Methods for spatial domain identification based on data interpolation and cell type deconvolution

Abstract

This invention proposes a method for spatial domain identification based on data interpolation and cell type deconvolution, belonging to the field of bioinformatics. This invention addresses the problems of traditional methods failing to utilize spatial gap information between adjacent points and insufficient integration of prior information about cell types at the tissue spatial structure level in low-resolution spatial transcriptome data. The method includes: acquiring spatial transcriptome datasets and single-cell RNA sequencing datasets, and preprocessing the acquired datasets; interpolating the preprocessed spatial transcriptome data and performing cell type deconvolution in conjunction with the single-cell RNA sequencing data; constructing a deep learning model based on a graph convolutional network; training the deep learning model using a self-supervised contrastive learning strategy based on gene expression information, spatial location information, and cell type information from the deconvolved spatial transcriptome data; and performing spatial domain identification on the test data based on the trained model.

Description

Technical Field

[0001] This invention relates to a method for spatial domain recognition based on data interpolation and cell type deconvolution, belonging to the field of bioinformatics technology. Background Technology

[0002] In recent years, spatial transcriptomics technology has developed rapidly, not only systematically characterizing cellular heterogeneity while preserving the spatial background of tissues, but also providing a powerful tool for accurately delineating spatial domains within tissues, thereby deepening our understanding of cell-environment interaction mechanisms. Currently, mainstream spatial transcriptomics technologies are mainly divided into two categories: (i) in situ hybridization or in situ sequencing methods: such as seqFISH, MERFISH, STARmap, and FISSEQ. These methods achieve cellular and even subcellular resolution by directly detecting pre-defined RNA targets in situ, but their multiplexing capabilities are limited, typically only detecting hundreds to thousands of genes; (ii) in situ capture techniques: such as spatial transcriptomics (ST), SLIDE-seq, ZipSeq, and HDST. These methods achieve unbiased analysis of the entire transcriptome by capturing transcripts in situ within tissues and then sequencing them in vitro. However, this point-based capture strategy has limitations: platforms like the 10X Visium have 4992 dots with a diameter of 55 μm printed on the slide, with a center-to-center distance of about 100 μm between the dots, resulting in approximately 70% of the regional gene expression data being lost. There is a significant information gap due to the lack of measurement. Traditional spatial domain identification methods mainly rely on the spatial proximity assumption, that is, dividing spatial domains by comparing the overall expression patterns between adjacent spots. Non-spatial clustering methods (such as K-means, Louvain, and Seurat) often fail to capture the continuity of tissues, and therefore often fail to fully reflect the actual spatial patterns of tissues. These methods rely solely on gene expression information, ignoring the spatial relationships between points, and cannot effectively identify the intrinsic structural features of tissues. To address this limitation, several improved methods incorporating spatial information have been proposed in recent years. For example, Giotto uses a Hidden Markov Random Field (HMRF) model to identify spatial domains with coherent gene expression patterns by fully utilizing the spatial dependencies between points; BayesSpace uses Bayesian statistical methods to integrate spatial neighborhood information to optimize clustering results; STAGATE and GAADE combine the GraphAttention Auto-Encoder framework to further integrate spatial information with gene expression profiles. In addition, studies have shown that histopathological images can effectively predict gene expression. Therefore, some methods (such as SpaGCN and stLearn) have improved spatial domain recognition by integrating neighborhood information and morphological features. DeepST goes a step further by combining image features, gene expression, and spatial localization.

[0003] Despite significant progress, these methods still face the challenge of fully exploiting the topological features of spatial relationships between cells or points in unlabeled data. To address this, some methods have introduced contrastive learning, which learns low-dimensional representations by leveraging the inherent structure and properties of the data, thus mitigating the challenges posed by limited labeled data. For example, conST integrates multimodal data for contrastive learning at the point, subgroup, and global levels; ConGI designs a joint learning strategy that includes gene expression, images, and contrastive loss functions between them; GraphST integrates spatial information with gene expression through graph self-supervised contrastive learning; MuCoST enhances point dependencies by fusing gene expression correlation and spatial proximity; and recently, STAIG combines graph contrastive learning and high-performance feature extraction to integrate gene expression, spatial coordinates, and histological images. However, while these methods have achieved breakthroughs in combining spatial information, histological images, and gene expression data, they still primarily rely on the spatial proximity assumption, which assumes that adjacent spots have similar gene expression patterns. While this assumption reveals to some extent the spatial distribution of gene expression and the regional characteristics of cell types within tissues, it does not fully utilize the cellular composition information within the gene expression data. Therefore, these methods may not be able to accurately capture the spatial distribution and functional characteristics of functionally relevant cell populations driven by specific biological processes, thus limiting the precise characterization of spatial domain details and biological functions. Summary of the Invention

[0004] To address the problem that traditional methods cannot utilize the spatial information between adjacent points and fail to fully integrate prior information about cell types at the tissue spatial structure level in low-resolution spatial transcriptomics, this invention proposes a method for spatial domain identification based on data interpolation and cell type deconvolution.

[0005] The technical solution adopted by the present invention to solve the above problems is as follows: The present invention includes the following steps:

[0006] Step 1: Obtain the spatial transcriptome dataset and single-cell RNA sequencing dataset, and perform data preprocessing on the obtained datasets;

[0007] Step 2: Perform data interpolation on the preprocessed spatial transcriptome data and combine it with single-cell RNA sequencing data to perform cell type deconvolution;

[0008] Step 3: Construct a deep learning model based on graph convolutional networks;

[0009] Step 4: Input the preprocessed spatial transcriptome data and single-cell RNA sequencing dataset into the deep learning model, and use a self-supervised contrastive learning strategy to train the deep learning model based on the cell type information obtained by cell type deconvolution combined with the gene expression information and spatial location information of the spatial transcriptome data.

[0010] Step 5: Perform spatial domain identification on the test data based on the trained model.

[0011] Furthermore, in step 1, the spatial transcriptome dataset includes the gene expression data of spot and the spatial location information of spot, and the single-cell RNA sequencing dataset includes the gene expression data of cell and cell type information. The gene expression data matrix in the spatial transcriptome dataset has a dimension of spot×gene and is arranged in a regular honeycomb pattern on the tissue slice, while the gene expression data matrix in the single-cell RNA sequencing dataset has a dimension of cell×gene.

[0012] Furthermore, the data preprocessing in step 1 specifically includes:

[0013] Step 1.1: Perform logarithmic transformation and normalization on the gene expression data of the spatial transcriptome dataset;

[0014] Step 1.2: Select a predetermined number of highly variable genes from the normalized spatial transcriptome data and single-cell RNA sequencing data to complete the preprocessing of the spatial transcriptome dataset and single-cell RNA sequencing dataset.

[0015] Furthermore, step 2 specifically includes:

[0016] Step 2.1: Divide the raw gene expression data in the preprocessed spatial transcriptome data into rows horizontally, and then divide each row according to the diagonal, dividing the complete gap region between the center point, left vertex and right vertex;

[0017] Step 2.2: Based on the original spatial coordinate data of gene expression data in the gap regions after division, calculate the centroid coordinates of the corresponding gap regions. The centroid coordinates are used as the spatial coordinates of the void region;

[0018] Step 2.3: Based on the original gene expression data and the spatial location information of the corresponding cells, the gene expression data of the divided gap regions are interpolated using the neighborhood interpolation method;

[0019] Step 2.4: Combine the preprocessed single-cell RNA sequencing data with cell type deconvolution of the interpolated gene expression data to infer the cell composition of each gene expression data sampling point;

[0020] Centroid coordinates of the void region The calculation formula is:

[0021]

[0022] In formulas (1) and (2), These are the coordinates of points within each segmented region. N The total number of points within the region;

[0023] The expression for gene expression data interpolation is:

[0024]

[0025] In formula (3), Spatial coordinates Gene expression at the site, These are coefficients calculated based on weighting factors such as spatial distance. This represents the gene expression values ​​at other points in the neighborhood.

[0026] Furthermore, the deep learning model based on graph convolutional networks constructed in step 3 includes:

[0027] The module includes a spatial adjacency graph construction module, a gene expression random graph construction module, an encoder-decoder module, and a self-supervised contrastive learning module.

[0028] The spatial adjacency graph construction module is used to construct a weighted adjacency graph from gene expression data after deconvolution based on cell type, and obtain spatial distance information between gene expression data and cell type similarity measurement;

[0029] The gene expression random graph construction module constructs a random graph by generating a raw neighborhood graph based on a weighted adjacency graph. G=(V,E) Add perturbations to generate a random gene expression graph. G'=(V',E') ;

[0030] The GCN-based encoder module is used to process the original neighborhood graph. G=(V,E) and randomized gene expression graph G'=(V',E') Modeling based on gene expression data;

[0031] The self-supervised learning comparison module is used to compare the original neighborhood graph. G=(V,E) and randomized gene expression graph G'=(V',E') Generate positive and negative sample pairs to optimize node representations and train the deep learning model.

[0032] Furthermore, step 4, training the deep learning model, includes the following steps:

[0033] Step 4.1: Input the deconvolutioned gene expression data of cell types into the spatial adjacency graph construction module to obtain the spatial distance information between gene expression data and cell type similarity measurement;

[0034] Step 4.2: Calculate the original neighborhood graph generated from the weighted adjacency graph. G=(V,E) In the random gene expression graph construction module, add perturbations to generate a random gene expression graph. G'=(V',E') ;

[0035] Step 4.3: Convert the original neighborhood graph G=(V,E) and randomized gene expression graph G'=(V',E') Input a GCN-based encoder module to model the gene expression data in the input graph;

[0036] Step 4.4: Convert the original neighborhood graph G=(V,E) and randomized gene expression graph G'=(V',E') Input the self-supervised learning contrastive module and use the self-supervised contrastive learning strategy to train the deep model.

[0037] Furthermore, step 4.1 specifically includes:

[0038] Step 4.1.1: Convert the deconvolutioned gene expression data of cell types into an undirected graph, which includes several nodes. V and edge E , among which, nodes V For all sampling points, the edges E This refers to the connection relationships between points;

[0039] Step 4.1.2: Calculate the Euclidean distance between sampling points based on the spatial coordinates of each point in the undirected graph to obtain the distance matrix;

[0040] Step 4.1.3: Using the K-nearest neighbor strategy, select the K nearest sampling points by Euclidean distance as its neighbor nodes to obtain a preliminary adjacency matrix. Where, if the points in the preliminary adjacency matrix... j It is a point i The value is 1 if it is a neighbor, and 0 if it is not, and the point is set to... j The original adjacency matrix is ​​added to its transpose, and values ​​greater than 1 are truncated to make all sampling point connections binary.

[0041] Step 4.1.4: Introduce cell type information, calculate the cell type similarity of any connected point pair in the preliminary adjacency matrix, and calculate the edge weight based on the cell type similarity and the basic weight of spatial information.

[0042] Furthermore, step 4.3 specifically includes:

[0043] Step 4.3.1: Based on the graph convolutional network and combined with the weighted adjacency graph, construct the gene expression data into a graph. ,in, V For a set of nodes, E For edge set, node set V Includes all sampling points, edge set E Indicates the connection relationship between nodes;

[0044] Step 4.3.2: Setting up the diagram middle Let be the node feature matrix, where For the number of nodes, d For each node's feature dimensions, A binary adjacency matrix between nodes;

[0045] Step 4.3.3: Normalize the adjacency matrix A to obtain the normalized adjacency matrix. ,in D Given a diagonal matrix, the diagonal elements are... , For nodes i The degree;

[0046] Step 4.3.4: Convert the normalized adjacency matrix The trainable weight matrix and bias terms of the corresponding layer of the graph convolutional network are input into the encoder to obtain the latent representation of the sampling points. ;

[0047] Step 4.3.5: Latent representation of the sampling points The input is fed into the decoder, and through the decoding process, it is mapped back to the original gene expression space by minimizing the self-reconstruction loss function of gene expression. Optimize the parameters of the deep learning model.

[0048] Furthermore, step 4.4 specifically includes:

[0049] Step 4.4.1: Create the original gene expression map G and random gene expression map G′ As input, two representation matrices are generated by an encoder based on a graph convolutional network. and ， The embedding representation of each node is generated by aggregating the neighbor information of each node in the matrix, thus obtaining the local context vector of each node. ;

[0050] Step 4.4.2: Construct positive and negative sample pairs, where the sampling points i matrix Its local context vector Forming positive sample pairs in a random graph G′ In the middle, sampling points i The representation of Its context vector Forming negative sample pairs;

[0051] Step 4.4.3: Construct the positive sample contrast loss function and negative sample loss function The quality of node representation is optimized by maximizing the mutual information between positive sample pairs and minimizing the mutual information between negative sample pairs.

[0052] Step 4.4.4: Loss function based on positive samples Negative sample loss function and self-reconstruction loss function Overall training loss function for constructing deep learning models Based on the overall training loss function The deep learning model is trained to obtain the latent representation of all sampling points.

[0053] Furthermore, step 5 specifically includes:

[0054] The mclust method is used to perform cluster analysis on the latent representations of all points generated during training to obtain spatial domain identification results for different cell tissues.

[0055] The beneficial effects of this invention are:

[0056] 1. This invention assumes that spatially adjacent spots with similar cell types are more likely to belong to the same spatial domain. It uses interpolation techniques to estimate the gene expression data of cells in the gaps between spots, and further extracts cell type information inside and between spots through deconvolution. SGCD (i.e. the method proposed in this invention) achieves accurate division of complex spatial domains through graph comparison learning based on the fusion of gene expression data, cell type information and spatial location information.

[0057] 2. This invention demonstrates that the deep learning model achieves excellent spatial domain recognition performance on both spatial transcriptome and single-cell RNA sequencing datasets. Compared with other state-of-the-art spatial domain recognition methods, the deep learning model exhibits superior performance across multiple evaluation metrics.

[0058] 3. This invention addresses the problem that traditional methods cannot utilize the gap information between adjacent points in low-resolution spatial transcriptomics (such as 10X Visium). It introduces a reasonable interpolation method to fill in the gene expression data in missing regions, improving the continuity of data acquisition. To address the issue that traditional methods fail to fully utilize cellular composition information, the interpolated data is fused and deconvolved to fully extract the cellular composition information of each point. This information is then combined with spatial location information to construct a weighted adjacency matrix. Furthermore, the parameter γ is used to dynamically adjust the weight balance between spatial proximity and cell type similarity, achieving modeling of their dual dependency relationship. Attached Figure Description

[0059] Figure 1 This is a flowchart of a method for spatial domain identification based on data interpolation and cell type deconvolution;

[0060] Figure 2 It is a diagram of a deep learning model;

[0061] Figure 3 This is a comparison and visualization of the spatial domain recognition performance of the proposed method on the DLPFC dataset with different methods. Figure 3 In the diagram, (A) is a comparison of the spatial domain accuracy of the proposed method with other methods in all 12 slices of the DLPFC dataset; (B) is a schematic diagram of the structural domain of the 10X Visium dataset; (C) is a schematic diagram of the spatial domain identification results of slice #151673; and (D) is a schematic diagram of the T-test statistical results of the proposed method with other methods.

[0062] Figure 4 This is a graph comparing the spatial domain recognition performance of the method proposed in this invention with that of different methods on a mouse brain dataset, and showing its application. Figure 4In the diagram, (A) is a schematic diagram of the distribution of the mouse brain dataset, and (B), (C) and (D) are schematic diagrams of the comparative test results of the method proposed in this invention with other methods;

[0063] Figure 5 This is a comparison and application graph showing the spatial domain recognition performance of the proposed method on the PDAC dataset with different methods. Figure 5 In the diagram, (A) is a schematic diagram of the analysis of PDAC-A slices, (B) is a schematic diagram of the cell type information extraction results of the PDAC dataset, and (C) is a schematic diagram comparing the clustering results of the method proposed in this invention with other methods.

[0064] Figure 6 This is a schematic diagram illustrating the spatial domain identification results of the method proposed in this invention in complex and heterogeneous cancer tissues with functional spatial domains. Figure 6 In the diagram, (A) is a schematic diagram of histological image segmentation; (B) is a test result diagram of the method proposed in this invention on 10X Visium data of human breast cancer tissue; (C) is a schematic diagram comparing the spatial domain identification results of the method proposed in this invention with other methods; (D) is a schematic diagram comparing the Jaccard index calculated by this invention with other methods in the IDC_5 region; (E) is a schematic diagram of the point distribution pattern of each spatial domain; (F) is a schematic diagram comparing the first five differentially expressed genes between spatial domain 5 representing the IDC region and spatial domain 8 representing the DCIS / LCIS region; (G) is [image of gene detection]; (H) is a schematic diagram of gene detection between spatial domain 5 and spatial domain 8; and (I) is a schematic diagram of the T-test statistical results of spatial domain identification of the method proposed in this invention with other methods. Detailed Implementation

[0065] Combination Figure 1-2 This implementation method is described as follows: Figure 1 As shown, the steps of the spatial domain recognition method based on data interpolation and cell type deconvolution described in this embodiment include:

[0066] The method for spatial domain identification using cell type deconvolution includes the following steps:

[0067] S1: Obtain spatial transcriptome datasets and single-cell RNA sequencing datasets;

[0068] In this embodiment, the acquisition of spatial transcriptome datasets and single-cell RNA sequencing datasets includes multiple spatial transcriptome datasets (such as spatial transcriptome datasets obtained through the 10X Visium platform) and single-cell RNA sequencing datasets. These datasets contain gene expression information, cell type information, and spatial location information, serving as the basis for subsequent analysis and model training. Spatial transcriptome data provides spatial information for spatial domain identification, while single-cell RNA sequencing data provides cell type information; the combination of the two supports accurate spatial domain delineation.

[0069] S2: Perform data preprocessing on the dataset;

[0070] In this embodiment, data preprocessing of the dataset includes data cleaning and standardization: gene expression data in the spatial transcriptome data are standardized to ensure that the expression level of each gene has the same scale. This includes logarithmic transformation and data normalization of the gene expression data. Gene screening: To improve computational efficiency and remove noise, only highly variable genes are retained. Specifically, the top 3000 highly variable genes (HVGs) are screened from the spatial transcriptome data for subsequent analysis.

[0071] S3: Perform data interpolation on spatial transcriptome data, and further combine it with single-cell RNA sequencing data to perform cell type deconvolution.

[0072] S301: Data interpolation;

[0073] Because gaps exist between sampling points in spatial transcriptome data, interpolation techniques can fill these gaps and restore data continuity. Interpolation methods utilize the spatial coordinates and gene expression relationships between spatial points to fill in missing regions by estimating the gene expression values ​​of neighboring spatial points. The specific steps are as follows:

[0074] SGCD is based on data from the 10X Visium platform, whose sampling points are arranged in a regular honeycomb pattern. Therefore, in this embodiment, the entire tissue slice is divided horizontally row by row, and each row is further divided according to the left and right 45-degree diagonals, thereby defining complete void regions between three points. Subsequently, based on the original spatial coordinate data, the centroid of the divided image is used as the new spatial coordinates of the void regions. The centroid coordinate calculation method is as follows:

[0075]

[0076]

[0077] In formulas (1) and (2), These are the coordinates of points within each segmented region. N This represents the total number of points within the region.

[0078] Gene expression data are interpolated within defined gap regions using the latent relationship between gene expression and spatial coordinates in the original spatial transcriptome data. The neighborhood-based interpolation method is represented as follows:

[0079]

[0080] In formula (3), spatial coordinates Gene expression at the location is , The coefficients are calculated based on weighting factors such as spatial distance. It represents the gene expression values ​​at other points in the neighborhood.

[0081] S302: Cell type deconvolution;

[0082] By combining single-cell RNA sequencing data with cell type deconvolution, the cell composition of each sampling point can be inferred, enhancing the cell type information of spatial data and making spatial domain identification more accurate.

[0083] S4: Construct a deep learning model based on graph convolutional networks;

[0084] The constructed deep learning model based on graph convolutional networks is as follows: Figure 2 As shown, it includes:

[0085] The module includes a data interpolation module, a spatial adjacency graph construction module, a random gene expression graph construction module, an encoder-decoder module, and a self-supervised contrastive learning module.

[0086] The data interpolation module is used to interpolate gene expression data and perform cell type deconvolution using spatial coordinate information;

[0087] The spatial adjacency graph construction module is used to construct a weighted adjacency graph from gene expression data after deconvolution based on cell type, and obtain spatial distance information between gene expression data and cell type similarity measurement;

[0088] The gene expression random graph construction module constructs a random graph by generating a raw neighborhood graph based on a weighted adjacency graph. G=(V,E) Add perturbations to generate a random gene expression graph. G'=(V',E') ;

[0089] The GCN-based encoder module is used to process the original neighborhood graph. G=(V,E) and randomized gene expression graph G'=(V',E') Modeling based on gene expression data;

[0090] The unique advantage of spatial transcriptomics lies in its ability to capture not only gene expression information but also spatial location information, thereby enabling the identification of spatially adjacent regions with similar cellular states and the subdivision of tissues into smaller domains. To fully utilize this spatial information, this implementation converts the data into an undirected graph, where nodes... V Represents all sampling points (spots), edges E This indicates the connection between points.

[0091] First, the Euclidean distance between each point and other points is calculated based on the spatial coordinates of each point, resulting in a distance matrix reflecting the degree of physical proximity. Then, using... K The nearest neighbor strategy (typically set in the experiments of this implementation) K =4), select the nearest point for each point. K Each point is considered as its neighbor, thus constructing a preliminary adjacency matrix, where if a point... j It is a point i If the neighbor is a neighbor, then the neighbor is a neighbor; otherwise, the neighbor is a neighbor. To ensure the symmetry of the graph, this implementation adds the original adjacency matrix to its transpose and truncates values ​​greater than 1, so that all connections are binarized (0 or 1).

[0092] To further enhance the expressive power of the graph model, this implementation introduces cell type information. This implementation assumes that spatially adjacent points with similar cell types are more likely to belong to the same spatial domain, and therefore assigns higher weights to existing edges. Specifically, for any pair of connected points, this implementation first calculates the cell type similarity between the spots, and then obtains the similarity measure according to formula (1).

[0093]

[0094] In formula (4), JSD It is Korbeck-Leibler ( KL Symmetric form of divergence. P and Q These represent the probability distributions of the predicted results and the actual results, respectively. JSD It can be represented as:

[0095]

[0096] In formula (5), For Kulbeck-Leibler ( KL) Divergence can be expressed as:

[0097]

[0098] Finally, the edge weights are updated to formula (7):

[0099]

[0100] In formula (7), γ As the fundamental weights for spatial information, even when cell type similarity is low, the edge weights can be guaranteed to be no less than [a certain value]. γ This implementation method is assumed to be... γ The value is 0.2. This weighting mechanism, based on the dual constraints of spatial distance and cell type similarity, enables graph neural networks to better integrate the two types of information during feature extraction. This makes spatially adjacent points with similar cell types closer together in the feature space, thereby achieving a more accurate spatial domain partitioning.

[0101] The encoder-decoder module is used to process the original neighborhood graph. G=(V,E) and randomized gene expression graph G'=(V',E') Modeling based on gene expression data;

[0102] In SGCD, this implementation adopts the basic principle of Graph Convolutional Networks (GCN) to model the data as a graph. , where the node set V It contains all the sampling points (spots), while the edge set E This represents the connection relationship between nodes. Specifically, let... Let be the node feature matrix, where It is the number of nodes. d It is the feature dimension of each node; and This represents the adjacency matrix between nodes, which can typically be a weighted matrix or a binary adjacency matrix, reflecting the connection relationships between nodes.

[0103] To perform graph convolution operations, this implementation first processes the adjacency matrix. A The adjacency matrix is ​​obtained by normalization. ,in D It is a diagonal matrix, and the diagonal elements of the matrix are... Represents a node i The degree of the nodes. This normalization step helps to avoid the effects of differences in node degree during graph convolution operations.

[0104] In the l In a graph convolutional network with multiple layers, the update rule for node features can be expressed as:

[0105]

[0106] In formula (8), It is the first l-1 The output of the layer (i.e., the node feature representation of the previous layer). and They are the firstl The trainable weight matrix and bias terms of the layer, It is a non-linear activation function (such as ReLU). Ultimately, Indicates the first l The output of the layer, and this implementation will Set as the original input gene expression matrix This implementation method will As the final output of the encoder, each line Indicates sampling point i The potential representation of .

[0107] Next, the potential representation The data is input into the decoder, and the decoding process maps it back to the original gene expression space. Unlike the encoder, the decoder uses a symmetrical structure to reconstruct the gene expression profile. The specific structure of the decoder is as follows:

[0108]

[0109] In formula (9), It is in the t Reconstructed gene expression profiles generated by layers. Initialized as the encoder output . and These are the trainable weight matrix and bias vector in the decoder, respectively, and they are shared among all nodes.

[0110] To effectively train the model, this implementation optimizes the model parameters by minimizing the self-reconstruction loss of gene expression. The specific reconstruction loss function is defined as follows:

[0111]

[0112] In formula (10), and These are nodes i The model obtains the original normalized gene expression levels and the gene expression levels reconstructed by the decoder. By minimizing this loss function, the model can learn effective node representations and reconstruct a spectrum close to the original gene expression from the latent space.

[0113] In this way, SGCD can effectively model gene expression data and improve the model's performance in gene expression reconstruction tasks by utilizing the structural information of graph convolutional networks.

[0114] Gene expression random graphs are used to generate original neighborhood graphs based on a constructed weighted adjacency graph. G=(V,E) Add perturbations to generate a random gene expression graph. G'=(V',E') .

[0115] Construct a random graph of gene expression. Given the original neighborhood graph. G=(V,E) and normalized gene expression matrix X This implementation method employs technical perturbation to generate a random graph. G'=(V',E') and its corresponding expression matrix X' At the same time, the topology of the graph remains unchanged.

[0116] S5: Refine the training using the self-supervised contrastive learning module;

[0117] To enhance the representational power of graph embeddings and improve the discriminability between nodes, this implementation proposes a strategy based on self-supervised contrastive learning (SCL), aiming to extract richer feature information from the local spatial context. The core idea of ​​this strategy is to optimize the representational power of the graph neural network (GCN) through self-supervised learning, so that the representation of each node can better capture its neighborhood structure and micro-environment features.

[0118] Specifically, this implementation method utilizes the original drawing. G and random graph G′ As input, two representation matrices are generated by a graph convolutional network (GCN) encoder. and The task of the GCN encoder is to generate the embedding representation of each node by aggregating the neighbor information of each node, thereby forming the local context vector of each node. This vector represents the semantic information of a node in its local environment and is obtained by aggregating the representations of its neighboring nodes.

[0119] Under this design, this implementation constructs positive and negative sample pairs: points i The representation of Its local context Forming positive sample pairs, while in random graphs G′ midpoint i The representation of G' Its context These constitute negative sample pairs. Positive sample pairs aim to maintain the similarity between a node and its local environment, while negative sample pairs help the model distinguish the semantic differences between different nodes.

[0120] The goal of SCL is to optimize the quality of node representations by maximizing the mutual information between positive sample pairs and minimizing the mutual information between negative sample pairs. In this way, the representations of adjacent nodes will be more similar, while different nodes will be effectively distinguished. To achieve this goal, this implementation uses the binary cross-entropy (BCE) loss function to construct a contrastive loss function, which takes the following form:

[0121]

[0122] In formula (11), It is a discriminator D Its form is It is composed of dual neural networks and is used to distinguish between positive and negative sample pairs. Indicates positive sample pairs The probability score.

[0123] in, It is a discriminator composed of adversarial neural networks. This is used to distinguish between positive and negative sample pairs. Specifically, Indicates positive sample pairs The assigned probability score reflects the similarity of the sample pair in the model. To enhance model stability and reduce the impact of noise during training, this implementation further... G′ A symmetrical contrast loss is introduced. ， By introducing the construction of a random graph, the positive and negative sample pairs during training are more balanced, avoiding overfitting to certain graph topologies. The form of this contrastive loss is as follows:

[0124]

[0125] In this way, SCL can dynamically adjust the model's learning objectives during training, enabling the model to not only learn local relationships between nodes but also maintain robustness across different graph structures. Ultimately, the contrastive learning-based loss function helps the model gradually optimize node representations during the learning process, allowing it to better capture complex structural information in graph data.

[0126] For loss function This implementation method is divided into two main parts: self-reconstruction loss and contrastive loss. The self-reconstruction loss... The reconstruction error between the original gene expression data and the reconstructed data is compared with the loss. and These represent the losses for positive and negative sample pairs, respectively. This implementation trains the representation learning module for spatial transcriptome data by minimizing the self-reconstruction loss and the contrastive loss. Simply put, the overall training loss of this module is defined as:

[0127]

[0128] In formula (13), and This is a weighting factor that balances the impact of reconstruction loss and contrast loss. Empirically, this implementation method sets... and They are 10 and 1 respectively.

[0129] S5: After training the data into a latent representation using the model, the mclust method is used to perform cluster analysis on the latent representation to identify the spatial domains of different organizations.

[0130] Example

[0131] This embodiment primarily conducted experiments on four sets of spatial transcriptome datasets measured using the 10x Genomics Visium technique, comparing the proposed method with five other algorithms, as detailed below:

[0132] SGCD (the method proposed in this invention) can accurately resolve the spatial structure of DLPFC tissue. This embodiment first evaluates the representation learning performance of SGCD on a benchmark dataset. The 10XVisium dataset of the human dorsolateral prefrontal cortex (DLPFC) is a widely used benchmark dataset for evaluating spatial clustering performance. This dataset consists of 12 slices, each with 4 or 6 DLPFC layers and one white matter (WM) layer. These domains have been annotated using morphological features and marker genes, thus providing a valuable resource for evaluating the clustering accuracy of the learned representations. Figure 3B). This embodiment compares the spatial domain accuracy of SGCD with STAIG, MuCoST, GraphST, STAGATE, and SpaGCN across all 12 slices in the DLPFC dataset (Figure 3A) using the adjusted Rand index (ARI) and normalized mutual information (NMI). SGCD achieves the best performance, with the highest average ARI and NMI accuracies of 0.651 and 0.710, respectively. GraphST's average ARI and NMI accuracies are 0.547 and 0.670, respectively; STAIG's are 0.538 and 0.641, respectively; MuCoST's are 0.534 and 0.667, respectively; STAGATE's are 0.505 and 0.639, respectively; and SpaGCN's are 0.413 and 0.6557, respectively (Figure 3A). To demonstrate the detailed performance of SGCD, this embodiment provides the spatial domain recognition results for slice #151673, including six DLPFC layers and one WM layer. First, SGCD exhibits the strongest clustering performance, achieving high scores of ARI = 0.683 and NMI = 0.758. Figure 3 (C on). Its spatial domain identification results show that the boundaries between the six DLPFC layers and one WM layer are very clear, providing a solid foundation for subsequent spatial data analysis. Meanwhile, the UMAP plot shows that the data clusters are compactly distributed and clearly separated (C on). Figure 3 In C), the PAGA trajectory clearly shows the smooth evolution from layer 1 to layer 6 and then to the white matter layer. Figure 3 The statistical results of the T-test (C below) showed that SGCD was significantly superior to other comparison methods, with differences reaching significant levels of p<0.01 (**), p<0.001 (***), and p<0.0001 (****), respectively. Figure 3D). These findings fully demonstrate its advantages in clustering accuracy and hierarchical reasoning. STAIG also demonstrates certain capabilities in clustering and trajectory reasoning, achieving scores of ARI = 0.529 and NMI = 0.697. Its spatial domain identification results can reveal the spatial evolution of data from layer 1 to layer 6 and then to WM. Although the overall effect is not as good as SGCD, the distribution of each cluster in the UMAP diagram is relatively clear, and it still has reference value in capturing changes in spatial structure. MuCoST's clustering performance reaches scores of ARI = 0.577 and NMI = 0.706, showing good overall performance. However, in practical terms, it has certain shortcomings in the boundary separation of some layers (such as between layer 2 and layer 4), resulting in unclear hierarchical transitions displayed in the PAGA trajectory diagram. This blurred boundary affects the accurate characterization of the spatial evolution of the data to some extent. GraphST also exhibits high clustering accuracy (ARI = 0.640, NMI = 0.736), effectively dividing different spatial domains. However, it falls slightly short in boundary identification between certain levels, resulting in trajectory reasoning in PAGA graphs not reaching the clarity of SGCD. Overall, GraphST can reflect the general hierarchical structure, but it lacks the fine detail of SGCD. Compared to the aforementioned methods, STAGATE performs relatively poorly. The boundaries between different spatial domains in its clustering results are somewhat blurred, and the distribution of data points in the UMAP graph is quite mixed, failing to form a clear separation effect. Due to the unclear clustering separation, it is difficult to show a reasonable hierarchical structure in PAGA trajectory reasoning, leading to insufficient accuracy in the spatial evolution analysis of the data. SpaGCN's clustering results are the most chaotic, with low ARI and NMI indices. There is no clear separation between different levels during its spatial domain identification process. There is significant overlap between clusters in the UMAP graph, while in the PAGA trajectory graph, the data transitions show obvious blurring and chaos, making it difficult to reveal the true hierarchical evolution relationship. This results in significant shortcomings for SpaGCN in terms of fine-grained segmentation and accurate characterization of spatial structure. Overall, the analysis of various methods shows that SGCD leads in clustering accuracy, representation learning, and hierarchical inference, providing the clearest and most accurate spatial domain identification results, while other methods suffer from varying degrees of blurred boundaries or unclear hierarchical transitions. To compensate for the gaps between sampling points in the 10X Visium dataset, SGCD designed a unique data preprocessing and modeling workflow.First, the raw spatial transcriptome data is preprocessed using interpolation methods to make the data more spatially continuous. Then, the interpolated data is processed using deconvolution to extract more precise cell type information. Finally, the cell type results obtained from deconvolution are combined with the original spatial coordinates to construct a weighted adjacency graph, effectively capturing the interaction between spatial neighborhood relationships and cell characteristics. This series of meticulous processes enables SGCD to accurately distinguish the transitions and boundaries between layers in spatial structure analysis, clearly reflecting the hierarchical structure and spatial evolution patterns within the cerebral cortex, providing solid data support and theoretical basis for subsequent research on brain function regulation and related pathological mechanisms.

[0133] To further verify the superior performance of SGCD in accurately characterizing and resolving the structural features of complex spatial domains, this embodiment conducts an in-depth evaluation on structurally complex mouse brain slice data, further demonstrating its significant advantages in the segmentation of complex spatial domains. This dataset contains 3798 spots and 36601 genes, divided into 52 subregions (…). Figure 4 A). This embodiment compares SGCD with SpaGCN, STAGATE, GraphST, Mucost, and STAIG ( Figure 4 The results showed that SGCD performed best in clustering, with an ARI of 0.497 and an NMI of 0.727, which was 11% higher than the second-best algorithm, GraphST. The t-test results showed that SGCD's NMI score was not significantly better than STAGATE and STAIG, but its ARI score was significantly better than the other comparison methods, with differences reaching a significance level of p<0.01 (**) and p<0.0001 (****), respectively. Figure 4 D). Furthermore, SGCD demonstrated exceptional ability in analyzing complex spatial domains of the mouse brain. In the mouse brain, the CCu (striatum) is primarily involved in motor control, reward mechanisms, and habit formation, while the AcbC (nucleus accumbens core) is closely related to the regulation of executive function, decision-making, and reward mechanisms. Spatial domains 42 and 19 identified by SGCD correspond to CCu and AcbC, respectively. Figure 4 B), whose segmentation results are highly consistent with the actual anatomical structure; in contrast, although STAGATE can capture the general outline, it is not as good as SGCD in terms of detail, while other methods have the problem of multiple spatial domains mixed together, making it difficult to achieve accurate and detailed segmentation.

[0134] SGCD's superior performance is attributed not only to its deep fusion of cell type information when constructing weighted adjacency graphs, but also to its full utilization of the gap information between nodes. Specifically, SGCD combines the cellular composition information of each spot with its spatial coordinates, endowing each node in the graph with rich biological context. When determining the edge weights between nodes, it considers both spatial distance and cell type similarity. More importantly, for the gap regions formed by the regular honeycomb arrangement in the 10X Visium dataset, SGCD uses gene expression data interpolation technology to effectively compensate for the lack of spatial continuity information caused by sampling gaps in the original data, thus achieving a more refined characterization of spatial neighborhood relationships. This multi-dimensional information synergy significantly enhances the graph neural network's ability to capture local and global structural features, thus demonstrating unparalleled advantages in the detailed partitioning of complex spatial domains, providing solid data support for the accurate analysis of brain microstructure and its functional regulation mechanisms.

[0135] SGCD enables accurate parsing and annotation of unannotated spatial domain data. PDAC (pancreatic ductal adenocarcinoma) is one of the most deadly malignant tumors worldwide, with a complex and variable tumor microenvironment. This embodiment obtains baseline data for spatial clustering from the SDM Bench database, dividing the PDAC-A and PDAC-B data slices into five partitions each. This paper presents the analysis results of the PDAC-A slice (…). Figure 5 This embodiment employs multiple spatial clustering methods, including SpaGCN, STAIG, STAGATE, GraphST, MuCoST, and SGCD, for analysis. Figure 5A). It should be noted that SpaGCN and STAIG did not use histological images in their analysis due to compatibility issues with the PDAC dataset. Among all methods, SGCD performed best, with clear clustering results and high accuracy. Its ARI (Adjusted Rand Index) and NMI (Normalized Mutual Information) were 0.538 and 0.508, respectively, showing high clustering accuracy, especially in the separation of complex tissue regions. GraphST and SpaGCN also performed well, with ARIs of 0.478 and 0.434, respectively, effectively dividing large regions. In contrast, STAGATE's ARI was 0.335, showing some shortcomings in detail processing and noticeable erroneous merging of regions. STAIG had some advantages in overall clustering, but lower accuracy in detailed separation. MuCoST performed the worst among all methods, with an ARI of 0.031 and an NMI of 0.098, showing poor clustering results and failing to effectively capture the spatial structure of the data, thus it is not suitable for the PDAC dataset. The statistical results of the T-test showed that SGCD was significantly superior to all comparison methods, with all differences reaching a significant level of p<0.0001 (****). Figure 5 C). Furthermore, this embodiment employs a deconvolution method to combine single-cell RNA-seq data with interpolated spatial transcriptome data, thereby obtaining richer cell type information for each point (C). Figure 5(B) Based on the spatial distribution of cell types, this embodiment conducted a detailed analysis of different regions. In spatial domain 0, the high-density distribution of cancer cell clones A and B indicates that this region is a typical cancer area, reflecting the proliferation and malignant characteristics of tumor cells. In spatial domains 2, 3, and 4, the significant distribution of ductal cells did not show cancer cell proliferation, consistent with the characteristics of non-malignant ductal epithelial regions, indicating that these regions are normal or non-malignant ductal cell regions. In spatial domain 1, the high-density distribution of acinar cells indicates that this region is normal pancreatic tissue, with no cancer cell proliferation observed, consistent with the definition of pancreatic tissue regions. Through these analyses, the high accuracy of the SGCD method enabled this embodiment to accurately distinguish between cancerous regions, non-malignant ductal epithelial regions, and pancreatic tissue regions. It is worth noting that SGCD is particularly suitable for datasets with only spatial domain partitioning results but lacking real labels, helping this embodiment to effectively label these data correctly, providing a reliable spatial analysis tool for in-depth research on the tumor microenvironment of PDAC. Through SGCD, this embodiment not only obtained more accurate spatial domain partitioning but also, without real labels, combined with cell type and spatial information, inferred the biological characteristics of each region, thus providing important research basis for disease diagnosis and treatment. Based on the clustering results described above, this embodiment further analyzes the significant advantages of SGCD. First, by constructing a weighted adjacency graph that integrates cell type and spatial location information, SGCD effectively compensates for the lack of spatial continuity information caused by gaps between sampling points in the 10X Visium data, thus achieving a fine characterization of spatial neighborhood relationships. This integration of multidimensional information not only makes the method significantly superior to other methods in clustering metrics (such as ARI and NMI), but more importantly, it can automatically and accurately resolve and label complex spatial domain structures even in the absence of real labels. Therefore, SGCD demonstrates unparalleled accuracy in distinguishing cancerous regions, non-malignant ductal epithelial regions, and normal pancreatic tissue regions in PDAC, providing solid data support and theoretical basis for a deeper understanding of the tumor microenvironment and its functional regulatory mechanisms.

[0136] SGCD can reveal complex spatial patterns and gene expression heterogeneity in cancer tissues. Breast cancer exhibits high heterogeneity and a complex microenvironment, making it difficult to fully describe its spatial structural characteristics solely through manual annotation of tumor morphology. Therefore, Fu et al. divided histological images into four main morphological types and twenty subregions based on pathological features, including ductal carcinoma in situ / lobular carcinoma in situ (DCIS / LCIS), healthy tissue, invasive ductal carcinoma (IDC), and tumor margins with low-grade malignancy characteristics (…). Figure 5A). To verify the generalization ability of SGCD in cancer tissues with complex spatial patterns and heterogeneous gene expression, this example was tested on 10X Visium data from human breast cancer tissues ( Figure 5 B) and compared it with baseline methods such as STAIG, MuCoST, GraphST, STAGATE, and SpaGCN. Figure 6 C). The results show that SGCD achieved the highest accuracy in the spatial domain recognition task (ARI=0.59, NMI=0.7). The t-test results show that SGCD significantly outperformed other comparative methods, with differences reaching significant levels of p<0.01 (**) and p<0.0001 (****), respectively. Figure 6 I).

[0137] To reveal the complex spatial patterns and highly heterogeneous gene expression characteristics in breast cancer, this embodiment analyzes the spatial clustering results from two perspectives: spatial distance variation and spatially variable genes. First, based on Ripley's L-function, this embodiment calculates the point distribution pattern of each spatial domain (…). Figure 6 E), it was found that spatial domains 8 (DCIS / LCIS_3), 5 (IDC_4), and 13 (IDC_5) exhibited significant clustering characteristics compared to other dispersed regions. In the partitioning of the DCIS / LCIS_3 and IDC_4 regions, SGCD, STAIG, and GraphST all consistently identified a single spatial domain, consistent with the reference results, while other methods showed multiple spatial domains mixed together. In the partitioning of the IDC_5 region, although none of the methods were completely consistent with the reference, SGCD clearly outlined the general contour of the region and only mixed in one other spatial domain, while MuCoST and GraphST only identified partial contours and mixed in more than two spatial domains. Further calculation of the Jaccard index ( Figure 6 D) shows that SGCD has the highest recognition accuracy. To further explore the heterogeneity of gene expression, this embodiment compares the top five differentially expressed genes between spatial domain 5, representing the IDC region, and spatial domain 8, representing the DCIS / LCIS region. Figure 6 F), the results showed significant expression differences between different clusters, reflecting obvious heterogeneity among tumor tissues. To further reveal the heterogeneous characteristics within tumor tissues, this embodiment further performed differential expression analysis and gene ontology (GO) enrichment analysis. Between spatial domain 5 and spatial domain 8, a total of 359 significantly differentially expressed genes were detected (|log fold change| ≥ 2; P value < 0.05). Among the top five differentially expressed genes ( Figure 6(G) Significant differences were observed in the expression levels of CXCL4, CCND1, S100A11, DEGS1, and KRT18. High expression of CXCL4 has been confirmed as an independent biomarker for poor prognosis in breast cancer. CCND1 plays a crucial role in cell proliferation and is closely related to tumor invasion and metastasis. S100A11 may promote tumor growth and metastasis by regulating signaling pathways in the tumor microenvironment. DEGS1 participates in sphingolipid synthesis, and its metabolic abnormalities may accelerate tumor growth and spread by interfering with cell apoptosis and survival balance. High expression of KRT18 is associated with deep tumor invasion, invasiveness, and drug resistance, and has potential application value in the diagnosis and prognostic assessment of breast cancer. Furthermore, GO enrichment analysis of differentially expressed genes (G) Figure 5 The study revealed the main biological processes (BPs) involved by these genes. Upregulated BPs were primarily concentrated in processes such as immune regulation, cell proliferation, and macrophage activation, reflecting cellular responses to the tumor microenvironment or stress. Genes in spatial domain 5 were more significantly associated with tumor progression, invasion, and metastasis. Downregulated BPs, on the other hand, mainly involved immune responses, indicating a weakened cellular response to inflammatory stimuli in spatial domain 8, suggesting suppressed immune function in exogenous antigen processing, extracellular matrix remodeling, and inflammation regulation. This weakened immune function may lead to tumor immune escape, further promoting tumor cell survival and spread. This result is consistent with the findings of Xu et al., who suggested that under specific pathological conditions, weakened immune responses may suppress chronic inflammation and interfere with the body's effective response to exogenous pathogens, thereby accelerating disease progression or triggering immune system dysfunction. SGCD exhibits unique advantages in such a complex tumor microenvironment, primarily due to its unique modeling process. First, SGCD uses interpolation to process the original 10X Visium data, filling gaps between sampling points and making the data more spatially continuous. Next, deconvolution is used to extract more precise cell type information from the interpolated data, effectively separating mixed signals. Finally, the deconvolution results are combined with the original spatial coordinates to construct a weighted adjacency graph, fully capturing the spatial neighborhood relationships and feature similarities between cells, thus achieving a refined analysis of the complex spatial structure within the tumor. This series of meticulous processes enables SGCD to clearly distinguish different spatial domains in breast cancer tissue, accurately reflecting the highly heterogeneous biological phenomena within the tumor, providing solid data support and theoretical basis for further analysis of tumor biological mechanisms, immune escape, and precision treatment strategies. In summary, SGCD not only achieves the highest accuracy in spatial domain identification in breast cancer tissue but also effectively reveals complex spatial patterns and gene expression heterogeneity within the tumor through a series of innovative modeling strategies, fully demonstrating its unique advantages in handling highly heterogeneous and complex microenvironments, and providing a new perspective for a deeper understanding of tumor biology.

[0138] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A method for spatial domain recognition based on data interpolation and cell type deconvolution, characterized in that, include: Step 1: Obtain the spatial transcriptome dataset and single-cell RNA sequencing dataset, and perform data preprocessing on the obtained datasets; Step 2: Perform data interpolation on the preprocessed spatial transcriptome data and combine it with single-cell RNA sequencing data to perform cell type deconvolution; Step 2 specifically includes: Step 2.1: Divide the raw gene expression data in the preprocessed spatial transcriptome data into rows horizontally, and then divide each row according to the diagonal, dividing the complete gap region between the center point, left vertex and right vertex; Step 2.2: Based on the original spatial coordinate data of gene expression data in the gap regions after division, calculate the centroid coordinates of the corresponding gap regions. The centroid coordinates are used as the spatial coordinates of the void region; Step 2.3: Based on the original gene expression data and the spatial location information of the corresponding cells, the gene expression data of the divided gap regions are interpolated using the neighborhood interpolation method; Step 2.4: Combine the preprocessed single-cell RNA sequencing data with cell type deconvolution of the interpolated gene expression data to infer the cell composition of each gene expression data sampling point; Centroid coordinates of the void region The calculation formula is: In formulas (1) and (2), These are the coordinates of points within each segmented region. N The total number of points within the region; The expression for gene expression data interpolation is: In formula (3), Spatial coordinates Gene expression at the site, These are coefficients calculated based on weighting factors such as spatial distance. This represents the gene expression values ​​at other points within the neighborhood. Step 3: Construct a deep learning model based on graph convolutional networks; The deep learning model based on graph convolutional networks constructed in step 3 includes: The module includes a spatial adjacency graph construction module, a gene expression random graph construction module, an encoder-decoder module, and a self-supervised contrastive learning module. The spatial adjacency graph construction module is used to construct a weighted adjacency graph from the gene expression data after deconvolution according to cell type, and to obtain spatial distance information between gene expression data and cell type similarity measurement; The gene expression random graph construction module constructs a random graph based on an original neighborhood graph generated from a weighted adjacency graph. G=(V,E) Add perturbations to generate a random gene expression graph. G'=(V',E') ; The GCN-based encoder module is used to process the original neighborhood graph. G=(V,E) and randomized gene expression graph G'=(V',E') Modeling based on gene expression data; The self-supervised learning comparison module is used to compare the original neighborhood graph. G=(V,E) and randomized gene expression graph G'=(V',E') Generate positive and negative sample pairs to optimize node representations and train deep learning models; Step 4: Input the gene expression data of the spatial transcriptome data after cell type deconvolution into the deep learning model, and use a self-supervised contrastive learning strategy to train the deep learning model based on the cell type information obtained from cell type deconvolution combined with the gene expression information and spatial location information of the spatial transcriptome data. Step 5: Perform spatial domain identification on the test data based on the trained model.

2. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 1, characterized in that, In step 1, the spatial transcriptome dataset includes the gene expression data of spot and the spatial location information of spot, and the single-cell RNA sequencing dataset includes the gene expression data of cell and cell type information. The gene expression data matrix in the spatial transcriptome dataset has a dimension of spot×gene and is arranged in a regular honeycomb pattern on the tissue slice. The gene expression data matrix in the single-cell RNA sequencing dataset has a dimension of cell×gene.

3. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 1, characterized in that, Step 1 involves data preprocessing of the acquired dataset, specifically including: Step 1.1: Perform logarithmic transformation and normalization on the gene expression data of the spatial transcriptome dataset; Step 1.2: Select a predetermined number of highly variable genes from the normalized spatial transcriptome data and single-cell RNA sequencing data to complete the preprocessing of the spatial transcriptome dataset and single-cell RNA sequencing dataset.

4. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 1, characterized in that, Step 4, training the deep learning model, includes the following steps: Step 4.1: Input the deconvolutioned gene expression data of cell types into the spatial adjacency graph construction module to obtain the spatial distance information between gene expression data and cell type similarity measurement; Step 4.2: Calculate the original neighborhood graph generated from the weighted adjacency graph. G=(V,E) In the random gene expression graph construction module, add perturbations to generate a random gene expression graph. G'=(V',E') ; Step 4.3: Convert the original neighborhood graph G=(V,E) and randomized gene expression graph G'=(V',E') Input a GCN-based encoder module to model the gene expression data in the input graph; Step 4.4: Convert the original neighborhood graph G=(V,E) and randomized gene expression graph G'=(V',E') Input the self-supervised learning contrastive module and use the self-supervised contrastive learning strategy to train the deep model.

5. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 4, characterized in that, Step 4.1 specifically includes: Step 4.1.1: Convert the deconvolutioned gene expression data of cell types into an undirected graph, which includes several nodes. V and edge E , where nodes V For all sampling points, the edges E This refers to the connection relationships between points; Step 4.1.2: Calculate the Euclidean distance between sampling points based on the spatial coordinates of each point in the undirected graph to obtain the distance matrix; Step 4.1.3: Using the K-nearest neighbor strategy, select the K nearest sampling points by Euclidean distance as its neighbor nodes to obtain a preliminary adjacency matrix. Where, if the points in the preliminary adjacency matrix... j It is a point i The value is 1 if it is a neighbor, and 0 if it is not, and the point is set to... j The original adjacency matrix is ​​added to its transpose, and values ​​greater than 1 are truncated to make all sampling point connections binary. Step 4.1.4: Introduce cell type information, calculate the cell type similarity of any connected point pair in the preliminary adjacency matrix, and calculate the edge weight based on the cell type similarity and the basic weight of spatial information.

6. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 4, characterized in that, Step 4.3 specifically includes: Step 4.3.1: Based on the graph convolutional network and combined with the weighted adjacency graph, construct the gene expression data into a graph. ,in, V For a set of nodes, E For edge set, node set V Includes all sampling points, edge set E Indicates the connection relationship between nodes; Step 4.3.2: Setting up the diagram middle Let be the node feature matrix, where For the number of nodes, d For each node's feature dimensions, A binary adjacency matrix between nodes; Step 4.3.3: Normalize the adjacency matrix A to obtain the normalized adjacency matrix. ,in D Given a diagonal matrix, the diagonal elements are... , For nodes i The degree; Step 4.3.4: Convert the normalized adjacency matrix The trainable weight matrix and bias terms of the corresponding layer of the graph convolutional network are input into the encoder to obtain the latent representation of the sampling points. ; Step 4.3.5: Latent representation of the sampling points The input is fed into the decoder, and through the decoding process, it is mapped back to the original gene expression space by minimizing the self-reconstruction loss function of gene expression. Optimize the parameters of the deep learning model.

7. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 4, characterized in that, Step 4.4 specifically includes: Step 4.4.1: Create the original gene expression map G and random gene expression map G′ As input, two representation matrices are generated by an encoder based on a graph convolutional network. and ， The embedding representation of each node is generated by aggregating the neighbor information of each node in the matrix, thus obtaining the local context vector of each node. ; Step 4.4.2: Construct positive and negative sample pairs, where the sampling points i matrix Its local context vector Forming positive sample pairs in a random graph G′ In the middle, sampling points i The representation of Its context vector Forming negative sample pairs; Step 4.4.3: Construct the positive sample contrast loss function and negative sample loss function The quality of node representation is optimized by maximizing the mutual information between positive sample pairs and minimizing the mutual information between negative sample pairs. Step 4.4.4: Loss function based on positive samples Negative sample loss function and self-reconstruction loss function Overall training loss function for constructing deep learning models Based on the overall training loss function The deep learning model is trained to obtain the latent representation of all sampling points.

8. The method for spatial domain recognition based on data interpolation and cell type deconvolution according to claim 1, characterized in that, Step 5 specifically includes: The mclust method is used to perform cluster analysis on the latent representations of all points generated during training to obtain spatial domain identification results for different cell tissues.

Images