Spatial transcriptomics data clustering method based on multi-view information rebalance learning

Abstract

The invention provides a spatial transcriptome data clustering method based on multi-view information rebalance learning, and relates to the technical field of biological information, and the method comprises the steps: carrying out the preprocessing of spatial transcriptome data, and constructing a spatial adjacency matrix; constructing a modal specific hypergraph to obtain hypergraph features, performing adaptive fusion on the hypergraph features and spatial neighborhood features extracted by the graph convolutional network to obtain fusion features, and realizing information rebalance in the modal; constructing a cross-modal heterogeneous graph, and adaptively optimizing heterogeneous edge connection by using pseudo tags and fusion features to realize information weight balance between modals; and reconstructing gene expression data based on a ZINB loss function, reconstructing a spatial adjacency matrix by using a graph auto-encoder, and finishing spatial transcriptomics data clustering in combination with a joint loss function training model. According to the method and the device, the problem that node information in space transcriptomics data views and among the views is unbalanced is solved.

Description

Technical Field

[0001] This application relates to the field of bioinformatics, and in particular to a spatial transcriptomics data clustering method based on multi-view information rebalancing learning. Background Technology

[0002] Spatial domain identification is a fundamental task in spatial transcriptome data analysis, aiming to identify regions with spatial consistency in gene expression patterns and tissue structure. Early identification methods, such as k-means, Louvain, and Leiden, primarily relied on non-spatial clustering strategies. These methods, based solely on gene expression data, did not adequately consider the spatial relationships between adjacent cells, potentially leading to discontinuous identified domains. In contrast, spatial clustering methods incorporate the spatial dependence of gene expression, enabling a more accurate characterization of the relationship between gene expression and spatial location. Therefore, designing robust spatial clustering methods is crucial for a deeper understanding of how the spatial environment regulates gene expression within tissues.

[0003] In recent years, researchers have developed various spatial clustering methods, such as CCST, STAGATE, and GraphST, to improve spatial domain recognition performance. While these methods effectively combine spatial location and gene expression data to enhance recognition, they often overlook the complex texture information contained in histological images. Visual details such as cell arrangement and tissue structure can provide important supplementary information for spatial transcriptome data analysis. To fully utilize the multimodal information contained in spatial transcriptome data, some methods attempt to incorporate histological image features into the learning process, such as SpaGCN, DeepST, stLearn, and MUSE. By integrating histological images, these methods enhance the understanding of interactions between adjacent cells, thereby improving the interpretability of cell spatial distribution.

[0004] Despite some progress in spatial domain identification for multimodal fusion, several key issues remain to be addressed: whether consistent core information can be ensured among cell clusters within different modalities; and whether erroneous connections introduced by sensor errors or other factors can be effectively eliminated during cross-modal integration. Clearly, information flow remains unbalanced, both within and across modal nodes. Specifically, when points in a particular modality suffer from insufficient sequencing depth or other issues affecting information representation, the transmission of effective information is hindered, thus impacting the extraction of meaningful latent representations from the overall data.

[0005] Most existing methods construct spatial graphs based on location information to establish connections between points. However, in complex tissue samples, cells of the same type are not necessarily spatially adjacent, while non-local nodes may exhibit similar gene expression patterns. Therefore, constructing modality-specific feature maps provides a more robust framework for characterizing the relationships between points. However, gene nodes are susceptible to sequencing errors, while image nodes are often affected by noise such as optical distortion and staining artifacts. Traditional graph structures typically focus on point-to-point relationships, neglecting the interdependencies between multiple points. When a point experiences information imbalance due to insufficient sequencing depth, this imbalance impairs the representation of its neighboring nodes and propagates throughout the network. Furthermore, gene modalities possess high dimensionality, while image modalities contain rich spatial and morphological information. When attempting to fuse features from these two modalities, information imbalance at a point in one modality can affect the representation of the corresponding point in the other modality, thus hindering effective cross-modal learning and information complementarity, and impacting the accuracy of feature fusion. Summary of the Invention

[0006] The purpose of this invention is to address the problem of imbalanced node information within and between views in spatial transcriptomics data using existing spatial clustering methods, and to provide a spatial transcriptomics data clustering method based on multi-view information rebalancing learning.

[0007] The above-mentioned objective of this application is achieved through the following technical solution: S1: Preprocess the acquired spatial transcriptome data to extract histological image features and gene expression features, and construct a spatial adjacency matrix; the spatial transcriptome data includes histological images and gene expression data. S2: For the two modalities of gene expression and histological images, modality-specific hypergraphs are constructed respectively; hypergraph neural networks are used to learn hypergraph features within the modality, and hypergraph features are adaptively fused with spatial neighborhood features extracted by graph convolutional networks to obtain fused features, thereby achieving information rebalancing within the modality; S3: Construct a cross-modal heterogeneous graph, utilize pseudo-labels and fusion features to adaptively optimize heterogeneous edge connections, and achieve intermodal information rebalancing; S4: Reconstruct gene expression data based on the ZINB loss function, reconstruct the spatial adjacency matrix using a graph autoencoder, and train the model using a joint loss function to complete spatial transcriptomics data clustering.

[0008] Optionally, step S1 includes: The preprocessing steps include: Histological images are extracted using a pre-trained visual transformer to obtain the feature matrix. ,in Number of spots Image feature dimension; Seurat V3 was used to filter and normalize the gene expression data to obtain the feature matrix. ,in The number of highly variable genes; Constructing a spatial adjacency matrix ,in , For point With point The Euclidean distance between them This is a hyperparameter used to control weight decay.

[0009] Optionally, step S2 includes: For each point in each mode Identify the nearest neighbor in the original feature space. Find a neighboring point and connect it to its neighboring points to form a corresponding hyperedge. The resulting mode-specific hypergraph is derived from the adjacency matrix. express:

[0010] In the formula Point Belongs to the super-edge ; This represents the i-th point; Two hypergraph convolutional encoders are used to capture latent hypergraph features of two modalities. The learned gene expression hypergraph features and pixel hypergraph features are denoted as follows: and ; Adaptive fusion of modality-specific spatial features and hypergraph features was performed, as follows:

[0011]

[0012] in and These are learnable parameters used to balance spatial features and hypergraph features; and Spatial features representing modality-specific spatial characteristics; the final fusion features of gene modality and histological image modality are respectively and .

[0013] Optionally, step S3 includes: S31: Using two fully connected networks to fuse gene modalities Fusion features with histological image modalities Mapping to a shared space yields hierarchical representations of the two modalities. and The details are as follows:

[0014] in Corresponding to gene characteristics or pixel features That is, fusion features; and Parameters representing a fully connected network, correspond and ; S32: Application of fully connected neural network fusion and The details are as follows:

[0015] In the formula, and For parameters of a fully connected network, It represents a comprehensive feature representation of gene modalities and image modalities; Apply the following constraints to enhance and Consistency between them:

[0016] In the formula, and They represent and The normalized matrix; Represents the consistency loss function; S33: Capture using the structure-aware module , and latent low-frequency information to obtain features , and The details are as follows:

[0017] in It is a scaling factor. , and It is the weight transformation matrix; Corresponding features , and ; Corresponding features , and ; S34: Features , and The integration is performed to obtain the integrated feature Q; S35: Utilize the clustering assignment results of the integrated feature Q as pseudo-labels for samples to optimize edge connections in heterogeneous graphs and eliminate spurious negative pairs, as detailed below:

[0018] in, This represents the cosine distance between two samples from two views; Representing gene modality The Each sample represents; Representing image modalities The Each sample represents;

[0019] in The loss function for adaptive heterogeneous graph reconstruction; This represents a temperature parameter used to adjust the distribution of similarity scores. Represents the similarity function; Indicates an indicator function; This represents a contrast loss term calculated between two modes; This indicates the number of spots in the gene modality and the image modality. Indicates the number of different modes; the first The first sample and the first The pseudo-labels for each sample are as follows: and .

[0020] Optionally, step S4 includes: Reconstruct gene expression data using the ZINB loss function:

[0021] in, These represent the mean, dispersion, and dropout probability of gene expression data, respectively. This represents the input gene expression data; Represents the ZINB loss function; ) represents the logarithmic function.

[0022] Optionally, step S4 may also include: A graph autoencoder is used to reconstruct the spatial adjacency matrix. A function is designed to calculate the regularization loss between the reconstructed matrix and the adjacency matrix.

[0023] in, Represents the first element in the original spatial adjacency matrix. Line number Column elements; This represents the transpose of the integrated feature Q; Indicates the activation function; This represents the reconstructed spatial adjacency matrix; Represents the first element in the reconstructed spatial adjacency matrix. Line number Column elements; Indicates the number of samples; Indicates the regularization loss; The joint loss function for:

[0024] in , , and These are the weighting coefficients; Represents the consistency loss function; The loss function represents the adaptive heterogeneous graph reconstruction. This represents the ZINB loss function.

[0025] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a spatial transcriptomics data clustering method based on multi-view information rebalancing learning.

[0026] A computer-readable storage medium storing instructions that, when executed, perform a spatial transcriptomics data clustering method based on multi-view information rebalancing learning.

[0027] The beneficial effects of the technical solution provided in this application are: Clustering of spatial transcriptome data is achieved through multi-view information rebalancing learning. Modality-specific hypergraphs are constructed for both histological images and gene expression modalities. To capture the similarity between non-local points, a hypergraph neural network is used to model complex intra-modal structures, thereby alleviating the imbalance in information propagation in traditional feature maps. Simultaneously, to more comprehensively capture the relationships between points, graph convolutional networks are used to learn modality-specific spatial features, which are adaptively fused with the hypergraph representation to achieve intra-modal information fusion, thus completing cross-modal integration of image and gene data. A cross-modal adaptive heterogeneous graph is reconstructed, removing unnecessary or erroneous negative edges during the reconstruction process. This process effectively blocks the propagation of imbalanced information related to pseudo-labels, preventing the model from learning erroneous heterogeneous negative samples due to noise. Therefore, this model can more closely correspond similar biological information in the feature space, ensuring a more balanced information flow and achieving more robust cross-modal integration, thereby solving the potential imbalance problem caused by sequencing depth and modality differences. Attached Figure Description

[0028] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a model structure diagram in the embodiments of this application; Figure 2 This is a schematic diagram of AHGR in an embodiment of this application; Figure 3 This is a spatial domain detection result diagram from an embodiment of this application; Figure 4 These are experimental results figures from embodiments of this application; Figure 5 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0029] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0030] Embodiments of this application provide a spatial transcriptomics data clustering method based on multi-view information rebalancing learning.

[0031] Please refer to Figure 1 , Figure 1 This is a model structure diagram of a spatial transcriptomics data clustering method based on multi-view information rebalancing learning, as described in an embodiment of this application, including: S1: Preprocess the acquired spatial transcriptome data to extract histological image features and gene expression features, and construct a spatial adjacency matrix; the spatial transcriptome data includes histological images and gene expression data. S2: For the two modalities of gene expression and histological images, modality-specific hypergraphs are constructed respectively; hypergraph neural networks are used to learn hypergraph features within the modality, and hypergraph features are adaptively fused with spatial neighborhood features extracted by graph convolutional networks to obtain fused features, thereby achieving information rebalancing within the modality; S3: Construct a cross-modal heterogeneous graph, utilize pseudo-labels and fusion features to adaptively optimize heterogeneous edge connections, and achieve intermodal information rebalancing; S4: Reconstruct gene expression data based on the ZINB loss function, reconstruct the spatial adjacency matrix using a graph autoencoder, and train the model using a joint loss function to complete spatial transcriptomics data clustering.

[0032] Step S1 includes: The preprocessing steps include: Histological images are extracted using a pre-trained visual transformer to obtain the feature matrix. ,in Number of spots Image feature dimension; Seurat V3 was used to filter and normalize the gene expression data to obtain the feature matrix. ,in The number of highly variable genes; Constructing a spatial adjacency matrix ,in , For point With point The Euclidean distance between them This is a hyperparameter used to control weight decay.

[0033] In one specific implementation of this application, in exploring the rebalancing of intra-view and inter-view information in spatial transcriptome data, complex biological patterns are refined and represented through preprocessing of histological images and gene expression data. For histological images, a pre-trained Vision Transformer (ViT) model extracts rich contextual features to generate a feature matrix. For gene expression data, Seurat V3 filters low-variance genes and retains high-variance genes to obtain better discriminative ability. This preprocessing reduces noise and enhances the stability of the data and the reliability of the analysis. Secondly, spatial location information is encoded by constructing a positional encoding matrix to represent the spatial adjacency between points. Assuming that closer points have higher similarity, the adjacency matrix transforms spatial proximity into the model's input, enhancing the understanding of spatial structure. and points The relationships between them are modeled as undirected weighted graphs, which serve as spatial inputs for subsequent learning. In one specific implementation of this application, meticulous preprocessing of tissue images and gene expression data extracts basic image features, constructing a robust gene feature matrix. Regarding spatial information processing, spatial location information is encoded to construct a corresponding adjacency matrix, explicitly capturing the spatial adjacency relationships between points. These preprocessing steps provide a solid foundation for subsequent intra- and inter-pattern information rebalancing learning, improving the model's performance in handling complex spatial transcriptome data.

[0034] Step S2 includes: For each point in each mode Identify the nearest neighbor in the original feature space. Find a neighboring point and connect it to its neighboring points to form a corresponding hyperedge. The resulting mode-specific hypergraph is derived from the adjacency matrix. express:

[0035] In the formula Point Belongs to the super-edge ; This represents the i-th point; Two hypergraph convolutional encoders are used to capture latent hypergraph features of two modalities. The learned gene expression hypergraph features and pixel hypergraph features are denoted as follows: and ; In one specific implementation of this application, in order to more effectively capture the complex relationships between spatial points in each modality and optimize the global coordination of points, modality-specific spatial features and hypergraph features are adaptively fused.

[0036] Adaptive fusion of modality-specific spatial features and hypergraph features was performed, as follows:

[0037]

[0038] in and These are learnable parameters used to balance spatial features and hypergraph features; and Spatial features representing modality-specific spatial characteristics; the final fusion features of gene modality and histological image modality are respectively and .

[0039] As one example, in spatial transcriptome data analysis, a spatial domain refers to a region exhibiting similar gene expression patterns across space. These regions are often associated with specific tissue types, cell populations, or biological functions. Traditional spatial domain identification tasks primarily focus on gene expression information at individual spatial points within a gene expression matrix, without explicitly considering the relative spatial positions between points. To further explore the spatial heterogeneity of spatial transcriptome data and considering that GCNs can capture complex spatial relationships between nodes, after constructing the spatial relationships in the spatial transcriptome data, a GCN encoder is used to propagate and aggregate gene expression and pixel features.

[0040] As one example, while spatial locations in spatial transcriptomics data effectively capture adjacency relationships between points, non-local points may have similar gene expression profiles due to interactions of functional state, transcriptomic features, and microenvironment. Furthermore, traditional modality feature maps focus on pairwise relationships, often neglecting the coordinated interactions between multiple points. Issues such as insufficient ranking depth at a single point can disrupt the representation of neighboring points and propagate information imbalance. Therefore, to mitigate the impact of local information imbalance on global information propagation, a hypergraph combining gene expression data and histological image features was constructed. To more effectively capture the complex relationships between spatial points in each modality and optimize global coordination of points, adaptive fusion of modality-specific spatial features and hypergraph features was performed.

[0041] As one example, by constructing modal hypergraphs for gene and tissue image modalities and learning their respective hypergraph structural features, followed by adaptive fusion with spatial features, high-order relationships and complex interactions between points are effectively captured. This method alleviates the typical point-pair information imbalance problem in traditional modal feature learning. In summary, the intramodal information rebalancing learning module refines the detection of spatial heterogeneity and complex relationships in multimodal data, thereby significantly improving the accuracy and robustness of spatial domain recognition.

[0042] Step S3 includes: S31: Using two fully connected networks to fuse gene modalities Fusion features with histological image modalities Mapping to a shared space yields hierarchical representations of the two modalities. and The details are as follows:

[0043] in Corresponding to gene characteristics or pixel features That is, fusion features; and Parameters representing a fully connected network, correspond and ; S32: Application of fully connected neural network fusion and The details are as follows:

[0044] In the formula, and For parameters of a fully connected network, It represents a comprehensive feature representation of gene modalities and image modalities; Apply the following constraints to enhance and Consistency between them:

[0045] In the formula, and They represent and The normalized matrix; Represents the consistency loss function; S33: Capture using the structure-aware module , and latent low-frequency information to obtain features , and The details are as follows:

[0046] in It is a scaling factor. , and It is the weight transformation matrix; Corresponding features , and ; Corresponding features , and ; As one example, by using structure-aware learning in the SM module, three different levels of latent representations are obtained, thereby enhancing the ability of modality to capture long-range feature dependencies.

[0047] S34: Features , and The integration is performed to obtain the integrated feature Q; S35: Utilize the clustering assignment results of the integrated feature Q as pseudo-labels for samples to optimize edge connections in heterogeneous graphs and eliminate spurious negative pairs, as detailed below:

[0048] in, This represents the cosine distance between two samples from two views; Representing gene modality The Each sample represents; Representing image modalities The Each sample represents;

[0049] in The loss function for adaptive heterogeneous graph reconstruction; This represents a temperature parameter used to adjust the distribution of similarity scores. Represents the similarity function; Indicates an indicator function; This represents a contrast loss term calculated between two modes; This indicates the number of spots in the gene modality and the image modality. Indicates the number of different modes; the first The first sample and the first The pseudo-labels for each sample are as follows: and .

[0050] As one example, after extracting fusion features from gene modalities and image modalities and performing cross-modal integration, the relationships between modalities are established as edges in a heterogeneous graph. Cluster assignments of the fusion features are used as pseudo-labels to optimize these edges. Figure 2As shown, when point A1 in the gene modality faces the problem of low sequencing depth, integrating it with point A2 in the imaging modality can minimize the distance between them and maximize their similarity. This allows the rich spatial and morphological information of point A2 to supplement the missing information of point A1, thereby enhancing the feature representation of point A1 and reducing the propagation of imbalanced information. When constructing heterogeneous negative sample edges with different instance points, the distance between different instance points is maximized, reducing the similarity between negative sample pairs. This method ensures that during the integration process, the information imbalance of the gene modality at site A1 will not lead to the incorrect construction of positive edges with site B2. However, since points A1 and C1 in the gene modality are highly similar due to their connection via hyperedges, they are likely to belong to the same cluster in the final clustering result. Therefore, if A1 and C2 are incorrectly selected as negative sample pairs, the model may push the original positive sample pairs further apart, potentially leading to a trivial solution. This will hinder the modality's ability to capture similar features within the same cluster, ultimately reducing cluster performance. Therefore, cluster assignments derived from the integrated feature Q are used as pseudo-labels for the samples. By iteratively eliminating false negative pairs with the same spatial clustering assignment, and continuously optimizing the selection of negative edges in cross-modal heterogeneous graphs, it is beneficial to learn more discriminative features.

[0051] As one example, the AHGR method is introduced to alleviate information imbalance during the integration of genetic and image modal features in multimodal information rebalancing. This method constructs a heterogeneous graph and utilizes pseudo-labels to guide the optimization of heterogeneous edges within the graph, ensuring that information from different modalities is globally coordinated and complementary. Ultimately, this multimodal information rebalancing method enhances the ability of modalities to capture cross-modal features, providing a more accurate and stable framework for multimodal data analysis tasks.

[0052] Step S4 includes: Reconstruct gene expression data using the ZINB loss function:

[0053] in, These represent the mean, dispersion, and dropout probability of gene expression data, respectively. This represents the input gene expression data; Represents the ZINB loss function; ) represents the logarithmic function.

[0054] Step S4 also includes: A graph autoencoder is used to reconstruct the spatial adjacency matrix. A function is designed to calculate the regularization loss between the reconstructed matrix and the adjacency matrix.

[0055] in, Represents the first element in the original spatial adjacency matrix. Line number Column elements; This represents the transpose of the integrated feature Q; Indicates the activation function; This represents the reconstructed spatial adjacency matrix; Represents the first element in the reconstructed spatial adjacency matrix. Line number Column elements; Indicates the number of samples; Indicates the regularization loss; The joint loss function for:

[0056] in , , and These are the weighting coefficients; Represents the consistency loss function; The loss function represents the adaptive heterogeneous graph reconstruction. This represents the ZINB loss function.

[0057] In one exemplary embodiment, the implementation details of the overall model are described as follows: The method was evaluated on 15 widely used spatial transcriptomics datasets. The first dataset is the Visium dataset, a 10x dataset of human breast cancer, containing 20 regions and 36,601 genes. The second dataset is mouse anterior brain tissue, annotated with 52 regions and containing 32,285 genes. The remaining dataset is from the LIBD human dorsolateral prefrontal cortex (DLPFC), consisting of 6 slices, each containing 33,538 genes and 5–7 regions.

[0058] Experiments were conducted on a PyTorch platform using a single NVIDIA GeForce RTX 3090 to evaluate the proposed MIRL. To validate the effectiveness of the clustering, ARI and NMI were used as evaluation metrics. The clustering results are as follows: Figure 4 As shown.

[0059] The layered organization of the cerebral cortex has complex relationships with its spatial structure and function, exhibiting significant differences in gene expression, morphology, and neural connectivity. To investigate this, a dataset consisting of 12 slices of the dorsolateral prefrontal cortex (DLPFC), annotated by Maynard et al., including six cortical layers and white matter (WM), was used. The performance of MIRL versus advanced spatial domain recognition methods was evaluated using this dataset.

[0060] First, MIRL was compared with 11 other representative methods based on the ARI and NMI values ​​of 12 slices, as shown in Table 1. MIRL outperformed all methods, achieving the highest or second-highest ARI and NMI in most slices. The performance varied significantly across different slices, indicating the impact of slice structural complexity on clustering results. Slice 151671, with five cortical layers, showed the best clustering performance. SCANPY, which relies solely on gene expression data while ignoring spatial information, performed the worst. In contrast, MIRL, DeepST, MAFN, and stMMR consistently demonstrated superior performance in spatial domain identification and were more sensitive to changes in specific domains.

[0061] To provide a more comprehensive evaluation of the clustering performance of different methods, a detailed analysis was conducted on the first slice (151507), as shown in Figure 1. The results show that MIRL and Spatial-MGCN outperform other methods in accurately depicting cortical boundaries and region alignment, achieving the highest ARI and NMI values. In contrast, K-means, CellPLM, and ScanPY struggle with layer boundaries, resulting in poor performance. stLearn, GraphST, and SCGDL show similar ARI and NMI scores but exhibit inconsistencies, particularly in layers 1, 3, and 4. DeepST and stMMR have difficulty distinguishing points between layers 4 and 5. While Spatial-MGCN aligns well with annotations, some outliers still exist. These findings were further validated through UMAP visualization, as shown in Figure 2. Figure 3 As shown in (D). It is worth noting that MIRL has distinct distinguishing points in different cortical layers, clearly reflecting the developmental sequence of these cortical layers and accurately reflecting the tissue structure and boundaries of the region.

[0062] In SRT data analysis, deciphering spatial gene expression patterns is crucial for understanding tissue function. High noise levels in the raw data make identifying layer-specific genes challenging. To evaluate the ability of MIRL to maintain layer marker gene consistency in DLPFC, the expression patterns of six layer-specific genes in slice 151507 before and after MIRL reconstruction were compared. Figure 3 (C)). Unlike the blurred patterns in the original data, MIRL shows clear spatial boundaries. Violin plot ( Figure 3 (E)) further demonstrates the significantly improved spatial expression pattern, highlighting the effectiveness of MIRL in depicting spatial domains and decoding gene expression.

[0063] In one specific embodiment of this application, the present invention proposes a spatial transcriptomics data clustering algorithm based on multi-view information rebalancing learning to solve the problem of imbalance of node information within and between views in spatial transcriptomics data. The main contributions of this research are summarized as follows: 1. A MIRL for SRT data clustering is proposed, which integrates intra-modal and inter-modal features end-to-end for spatial domain identification. To the best of our knowledge, the technical solution in this application is the first to simultaneously consider intra- and inter-modal information imbalances in SRT data analysis and introduce spatial clustering methods to address these challenges.

[0064] 2. An Adaptive Heterogeneous Graph Reconstruction (AHGR) method is proposed. This method establishes multimodal transport relationships by constructing a heterogeneous graph and uses pseudo-labels to adaptively optimize heterogeneous edge connections, thereby achieving a rebalancing of multimodal transport information flow. From this perspective, this work may provide new insights for the development of bioinformatics tasks in SRT data analysis.

[0065] 3. A series of experiments were conducted on various benchmark datasets to verify the effectiveness of the proposed method. The experimental results demonstrate the superiority of the proposed method.

[0066] This application also discloses an electronic device. (See reference...) Figure 5 , Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0067] The communication bus 502 is used to enable communication between these components.

[0068] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0069] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0070] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the above-described spatial transcriptomics data clustering method based on multi-view information rebalancing learning.

[0071] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0072] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A spatial transcriptomic data clustering method based on multi-view information rebalanced learning, characterized in that, The method comprises the following steps: S1: preprocessing the obtained spatial transcriptome data, extracting histological image features and gene expression features, and constructing a spatial adjacency matrix; the spatial transcriptome data comprises histological images and gene expression data; S2: constructing a modality-specific hypergraph for each of the gene expression and histological image modalities; learning hypergraph features within the modalities using a hypergraph neural network, adaptively fusing the hypergraph features with spatial neighborhood features extracted by a graph convolution network to obtain fused features, and achieving information rebalancing within the modalities; S3: constructing a cross-modality heterogeneous graph, adaptively optimizing heterogeneous edge connections using pseudo labels and fused features, and achieving information rebalancing between modalities; S4: reconstructing gene expression data based on a ZINB loss function, reconstructing a spatial adjacency matrix using a graph autoencoder, training the model using a joint loss function, and completing spatial transcriptome data clustering.

2. The method of claim 1, wherein the method is based on multi-view information rebalancing learning. Step S1 comprises: The preprocessing step comprises: extracting histological images using a pre-trained visual transformer to obtain a feature matrix wherein is the number of spots, is the image feature dimension; Filter and normalize gene expression data using Seurat V3 to obtain a feature matrix wherein is the number of high-variant genes; Constructing a spatial adjacency matrix where , is the Euclidean distance between points and points , is a hyperparameter that controls the weight decay.

3. The method of claim 1, wherein the method is based on multi-view information rebalancing learning. Step S2 comprises: For each point in each modality , the k nearest neighbors in the original feature space are identified , and the point is connected with the neighbors to form the corresponding hyperedge , resulting in a modality-specific hypergraph represented by the adjacency matrix ​ In the formula representing points belongs to the hyper-edge ; representing the i-th point; Two supergraph convolutional encoders are utilized to capture the latent supergraph features of two modalities, the learned gene expression supergraph features and the pixel supergraph features are denoted as and respectively; The spatial features and hypergraph features specific to the modalities are adaptively fused as follows: wherein and are learnable parameters for balancing spatial features and hypergraph features; and denote spatial features representing modality-specific spatial features; the final fused features for the gene modality and the histology image modality are and respectively.

4. The method of claim 1, wherein the method is based on multi-view information rebalancing learning. Step S3 comprises: S31: map the fused features of the gene modality and the histology image modality to a shared space using two fully connected networks, obtaining hierarchical representations of both modalities and and as follows:​ wherein corresponding to the gene feature or the pixel feature i.e. the fused feature; and representing the parameters of the fully connected network, corresponding to and ; S32: apply the fully connected neural network fusion and , as follows: wherein, and are parameters of the fully connected network, is a comprehensive feature representation of the gene modality and the image modality; The following constraints are applied to enhance the consistency between and ​ wherein and respectively represent and normalization matrices; represents a consistency loss function; S33: Capture the potential low frequency information of the structure using the structure perception module to obtain features , and , , and , in particular as follows: wherein is a scaling factor, , and are weight transformation matrices; corresponding features , and ; corresponding features , and ; S34: integrate the features , and to obtain an integrated feature Q; S35: using the clustering assignment results of the integrated features Q as the pseudo labels of the samples, optimizing the heterogeneous graph edge connections, and eliminating false negatives, specifically as follows: wherein, denotes the cosine distance between two samples from two views; denotes the gene modality th sample representation of the th sample representation of the denotes the image modality th sample representation of the th sample representation of the wherein is a loss function for adaptive heterogeneous graph reconstruction; denotes a temperature parameter for adjusting the distribution of similarity scores; denotes a similarity function; denotes an indicator function; denotes a contrastive loss term computed between two modalities; denotes the number of spots in the gene modality and the image modality, denotes the number of different modalities; the th sample and the th sample have pseudo labels and , respectively.

5. The method of claim 4, wherein the method is based on multi-view information rebalancing learning. Step S4 comprises: Using a ZINB loss function to reconstruct gene expression data: wherein, denote the mean, dispersion, and dropout probability of the gene expression data, respectively; denote the input gene expression data; denote the ZINB loss function; ) denotes the logarithm function.

6. The method of claim 4, wherein the method is based on multi-view information rebalancing learning. Step S4 further comprises: Using a graph autoencoder to reconstruct the spatial adjacency matrix, a function is designed to calculate the regularization loss between the reconstructed matrix and the adjacency matrix: wherein, represents the element in the original spatial adjacency matrix in the i-th row and j-th column; represents the element in the original spatial adjacency matrix in the i-th row and j-th column; represents the element in the original spatial adjacency matrix in the i-th row and j-th column; represents the transpose of the integrated feature Q; represents the activation function; represents the reconstructed spatial adjacency matrix; represents the element in the reconstructed spatial adjacency matrix in the i-th row and j-th column; represents the element in the reconstructed spatial adjacency matrix in the i-th row and j-th column; represents the element in the reconstructed spatial adjacency matrix in the i-th row and j-th column; represents the number of samples; represents the regularization loss; The joint loss function is: wherein 、 、 and are weight coefficients; denotes a consistency loss function; denotes a loss function for adaptive heterogeneous graph reconstruction; denotes a ZINB loss function.

7. An electronic device, comprising: The electronic device comprises a processor, a memory, a user interface, and a network interface, the memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform the spatial transcriptome data clustering method based on multi-view information rebalancing learning according to any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions, when the instructions are executed by a computer, the method for spatial transcriptome data clustering based on multi-view information rebalancing learning according to any one of claims 1-6 is executed.

Images