A method and system for dual self-supervised clustering analysis of spatial transcriptomes

Abstract

The application provides a spatial transcriptome double self-supervision clustering analysis method and system, relates to the field of bioinformatics, and performs pretreatment on spatial transcriptome data to be analyzed to obtain sample data composed of gene expression data, cell image features and an adjacency matrix; the sample data is input into a trained double self-supervision model to perform clustering, and the clustering distribution of the sample data is obtained; according to the clustering distribution, a cell label is obtained as a final recognized cell type; the double self-supervision is self-supervision training of a linear autoencoder and training of a global target distribution supervised graph convolutional neural network encoder; through self-supervision training of the linear autoencoder and training of the global target distribution supervised graph convolutional neural network encoder, the training effect of the encoder is improved in a double self-supervision manner, and efficient and accurate classification and recognition of spatial transcriptome cells are realized.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and relates to a dual self-supervised clustering analysis method and system for spatial transcriptomes. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Spatial transcriptomics, as a pioneering technology, can not only obtain all gene activities in tissue samples, but also locate the coordinates of gene activities, which helps to better understand biological processes and diseases; among them, the discovery and determination of spatial transcriptomic cell types is one of the key issues in the study and analysis of life processes.

[0004] Traditional statistical models struggle to handle multimodal data, while existing machine learning methods do not effectively utilize spatial information. The emergence of graph neural networks has advanced network analysis, but their application in spatial transcriptomics is still in the exploratory stage, and currently cannot effectively utilize the high-dimensional information of each cell to accurately identify and analyze cell types.

[0005] Therefore, developing an efficient and accurate spatial transcriptome clustering analysis method has become a worthy research topic. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a dual self-supervised clustering analysis method and system for spatial transcriptomes. By using self-supervised training of a linear autoencoder and training of a global target distribution supervised graph convolutional neural network encoder, the training effect of the encoder is improved through dual self-supervision, thereby achieving efficient and accurate classification and identification of spatial transcriptome cells.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] The first aspect of this invention provides a dual self-supervised clustering analysis method for spatial transcriptomes;

[0009] A dual self-supervised clustering analysis method for spatial transcriptomics includes:

[0010] The spatial transcriptome data to be analyzed is preprocessed to obtain sample data consisting of gene expression data, cell image features, and adjacency matrix;

[0011] The sample data is input into a trained dual self-supervised model for clustering to obtain the cluster distribution of the sample data.

[0012] Based on the cluster distribution, cell tags are obtained as the final cell types to be identified;

[0013] The aforementioned dual self-supervision refers to the self-supervised training of the linear autoencoder and the training of the convolutional neural network encoder that constructs a global target distribution supervised graph.

[0014] Furthermore, the linear autoencoder takes cell gene expression and cell image features as input and outputs a first clustering distribution Q(z). Specifically, it uses the trained encoder to learn cell gene expression and cell image features to obtain fused coding features z, and calculates the first clustering distribution Q(z) based on the fused coding features z.

[0015] Furthermore, the graph convolutional neural network encoder takes the gene expression matrix after PCA dimensionality reduction, the adjacency matrix, and the fusion coding features learned by the linear autoencoder as input, and outputs the second clustering distribution Q(h). Specifically, the trained graph convolutional network is used to learn the adjacency matrix and the fusion coding features to obtain the graph embedding feature h, and the second clustering distribution Q(h) is calculated based on the graph embedding feature h.

[0016] Furthermore, the step of using the gene expression matrix, adjacency matrix, and fused coding features learned by the linear autoencoder after PCA dimensionality reduction as input involves connecting each layer of the linear autoencoder to the graph convolutional neural network encoder. The fused coding features learned by the linear autoencoder are then input into the graph convolutional neural network encoder for further representation learning.

[0017] Furthermore, the first clustering distribution Q(z) and the second clustering distribution Q(h) are calculated in the same way, both based on features, using Student's T distribution as the kernel function to calculate the similarity between features and cluster centers;

[0018] Based on the second cluster distribution Q(h), the second target distribution P(h) is calculated.

[0019] Furthermore, the self-supervised training of the linear autoencoder involves using the first KL divergence loss function, constructed from the reconstruction loss function and the first target distribution P(z), as the loss function of the linear autoencoder to supervise the training process of the linear autoencoder.

[0020] Furthermore, the training of the global target distribution supervised graph convolutional neural network encoder involves concatenating the first target distribution P(z) and the second target distribution P(h) to obtain the global target distribution P, and using the global target distribution P to construct the second KL divergence loss function, thus supervising the training process of the graph convolutional neural network encoder.

[0021] A second aspect of the present invention provides a dual self-supervised clustering analysis system for spatial transcriptomes.

[0022] A dual self-supervised clustering analysis system for spatial transcriptomics includes a data processing module, a distribution clustering module, and a type identification module.

[0023] The data processing module is configured to preprocess the spatial transcriptome data to be analyzed to obtain sample data consisting of gene expression data, cell image features, and adjacency matrix.

[0024] The distribution clustering module is configured to input sample data into a trained dual self-supervised model for clustering to obtain the cluster distribution of the sample data.

[0025] The type identification module is configured to obtain cell labels based on cluster distribution, which are then used as the final cell types to be identified.

[0026] The aforementioned dual self-supervision refers to the self-supervised training of a linear autoencoder and the training of a convolutional neural network encoder that constructs a global target distribution supervised graph.

[0027] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of a dual self-supervised clustering analysis method for spatial transcriptomes as described in the first aspect of the present invention.

[0028] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a dual self-supervised clustering analysis method for spatial transcriptomes as described in the first aspect of the present invention.

[0029] The above one or more technical solutions have the following beneficial effects:

[0030] This invention provides a dual self-supervised clustering analysis method and system for spatial transcriptomes. By using self-supervised training of a linear autoencoder and training of a convolutional neural network encoder that constructs a global target distribution supervised graph, the training effect of the encoder is improved through dual self-supervision, thereby achieving efficient and accurate classification and identification of spatial transcriptome cells.

[0031] The reconstruction loss function and the divergence loss function constructed from the target distribution are used to supervise the training of the linear autoencoder. The target distribution of the linear autoencoder and the target distribution of the graph convolutional neural network encoder are used to supervise the training of the graph convolutional neural network encoder, thereby improving the training efficiency and accuracy and improving the classification and recognition effect of spatial transcriptome cells as a whole.

[0032] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0033] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0034] Figure 1 This is a flowchart of the method in the first embodiment.

[0035] Figure 2 This is a spatial transcriptome organization image of the first embodiment.

[0036] Figure 3 This is a structural diagram of the dual self-supervised model for the first embodiment.

[0037] Figure 4 The diagram shows the recognition results of the five methods in the first embodiment.

[0038] Figure 5 This is a system structure diagram of the second embodiment. Detailed Implementation

[0039] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0040] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the invention; unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0042] Example 1

[0043] This embodiment discloses a dual self-supervised clustering analysis method for spatial transcriptomics;

[0044] like Figure 1 As shown, a dual self-supervised clustering analysis method for spatial transcriptomics includes:

[0045] Step S101: Preprocess the spatial transcriptome data to be analyzed to obtain sample data consisting of gene expression data, cell image features, and adjacency matrix.

[0046] Spatial transcriptome data is obtained by laying animal tissue flat on slides to obtain tissue images, such as... Figure 2As shown, information for each cell in the image is identified, including gene expression data, location information, and cell image. Preprocessing of the spatial transcriptome data includes:

[0047] (1) Processing gene expression data

[0048] Gene expression data is subjected to gene screening, regularization, and dimensionality reduction. Data with insufficient cell count is deleted, and then the gene expression data is regularized to obtain processed data X. PCA is used to reduce the dimensionality of X to obtain dimensionality-reduced data X'.

[0049] (2) Extracting cell image features

[0050] For cell groups within the same batch, given cells The two-dimensional spatial coordinates are (l xi ,l yi In images of highly expressed tissues, with (l xi ,l yi A 40x40 pixel region centered on the cell is used as the image data Y corresponding to each cell. i Select a suitable feature extractor, such as the Vision-Transformer (ViT) model, to extract image features U. i =VIT(Y) i ).

[0051] (3) Construct the adjacency matrix

[0052] Based on cell location information in spatial transcriptome data, the distance between each pair of cells is calculated, i.e., the Euclidean distance:

[0053]

[0054] When constructing an adjacency matrix, cells that are close to each other should have a larger weight, while cells that are far apart should have a smaller weight. Therefore, the formula for an adjacency matrix is:

[0055]

[0056] Among them, the hyperparameter t is introduced to set the degree to which the adjacency matrix is ​​clustered to the diagonal.

[0057] Step S102: Input the sample data into the trained dual self-supervised model for clustering to obtain the cluster distribution of the sample data;

[0058] like Figure 3 As shown, the constructed dual self-supervised model includes a graph convolutional neural network encoder (GAE) and a linear autoencoder (AE). The graph encoder contains two layers of GCN, and the AE contains an encoder with two linear layers and a decoder with two linear layers.

[0059] A linear autoencoder (AE) takes cell gene expression X and cell image features U as input and outputs the first cluster distribution Q(z). Specifically, it uses a trained encoder to learn the fused encoded features z from the cell gene expression X and cell image features U.

[0060]

[0061] Based on the fusion coding feature z, the first cluster distribution Q(z) is calculated.

[0062] A linear autoencoder (AE) concatenates gene expression data X with cell image features U, and then inputs this concatenation along with the adjacency matrix A into an encoder to obtain fused encoded features Z, i.e.:

[0063]

[0064] The fused encoded feature Z is input into the Decoder to reconstruct the original data, i.e.

[0065]

[0066] The Graph Convolutional Neural Network Encoder (GAE) takes the gene expression matrix after PCA dimensionality reduction, the adjacency matrix A, and the fused encoding features z learned by the linear autoencoder as input, and outputs the second clustering distribution Q(h). Specifically, it uses the trained graph convolutional network to learn the adjacency matrix A and the fused encoding features z to obtain the graph embedding features h, i.e.:

[0067]

[0068] The second clustering distribution Q(h) is calculated based on the graph embedding feature h.

[0069] Each layer of the linear autoencoder is connected to the graph convolutional neural network encoder. The fused encoding features learned by the linear autoencoder are added to the output of each convolutional layer of the graph convolutional neural network encoder and used as the input of the next convolutional layer.

[0070] The first cluster distribution Q(z) and the second cluster distribution Q(h) are calculated in the same way, both based on features, using Student's T distribution as the kernel function, and calculating the features and cluster centers μ. j The similarity between them; taking the graph embedding feature h as an example, the formula for the second cluster distribution Q(h) is:

[0071]

[0072] Where, μ j It can be initialized using the K-means algorithm.

[0073] Based on the first cluster distribution Q(z) and the second cluster distribution Q(h), the first target distribution P(z) and the second target distribution P(h) are calculated respectively. Taking the graph embedding feature h as an example, the second target distribution P(h) is:

[0074]

[0075] Using the same method, calculate the first cluster distribution Q(z) and the first target distribution P(z) of the fused coding feature z.

[0076] A dual self-supervised mechanism was designed. The first layer is a linear autoencoder (AE) that completes self-supervised training based on the first clustering distribution Q(z) and the first target distribution P(z). The second layer is the training of a global target distribution P-supervised graph convolutional neural network encoder (GAE).

[0077] Self-supervised training of a linear autoencoder involves applying the reconstruction loss function L... res The first KL divergence loss function L is constructed with the first target distribution P(z). zkl L, as the loss function of a linear autoencoder z The training process of a linear autoencoder is supervised.

[0078] The first KL divergence loss function is defined as the Kullback-Leibler (KL) divergence loss, that is:

[0079]

[0080] Reconstruction loss function L res The mean square error (MSE) is used, i.e.:

[0081]

[0082] Loss function L z The final result is: L z =L res +L zkl .

[0083] The global target distribution P trained by the supervised graphical convolutional neural network encoder (GAE) is obtained by concatenating the first target distribution P(z) and the second target distribution P(h). The second KL divergence loss function L is then constructed using the global target distribution P. hkl The training process of the supervised graph convolutional neural network encoder is as follows:

[0084] Construct a concatenation function to obtain the global target distribution P:

[0085] P = s(P(z), P(h))

[0086] Here, the concatenation function used is the mean function, and the second KL divergence loss function is L. hkl Defined as the Kullback-Leibler (KL) divergence loss, i.e.:

[0087]

[0088] Step S103: Based on the cluster distribution, obtain cell tags as the final cell types to be identified;

[0089] Gene expression sample data is input into a trained dual self-supervised model to calculate the target distribution of each cell, and cell labels are obtained through the softmax activation function to ultimately identify cell types.

[0090] The effectiveness of the analytical method in this embodiment was verified through comparative experiments. As comparative examples, existing methods used in this experiment included BayesSpace, SEDR, SpaGCN, and STAGATE.

[0091] BayesSpace: Employs a fully Bayesian statistical approach, using spatial priors to promote the belonging of adjacent cells to the same cluster.

[0092] SEDR: Employs deep autoencoder networks and graph autoencoders to embed spatial information.

[0093] SpaGCN: Improves clustering performance by combining gene expression, spatial location, and histological data to generate undirected weighted graphs that capture spatial dependencies.

[0094] STAGATE: Uses a graph attention autoencoder framework to perform spatial transcriptome data analysis by integrating spatial information and gene expression profiles.

[0095] The comparative experiment was conducted as follows: The four methods described above and the method described in this embodiment, STJL, were used to identify cell types in the same dataset. The adjusted Land coefficient (ARI) and normalized mutual information (NMI) were calculated. The experimental results are shown in the table below.

[0096] Table 1 Comparison of Clustering Results

[0097]

[0098] As shown in Table 1, STJL achieved the highest ARI and NMI values ​​among all methods, exhibiting the best clustering results; the recognition performance of the five methods is as follows: Figure 4 As shown, STJL clustering exhibits smoother boundaries, while clustering obtained using other comparison methods is more fragmented and contains significant noise.

[0099] Example 2

[0100] This embodiment discloses a dual self-supervised clustering analysis system for spatial transcriptomics;

[0101] like Figure 5 As shown, a dual self-supervised clustering analysis system for spatial transcriptomics includes a data processing module, a distribution clustering module, and a type identification module:

[0102] The data processing module is configured to preprocess the spatial transcriptome data to be analyzed to obtain sample data consisting of gene expression data, cell image features, and adjacency matrix.

[0103] The distribution clustering module is configured to input sample data into a trained dual self-supervised model for clustering to obtain the cluster distribution of the sample data.

[0104] The type identification module is configured to obtain cell labels based on cluster distribution, which are then used as the final cell types to be identified.

[0105] The aforementioned dual self-supervision refers to the self-supervised training of a linear autoencoder and the training of a convolutional neural network encoder that constructs a global target distribution supervised graph.

[0106] Example 3

[0107] The purpose of this embodiment is to provide a computer-readable storage medium.

[0108] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps in a dual self-supervised clustering analysis method for spatial transcriptomes as described in Embodiment 1 of this disclosure.

[0109] Example 4

[0110] The purpose of this embodiment is to provide an electronic device.

[0111] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a dual self-supervised clustering analysis method for spatial transcriptomics as described in Embodiment 1 of this disclosure.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dual self-supervised clustering analysis method for spatial transcriptomics, characterized in that, include: The spatial transcriptome data to be analyzed is preprocessed to obtain sample data consisting of gene expression data, cell image features, and adjacency matrix; The sample data is input into a trained dual self-supervised model for clustering to obtain the cluster distribution of the sample data; Based on the cluster distribution, cell tags are obtained, which serve as the final cell types for identification; The aforementioned dual self-supervision refers to the self-supervised training of the linear autoencoder and the training of the convolutional neural network encoder that constructs a global target distribution supervision graph. The linear autoencoder takes cell gene expression and cell image features as input and outputs a first clustering distribution Q(z). Specifically, it uses the trained encoder to learn cell gene expression and cell image features to obtain fused coding features z, and calculates the first clustering distribution Q(z) based on the fused coding features z. The graph convolutional neural network encoder takes the gene expression matrix after PCA dimensionality reduction, the adjacency matrix, and the fusion coding features learned by the linear autoencoder as input, and outputs the second clustering distribution Q(h). Specifically, it uses the trained graph convolutional network to learn the adjacency matrix and the fusion coding features to obtain the graph embedding feature h, and calculates the second clustering distribution Q(h) based on the graph embedding feature h. The self-supervised training of the linear autoencoder involves using the first KL divergence loss function, constructed from the reconstruction loss function and the first target distribution P(z), as the loss function of the linear autoencoder to supervise the training process of the linear autoencoder. The training of the supervised graph convolutional neural network encoder for constructing the global target distribution involves concatenating the first target distribution P(z) and the second target distribution P(h) to obtain the global target distribution P, and then using the global target distribution P to construct the second KL divergence loss function, which is the training process of the supervised graph convolutional neural network encoder.

2. The method of double self-supervised clustering analysis of spatial transcriptome according to claim 1, wherein, The process of using the gene expression matrix, adjacency matrix, and fused coding features learned by the linear autoencoder after PCA dimensionality reduction as input involves connecting each layer of the linear autoencoder to the graph convolutional neural network encoder. The fused coding features learned by the linear autoencoder are then input into the graph convolutional neural network encoder for further representation learning.

3. The method of double self-supervised clustering analysis of spatial transcriptome according to claim 1, wherein, The first cluster distribution Q(z) and the second cluster distribution Q(h) are calculated in the same way. They are both based on features and use the Student's T distribution as the kernel function to calculate the similarity between features and cluster centers. Based on the first cluster distribution Q(z) and the second cluster distribution Q(h), the first target distribution P(z) and the second target distribution P(h) are calculated respectively.

4. A system for dual self-supervised clustering analysis of spatial transcriptomes, characterized in that, It includes a data processing module, a distribution clustering module, and a type recognition module: The data processing module is configured to preprocess the spatial transcriptome data to be analyzed to obtain sample data consisting of gene expression data, cell image features, and adjacency matrix. The distribution clustering module is configured to input sample data into a trained dual self-supervised model for clustering to obtain the cluster distribution of the sample data. The type identification module is configured to obtain cell labels based on cluster distribution, which are then used as the final cell types to be identified. The aforementioned dual self-supervision refers to the self-supervised training of the linear autoencoder and the training of the convolutional neural network encoder that constructs a global target distribution supervision graph. The linear autoencoder takes cell gene expression and cell image features as input and outputs a first clustering distribution Q(z). Specifically, it uses the trained encoder to learn cell gene expression and cell image features to obtain fused coding features z, and calculates the first clustering distribution Q(z) based on the fused coding features z. The graph convolutional neural network encoder takes the gene expression matrix after PCA dimensionality reduction, the adjacency matrix, and the fusion coding features learned by the linear autoencoder as input, and outputs the second clustering distribution Q(h). Specifically, it uses the trained graph convolutional network to learn the adjacency matrix and the fusion coding features to obtain the graph embedding feature h, and calculates the second clustering distribution Q(h) based on the graph embedding feature h. The self-supervised training of the linear autoencoder involves using the reconstruction loss function and the first target distribution P(z) to construct a first KL divergence loss function as the loss function of the linear autoencoder, thereby supervising the training process of the linear autoencoder. The training of the global target distribution supervised graph convolutional neural network encoder involves concatenating the first target distribution P(z) and the second target distribution P(h) to obtain the global target distribution P, and using the global target distribution P to construct a second KL divergence loss function, thereby supervising the training process of the graph convolutional neural network encoder.

5. A computer-readable storage medium having stored thereon a program, characterized in that, When executed by a processor, the program implements the steps in a dual self-supervised clustering analysis method for spatial transcriptomics as described in any one of claims 1-3.

6. An electronic device comprising a memory, a processor, and a program stored in the memory and capable of running on the processor, characterized by When the processor executes the program, it implements the steps in the dual self-supervised clustering analysis method for spatial transcriptomics as described in any one of claims 1-3.

Images