A spatial transcriptome spatial domain identification method based on relation graph attention

Abstract

The application discloses a spatial transcriptome spatial domain identification method based on a relation graph attention, comprising the following steps: S1, collecting gene expression, histological image data and spatial position information through stRGAT; S2, performing dimension reduction and normalization on the data, and constructing an undirected graph by using a standardized Euclidean distance to measure the similarity between nodes; S3, inputting the pretreated data into a relation graph attention network, and extracting final potential embedding; and S4, performing spatial domain identification by using an unsupervised clustering algorithm. The application solves the problem of low accuracy of the existing spatial domain identification algorithm, and cannot effectively identify the expected cortical layer structure.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence image processing technology, specifically, it relates to a spatial transcriptome spatial domain recognition method based on relation graph attention. Background Technology

[0002] In specific tissues of multicellular organisms, there are many different types of cells, each with a specific function, and their life cycle and behavior are related to their interaction with the surrounding environment. In particular, cell location information and spatial organization methods within tissues can effectively help understand the biological functions of tissues and aid in disease research. Single-cell RNA sequencing technology (scRNA-seq) has become one of the core technologies in the field of genomics with impressive high resolution and plays an important role in immunology, biomedicine, and omics analysis. However, during sequencing, scRNA-seq technology mixes tissue cells together, which leads to a lack of cell location information crucial for understanding tissue function. The emergence of spatial transcriptomics technology has solved this problem to some extent. There are many spatial transcriptomics sequencing technologies, such as 10x visium, Slide-seq, and others. In 2020, spatial transcriptomics was named one of Nature's Techniques of the Year, providing effective technical support for accurately understanding the histological functions of cells in the spatial dimension.

[0003] Identifying spatial domains (i.e., regions where gene expression locations and histological locations are identical in spatial dimensions) is one of the most fundamental and important research topics in spatial transcriptomics. However, accurately identifying biologically significant spatial domains remains a significant challenge. Spatial domain identification methods can be divided into spatial clustering and non-spatial clustering methods. Non-spatial clustering methods mainly employ traditional clustering methods, such as k-means and Louvain algorithms, but these methods only consider sparse gene expression data, making it difficult to correlate clustering results with tissue sections. Spatial clustering methods, by combining gene expression, spatial location information, and morphological images, fully consider spatial dimensional information, thereby obtaining more interpretable analytical results.

[0004] For example, BayesSpace draws primarily on image processing techniques, using Markov Random Fields (MRFs) to implement Bayesian models that encourage points within the same cluster to be close to each other. StLearn extracts features from tissue images and normalizes gene expression data by the distance between spatially adjacent points before clustering. The updated data is then used to determine cell types. The StLearn method involves utilizing pseudo-spatiotemporal (PST) distances to compute transcriptional states and reconstruct spatial gradients. This is accomplished by considering connections within and between cell types using a directed minimum spanning tree optimization technique. SpaGCN uses graph convolutional networks (GCNs) to process gene expression, spatial location, and morphological data. It aggregates gene expression information from neighboring points by generating undirected weighted graphs, and finally uses an unsupervised clustering algorithm to cluster them, thereby identifying spatial domains. SEDR uses a deep autoencoder to construct latent gene representations in a low-dimensional latent space, and then embeds the corresponding spatial information using a variational graph autoencoder. RESEPT preserves the 3D embeddings based on graph neural networks and maps them to RGB image color channels. Then, using the idea of ​​image segmentation, a supervised convolutional neural network was used to segment the RGB image, thus achieving the goal of spatial domain recognition.

[0005] While these methods can consider the spatial features of spatial transcriptomics, the calculation of node similarity is predetermined before training, preventing adaptive learning and insufficient exploration of the spatial structural features of spatial transcriptomics data. The algorithms also exhibit poor scalability, with different datasets leading to even worse performance. Furthermore, most algorithms focus on identifying accurate spatial domains but lack research on fine-grained identification of specific tissues. Many methods mention the similarity of neighboring nodes. For neural network models, over-considering the similarity of neighboring nodes may cause the model to ignore these relationships at different levels. More importantly, for deep learning-based methods, extending to multiple contiguous parts has greater algorithmic optimization and biomedical value. Summary of the Invention

[0006] To address the issue of low accuracy in existing spatial domain identification algorithms, this invention provides a spatial transcriptome spatial domain identification method based on relation graph attention.

[0007] To achieve the above-mentioned technical objectives, the technical solution adopted by the present invention is as follows:

[0008] A spatial domain identification method for spatial transcriptomics based on relation graph attention includes the following steps:

[0009] S1. Gene expression, histological image data, and spatial location information were acquired using stRGAT;

[0010] S2. Reduce and normalize the data, and construct an undirected graph using standardized Euclidean distance to measure the similarity between nodes.

[0011] S3. Input the preprocessed data into the relational graph attention network to extract the final latent embeddings;

[0012] S4, and then use an unsupervised clustering algorithm for spatial domain identification.

[0013] Furthermore, in step S2, the data is reduced in dimensionality and the gene expression data is stored in an N×D matrix, where N represents the number of spots and D represents the number of genes; the coordinates of each point in a two-dimensional space (x, y) are also included, and genes expressed at fewer than three points are excluded.

[0014] Further, in step S2, the data is normalized: the UMI count of each gene is divided by the total UMI count of all genes at that point, multiplied by 10000, and converted to the natural logarithm. Using gene expression data and histological images as input data, the data is read and converted into matrix format, and then processed.

[0015] Furthermore, in step S2, after data preprocessing, the gene expression and histological image data need to be converted into graph structure data;

[0016] The input data is converted into an unweighted directed graph. In spatial transcriptome data, the physical location and histological information of the two fixed points are equally important. First, the physical similarity of tissue slices is calculated. To improve the identification of boundary regions, a Gaussian smoothing method is used to calculate the color value of the current node to avoid the influence of a single maximum outlier.

[0017]

[0018] In the formula, (xp, yp) is the coordinate of a pixel, and σ is the standard deviation. For an undirected graph G(V, E), we can calculate the similarity between two points. Due to the different distributions of tissue histological images, gene expression matrices, and location coordinates, we use standardized Euclidean distance to calculate the distance V between different nodes.

[0019]

[0020] `ui` is the mean of all data in this dimension, and `si` is the corresponding standard deviation. The specific operation involves calculating the standardized Euclidean distance between points in `stRGAT`. Each vertex in the graph represents a point, and edges with specified weights are used to connect each pair of points. `stRGAT` uses Principal Component Analysis (PCA) to reduce the dimensionality of the preprocessed gene expression matrix. Using the first 50 principal components as input, it is suitable for all datasets in the recognition spatial domains analyzed in this paper.

[0021] The distance between two nodes u and v is calculated using the following formula.

[0022]

[0023] Further, in step S3, the graph attention calculation process is as follows: The graph attention network takes normalized gene expression as input, with the input set being h = {h1, hi+1, ..., hN}, i ∈ {1, 2, 3, 4, ..., N}, where N represents the number of nodes, and the output node is h' = {h1', hi'+1, ..., hN'}, i ∈ {1, 2, 3, 4, ..., N};

[0024] Using a common linear transformation, it is defined by a weight matrix, denoted as: F is the number of features in each node. Self-attention is applied to each node separately, and then a shared attention mechanism is implemented on each node. The attention coefficient is calculated as follows:

[0025]

[0026]

[0027] The intermediate representation of node i (left red rectangle) is combined with the intermediate representation of the neighboring nodes (blue rectangle) under each relation r to form each logit. For each logit matrix of each relation type, a softmax is taken to form attention coefficients ai,j(R); these attention coefficients construct a weighted sum of the neighborhood nodes of each relation (black rectangle), and these are aggregated and nodes are generated non-linearly (right red rectangle).

[0028] Furthermore, under each relation r, the intermediate representation of node i (red rectangle on the left) is combined with the intermediate representations of its neighboring nodes (blue rectangles) to form each logit. For each logit matrix of each relation type, a softmax function is applied to form the attention coefficients. Attention coefficients are used to construct a weighted sum of neighborhood nodes for each relation (black rectangle), and then these are aggregated and passed through a non-linear function to generate an updated representation of node i (right red rectangle);

[0029] The similarity weights were normalized using the softmax function.

[0030]

[0031] To adaptively learn the similarity between neighboring points, a self-attention mechanism widely used in graph neural networks is employed. The attention mechanism g is a single-layer feedforward neural network where parameters are shared among nodes and weighted by a weight vector. The unexpanded attention behavior is as follows:

[0032]

[0033] T represents transposition, ∥ is the connection operation, and the normalized attention coefficients are used to calculate the linear combination of the corresponding features. The characteristics of these combinations are selected as the final output features of each node.

[0034]

[0035] NL represents a nonlinear relationship. Based on the output of the graph attention network, stRGAT uses an unsupervised clustering algorithm to cluster points to identify differences in the spatial domain. The specific clustering method used is the Louvain method.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] stRGAT enhances the ability to identify layer formation in human dorsolateral prefrontal cortex (DLPFC) tissue; compared to existing methods, stRGAT can effectively identify expected cortical layer structures. Among the five models, stRGAT's identified spatial domains clearly describe layer boundaries and achieve the highest clustering accuracy (ARI = 0.537). Its overall results are also closer to those of manual annotation. In contrast, while the Bayesian spatial identification results are relatively consistent, they are also affected by their algorithms, resulting in unclear cortical layering and more outliers that differ significantly from the actual cortical layering. SpaGCN benefits from the structure of the graph convolutional neural network used, and the experimental results are ideal, but there are still some regions with unreasonable layering. SpaGCN even identifies fewer spatial domains. STAGATE also uses an attention network graph to model information from neighboring nodes. Although it achieved a high ARI score (0.57), it does not use tissue image information and does not consider neighboring node information, ignoring information from long-distance nodes, which also contributes to layering errors. Attached Figure Description

[0038] Figure 1This is an overall flowchart of a spatial transcriptome spatial domain identification method based on relation graph attention in an embodiment of the present invention;

[0039] Figure 2 This is a comparison chart of experimental results between five existing algorithms and the method of this application in the embodiments of the present invention. Detailed Implementation

[0040] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0041] like Figure 1 As shown, this embodiment provides a spatial transcriptome spatial domain identification method based on relation graph attention, including the following steps:

[0042] S1. Gene expression, histological image data, and spatial location information were acquired using stRGAT;

[0043] S2. Reduce and normalize the data, and construct an undirected graph using standardized Euclidean distance to measure the similarity between nodes.

[0044] S3. Input the preprocessed data into the relational graph attention network to extract the final latent embeddings;

[0045] S4, and then use an unsupervised clustering algorithm for spatial domain identification.

[0046] In step S2, the data is reduced in dimensionality and the gene expression data is stored in an N×D matrix, where N represents the number of spots and D represents the number of genes; the coordinates of each point in a two-dimensional space (x, y) are also included, and genes expressed at fewer than three points are excluded.

[0047] In step S2, the data is normalized: the UMI count of each gene is divided by the total UMI count of all genes at that point, multiplied by 10000, and converted to the natural logarithm. Using gene expression data and histological images as input data, the data is read and converted into matrix format, and then processed.

[0048] After data preprocessing in step S2, gene expression and histological image data need to be converted into graph structure data;

[0049] The input data is converted into an unweighted directed graph. In spatial transcriptome data, the physical location and histological information of the two fixed points are equally important. First, the physical similarity of tissue slices is calculated. To improve the identification of boundary regions, a Gaussian smoothing method is used to calculate the color value of the current node to avoid the influence of a single maximum outlier.

[0050]

[0051] In the formula, (xp, yp) is the coordinate of a pixel, and σ is the standard deviation. For an undirected graph G(V, E), we can calculate the similarity between two points. Due to the different distributions of tissue histological images, gene expression matrices, and location coordinates, we use standardized Euclidean distance to calculate the distance V between different nodes.

[0052]

[0053] `ui` is the mean of all data in this dimension, and `si` is the corresponding standard deviation. The specific operation involves calculating the standardized Euclidean distance between points in `stRGAT`. Each vertex in the graph represents a point, and edges with specified weights are used to connect each pair of points. `stRGAT` uses Principal Component Analysis (PCA) to reduce the dimensionality of the preprocessed gene expression matrix. Using the first 50 principal components as input, it is suitable for all datasets in the recognition spatial domain analyzed in this paper. The distance between two nodes `u` and `v` is calculated using the following formula.

[0054]

[0055] Step S3: Graph attention calculation process: The graph attention network takes normalized gene expression as input, with the input set h = {h1, hi+1, ..., hN}, i ∈ {1, 2, 3, 4, ..., N}, where N represents the number of nodes. The output nodes are h' = {h1', hi'+1, ..., hN'}, i ∈ {1, 2, 3, 4, ..., N}.

[0056] Using a common linear transformation, it is defined by a weight matrix, denoted as: F is the number of features in each node. Self-attention is applied to each node separately, and then a shared attention mechanism is implemented on each node. The attention coefficient is calculated as follows:

[0057]

[0058]

[0059] The intermediate representation of node i (left red rectangle) is combined with the intermediate representation of the neighboring nodes (blue rectangle) under each relation r to form each logit. For each logit matrix of each relation type, a softmax is taken to form attention coefficients ai,j(R); these attention coefficients construct a weighted sum of the neighborhood nodes of each relation (black rectangle), and these are aggregated and nodes are generated non-linearly (right red rectangle).

[0060] Under each relation r, the intermediate representation of node i (red rectangle on the left) is combined with the intermediate representations of its neighboring nodes (blue rectangles) to form each logit. For each logit matrix of each relation type, a softmax function is applied to form the attention coefficients. Attention coefficients are used to construct a weighted sum of neighborhood nodes for each relation (black rectangle), and then these are aggregated and passed through a non-linear function to generate an updated representation of node i (right red rectangle);

[0061] The similarity weights were normalized using the softmax function.

[0062]

[0063] To adaptively learn the similarity between neighboring points, a self-attention mechanism widely used in graph neural networks is employed. The attention mechanism g is a single-layer feedforward neural network where parameters are shared among nodes and weighted by a weight vector. The unexpanded attention behavior is as follows:

[0064]

[0065] T represents transposition, ∥ is the connection operation, and the normalized attention coefficients are used to calculate the linear combination of the corresponding features. The characteristics of these combinations are selected as the final output features of each node.

[0066]

[0067] NL represents a nonlinear relationship. Based on the output of the graph attention network, stRGAT uses an unsupervised clustering algorithm to cluster points to identify differences in the spatial domain. The specific clustering method used is the Louvain method.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] First, we tested on the 10x Visium dataset, which contains 12 human dorsolateral prefrontal cortex (DLPFC) images. Using the adjusted RAND index as a measure of clustering accuracy, we compared it with five mainstream spatial methods (SpaGCN, stLearn, BayesSpace, SEDR, and STAGATE). Since most past algorithms have demonstrated that spatial domain recognition algorithms generally outperform non-spatial methods, we only compared more established spatial methods. We annotated the cortical layers (L1-L6) and white matter (WM) of the 12 DLPFC images using genetic markers and cellular structures on a publicly available dataset called the 10x Visium ST benchmark dataset, as shown below. Figure 2As shown in a. Figure 2 a is the fundamental real point of the cortex and white matter (WM) in DLPFC part 151673. Figure 2 b is the ARI box plot of the 12 datasets included in DLPFC.

[0070] Figure 2 c is the experimental result based on dataset 151676, and the algorithms compared are SpaGCN, stLearn, BayesSpace, STAGATE and SEDR.

[0071] In the human dorsolateral prefrontal cortex dataset, we first compared the performance of each algorithm on 151,673 datasets. To ensure fairness, we did not use default parameters when running other algorithm models, as default parameters do not guarantee optimal experimental results. We made appropriate adjustments to ensure that the experimental results of the compared algorithms reached or approached the optimal results shown in their papers. Figure 2 c shows the spatial domain identification results of dataset 151673 using space, SEDR, stLearn, SpaGCN, STAGATE, and stRGAT.

[0072] stRGAT enhances the ability to identify layer formation in human dorsolateral prefrontal cortex (DLPFC) tissue; compared to existing methods, stRGAT can effectively identify expected cortical layer structures. Among the five models, stRGAT's identified spatial domains clearly describe layer boundaries and achieve the highest clustering accuracy (ARI = 0.537). Its overall results are also closer to those of manual annotation. In contrast, while the Bayesian spatial identification results are relatively consistent, they are also affected by their algorithms, resulting in unclear cortical layering and more outliers that differ significantly from the actual cortical layering. SpaGCN benefits from the structure of the graph convolutional neural network used, and the experimental results are ideal, but there are still some regions with unreasonable layering. SpaGCN even identifies fewer spatial domains. STAGATE also uses an attention network graph to model information from neighboring nodes. Although it achieved a high ARI score (0.57), it does not use tissue image information and does not consider neighboring node information, ignoring information from long-distance nodes, which also contributes to layering errors.

[0073] The foregoing has provided a detailed description of a spatial transcriptome spatial domain identification method based on relation graph attention provided in this application. The specific embodiments described are merely for the purpose of helping to understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A spatial transcriptome spatial domain identification method based on relation graph attention, characterized in that, Including the following steps: S1. Gene expression, histological image data, and spatial location information were acquired using stRGAT; S2. Reduce and normalize the data, and construct an undirected graph using standardized Euclidean distance to measure the similarity between nodes. S3. Input the preprocessed data into the relational graph attention network to extract the final latent embeddings; S4, and use unsupervised clustering algorithm for spatial domain identification; After data preprocessing in step S2, gene expression and histological image data need to be converted into graph structure data; The input data is converted into an unweighted directed graph. In the spatial transcriptome data, for two fixed points, the physical location similarity of tissue slices is calculated. Gaussian smoothing is used to calculate the color value of the current node to avoid the influence of a single maximum outlier. In the formula, (x p y p () is the coordinate of a pixel. It is the standard deviation. For an undirected graph G(V, E), the similarity between two points is calculated. Due to the different distributions of tissue histological images, gene expression matrices, and location coordinates, the standardized Euclidean distance is used to calculate the distance V between different nodes. u i s is the average of all data in this dimension. i To determine the corresponding standard deviation, the specific operation involves calculating the standardized Euclidean distance between points in stRGAT. Each vertex in the graph represents a point, and edges with specified weights are used to connect each pair of points. stRGAT uses principal component analysis to reduce the dimensionality of the preprocessed gene expression matrix, using the first 50 principal components as input. For two nodes u and v, the distance calculation formula is as follows: ； Step S3: Graph attention calculation process: The graph attention network takes normalized gene expression as input, with the input set h={ℎ1, ℎi+1, ..., ℎN}, i∈{1,2,3,4, ...,N}, where N represents the number of nodes. The output nodes are h'={ℎ1', ℎi'+1, ..., ℎN'}, i∈{1,2,3,4, ...,N}. A linear transformation, defined by a weight matrix, is used, denoted as W∈ℝ F'×F, where F is the number of features in each node. Self-attention is applied to each node separately, implementing a shared attention mechanism at each node. The attention coefficients are calculated as follows: The intermediate representation region of node i is combined with the intermediate representation regions of the neighboring nodes under each relation r to form each logit. For each logit matrix of each relation type, a softmax function is applied to form the attention coefficient. ; Note that the coefficients construct a weighted sum of the neighborhood nodes for each relation, aggregate these, and generate node i through a non-linear process.

2. The spatial transcriptome spatial domain identification method based on relation graph attention according to claim 1, characterized in that, In step S2, the data is reduced in dimensionality and the gene expression data is stored in an N×D matrix, where N represents the number of nodes and D represents the number of genes; the coordinates of each point in a two-dimensional space (x, y) are also included, and genes expressed at fewer than three points are excluded.

3. The spatial transcriptome spatial domain identification method based on relation graph attention according to claim 2, characterized in that, In step S2, the data is normalized: the UMI count of each gene is divided by the total UMI count of all genes at that point, multiplied by 10000, and converted to the natural logarithm. Using gene expression data and histological images as input data, the data is read and converted into matrix format, and then processed.

4. The spatial transcriptome spatial domain identification method based on relation graph attention according to claim 3, characterized in that, Under each relation r, the intermediate representation region of node i is combined with the intermediate representation regions of its neighboring nodes to form each logit. For each logit matrix of each relation type, softmax is applied to form the attention coefficients. Attention coefficients are used to construct a weighted sum of neighboring nodes for each relation region, and then these are aggregated and passed through a nonlinear function to generate the updated representation region of node i. The similarity weights were normalized using the softmax function. To adaptively learn the similarity between neighboring points, a self-attention mechanism used in graph neural networks is employed. The attention mechanism g is a single-layer feedforward neural network where parameters are shared among nodes and weighted by a weight vector. The unexpanded attention behavior is as follows: T represents transposition, || is the join operation, and the normalized attention coefficients are used to calculate the linear combination of the corresponding features. The characteristics of these combinations are selected as the final output features of each node. NL represents a nonlinear relationship. Based on the output of the graph attention network, stRGAT uses an unsupervised clustering algorithm to cluster points to identify differences in the spatial domain. The specific clustering method used is the Louvain method.

Images