C, and TF1; while spatial domain 9 was primarily enriched in extracellular matrix-related genes such as COL1A1, COL1A2, and POSTN, indicating that this method can effectively distinguish the molecular expression characteristics corresponding to different pathological regions. Further pathway enrichment analysis showed that spatial domain 7 was mainly enriched in metabolic pathways such as branched-chain amino acid degradation, fatty acid metabolism, and mitochondrial function; spatial domain 9 was mainly enriched in biological processes related to the tumor microenvironment, such as extracellular matrix remodeling, angiogenesis, and integrin signaling pathways. Different spatial domains exhibited significantly different functional enrichment characteristics, showing good consistency with their corresponding pathological tissue regions. These results demonstrate that STWLCG can not only accurately identify spatial domain structures in breast cancer tissues but also further reveal the molecular functional characteristics and tumor microenvironment heterogeneity corresponding to different spatial domains, providing reliable technical support for functional analysis of spatial transcriptomic data and research on tumorigenesis and development mechanisms.

[0053] 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.

[0054] Therefore, this invention employs the aforementioned spatial transcriptome domain identification method that combines spatial relationship perception with cross-view semantic collaboration. It enhances spatial neighborhood relationship perception through local representation learning, improves multi-view information fusion through cross-view semantic collaboration, and jointly optimizes to obtain node representations that integrate spatial structure and semantic information. This effectively supports spatial domain identification and downstream analysis tasks, and improves identification accuracy.

[0055] 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 spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration, characterized in that, Includes the following steps: Step S1: Obtain spatial transcriptome data, construct the graph structure of structural view and semantic view and learn the node representation of each view. Obtain the initial representation that integrates spatial structure and gene expression information by reconstructing constraints. The structural view includes undirected spatial structure graph and enhanced spatial structure graph. Step S2: Based on the node representation of the structure view, construct a spatial relationship-aware local representation learning, and use spatial neighborhood relationships to enhance the consistency of local representations within and between the structure view to obtain the local representation learning target; Step S3: Based on the node representations of the structure view and semantic view, construct cross-view semantic collaborative optimization, establish global semantic consistency constraints, and collaboratively optimize the representations of the structure view and semantic view to obtain the cross-view semantic collaborative optimization objective. Step S4: Jointly optimize the reconstruction constraints, local representation learning objectives, and cross-view semantic collaborative optimization objectives to obtain the final node representation, which is used for spatial domain recognition and downstream analysis tasks.

2. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 1, characterized in that, Step S1 specifically includes: Step S11: Based on the spatial coordinate information in the spatial transcriptome data, calculate the spatial distance between any two nodes using Euclidean distance, and then utilize... - The nearest neighbor strategy is used to construct spatial adjacency relationships and obtain an initial adjacency matrix; the initial adjacency matrix is ​​then symmetricized to obtain an undirected spatial structure graph; Step S12: Perform a random edge perturbation operation on the initial adjacency matrix, randomly delete some edge connections according to a preset edge discard ratio, generate an enhanced adjacency matrix, and construct an enhanced spatial structure graph based on the enhanced adjacency matrix; Step S13: Dimensionality reduction of node expression features based on gene expression matrix, calculation of expression similarity between nodes, and selection of candidate neighbors using k-nearest neighbor strategy; similarity threshold is introduced to retain only connections with similarity exceeding the similarity threshold, resulting in a semantic graph, which is then used as a semantic view. Step S14: Use a multi-view graph encoding mechanism to perform differentiated modeling of the structural view and the semantic view: input the undirected spatial structural graph and the enhanced spatial structural graph into the shared encoder respectively to generate the spatial structural graph node representation and the enhanced spatial structural graph node representation; input the semantic graph into the independent encoder to generate the semantic graph node representation; Step S15: Based on the spatial structure graph node representation, introduce reconstruction constraints into the undirected spatial structure graph branch, perform feature reconstruction and adjacency matrix reconstruction, and define a weighted reconstruction loss function.

3. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 2, characterized in that, In step S14, both the shared encoder and the independent encoder are implemented using a single-layer graph convolutional network, and the propagation rule is defined as follows: ； in, This is the node representation matrix output by the graph convolutional network encoder. The ReLU activation function is used. To add the adjacency matrix after adding self-loops, For degree matrix, This is the original input gene expression matrix. This is a trainable parameter matrix.

4. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 3, characterized in that, In step S15, the weighted reconstruction loss function is calculated according to the following formula: ； in, To reconstruct the weighted loss function, The total number of nodes. For node indexing, The parameters that contribute to balancing feature reconstruction and structural reconstruction. The initial adjacency matrix, , These are the node feature matrix and the structural adjacency matrix reconstructed by the decoder, respectively.

5. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 1, characterized in that, Step S2 specifically includes: Step S21: Based on the spatial structure graph node representation and the enhanced spatial structure graph node representation, construct a shared projection network to map the node representations of different views to a unified representation space; Step S22: Normalize the projection representation and calculate the similarity between nodes using temperature-scaled cosine similarity; calculate the Euclidean distance between any two nodes based on the spatial coordinates of the points, construct a continuous spatial relationship weight function, and combine it with the adjacency relationship of the spatial structure graph to obtain a spatial relationship weighted adjacency matrix, providing spatial relationship constraints for local representation learning; Step S23: Based on the spatial structure graph and the enhanced spatial structure graph, construct multiple positive sample relationships for each node, including: cross-view self-consistency, intra-view neighborhood consistency, and cross-view neighborhood consistency; except for the above positive sample relationships, all other nodes participate in local representation learning as negative samples; Step S24: Based on the above spatial relationship constraints and multiple positive sample relationships, construct a local representation learning objective function, and introduce a normalization factor to modify the local representation learning objective. Perform symmetrical optimization on the objective functions in both directions of the spatial structure graph and the enhanced spatial structure graph to obtain the final spatial relationship perception local representation learning objective.

6. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 5, characterized in that, In step S22, the continuous spatial relation weight function Calculate using the following formula: ； in, For nodes With nodes The Euclidean distance between them Set to the average of the neighborhood distances.

7. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic coordination according to claim 5, characterized in that, In step S23, the multiple positive sample relationship specifically includes: Cross-view consistency: Positive sample relationships formed between corresponding nodes in the spatial structure graph and the enhanced spatial structure graph; Intra-view neighborhood consistency: Positive sample relationships formed between a node and its spatial neighboring nodes in the same spatial structure graph; Cross-view neighborhood consistency: Positive sample relationships formed between nodes in the spatial structure graph and their corresponding neighboring nodes in the enhanced spatial structure graph.

8. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 1, characterized in that, Step S3 specifically includes: Step S31: Normalize the node representations of the spatial structure graph and the semantic graph. Constrain the representation differences of the same node in different views through semantic consistency, so that the spatial structure graph and the semantic graph remain consistent in the shared semantic space. Step S32: Introduce variance constraints to each dimension of features; Step S33: Introduce covariance constraints to each dimension of features; Step S34: Integrate semantic consistency constraints, variance constraints, and covariance constraints to construct a cross-view global semantic alignment objective function, and introduce a graph structure smoothing regularization term on the semantic view. Apply smoothing constraints to the representation of adjacent nodes based on the Laplacian matrix corresponding to the semantic graph. Step S35: Jointly optimize the cross-view global semantic alignment objective function with the graph structure smoothing constraint to obtain the cross-view semantic collaborative optimization objective.

9. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 8, characterized in that, In step S34, the cross-view global semantic alignment objective function as follows: ； in, For invariance loss, For variance loss, For covariance loss, , , Hyperparameters are used to balance various losses.

10. The spatial transcriptome domain identification method based on spatial relationship perception and cross-view semantic collaboration according to claim 1, characterized in that, Step S4 specifically includes: Step S41: Jointly optimize the reconstruction loss Local representation of learning objectives and cross-view semantic collaborative optimization objectives The overall objective function is calculated using the following formula: ； in, , , To weigh the weighting factors for each loss term; Step S42: The gradient optimization algorithm is used to iteratively optimize the overall objective function to obtain a low-dimensional node representation that integrates spatial structure information and global semantic information, and the low-dimensional node representation is applied to spatial domain identification and downstream analysis of spatial transcriptomics.

Images