A single cell identification method and device based on graph neural network self-supervised clustering

Abstract

The present invention discloses a single-cell identification method and device based on graph neural network self-supervised clustering, the method comprising: obtaining single cells from different batches, sequencing the single cells to obtain a cell gene count matrix, constructing a cell-gene graph network with the received single-cell transcriptome data; constructing multiple cell subgraphs, and merging the cell subgraphs to obtain a cell interaction graph network; obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, optimizing to obtain the optimal cell Gaussian latent variable representation, and performing self-supervised single-cell clustering to complete single-cell type identification. The present invention constructs a gene-cell interaction relationship network and cell representation and clusters the single-cell transcriptome data across batches, thereby achieving accurate identification of cell types and optimizing for issues such as batch effects and gene deletions, providing a guarantee for the accuracy of downstream analysis of single cells.

Description

Technical Field

[0001] The present invention relates to the field of novel biotechnology, and in particular to a single-cell identification method, system, terminal, and computer-readable storage medium based on graph neural network self-supervised clustering. Background Art

[0002] During the single-cell sequencing process, due to the low initial amount of RNA obtained from a single cell, single-cell sequencing has the characteristics of high noise and multiple zero values compared with existing sequencing methods. At the same time, single-cell sequencing technology and the biological characteristics of cells jointly generate a large amount of noise, further increasing the difficulty of cell type identification. In addition, the batch effect caused by the technical noise generated by the experimental design will confuse the biological differences between samples, complicate downstream analysis tasks, and have a great impact on the accuracy of downstream analysis tasks, thereby leading to incorrect interpretation of the analysis results. Faced with the missing values and batch effect characteristics of single-cell data, most of the existing development technologies are single-module solutions, and are not designed and optimized for clustering. In data filling work, some commonly used methods have difficulty handling data that deviate significantly from the multivariate normal distribution or non-uniform distribution. In addition, these methods all rely on pre-existing cell-cell or gene-gene correlations in single-cell data to calculate reasonable values.

[0003] The performance of existing batch effect processing methods is mostly affected by the processing order of multiple batches. Therefore, how to explore the optimization methods of data completion and error correction in clustering algorithms is the basis for achieving accurate single-cell annotation; and problems such as batch effects and gene deletions have a great impact on the accuracy of cell identification and cannot provide guarantees for accurate analysis of downstream tasks. Therefore, how to overcome the problems in existing technologies such as the great influence of different batches of experimental data on clustering accuracy, overfitting in fine-grained applications, and low clustering accuracy of unbalanced sample data, which make it impossible to guarantee the accuracy of downstream analysis of single cells, is an issue that needs to be urgently addressed in the current environment.

[0004] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0005] The main purpose of the present invention is to provide a single-cell identification method and equipment based on self-supervised clustering of graph neural networks, aiming to solve the problems in the existing technology that different batches of experimental data have a significant impact on clustering accuracy, overfitting exists in fine-grained applications, and the clustering accuracy of unbalanced sample data is low, resulting in the inability to guarantee the downstream analysis accuracy of single cells.

[0006] To achieve the above objectives, the present invention provides a single cell identification method based on graph neural network self-supervised clustering, which comprises the following steps:

[0007] Obtaining different batches of single cells, sequencing the single cells to obtain a cell gene count matrix, receiving single cell transcriptome data based on the cell gene count matrix, and constructing a cell-gene graph network based on the single cell transcriptome data;

[0008] constructing a plurality of cell subgraphs based on the cell-gene graph network, and merging the cell subgraphs to obtain a cell interaction graph network of a single cell;

[0009] Obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data;

[0010] The cell Gaussian latent variable representation is optimized to obtain an optimal cell Gaussian latent variable representation, and self-supervised single-cell clustering is performed on single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification.

[0011] Optionally, the single-cell identification method based on graph neural network self-supervised clustering, wherein the step of constructing multiple cell subgraphs based on the cell-gene graph network and merging the cell subgraphs to obtain a cell interaction graph network of a single cell, specifically includes:

[0012] When there are two cells in the cell-gene graph network that are both connected by edges to the associated genes, the weights of the two cells and the associated genes are summed and averaged to obtain the weights of the edges of the two cells, and a cell subgraph of the total number of genes in the cell-gene graph network is obtained based on the weights of the edges;

[0013] If the two cells are connected by edges in multiple cell subgraphs, the multiple cell subgraphs are merged, and the weights of the edges in the merged cell subgraphs are summed and averaged to obtain the weights of the edges of the two merged cells, and the cell interaction graph network of the single cell is obtained based on the weights of the merged edges.

[0014] Optionally, in the single-cell recognition method based on graph neural network self-supervised clustering, the variational autoencoder includes an encoder and a decoder; wherein the encoder and the decoder are both deep convolutional neural networks, and the network parameters of the deep convolutional neural network are calculated by a minimization formula; wherein the minimization formula is:

[0015] L(q)=E q [ln((|,ω))-(q(z|,ω)||(z,ω))];

[0016] Among them, L(q) is the loss function of the encoder, E qis the expectation of the encoder, x is the single-cell sequencing data, z is the low-dimensional latent variable, ω is the training parameter, p(x|z,ω) is to decode the low-dimensional latent variable and reconstruct the single-cell sequencing data, and KL((|,ω)||(z,ω)) is to encode the single-cell sequencing data into a low-dimensional latent variable.

[0017] Optionally, the single-cell identification method based on graph neural network self-supervised clustering, wherein the low-dimensional single-cell sequencing data of the cell interaction graph network is obtained, the low-dimensional single-cell sequencing data is input into a variational autoencoder for processing, and the cell Gaussian latent variable representation of the single-cell transcriptome data is obtained, specifically comprising:

[0018] Obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, and extracting low-dimensional latent variables of the low-dimensional single-cell sequencing data;

[0019] Modeling the conditional distribution of the low-dimensional latent variable based on the encoder to obtain a data model, and reconstructing the data model based on the decoder to obtain the distribution of single-cell sequencing data;

[0020] A confidence interval is calculated based on the distribution, and a cell Gaussian latent variable representation of the single-cell transcriptome data is obtained based on the confidence interval.

[0021] Optionally, the single-cell identification method based on graph neural network self-supervised clustering, wherein the cell Gaussian latent variable representation is optimized to obtain an optimal cell Gaussian latent variable representation, and the single cells are self-supervised clustered based on the optimal cell Gaussian latent variable representation to complete the single cell type identification, specifically includes:

[0022] Obtaining a gene feature matrix of a single cell in the cell graph network based on the cell Gaussian latent variable representation, and decomposing the feature matrix to obtain a plurality of independent graph signals, wherein the graph signal is an m-dimensional attribute signal describing the smoothness of the graph;

[0023] Calculating the smoothness of the graph signal according to a first formula, optimizing the smoothness according to a second formula, and calculating the smoothness of the characteristic vector in the cell graph network according to a Laplacian-Beltrami formula;

[0024] Obtaining an optimal cell Gaussian latent variable representation based on the optimized smoothness and the smoothness of the eigenvector, and clustering single cells based on the optimal cell Gaussian latent variable representation;

[0025] The loss of the clustering is calculated according to the third formula of the self-supervised method, the clustering is optimized based on the loss, and the type of the single cell is accurately identified based on the optimized clustering.

[0026] Optionally, in the single cell recognition method based on graph neural network self-supervised clustering, the first formula is:

[0027]

[0028] Among them, G is the cell graph network, L is the Laplacian matrix, and v i is the i-th attribute graph node, v j is the jth attribute graph node, f() is the output method, and ε is the edge set of the graph adjacency matrix;

[0029] The second formula is:

[0030]

[0031] Where X is the cell-based feature matrix, is the filtered version of the j-th image signal from X, U p is the eigenvector, Λ is the diagonal matrix of eigenvalues, m is a constant, p(λ i ) is obtained by scaling e i The value of e is used to retain the low-frequency signal and remove the high-frequency signal, i is the coefficient of the attribute graph signal, u i is the attribute graph signal;

[0032] The Laplacian-Beltrami formula is:

[0033]

[0034] Among them, Ω(u q ) is the eigenvector u q The smoothness of , ∑ is the summation formula, is the eigenvector u q The i-th element of is the eigenvector u q The jth element of a i,j is the i-th row and j-th column element of the graph adjacency matrix, d i is the dimension of the i-th row, d j is the dimension of the jth column, T is the matrix transpose, L s is the sth Laplace matrix, λ q is the characteristic value.

[0035] Optionally, in the single-cell recognition method based on graph neural network self-supervised clustering, the third formula is:

[0036]

[0037] in, is the clustering loss, λ tig is the cluster-to-cluster adversarial training parameter, λ seq is the clustering adversarial training parameter between clusters, is the intra-cluster loss, is the loss between clusters;

[0038] The loss within the cluster is:

[0039]

[0040] Among them, C is all clusters after clustering, c is a cluster among all clusters after clustering, v i is the i-th attribute graph node, v j is the jth attribute graph node, is the feature learned from the original data of row i, is the feature learned from the jth column of original data;

[0041] The loss between clusters is:

[0042]

[0043] Where n is a constant.

[0044] In addition, to achieve the above-mentioned purpose, the present invention further provides a single-cell identification system based on graph neural network self-supervised clustering, wherein the single-cell identification system based on graph neural network self-supervised clustering includes:

[0045] a cell-gene graph network construction module, configured to obtain single cells from different batches, sequence the single cells to obtain a cell gene count matrix, receive single-cell transcriptome data based on the cell gene count matrix, and construct a cell-gene graph network based on the single-cell transcriptome data;

[0046] A cell interaction graph network construction module is used to construct multiple cell subgraphs based on the cell and gene graph network, and merge the cell subgraphs to obtain a cell interaction graph network of a single cell;

[0047] a cell Gaussian latent variable representation generation module, configured to obtain low-dimensional single-cell sequencing data of the cell interaction graph network, input the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtain a cell Gaussian latent variable representation of the single-cell transcriptome data;

[0048] The self-supervised clustering single-cell module is used to optimize the cell Gaussian latent variable representation to obtain the optimal cell Gaussian latent variable representation, and perform self-supervised single-cell clustering on the single cells based on the optimal cell Gaussian latent variable representation to complete the single-cell type identification.

[0049] In addition, to achieve the above-mentioned purpose, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and a single-cell recognition program based on graph neural network self-supervised clustering stored on the memory and runnable on the processor, wherein the single-cell recognition program based on graph neural network self-supervised clustering implements the steps of the single-cell recognition method based on graph neural network self-supervised clustering as described above when executed by the processor.

[0050] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a single-cell recognition program based on graph neural network self-supervised clustering, and when the single-cell recognition program based on graph neural network self-supervised clustering is executed by a processor, the steps of the single-cell recognition method based on graph neural network self-supervised clustering as described above are implemented.

[0051] In the present invention, different batches of single cells are obtained, and the single cells are sequenced to obtain a cell gene count matrix, and single cell transcriptome data is received based on the cell gene count matrix, and a cell-gene graph network is constructed based on the single cell transcriptome data; multiple cell subgraphs are constructed based on the cell-gene graph network, and the cell subgraphs are merged to obtain a cell interaction graph network of a single cell; low-dimensional single-cell sequencing data of the cell interaction graph network is obtained, and the low-dimensional single-cell sequencing data is input into a variational autoencoder for processing to obtain a cell Gaussian latent variable representation of the single cell transcriptome data; the cell Gaussian latent variable representation is optimized to obtain an optimal cell Gaussian latent variable representation, and self-supervised single cell clustering is performed on the single cell based on the optimal cell Gaussian latent variable representation to complete the type identification of the single cell. The present invention constructs a gene-cell interaction relationship network through an algorithm model for single cell transcriptome data across batches, constructs a cell representation and clusters it, thereby achieving accurate identification of cell types and optimizing for batch effects, gene deletions and other problems, providing a guarantee for the downstream analysis accuracy of single cells. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of a preferred embodiment of the single cell recognition method based on graph neural network self-supervised clustering of the present invention;

[0053] Figure 2 Flowchart of step S20 in a preferred embodiment of the single cell recognition method based on graph neural network self-supervised clustering of the present invention;

[0054] Figure 3 Flowchart of step S30 in a preferred embodiment of the single cell recognition method based on graph neural network self-supervised clustering of the present invention;

[0055] Figure 4Flowchart of step S40 in a preferred embodiment of the single cell recognition method based on graph neural network self-supervised clustering of the present invention;

[0056] Figure 5 This is an overall flow chart of a preferred embodiment of the single cell identification method based on graph neural network self-supervised clustering of the present invention;

[0057] Figure 6 This is a schematic diagram of the principle of a preferred embodiment of the single-cell recognition system based on graph neural network self-supervised clustering of the present invention;

[0058] Figure 7 Schematic diagram of the operating environment of a preferred embodiment of the terminal of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clear and distinct, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0060] The single cell recognition method based on graph neural network self-supervised clustering described in the preferred embodiment of the present invention is as follows: Figure 1 As shown, the single cell recognition method based on graph neural network self-supervised clustering includes the following steps:

[0061] Step S10: Obtain single cells from different batches, sequence the single cells to obtain a cell gene count matrix, receive single cell transcriptome data based on the cell gene count matrix, and construct a cell-gene graph network based on the single cell transcriptome data.

[0062] Specifically, different batches of single cells are obtained, and the cell gene count matrix X_k(m×n) is obtained by sequencing the single cells in different batches, where the columns (cell) of the cell gene count matrix represent different cells, and the rows (gene) of the matrix represent the gene expression levels of the cells. In the sequencing data matrix of k batches, m is the total number of genes in the sequencing data, n is the total number of cells in the sequencing data, and m, n, k∈R, where R is a real number. For example, pbmc_10k_v3 and the GSE115189 dataset of the GEO open source database are selected as test data for analysis; biological data are input, where pbmc_10k_v3 contains 11769 columns (total number of cells). , 23036 rows (total number of genes), i.e. 11769 samples, 23036 features; while the GSE115189 dataset contains 3372 columns (total number of cells), 24654 rows (total number of genes), i.e. 3372 samples, 24654 features. This example only selects two human peripheral blood datasets. In actual applications, multiple datasets can be added as needed; and single-cell transcriptome data are input to the model using the cell gene count matrix method; if a gene in the single-cell transcriptome data is expressed in a cell, then an edge will be generated between the gene and the cell, and the weight of each edge is determined by the expression level of the gene in the cell, and a cell-gene graph network is constructed based on the expression level in the cell.

[0063] Step S20: construct multiple cell subgraphs based on the cell-gene graph network, and merge the cell subgraphs to obtain a cell interaction graph network of a single cell.

[0064] For detailed process, please refer to Figure 2 , which is a flowchart of step S20 in single-cell recognition based on graph neural network self-supervised clustering provided by the present invention.

[0065] like Figure 2 As shown, step S20 includes:

[0066] Step S21: When there are two cells in the cell-gene graph network that are both connected to the associated genes by edges, the weights of the two cells and the associated genes are summed and averaged to obtain the weights of the edges of the two cells, and a cell subgraph of the total number of genes in the cell-gene graph network is obtained based on the weights of the edges;

[0067] Step S22: If the two cells are connected by edges in multiple cell subgraphs, the multiple cell subgraphs are merged, the weights of the edges in the merged cell subgraphs are summed and averaged to obtain the weights of the edges of the two merged cells, and a cell interaction graph network of a single cell is obtained based on the weights of the merged edges.

[0068] Specifically, in the cell-gene graph network, if two cells are connected by edges to a certain gene (associated gene), then there is an edge connection between the two cells in the subgraph, and the weight of the edge is the summed average of the weights of the two cells and the associated genes in the cell-gene graph network. The cell subgraphs with a total number of genes, m, in the cell-gene graph network are obtained through the weight of the edge, and then the cell subgraphs are merged. The merging method is: if two cells are connected by edges in multiple cell subgraphs, then the multiple cell subgraphs are merged, and the weights of the edges in the merged cell subgraphs are summed and averaged to obtain the weights of the edges of the two cells after the merger, and based on the weights, a cell interaction graph network based on the gene expression relationship between cells is obtained.

[0069] Step S30: obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data.

[0070] For detailed process, please refer to Figure 3 , which is a flowchart of step S30 in single-cell recognition based on graph neural network self-supervised clustering provided by the present invention.

[0071] like Figure 3 As shown, step S30 includes:

[0072] Step S31, obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, and extracting low-dimensional latent variables of the low-dimensional single-cell sequencing data;

[0073] Step S32: Modeling the conditional distribution of the low-dimensional latent variable based on the encoder to obtain a data model, and reconstructing the data model based on the decoder to obtain the distribution of single-cell sequencing data;

[0074] Step S33: Calculate a confidence interval based on the distribution, and obtain a cell Gaussian latent variable representation of the single-cell transcriptome data based on the confidence interval.

[0075] Specifically, in the representation of Gaussian latent variables, it is assumed that each cell data in single-cell sequencing, that is, each graph node in the cell interaction graph network, has a low-dimensional latent variable z. The low-dimensional latent variable z has a pre-assumed Gaussian prior distribution and is independent of each other. The conditional distribution q(z|x) of the low-dimensional latent variable z based on the observed data X (single-cell sequencing data) is modeled through the encoder; and the distribution of the original data (i.e., single-cell transcriptome gene expression data) is reconstructed from the low-dimensional latent variable z through the decoder, i.e., p(x|z); wherein, in the deep variational autoencoder, the encoder and the decoder are both deep convolutional neural networks, and the parameters in the network are obtained by minimizing the evidence lower bound of the formula; wherein, the minimization formula is:

[0076] L(q)=E q [ln((|,ω))-(q(z|,ω)||(z,ω))]; where L(q) is the loss function of the encoder, E q is the expectation of the encoder, x is the single-cell sequencing data, z is the low-dimensional latent variable, ω is the training parameter, p(x|z,ω) is to decode the low-dimensional latent variable and reconstruct the single-cell sequencing data, KL((|,ω)||(z,ω)) is to encode the single-cell sequencing data into a low-dimensional latent variable; for the data distribution reconstructed by the decoder, the confidence interval can be calculated, and the low-dimensional latent variable z is used as the input of clustering. By fitting the distribution, similar cells can be placed at the most similar cluster centroid to balance the biological and technical differences between clusters; the distribution of similar cells is adjusted by back propagation to eliminate the batch effect, and the generation model is relied upon to fill in the missing values in the hidden layer to optimize the batch effect and missing value problems.

[0077] Step S40: Optimize the cell Gaussian latent variable representation to obtain an optimal cell Gaussian latent variable representation, and perform self-supervised single-cell clustering on the single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification.

[0078] For detailed process, please refer to Figure 4 , which is a flowchart of step S40 in single-cell recognition based on graph neural network self-supervised clustering provided by the present invention.

[0079] like Figure 4 As shown, the step S40 includes:

[0080] Step S41: obtaining a gene feature matrix of a single cell in the cell graph network based on the cell Gaussian latent variable representation, and decomposing the feature matrix to obtain a plurality of independent graph signals, wherein the graph signal is an m-dimensional attribute signal describing the smoothness of the graph;

[0081] Step S42: Calculate the smoothness of the graph signal according to the first formula, optimize the smoothness according to the second formula, and calculate the smoothness of the characteristic vector in the cell graph network according to the Laplacian-Beltrami formula;

[0082] Step S43: obtaining an optimal cell Gaussian latent variable representation based on the optimized smoothness and the smoothness of the eigenvector, and clustering single cells based on the optimal cell Gaussian latent variable representation;

[0083] Step S44: Calculate the loss of the clustering according to the third formula of the self-supervision method, optimize the clustering based on the loss, and accurately identify the type of the single cell based on the optimized clustering.

[0084] Specifically, the traditional graph convolution algorithm uses the estimated order for convolution, which makes the node representation over-smooth or under-smooth, affecting the performance of node clustering; in the algorithm of the present invention, for the cell graph network G, X is the gene feature matrix of the cell where x1,...,x n are all gene characteristics of cells, T is matrix transposition, is a real number, m×n is the dimension of the cell; the gene feature matrix of the cell can be decomposed into n independent graph signals, and a graph signal is an m-dimensional attribute signal describing the smoothness of the graph, that is, Among them, X is the gene feature matrix of the cell, u i is the graph signal, e i is the attribute graph signal u i The coefficient of |e i The size of | and the graph signal u i The smoothness of the graph signal can be quantified by the Laplace quadratic form of the graph: Among them, G is the cell graph network, L is the Laplacian matrix, and v i is the i-th attribute graph node, v j is the jth attribute graph node, f() is the output method, and ε is the edge set of the graph adjacency matrix; and the smoothness of the eigenvector in the cell graph network is calculated according to the Laplacian-Beltrami formula, and the Laplacian-Beltrami formula is: Among them, Ω(u q ) is the eigenvector u q The smoothness of , ∑ is the summation formula, is the eigenvector u q The i-th element of is the eigenvector u q The jth element of ai,j is the i-th row and j-th column element of the graph adjacency matrix, d i is the dimension of the i-th row, d j is the dimension of the jth column, T is the matrix transpose, L s is the sth Laplace matrix, λ q is the characteristic value.

[0085] Among them, the more orders of low-pass filtering are used to convolve the image signal, the smoother the image signal is, and the smoothness of the image signal is optimized, that is:

[0086] Where X is the cell-based feature matrix, is the filtered version of the j-th image signal from X, U p is the eigenvector, Λ is the diagonal matrix of eigenvalues, m is a constant, p(λ i ) is obtained by scaling e i The value of e is used to retain the low-frequency signal and remove the high-frequency signal, i is the coefficient of the attribute graph signal, u i is the attribute graph signal; based on the above steps, the order of the graph convolutional neural network can be measured and optimally selected by the smoothness of the graph signal, and this feature can be used to eliminate under-smoothing and over-smoothing in the graph convolution algorithm, so that single-cell data can be better represented and clustered. The main function of the obtained single-cell Gaussian latent variable representation is to remove the distribution influence of different batches on single-cell data and characterize single-cell data; then, combined with the upstream variational autoencoder module, the low-dimensional latent variable Gaussian distribution of the cell is dynamically adjusted through back propagation, thereby optimizing the batch effect problem of single-cell data.

[0087] Furthermore, since the clustering algorithm preferentially divides data into clusters of similar size, and "node pairs" with a long distance may belong to the same cluster, while "node pairs" with a short distance may belong to different clusters; in the single-cell clustering problem, it makes it difficult for cell types with a small number of cells to form independent clusters, but they are misclassified into other cell type clusters with a larger number; in addition, the distance is not the only criterion for measuring whether cells are of the same type, and other strategies are needed to provide additional supervision for cell clustering; in response to the above problems, the present invention adopts a self-supervised approach to improve clustering accuracy, obtains the loss of the cluster according to the self-supervised approach, optimizes the cluster based on the loss, and accurately identifies the type of single cells based on the optimized cluster, wherein the loss is calculated as follows:

[0088] in, is the clustering loss, λ tig is the cluster-to-cluster adversarial training parameter, λ seqis the clustering adversarial training parameter between clusters, is the intra-cluster loss, is the loss between clusters; the loss within the cluster is:

[0089] Among them, C is all clusters after clustering, c is a cluster among all clusters after clustering, v i is the i-th attribute graph node, v j is the jth attribute graph node, is the feature learned from the original data of row i, is the feature learned from the jth column of original data; the loss between clusters is:

[0090] Where n is a constant; the algorithm in the present invention clusters single cells, grouping single cells of the same type into a cluster, thereby achieving single cell identification. To a certain extent, it ensures that the distance between each cluster is maximized and the distance between each node within a cluster is minimized, so that the cells in each cluster are considered to belong to the same cell type.

[0091] Furthermore, if Figure 5 The overall flow chart of the single-cell identification method based on graph neural network self-supervised clustering is shown in the figure. Specifically, the single-cell transcriptome gene expression data is first input, that is, in the cell gene count matrix obtained by single-cell sequencing in different batches, the columns of the matrix represent different cells, and the rows of the matrix represent the gene expression levels of the cells; secondly, the cell interaction graph network is constructed: (1) Construct a cell-gene graph network. If a gene is expressed in a cell, then an edge will be generated between the gene and the cell. The weight of each edge is determined by the expression level of the gene in the cell; (2) Construct a cell subgraph: In the cell-gene graph network, if two cells are connected to a gene (associated gene), then there is an edge connection between the two cells in the subgraph. The weight of the edge is the sum of the weights of the two cells and the associated gene in the cell-gene graph network. and average; (3) cells that are connected by edges in multiple subgraphs are merged to obtain a cell interaction graph network based on the gene expression relationship between cells; then, the generative model is automatically optimized and adjusted during the back propagation process to remove batch effects and automatically generate missing values to obtain the optimal single-cell Gaussian latent variable representation of single-cell data; finally, based on the optimal single-cell Gaussian latent variable representation, self-supervised single-cell clustering is performed, that is, the model optimizes the association relationship in the single-cell relationship graph through a graph smoothing expression method based on graph signals; self-supervised clustering is performed on the cells in the single-cell relationship graph, and the algorithm in the model ensures that the distance between each cluster is maximized and the distance between each node in the cluster is minimized to a certain extent. The cells in each cluster are considered to belong to the same cell type, which provides a guarantee for the accuracy of downstream analysis of single cells.

[0092] Furthermore, if Figure 6 As shown, based on the above-mentioned single-cell identification method based on graph neural network self-supervised clustering, the present invention also provides a single-cell identification system based on graph neural network self-supervised clustering, wherein the single-cell identification system based on graph neural network self-supervised clustering includes:

[0093] a cell-gene graph network construction module 51 for obtaining single cells from different batches, sequencing the single cells to obtain a cell gene count matrix, receiving single cell transcriptome data based on the cell gene count matrix, and constructing a cell-gene graph network based on the single cell transcriptome data;

[0094] A cell interaction graph network construction module 52 is used to construct multiple cell subgraphs based on the cell and gene graph network, and merge the cell subgraphs to obtain a cell interaction graph network of a single cell;

[0095] a cell Gaussian latent variable representation generation module 53, configured to obtain low-dimensional single-cell sequencing data of the cell interaction graph network, input the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtain a cell Gaussian latent variable representation of the single-cell transcriptome data;

[0096] The self-supervised single-cell clustering module 54 is used to optimize the cell Gaussian latent variable representation to obtain the optimal cell Gaussian latent variable representation, and perform self-supervised single-cell clustering on the single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification.

[0097] Furthermore, if Figure 7 As shown, based on the above-mentioned single-cell identification method and system based on graph neural network self-supervised clustering, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 7 Only some of the components of the terminal are shown, but it should be understood that implementation of all of the shown components is not required, and more or fewer components may be implemented instead.

[0098] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory of the terminal. In other embodiments, the memory 20 may also be an external storage device of the terminal, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the terminal. Furthermore, the memory 20 may also include both an internal storage unit of the terminal and an external storage device. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code of the installation terminal. The memory 20 may also be used to temporarily store data that has been output or is to be output. In one embodiment, a single-cell recognition program 40 based on graph neural network self-supervised clustering is stored on the memory 20, and the single-cell recognition program 40 based on graph neural network self-supervised clustering can be executed by the processor 10, thereby realizing the single-cell recognition method based on graph neural network self-supervised clustering in the present application.

[0099] In some embodiments, the processor 10 can be a central processing unit (CPU), a microprocessor or other data processing chip, used to run the program code or process data stored in the memory 20, such as executing the single-cell recognition method based on graph neural network self-supervised clustering.

[0100] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other via a system bus.

[0101] In one embodiment, when the processor 10 executes the single cell recognition program 40 based on graph neural network self-supervised clustering in the memory 20, the following steps are implemented:

[0102] Obtaining different batches of single cells, sequencing the single cells to obtain a cell gene count matrix, receiving single cell transcriptome data based on the cell gene count matrix, and constructing a cell-gene graph network based on the single cell transcriptome data;

[0103] constructing a plurality of cell subgraphs based on the cell-gene graph network, and merging the cell subgraphs to obtain a cell interaction graph network of a single cell;

[0104] Obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data;

[0105] The cell Gaussian latent variable representation is optimized to obtain an optimal cell Gaussian latent variable representation, and self-supervised single-cell clustering is performed on single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification.

[0106] The step of constructing multiple cell subgraphs based on the cell-gene graph network and merging the cell subgraphs to obtain a cell interaction graph network of a single cell specifically includes:

[0107] When there are two cells in the cell-gene graph network that are both connected by edges to the associated genes, the weights of the two cells and the associated genes are summed and averaged to obtain the weights of the edges of the two cells, and a cell subgraph of the total number of genes in the cell-gene graph network is obtained based on the weights of the edges;

[0108] If the two cells are connected by edges in multiple cell subgraphs, the multiple cell subgraphs are merged, and the weights of the edges in the merged cell subgraphs are summed and averaged to obtain the weights of the edges of the two merged cells, and the cell interaction graph network of the single cell is obtained based on the weights of the merged edges.

[0109] The variational autoencoder includes an encoder and a decoder; wherein the encoder and the decoder are both deep convolutional neural networks, and the network parameters of the deep convolutional neural network are calculated by a minimization formula; wherein the minimization formula is:

[0110] L(q)=E q [ln((|,ω))-(q(z|,ω)||(z,ω))];

[0111] Among them, L(q) is the loss function of the encoder, E q is the expectation of the encoder, x is the single-cell sequencing data, z is the low-dimensional latent variable, ω is the training parameter, p(x|z,ω) is to decode the low-dimensional latent variable and reconstruct the single-cell sequencing data, and KL((|,ω)||(z,ω)) is to encode the single-cell sequencing data into a low-dimensional latent variable.

[0112] The step of obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data specifically includes:

[0113] Obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, and extracting low-dimensional latent variables of the low-dimensional single-cell sequencing data;

[0114] Modeling the conditional distribution of the low-dimensional latent variable based on the encoder to obtain a data model, and reconstructing the data model based on the decoder to obtain the distribution of single-cell sequencing data;

[0115] A confidence interval is calculated based on the distribution, and a cell Gaussian latent variable representation of the single-cell transcriptome data is obtained based on the confidence interval.

[0116] The step of optimizing the cell Gaussian latent variable representation to obtain an optimal cell Gaussian latent variable representation, and performing self-supervised single-cell clustering on single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification, specifically includes:

[0117] Obtaining a gene feature matrix of a single cell in the cell graph network based on the cell Gaussian latent variable representation, and decomposing the feature matrix to obtain a plurality of independent graph signals, wherein the graph signal is an m-dimensional attribute signal describing the smoothness of the graph;

[0118] Calculating the smoothness of the graph signal according to a first formula, optimizing the smoothness according to a second formula, and calculating the smoothness of the characteristic vector in the cell graph network according to a Laplacian-Beltrami formula;

[0119] Obtaining an optimal cell Gaussian latent variable representation based on the optimized smoothness and the smoothness of the eigenvector, and clustering single cells based on the optimal cell Gaussian latent variable representation;

[0120] The loss of the clustering is calculated according to the third formula of the self-supervised method, the clustering is optimized based on the loss, and the type of the single cell is accurately identified based on the optimized clustering.

[0121] Among them, the first formula is:

[0122]

[0123] Among them, G is the cell graph network, L is the Laplacian matrix, and v i is the i-th attribute graph node, v j is the jth attribute graph node, f() is the output method, and ε is the edge set of the graph adjacency matrix;

[0124] The second formula is:

[0125]

[0126] Where X is the cell-based feature matrix, is the filtered version of the j-th image signal from X, U p is the eigenvector, Λ is the diagonal matrix of eigenvalues, m is a constant, p(λ i ) is obtained by scaling e i The value of e is used to retain the low-frequency signal and remove the high-frequency signal, i is the coefficient of the attribute graph signal, u i is the attribute graph signal;

[0127] The Laplacian-Beltrami formula is:

[0128]

[0129] Among them, Ω(u q ) is the eigenvector u q The smoothness of , ∑ is the summation formula, is the eigenvector u q The i-th element of is the eigenvector u q The jth element of a i,j is the i-th row and j-th column element of the graph adjacency matrix, d i is the dimension of the i-th row, d j is the dimension of the jth column, T is the matrix transpose, L s is the sth Laplace matrix, λ q is the characteristic value.

[0130] Wherein, the third formula is:

[0131]

[0132] in, is the clustering loss, λ tig is the cluster-to-cluster adversarial training parameter, λ seq is the clustering adversarial training parameter between clusters, is the intra-cluster loss, is the loss between clusters;

[0133] The loss within the cluster is:

[0134]

[0135] Among them, C is all clusters after clustering, c is a cluster among all clusters after clustering, v i is the i-th attribute graph node, v j is the jth attribute graph node, is the feature learned from the original data of row i, is the feature learned from the jth column of original data;

[0136] The loss between clusters is:

[0137]

[0138] Where n is a constant.

[0139] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a single-cell recognition program based on graph neural network self-supervised clustering, and when the single-cell recognition program based on graph neural network self-supervised clustering is executed by a processor, the steps of the single-cell recognition method based on graph neural network self-supervised clustering as described above are implemented.

[0140] In summary, the present invention provides a single-cell identification method and device based on graph neural network self-supervised clustering, the method comprising: obtaining single cells from different batches, sequencing the single cells to obtain a cell gene count matrix, receiving single-cell transcriptome data based on the cell gene count matrix, and constructing a cell-gene graph network based on the single-cell transcriptome data; constructing multiple cell subgraphs based on the cell-gene graph network, and merging the cell subgraphs to obtain a cell interaction graph network of single cells; obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data; optimizing the cell Gaussian latent variable representation to obtain an optimal cell Gaussian latent variable representation, and performing self-supervised single-cell clustering on the single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type recognition. The present invention constructs a gene-cell interaction relationship network through an algorithm model for single-cell transcriptome data across batches, constructs cell representation and clusters it, thereby achieving accurate identification of cell types and optimizing for batch effects, gene deletions and other problems.

[0141] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or terminal comprising the element.

[0142] Of course, those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware (such as a processor, controller, etc.) through a computer program. The program can be stored in a computer-readable storage medium that can be read by a computer. When the program is executed, it can include the processes in the above-described method embodiments. The computer-readable storage medium can be a memory, a magnetic disk, an optical disk, etc.

[0143] It should be understood that the application of the present invention is not limited to the above examples. For those skilled in the art, improvements or changes can be made based on the above description. All these improvements and changes should fall within the scope of protection of the claims attached to the present invention.

Claims

1. A single cell recognition method based on graph neural network self-supervised clustering, characterized by: The single cell recognition method based on graph neural network self-supervised clustering includes: Obtaining different batches of single cells, sequencing the single cells to obtain a cell gene count matrix, receiving single cell transcriptome data based on the cell gene count matrix, and constructing a cell-gene graph network based on the single cell transcriptome data; constructing a plurality of cell subgraphs based on the cell-gene graph network, and merging the cell subgraphs to obtain a cell interaction graph network of a single cell; The step of constructing a plurality of cell subgraphs based on the cell-gene graph network and merging the cell subgraphs to obtain a cell interaction graph network of a single cell specifically includes: When there are two cells in the cell-gene graph network that are both connected by edges to the associated genes, the weights of the two cells and the associated genes are summed and averaged to obtain the weights of the edges of the two cells, and a cell subgraph of the total number of genes in the cell-gene graph network is obtained based on the weights of the edges; If the two cells are connected by edges in multiple cell subgraphs, the multiple cell subgraphs are merged, the weights of the edges in the merged cell subgraphs are summed and averaged to obtain the weights of the edges of the two merged cells, and a cell interaction graph network of the single cell is obtained based on the weights of the merged edges; Obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data; The cell Gaussian latent variable representation is optimized to obtain an optimal cell Gaussian latent variable representation, and self-supervised single-cell clustering is performed on single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification.

2. The single cell recognition method based on graph neural network self-supervised clustering according to claim 1, characterized in that The variational autoencoder includes an encoder and a decoder; wherein the encoder and the decoder are both deep convolutional neural networks, and the network parameters of the deep convolutional neural network are calculated by a minimization formula; wherein the minimization formula is: L(q)=E q [ln(p(x|z,ω))-KL(q(z|x,ω)||p(z,ω))]; Among them, L(q) is the loss function of the encoder, E q is the expectation of the encoder, x is the single-cell sequencing data, z is the low-dimensional latent variable, ω is the training parameter, p(x|z,ω) is to decode the low-dimensional latent variable and reconstruct the single-cell sequencing data, and KL(q(z|x,ω)||p(z,ω)) is to encode the single-cell sequencing data into a low-dimensional latent variable.

3. The single cell recognition method based on graph neural network self-supervised clustering according to claim 2, characterized in that: The step of obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, inputting the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtaining a cell Gaussian latent variable representation of the single-cell transcriptome data specifically includes: Obtaining low-dimensional single-cell sequencing data of the cell interaction graph network, and extracting low-dimensional latent variables of the low-dimensional single-cell sequencing data; Modeling the conditional distribution of the low-dimensional latent variable based on the encoder to obtain a data model, and reconstructing the data model based on the decoder to obtain the distribution of single-cell sequencing data; A confidence interval is calculated based on the distribution, and a cell Gaussian latent variable representation of the single-cell transcriptome data is obtained based on the confidence interval.

4. The single cell identification method based on graph neural network self-supervised clustering according to claim 1, characterized in that The optimizing the cell Gaussian latent variable representation to obtain the optimal cell Gaussian latent variable representation, and performing self-supervised single-cell clustering on single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification, specifically includes: A gene feature matrix of a single cell in the cell interaction graph network is obtained based on the cell Gaussian latent variable representation, and the gene feature matrix is decomposed to obtain multiple independent graph signals, wherein the graph signal is an m-dimensional attribute signal describing the smoothness of the graph; Calculating the smoothness of the graph signal according to the first formula, optimizing the smoothness according to the second formula, and calculating the smoothness of the characteristic vector in the cell interaction graph network according to the Laplacian-Beltrami formula; Obtaining an optimal cell Gaussian latent variable representation based on the optimized smoothness and the smoothness of the eigenvector, and clustering single cells based on the optimal cell Gaussian latent variable representation; The loss of the clustering is calculated according to the third formula of the self-supervised method, the clustering is optimized based on the loss, and the type of the single cell is accurately identified based on the optimized clustering.

5. The single cell identification method based on graph neural network self-supervised clustering according to claim 4 is characterized in that: The first formula is: Among them, G is the cell graph network, L is the Laplacian matrix, and v i is the i-th attribute graph node, v j is the jth attribute graph node, f() is the output method, and ε is the edge set of the graph adjacency matrix; The second formula is: Among them, X is the gene feature matrix of the cell, is the filtered version of the j-th image signal from X, U p is the eigenvector, Λ is the diagonal matrix of eigenvalues, m is a constant, p(λ i ) is obtained by scaling e i The value of e is used to retain the low-frequency signal and remove the high-frequency signal, i is the coefficient of the attribute graph signal, u i is the attribute graph signal; The Laplacian-Beltrami formula is: Among them, Ω(u q ) is the eigenvector u q The smoothness of , ∑ is the summation formula, is the eigenvector u q The i-th element of is the eigenvector u q The jth element of a i,j is the i-th row and j-th column element of the graph adjacency matrix, d i is the dimension of the i-th row, d j is the dimension of the jth column, T is the matrix transpose, L s is the sth Laplace matrix, λ q is the characteristic value.

6. The single cell identification method based on graph neural network self-supervised clustering according to claim 4, characterized in that The third formula is: in, is the clustering loss, λ tig is the cluster-to-cluster adversarial training parameter, λ seq is the clustering adversarial training parameter between clusters, is the intra-cluster loss, is the loss between clusters; The loss within the cluster is: Among them, C is all clusters after clustering, c is a cluster among all clusters after clustering, v i is the i-th attribute graph node, v j is the jth attribute graph node, is the feature learned from the original data of row i, is the feature learned from the jth column of original data; The loss between clusters is: Where n is a constant.

7. A single cell recognition system based on graph neural network self-supervised clustering, characterized by: The single-cell recognition system based on graph neural network self-supervised clustering is applied to the single-cell recognition method based on graph neural network self-supervised clustering according to any one of claims 1 to 6, and the single-cell recognition system based on graph neural network self-supervised clustering comprises: a cell-gene graph network construction module, configured to obtain single cells from different batches, sequence the single cells to obtain a cell gene count matrix, receive single-cell transcriptome data based on the cell gene count matrix, and construct a cell-gene graph network based on the single-cell transcriptome data; A cell interaction graph network construction module is used to construct multiple cell subgraphs based on the cell and gene graph network, and merge the cell subgraphs to obtain a cell interaction graph network of a single cell; a cell Gaussian latent variable representation generation module, configured to obtain low-dimensional single-cell sequencing data of the cell interaction graph network, input the low-dimensional single-cell sequencing data into a variational autoencoder for processing, and obtain a cell Gaussian latent variable representation of the single-cell transcriptome data; The self-supervised single-cell clustering module is used to optimize the cell Gaussian latent variable representation to obtain the optimal cell Gaussian latent variable representation, and perform self-supervised single-cell clustering on single cells based on the optimal cell Gaussian latent variable representation to complete single-cell type identification.

8. A terminal, characterized in that: The terminal includes: a memory, a processor, and a single-cell recognition program based on graph neural network self-supervised clustering stored in the memory and runnable on the processor. When the single-cell recognition program based on graph neural network self-supervised clustering is executed by the processor, the steps of the single-cell recognition method based on graph neural network self-supervised clustering as described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a single-cell recognition program based on graph neural network self-supervised clustering. When the single-cell recognition program based on graph neural network self-supervised clustering is executed by a processor, the steps of the single-cell recognition method based on graph neural network self-supervised clustering as described in any one of claims 1 to 6 are implemented.

Images