Graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structures

Abstract

The invention provides a graph attention network algorithm with global feature fusion for interpreting a spatial transcriptome structure, which integrates local and global information by constructing a spatial neighbor graph and a gene expression similar graph, and performs feature extraction and clustering by using a graph neural network and an attention mechanism. And accurate identification of the spatial domain is realized. Firstly, a spatial neighbor graph is constructed according to spatial position information of points; meanwhile, the cosine similarity of each point is calculated according to gene expression, and a gene expression similarity graph is constructed. Taking the two images as the input of an automatic image attention encoder, and extracting a low-dimensional feature vector with spatial information and gene expression through the encoder; and giving special attention to low-dimensional embedded feature vectors of the two graphs through an attention mechanism, and carrying out linear combination on the low-dimensional embedded feature vectors according to a certain weight. The spatial context and the gene expression data are effectively balanced by fusing the features of the spatial diagram and the expression diagram. And decoding and reversing back to an original feature space, and reconstructing a gene expression matrix. Finally, the learned low-dimensional embedding can be used for a series of downstream tasks, such as spatial domain identification.

Description

Technical Field

[0001] The present invention relates to the fields of bioinformatics and computational biology, and in particular to a spatial domain identification method for spatial transcriptomics data. Background Art

[0002] Spatial transcriptomics (ST) technology can study the gene expression levels of a single or several cells at different spatial locations, providing an important means for understanding the structural characteristics and functions of biological tissues. However, existing spatial domain recognition methods mainly rely on the gene expression profiles of spatially adjacent cells, ignoring the gene expression similarity in non-local space, which limits the accuracy of spatial domain recognition. Therefore, it is of great significance to develop a spatial domain recognition method that can simultaneously consider local and global gene expression similarities.

[0003] Problems with existing technologies

[0004] At present, some methods for identifying spatial domains, such as PROST, STitch3D, stMDA, GraphST, etc., use the spatial information of points to construct an adjacency graph and combine it with the gene expression profile as the input of the graph neural network. However, these methods ignore the influence of neighboring spots on the gene expression of a given spot. AVGN and DeepST use enhanced gene expression data based on the similarity of neighboring points and combine point coordinates to construct the adjacency matrix. STAGATE constructs a spatial neighbor graph and identifies spatial domains based on the pre-clustering of the gene expression neighbor graph.

[0005] Existing methods only build graph structures based on the premise that spatially adjacent cells have more similar gene expression profiles. This causes the graph network to prioritize local spatial information while failing to capture global context, ultimately limiting the network's ability to identify broader biological structural features. Summary of the invention

[0006] The present invention proposes a graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure, which is a new method designed to capture local and global information for accurate identification of ST data. It first constructs an expression graph that combines local spatial proximity and global gene expression similarity. Then an encoder with adaptive attention is used to assign optimal weights to adjacent nodes, enhancing the network's ability to extract relevant features. By linearly combining the features of spatial and expression graphs, GATES effectively balances spatial background and gene expression data. Demonstrating consistently high accuracy on datasets from different platforms, GATES demonstrates its robustness and adaptability. This strong performance emphasizes its potential to reveal biologically meaningful patterns and insights in different spatial transcriptomics technologies. In order to achieve the above objectives, the present invention adopts the following technical solutions:

[0007] Step 1: Preprocess the spatial transcriptomics data.

[0008] Step 2: Construct a spatial neighbor graph based on the spatial location information of the points.

[0009] Step three: construct a gene expression similarity graph based on gene expression information.

[0010] In step 4, these two images are used as input to the graph attention autoencoder to extract low-dimensional feature vectors with spatial information and gene expression.

[0011] Step 5: Use the attention mechanism of the graph neural network to linearly combine the two low-dimensional embedding vectors according to certain weights.

[0012] Step 6: Decode the final graph structure and invert it back to the original feature space to reconstruct the gene expression matrix.

[0013] Step seven: Use clustering algorithm to cluster the low-dimensional embedding vector to obtain the recognition result in the spatial domain.

[0014] Step 8, testing phase: Apply this algorithm to the spatial transcriptome dataset to identify spatial domains. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Flowchart of a graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure proposed in the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] See also Figure 1 The present invention proposes a graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure, wherein the method comprises the following steps:

[0018] Step 1: Preprocess the spatial transcriptomics data.

[0019] Data points outside the main tissue regions were removed to reduce noise and interference, logarithmic transformation and normalization were performed to enhance the linearity and stability of the data, while the bias caused by different sequencing depths was eliminated, and the top 3,000 genes with the highest mutation rates were selected to reduce the complexity of the analysis.

[0020] Step 2: Construct a spatial neighbor graph based on the spatial location information of the points.

[0021] In order to combine the similarities of adjacent points, we construct a spatial neighbor graph G based on the predefined radius and spatial location information spatial (V, E), select 6 neighboring points for each defining point. Let A be a spatial adjacency matrix of size |V| × |V|, where A(i, j) = 1 indicates that vertices i and j are adjacent, which means that their Euclidean distance is less than r. To account for self-loops, we expand A with the identity matrix I to form A * =I+A This ensures that both neighbor connections and self features are captured.

[0022] Step three: construct a gene expression similarity graph based on gene expression information.

[0023] To obtain global information, we constructed an undirected neighbor graph G based on gene expression similarity. gene (V, E). For each point, we calculate its gene expression similarity with all other points (e.g., using cosine similarity) and select the k points with the highest similarity scores as nearest neighbors to form the gene expression similarity adjacency matrix B.

[0024] The size of matrix B is |V|×|V|. If the similarity score between point i and point j is within the top k, then matrix B(i,j)=1, indicating that there is an edge between point i and point j.

[0025] To account for self-loops, we expand the matrix to B * =I+B, where I is the identity matrix, representing a gene expression graph with self-connection.

[0026] In step 4, these two images are used as input to the graph attention autoencoder to extract low-dimensional feature vectors with spatial information and gene expression.

[0027] Let L denote the number of layers in the encoder and S denote the number of points in a slice. The encoder mainly extracts a low-dimensional embedding representation of the input graph structure as d. At the kth layer (k∈{1,2,…,L-1}), the kth encoder layer generates the embedding of point i (i∈{1,2,…,}) as follows:

[0028]

[0029] Among them, W k is the trainable weight matrix of the encoder, N i is the neighbor set of point i, σ is the nonlinear activation function, is the edge weight of the normalized edge between point i and point j in the graph attention layer. In our method, the L-th encoder layer has no attention mechanism and is expressed as:

[0030]

[0031] The output of the last layer of the encoder is taken as the final embedding.

[0032] Step 5: Use the attention mechanism of the graph neural network to linearly combine the two low-dimensional embedding vectors according to certain weights.

[0033] The attention mechanism operates as a single-layer feed-forward neural network, and the shared parameters between nodes are defined by the weight vector. In the k-th encoder layer, the edge weight from node i to the adjacent node j is calculated as follows:

[0034]

[0035] in and is the trainable edge weight vector in the kth layer, which is applied to the transformation embedding of point i (source) and point j (receiver) respectively, performing different operations. In order to facilitate the comparison of coefficients between nodes, we normalize the choice of j:

[0036]

[0037] We employ attention mechanisms for two different types of networks: spatial similarity and gene expression similarity. We linearly combine the information learned at each layer to obtain the final information for each layer:

[0038]

[0039] att ij represents the edge weight between the i-th and j-th points in the fusion attention layer output, and and denote the edge weights between the i-th and j-th points in the learned spatial neighbor graph and gene expression similarity graph outputs, respectively. α is a hyperparameter that controls the linear combination of these features, enabling the model to effectively balance spatial context and gene expression data by fusing these two types of information.

[0040] Step 6: Decode the final graph structure and invert it back to the original feature space to reconstruct the gene expression matrix.

[0041] The decoder is responsible for reconstructing the latent representation into the original expression data. The embedding of the k-th layer decoder reconstructing point i at the (k-1) layer is defined as:

[0042]

[0043] represents the edge weight of the decoder attention layer, Represents the trainable weight matrix of the decoder. Similar to the encoder, the last layer abandons the use of the attention mechanism:

[0044]

[0045] To alleviate overfitting in our method, we set and

[0046] Step seven: Use clustering algorithm to cluster the low-dimensional embedding vector to obtain the recognition result in the spatial domain.

[0047] The method is characterized in that clustering analysis is performed using clustering algorithms such as mclust or Louvain.

[0048] Step 8, testing phase. Apply this algorithm to the spatial transcriptome dataset, identify the spatial domain according to the steps and perform a series of downstream tasks.

[0049] This paper proposes a more precise framework to train algorithms on spatial transcriptomic data. It aims to capture structural details in spatial transcriptomic data. It first constructs an expression graph that integrates local and global information using spatial proximity and gene expression similarity. Then, the model uses a graph attention autoencoder to assign appropriate edge weights to adjacent nodes, enhancing the model's feature extraction ability. By fusing the features of the spatial graph and the expression graph, the spatial background and gene expression data are effectively balanced to accurately identify the spatial domain.

[0050] In the description of this specification, the description of reference terms such as "embodiment" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment are included in at least one implementation of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment. Moreover, the specific features, structures, materials or characteristics described can be combined in any suitable manner in one or more embodiments.

[0051] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure, characterized by The following steps are involved: Step 1: Preprocess the spatial transcriptomics data. Step 2: Construct a spatial neighbor graph based on the spatial location information of the points. Step three: construct a gene expression similarity graph based on gene expression information. In step 4, these two graphs are used as input to the graph attention autoencoder to extract low-dimensional feature vectors with spatial information and gene expression. Step 5: Use the attention mechanism of the graph neural network to linearly combine the two low-dimensional feature vectors according to certain weights. Step 6: Decode the final graph structure and invert it back to the original feature space to reconstruct the gene expression matrix. Step seven: Use clustering algorithm to cluster the low-dimensional embedding vector to obtain the recognition result in the spatial domain. Step 8, testing phase: Apply this algorithm to the spatial transcriptome dataset to identify spatial domains.

2. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 1, characterized in that In the step 1, the preprocessing includes the following contents: Data points outside the main tissue regions were removed to reduce noise and interference, logarithmic transformation and normalization were performed to enhance the linearity and stability of the data, while the bias caused by different sequencing depths was eliminated, and the top 3,000 genes with the highest mutation rates were selected to reduce the complexity of the analysis.

3. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 2, characterized in that In the step 2, constructing a spatial neighbor graph includes the following contents: In order to combine the similarities of adjacent points, we construct a spatial neighbor graph G based on the predefined radius r and spatial location information spatial (V, E), select 6 neighboring points for each defining point. Let A be a spatial adjacency matrix of size |V| × |V|, where A(i, j) = 1 indicates that vertices i and j are adjacent, which means that their Euclidean distance is less than r. To account for self-loops, we expand A with the identity matrix I to form A * =I+A This ensures that both neighbor connections and self features are captured.

4. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 3, characterized in that In the step 3, constructing a gene expression similarity graph includes the following contents: To obtain global information, we constructed an undirected neighbor graph G based on gene expression similarity. gene (V, E). For each point, we calculate its gene expression similarity with all other points (e.g., using cosine similarity) and select the k points with the highest similarity scores as nearest neighbors to form the gene expression similarity adjacency matrix B. The size of matrix B is |V×V|. If the similarity score between point i and point j is within the top k, then matrix B(i,j)=1, indicating that there is an edge between point i and point j. To account for self-loops, we expand the matrix to B * =I+B, where I is the identity matrix, representing a gene expression graph with self-connection.

5. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 4, characterized in that In step 4, the graph attention autoencoder is used for feature extraction. Includes the following: Let L denote the number of layers in the encoder and S denote the number of points in a slice. The encoder mainly extracts a low-dimensional embedding representation of the input graph structure as d. At the kth layer (k∈{1,2,…,L-1}), the kth encoder layer generates the embedding of point i (i∈{1,2,…,}) as follows: Among them, W k is the trainable weight matrix of the encoder, N i is the neighbor set of point i, σ is the nonlinear activation function, is the edge weight of the normalized edge between point i and point j in the graph attention layer. In our method, the L-th encoder layer has no attention mechanism and is expressed as: The output of the last layer of the encoder is taken as the final embedding.

6. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 5, characterized in that In step 5, the attention mechanism of the graph neural network is used to linearly combine the two low-dimensional embedding vectors according to certain weights. The attention mechanism includes the following: The attention mechanism operates as a single-layer feed-forward neural network, and the shared parameters between nodes are defined by the weight vector. In the k-th encoder layer, the edge weight from node i to the adjacent node j is calculated as follows: in and is the trainable edge weight vector in the kth layer, which is applied to the transformation embedding of point i (source) and point j (receiver) respectively, and performs different operations. In order to facilitate the comparison of coefficients between nodes, we normalize them: We adopt an attention mechanism for two different types of networks: spatial similarity and gene expression similarity, which linearly combines the information learned from the two networks at each layer to obtain the final information of each layer: att ij represents the edge weight between the i-th and j-th points in the fusion attention layer output, and and denote the edge weights between the i-th and j-th points in the learned spatial neighbor graph and gene expression similarity graph outputs, respectively. α is a hyperparameter that controls the linear combination of these features, enabling the model to effectively balance spatial context and gene expression data by fusing these two types of information.

7. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 6, characterized in that In step 6, the final graph structure is decoded and inverted back to the original feature space to reconstruct the gene expression matrix, where the decoder content includes the following steps: The decoder is responsible for reconstructing the latent representation into the original expression data. The embedding of the k-th layer decoder reconstructing point i at the (k-1) layer is defined as: represents the edge weight of the decoder attention layer, Represents the trainable weight matrix of the decoder. Similar to the encoder, the last layer abandons the use of the attention mechanism: To alleviate overfitting in our method, we set and 8. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 7, characterized in that In step seven, cluster analysis is performed using clustering algorithms such as mclust or Louvain.

9. A graph attention network algorithm with global feature fusion for interpreting spatial transcriptome structure according to claim 8, characterized in that In step eight, the algorithm is applied to the spatial transcriptome dataset, which includes the following: Taking the spatial transcriptome data as input, the algorithm flow is followed to compare and analyze the identified spatial domain with the real spatial domain.

Images