ScRNA-seq data clustering method, system and device based on ZINB distribution and graph attention

Abstract

The invention provides an scRNA-seq data clustering method, system and device based on ZINB distribution and graph attention. The scRNA-seq data clustering system mainly comprises three core modules: a ZINB auto-encoder, which is used for modeling scRNA-seq data based on zero-expansion negative binomial distribution, generating robust potential representation through a denoising auto-encoder, and accurately capturing sparsity, excessive discreteness and shedding events of gene expression; the residual image attention auto-encoder is used for constructing a cell relation graph by using a Pearson correlation coefficient, dynamically learning a neighborhood weight in combination with a multi-head attention mechanism, retaining original features through residual connection, and relieving the excessive smoothness problem of image convolution; and a deep clustering model is self-optimized: target distribution and soft label distribution are minimized through KL divergence, and end-to-end joint optimization of embedded learning and clustering is realized.

Description

Technical Field

[0001] The present application belongs to the field of single-cell sequencing analysis technology, and in particular relates to a scRNA-seq data clustering method, system and device based on ZINB distribution and graph attention. Background Art

[0002] Cell sequencing is a high-throughput technology that can simultaneously measure the expression levels of hundreds of genes in hundreds to thousands of individual cells. Single-cell RNA sequencing (scRNA-seq) can analyze gene expression heterogeneity at single-cell resolution, providing an important means to reveal the diversity of cell types and functions. However, scRNA-seq data presents the following key challenges: high dropout rate and data sparsity: Due to technical limitations (such as low mRNA capture efficiency), a large number of "false zeros" (genes that are actually expressed but not detected) exist in the gene expression matrix. Traditional methods (such as PCA, t-SNE, and UMAP) have difficulty effectively distinguishing true and false zeros and filling missing values. Inadequate modeling of cell relationships: Existing methods (such as KNN graphs based on Euclidean distance) can only capture linear similarities and cannot represent complex nonlinear relationships between cells. Graph convolutional networks (GCNs) suffer from over-smoothing due to repeated Laplace smoothing, which mixes features from different clusters. Embedding and clustering are separated: In traditional multi-step methods (such as dimensionality reduction followed by clustering), the embedded representation may not be suitable for the subsequent clustering task, resulting in suboptimal results.

[0003] To solve these problems, ZINB autoencoders and graph neural networks have partially alleviated the above problems, but they still have the following limitations:

[0004] Inaccurate data distribution modeling: The existing ZINB model has limited ability to jointly model sparsity and overdispersion; the graph structure construction is not robust: the KNN graph based on Euclidean distance cannot reflect the true biological association; end-to-end optimization is missing: embedding learning and clustering objectives are not jointly optimized, affecting the final performance.

[0005] Therefore, there is an urgent need for a new method that can simultaneously address data sparsity, complex relationship modeling, and embedding-clustering collaborative optimization to improve the accuracy and robustness of scRNA-seq data clustering.

[0006] The above statements are only used to provide background technical information related to this application. Unless otherwise indicated herein, the contents described in this section are not prior art for the contents of other parts of this application. Summary of the Invention

[0007] The present invention proposes a scRNA-seq data clustering method, system and device based on ZINB distribution and graph attention, and provides a new model of scRNA-seq data clustering (scZGA), which provides a high-precision, robust and interpretable clustering solution for single-cell data analysis and solves the problems of the existing technology to a certain extent.

[0008] The clustering results show that this application implements sCZGA to consistently achieve higher scores on six real-world SCRNA-seq datasets by using evaluation metrics such as normalized mutual information (NMI) and adjusted Rand index (ARI).

[0009] According to a first aspect of an embodiment of the present application, a scRNA-seq data clustering method based on ZINB distribution and graph attention is provided, comprising the following steps:

[0010] Preprocess the original scRNA-seq data matrix, retain highly variable genes and perform normalization;

[0011] The preprocessed data matrix is modeled by the ZINB autoencoder, which simultaneously captures the sparsity, overdispersion, and dropout events of gene expression, and a robust latent representation Z is generated by the denoising autoencoder.

[0012] The Pearson correlation coefficient between cells is calculated using the potential representation Z, and the cell relationship graph G is constructed based on the Pearson correlation coefficient to obtain the weighted adjacency matrix ā;

[0013] The latent representation Z and the weighted adjacency matrix ā are input into the residual graph attention autoencoder, the neighborhood weights are dynamically learned through the multi-head attention mechanism, and the residual connection is used to retain the original features;

[0014] A self-optimizing deep clustering model is used to jointly optimize the output of the residual graph attention autoencoder, achieving end-to-end embedding learning and clustering by minimizing the KL divergence of the target distribution and the soft label distribution.

[0015] In some embodiments of the present application, ZINB autoencoder modeling includes:

[0016] Construct an encoder network that maps the preprocessed data into a low-dimensional latent space through three fully connected layers with ReLU activation function;

[0017] Construct a decoder network and predict the mean, dispersion and dropout probability of the ZINB distribution through independent fully connected layers. The formula is:

[0018]

[0019] ZINB(Xraw |π,μ,θ)=πδ0(X raw )+(1-π)NB(X raw |μ,θ)

[0020] Among them, X raw is the raw count matrix, μ, θ, and π represent the mean, dispersion, and dropout probability, respectively;

[0021] Finally, the model is optimized by the negative log-likelihood loss function to fill in missing values and generate robust latent representations.

[0022] In some embodiments of the present application, the construction of the cell relationship diagram is specifically as follows:

[0023] First, define a grid in the matrix data as G = (V, E), where V = {v i} i＝1,…,n It consists of a set of nodes, E={e ij} is the set of edges between nodes, where e ij Represents the connection node v i ; to node v j Edge, X∈R n×m is the node feature matrix, where x i Indicates v i The eigenvector of

[0024] (a) Based on the potential representation of the ZINB autoencoder, the inter-cell Pearson correlation coefficient matrix S is calculated as follows:

[0025]

[0026] in, represents the average expression value of the i-th cell.

[0027] (b) Take the top k highest correlation coefficient values for each row to generate the preliminary adjacency matrix P, the formula is:

[0028]

[0029] Among them, if S ij Belong to the first k highest values of the i-th row in the similarity matrix S, and P ij Set to 1; otherwise, P ij Set to 0;

[0030] (c) Obtain the adjacency matrix A and generate the final weighted adjacency matrix ā through symmetry processing and Hadamard product. The formula is as follows:

[0031]

[0032] Here, ⊙ represents the Hadamard product.

[0033] In some embodiments of the present application, the residual graph attention autoencoder includes an encoder GE; the encoder GE includes a graph attention layer and a graph convolution layer, specifically:

[0034] The graph attention layer dynamically allocates neighborhood weights through multi-head attention coefficients and connects the residual terms to retain the original features;

[0035] The graph convolution layer propagates information through the normalized adjacency matrix and outputs the node feature matrix Z.

[0036] In some embodiments of the present application, the residual graph attention autoencoder includes a node feature decoder D X and a neighbor decoder D A , specifically:

[0037] Node feature decoder D X Reconstruct the feature matrix through two fully connected layers The formula is:

[0038] Adjacent decoder D A Reconstruct the adjacency matrix through the inner product, the formula is:

[0039] In some embodiments of the present application, the self-optimizing deep clustering model is implemented by the following steps:

[0040] (a) Initialize the cluster center and perform k-means clustering on the latent representation Z of the graph autoencoder to generate the initial cluster center {u u};

[0041] (b) Calculate the soft label distribution and use Student’s t distribution to calculate the similarity between the embedded point and the cluster center:

[0042]

[0043] Among them, z i is the embedding point, u u is the initial cluster center;

[0044] (c) Generate target distribution by normalizing soft labels and optimizing clustering confidence iu , the formula is:

[0045]

[0046] (c) End-to-end optimization, minimizing the KL divergence loss to optimize the embedding space.

[0047] In some embodiments of the present application, the method further includes evaluating and visualizing the model, specifically including:

[0048] Calculate NMI, ARI and CA indicators to verify clustering performance;

[0049] Use UMAP to reduce the dimensionality of the latent representation to 2D space, visualize the cluster structure and calculate the silhouette coefficient.

[0050] According to a second aspect of an embodiment of the present application, a scRNA-seq data clustering system based on ZINB distribution and graph attention is provided, comprising a preprocessing unit, a ZINB autoencoder unit, a cell relationship graph construction unit, a residual graph attention autoencoder unit, and an end-to-end learning and clustering unit; specifically:

[0051] The preprocessing unit is used to preprocess the original scRNA-seq data matrix, retain highly variable genes and perform normalization;

[0052] ZINB autoencoder unit, which is used to model the preprocessed data matrix through the ZINB autoencoder, while capturing the sparsity, overdispersion and dropout events of gene expression, and generating a robust latent representation through the denoising autoencoder;

[0053] A cell relationship graph construction unit, used to construct a cell relationship graph based on the Pearson correlation coefficient;

[0054] The residual graph attention autoencoder unit is used to input the latent representation and the weighted adjacency matrix into the residual graph attention autoencoder, dynamically learn the neighborhood weights through the multi-head attention mechanism, and use residual connections to preserve the original features;

[0055] An end-to-end learning and clustering unit is used to jointly optimize the output of the residual graph attention autoencoder using a self-optimizing deep clustering model, achieving end-to-end embedding learning and clustering by minimizing the KL divergence of the target distribution and the soft label distribution.

[0056] According to a third aspect of an embodiment of the present application, a scRNA-seq data clustering device based on ZINB distribution and graph attention is provided, comprising:

[0057] a storage unit for storing executable instructions; and

[0058] A processing unit is configured to be connected to the memory to execute executable instructions to thereby complete any of the above methods.

[0059] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored; the computer program is executed by a processor to implement an scRNA-seq data clustering method based on ZINB distribution and graph attention.

[0060] A single-cell RNA sequencing (scRNA-seq) data clustering method (scZGA) based on ZINB distribution and graph attention network is adopted in this application, aiming to solve the data sparsity caused by high dropout rate in the existing technology, insufficient modeling of nonlinear relationships between cells, and embedding and clustering separation problems.

[0061] The scRNA-seq data clustering method, system, and device based on ZINB distribution and graph attention in this application mainly include three core modules:

[0062] ZINB Autoencoder: Models scRNA-seq data based on the zero-inflated negative binomial distribution and generates robust latent representations through a denoising autoencoder, accurately capturing sparsity, overdispersion, and dropout events in gene expression.

[0063] Residual Graph Attention Autoencoder: This uses the Pearson correlation coefficient to construct a cell relationship graph, combines it with a multi-head attention mechanism to dynamically learn neighborhood weights, and uses residual connections to preserve the original features, alleviating the over-smoothing problem of graph convolution.

[0064] Self-optimizing deep clustering model: By minimizing the KL divergence of target distribution and soft label distribution, it achieves end-to-end joint optimization of embedding learning and clustering.

[0065] Experiments show that this method improves the normalized mutual information (NMI) and adjusted Rand index (ARI) by 15%-25% on six real-world scRNA-seq datasets, including 10x_PBMC and Zeisel, and significantly outperforms existing methods (such as scDeepCluster and Seurat) in clustering accuracy (CA). Furthermore, UMAP visualization reveals that the latent space has closer intra-cluster distances and greater inter-cluster distances, and the silhouette coefficient (Silhouette) is 30%-50% higher than traditional methods (PCA and t-SNE). This invention provides a highly accurate, robust, and interpretable clustering solution for single-cell data analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0067] Figure 1 : A schematic diagram of the steps of a scRNA-seq data clustering method based on ZINB distribution and graph attention according to an embodiment of the present application is shown;

[0068] Figure 2 : A schematic diagram of the steps of ZINB autoencoder modeling according to an embodiment of the present application is shown;

[0069] Figure 3 : shows a schematic diagram of constructing a cell relationship diagram according to an embodiment of the present application;

[0070] Figure 4 : shows a schematic diagram of the principle of the self-optimizing deep clustering model according to an embodiment of the present application;

[0071] Figure 5 : A schematic diagram of the workflow of the scRNA-seq data clustering method according to an embodiment of the present application is shown;

[0072] Figure 6 : shows a clustering performance comparison diagram of different clustering methods according to an embodiment of the present application;

[0073] Figure 7 ] shows the NMI score model performance graph for different numbers of attention heads according to an embodiment of the present application;

[0074] Figure 8 ] shows the NMI score model performance graph under different GCN layer numbers according to an embodiment of the present application;

[0075] Figure 9 and Figure 10 FIG2 shows NMI score diagrams of six samples in three ablation experiments according to an embodiment of the present application;

[0076] Figure 11 ] shows a visualization result diagram of different methods according to embodiments of the present application;

[0077] Figure 12 : A schematic diagram of the structure of a scRNA-seq data clustering system based on ZINB distribution and graph attention according to an embodiment of the present application is shown;

[0078] Figure 13 A structural schematic diagram of the scRNA-seq data clustering device based on ZINB distribution and graph attention according to an embodiment of the present application is shown in FIG. DETAILED DESCRIPTION

[0079] Regarding this application, cellular sequencing is a high-throughput technology capable of simultaneously measuring the expression levels of hundreds of genes in hundreds to thousands of individual cells. This cellular-level resolution opens up new possibilities for effectively revealing the heterogeneity and functional diversity of previously unknown cell populations. Therefore, clustering cell subpopulations from heterogeneous tissues has become a key task in analyzing scRNA-Seq data.

[0080] However, accurately clustering cells in scRNA-seq datasets remains a major challenge. Due to the low efficiency of mRNA capture, scRNA-seq has a high dropout rate, resulting in a highly sparse gene expression matrix.

[0081] To address this problem, early researchers proposed several dimensionality reduction methods. For example, principal component analysis (PCA), T-distributed stochastic neighbor embedding (tSNE), and unified manifold approximation and projection (UMAP) PCA uses linear transformations to preserve key information and perform denoising. In particular, Spearman subsampling-clustering-classification (SSCC) uses principal component analysis to project subsampled units into the feature space. t-SNE is a nonlinear method that aims to maintain pairwise similarities between data points when projecting data into a low-dimensional space. Similar to tSNE, UMAP also preserves similarity but provides a better balance between local and global structures.

[0082] All of these methods increase information density and alleviate the problem of data scarcity to some extent. Nevertheless, not all zeros in the expression matrix are "true" zeros, meaning that these genes are expressed but not captured by SCRNA-seq. To infer these missing values, SCRMD uses a robust matrix factorization method by minimizing an objective function that includes low-rank and sparsity penalties. sclmpute uses a Gamma-Normal mixed model to infer dropout events and applies Lasso penalized linear regression for attribution. CIDR uses a mixed model of expressed and unexpressed gene states to estimate the true gene expression level.

[0083] As sequencing data increases, the true distribution of single-cell data is gradually becoming apparent. This provides more prior knowledge about the distribution, making regularized autoencoders a popular tool for inferring scRNA-seq data. VASC uses a variational autoencoder to fit single-cell data to a normal distribution. SCGNN uses a left-sectioned mixture of Gaussians to model single-cell data. In recent years, autoencoder models based on the zero-inflated negative binomial distribution have gained increasing recognition among researchers. Methods such as scDeepCluster, scV, and SCTPC use the ZINB model to encode scRNA-seq data, achieving promising results.

[0084] To better describe the relationships between global cell populations, graph representations and emerging graph neural network (GNN) techniques are increasingly being used. Seurat constructs a k-nearest neighbor (KNN) graph of cells. scGNN uses a graph convolutional network (GCN) to learn cell representations using graph structural information. Similarly, scGAC designs a graph attention autoencoder to propagate information between cells with different weights.

[0085] Although researchers have made significant progress in clustering cell types from single-cell RNA sequencing data, many questions remain.

[0086] These issues include the B.ZINB autoencoder for effectively constructing cell relationship graphs, the over-smoothing problem of graph convolutional networks, and the separation between embedding learning and clustering modules. The first problem is mainly due to the use of Euclidean distance in k-nearest neighbor (KNN) graphs, which only measures the linear similarity between cells and cannot capture the underlying relationships. Second, since graph convolution is essentially a form of Laplace smoothing, repeatedly applying Laplace smoothing may mix the features of vertices from different clusters and make them indistinguishable. Finally, a disadvantage of sequential multi-step clustering methods is that the learned embeddings may not be the best match for the subsequent clustering task.

[0087] To address these issues, this paper proposes a new model called single-cell clustering based on the ZINB model and graph attention network (scZGA). The scZGA model mainly consists of three components: distribution fitting, representation learning, and cell clustering.

[0088] The main technical effects are summarized as follows:

[0089] 1) The ScZG A model in this application introduces the Pearson correlation coefficient into the KNN algorithm to more effectively measure the similarity between cells instead of using the Euclidean distance.

[0090] 2) To avoid the over-smoothing problem inherent in GCNs, this application not only uses a graph attention network to learn the weights of the edges, but also uses residual connections to preserve the original features of the data.

[0091] 3) This application adapts traditional joint deep clustering methods to graph neural networks to learn representations that are conducive to clustering.

[0092] This application proposes a SCRNA-seq data clustering (scZGA) method, system, and device based on ZINB distribution and graph attention network. The new model scZGA consists of three key modules: the first module uses the ZINB model to capture the global probability structure; the second module constructs a graph using the Pearson correlation coefficient and adopts a graph autoencoder with residual connections to learn important neighborhood relationships while preserving topological structure information; the last module performs deep clustering through a self-optimized embedding algorithm. With these improvements, clustering results show that sCZGA consistently achieves higher scores on six real-world SCRNA-seq datasets by using evaluation metrics such as normalized mutual information (NMI) and adjusted Rand index (ARI).

[0093] In order to make the technical solutions and advantages of the embodiments of the present application more clearly understood, the exemplary embodiments of the present application are further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, and are not an exhaustive list of all the embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other unless they conflict.

[0094] Example 1

[0095] Figure 1 A schematic diagram of the steps of a scRNA-seq data clustering method based on ZINB distribution and graph attention according to an embodiment of the present application is shown in FIG.

[0096] like Figure 1 As shown in Figure 2, the scRNA-seq data clustering method based on ZINB distribution and graph attention includes the following steps:

[0097] S1: Preprocess the original scRNA-seq data matrix, retain highly variable genes and perform normalization;

[0098] S2: The preprocessed data matrix is modeled by a ZINB autoencoder, which simultaneously captures the sparsity, overdispersion, and dropout events of gene expression, and a denoising autoencoder is used to generate a robust latent representation.

[0099] S3: Calculate the Pearson correlation coefficient between cells using the potential representation Z, construct the cell relationship graph G based on the Pearson correlation coefficient, and obtain the weighted adjacency matrix ā;

[0100] S4: The latent representation Z and the weighted adjacency matrix ā are input into the residual graph attention autoencoder, the neighborhood weights are dynamically learned through the multi-head attention mechanism, and the residual connection is used to retain the original features;

[0101] S5: Jointly optimize the output of the residual graph attention autoencoder using a self-optimizing deep clustering model, achieving end-to-end embedding learning and clustering by minimizing the KL divergence of the target distribution and the soft label distribution.

[0102] Step 1 reduces technical noise (such as sequencing depth differences) through standardization, but cannot solve data sparsity (false zero problem). Step 2's ZINB autoencoder further models sparsity and overdispersion, fills in missing values, and generates a more robust potential representation.

[0103] Step 2 captures the global gene expression distribution (such as inter-gene covariance) but ignores the local relationship between cells. Step 4 dynamically learns the cell neighborhood weights (based on the Pearson correlation coefficient) through the graph attention mechanism, combined with residual connections to retain global features and achieve global-local information fusion.

[0104] Steps 1 and 2 jointly address data noise and sparsity. Based on this, step 4 uses graph structure to enhance cell relationship modeling, forming a complementary approach.

[0105] Overall, at the data level, steps 1 and 2 provide denoised node features and graph structure for step 4. Functionally, step 2 addresses global data distribution issues, while step 4 enhances local relationship modeling. These two work together to improve clustering accuracy. Optimization-wise, through end-to-end training, the representation learning in steps 2 and 4 is deeply coupled with the clustering objective, avoiding the suboptimal issues of traditional multi-step approaches.

[0106] This application provides a new model for scRNA-seq data clustering (scZGA), which provides a high-precision, robust and interpretable clustering solution for single-cell data analysis and solves the problems of the existing technology to a certain extent.

[0107] Figure 2 Schematic diagram of the steps of ZINB autoencoder modeling according to an embodiment of the present application is shown in FIG.

[0108] like Figure 2 As shown, the specific implementation of ZINB autoencoder modeling includes:

[0109] S21: Construct an encoder network to map the preprocessed data into a low-dimensional latent space through three fully connected layers with ReLU activation function;

[0110] S22: Construct a decoder network to predict the mean, dispersion, and dropout probability of the ZINB distribution through independent fully connected layers. The formula is:

[0111]

[0112] ZINB(X raw |π,μ,θ)=πδ0(X raw )+(1-π)NB(X raw |μ,θ)

[0113] Among them, X raw is the raw count matrix, μ, θ, and π represent the mean, dispersion, and dropout probability, respectively;

[0114] S23: Finally, the model is optimized by the negative log-likelihood loss function to fill in missing values and generate robust latent representations.

[0115] Figure 3 Schematic diagram of constructing a cell relationship diagram according to an embodiment of the present application is shown in FIG.

[0116] like Figure 3As shown in the figure, the specific steps for constructing the cell relationship diagram are:

[0117] First, define a grid in the matrix data as G = (V, E), where V = {v i} i＝1,…,n It consists of a set of nodes, E={e ij} is the set of edges between nodes, where e ij Represents the connection node v i ; to node v j Edge, X∈R n×m is the node feature matrix, where x i Indicates v i The eigenvector of

[0118] S31: Based on the potential representation Z of the ZINB autoencoder, calculate the inter-cell Pearson correlation coefficient matrix S, the formula is:

[0119]

[0120] in, represents the average expression value of the i-th cell.

[0121] S32: Take the first k highest correlation coefficient values for each row to generate a preliminary adjacency matrix P, the formula is:

[0122]

[0123] Among them, if S ij Belong to the first k highest values of the i-th row in the similarity matrix S, and P ij Set to 1; otherwise, P ij Set to 0;

[0124] S33: Obtain the adjacency matrix A, and generate the final weighted adjacency matrix ā through symmetry processing and Hadamard product. The formula is as follows:

[0125]

[0126] Here, ⊙ represents the Hadamard product.

[0127] Next, regarding the residual graph attention autoencoder, it includes an encoder GE and two decoders.

[0128] Regarding the encoder GE, it includes a graph attention layer and a graph convolution layer, specifically described as follows:

[0129] The graph attention layer dynamically allocates neighborhood weights through multi-head attention coefficients and connects the residual terms to retain the original features;

[0130] The graph convolution layer propagates information through the normalized adjacency matrix and outputs the node feature matrix Z.

[0131] Regarding two decoders, a node feature decoder D X and a neighbor decoder D A , the specific description is as follows:

[0132] Node feature decoder D X Reconstruct the feature matrix through two fully connected layers The formula is:

[0133] Adjacent decoder D A Reconstruct the adjacency matrix through the inner product, the formula is:

[0134] Figure 4 Schematic diagram of the principle of the self-optimizing deep clustering model according to an embodiment of the present application is shown in FIG.

[0135] like Figure 4 As shown in Figure 2, the self-optimizing deep clustering model is implemented through the following steps:

[0136] S51: Initialize the cluster center, perform k-means clustering on the latent representation Z of the graph autoencoder, and generate the initial cluster center {u u};

[0137] S52: Calculate the soft label distribution and use Student's t distribution to calculate the similarity between the embedded point and the cluster center:

[0138]

[0139] Among them, z i is the embedding point, u u is the initial cluster center;

[0140] S53: Generate target distribution by normalizing soft labels and optimizing clustering confidence iu , the formula is:

[0141]

[0142] S54: End-to-end optimization, minimizing KL divergence loss to optimize the embedding space.

[0143] Finally, the scRNA-seq data clustering method of this application also includes model evaluation and visualization. The specific steps include: calculating the NMI, ARI, and CA indicators to verify the clustering performance; using UMAP to reduce the dimensionality of the latent representation to 2D space, visualizing the cluster structure and calculating the silhouette coefficient.

[0144] In order to better illustrate the technical solution of this application, the following further implementation and explanation are provided through specific data.

[0145] In this example, we selected six real single-cell RNA sequencing (scRNA-Seq) datasets to validate the performance of the scZGA model. These datasets include: the 10X PBMC dataset, constructed using the 10X Genomics Chromium platform; the Zeisel dataset, generated using STRT-seq (single-cell tag reverse transcription sequencing); the Mouse_bladder_cell dataset, derived from Microwell-seq (microplate single-cell sequencing); and the Human1 and Human3 datasets, generated using the inDrop (integrated microfluidic single-cell RNA sequencing) platform, representing different samples from the same research project.

[0146] Key Terms: STRT-seq: single-cell tagged reverse transcription sequencing; Microwell-seq: microplate-based single-cell sequencing technology; inDrop: integrated microfluidics single-cell RNA sequencing platform; 10X Genomics Chromium: 10X Genomics' high-throughput single-cell sequencing platform. Table 1 details the experimental parameters and metadata for the real-world scRNA-seq datasets described above.

[0147] Table 1 Details of the evaluation dataset

[0148]

[0149] Figure 5 Schematic diagram of the workflow of the scRNA-seq data clustering method according to an embodiment of the present application is shown in FIG.

[0150] like Figure 5 As shown in the figure, the scRNA-seq data clustering method of this application involves three models. The first model is a ZINB autoencoder, which is used to fit the data distribution for data filling; the second model is a graph autoencoder, which combines graph attention and convolutional layers; and the third model is a deep clustering model integrated with the graph autoencoder to learn the optimal embedding.

[0151] In summary, the overall workflow is as follows: scZGA consists of three core models: ZINB autoencoder, graph autoencoder, and deep clustering model. preThis is used as input to a ZINB autoencoder, which constructs a cell-cell graph with edge weights determined by the Pearson correlation coefficient. Once the graph is constructed, it is processed by a graph autoencoder with residual connections and uses the latent representation from the graph autoencoder to perform deep clustering.

[0152] 1. Data preprocessing.

[0153] First, the SCANPY toolkit was used to screen and retain the top 2000 highly variable genes as node features. Second, the data matrix was logarithmically transformed and normalized using Z-score standardization, which scales each gene to unit variance and zero mean. The preprocessing matrix X pre The expression is:

[0154] X pre =zscore(log(X raw +1))

[0155] Where Xraw and Xpre are the raw count matrix and preprocessed matrix, and zscore is the z-score normalization.

[0156] 2. ZINB autoencoder.

[0157] In order to accurately capture the characteristics of sCRNA-seq data, including its significant sparsity and variance exceeding the mean, a denoising autoencoder based on the ZINB distribution is used. The parameters of the ZINB are the mean μ, dispersion θ, and loss probability π of the negative binomial distribution. The expression is as follows:

[0158]

[0159] ZINB(X raw |π,μ,θ)=πδ0(X raw )+(1-π)NB(X raw |μ,θ)

[0160] where X raw is the raw count matrix.

[0161] A denoising autoencoder (DAE) is an autoencoder that processes input data points with the goal of reconstructing the original, uncorrupted data points as its output.

[0162] To learn robust latent feature representations, a denoising autoencoder is applied to project the preprocessed matrix into a latent space, which is then used to construct the graph.

[0163] Corruption is induced by first applying random Gaussian noise to the input data:

[0164] Xcorrupt =X pre +ε;

[0165] Here, ε represents random Gaussian noise.

[0166] Next, the encoder is constructed by stacking three fully connected neural networks with rectifier activation. Different from the encoder, three independent fully connected layers are added in the decoder to estimate the ZINB parameters. The encoder function is defined as X = encoder (X pre ), the first two fully connected layers of the decoder are D = g(X), and the estimation layer can be described as:

[0167] M=exp(W μ D)

[0168] Θ=exp(W θ D)

[0169] П=sigmoid(W π D):

[0170] Where W represents the learnable weights; M, Θ, and П are parameter matrices representing the dropout probability, mean, and dispersion of the network output, respectively. The choice of activation function depends on the range and definition of the parameters. Since the dropout probability ranges from 0 to 1, we choose the sigmoid function. Furthermore, due to the non-negative nature of the mean and dispersion, we use the exponential function. The negative log-likelihood of the ZINB distribution is used as the loss function for the autoencoder based on the ZINB model and is defined as follows:

[0171] L ZINB = -log(ZINB(X raw |π,μ,θ))

[0172] 3. Construction of cell relationship diagram.

[0173] Define a grid in the matrix data as G = (V, E), where V = {v i} i＝1,…,n It consists of a set of nodes, E={e ij} is the set of edges between nodes, where e ij Represents the connection node v i ; to node v j Edge, X∈R n×m is the node feature matrix, where x i Indicates v i The eigenvector of .

[0174] To characterize the relationship between cells, this application constructs a cell graph as input to the subsequent graph autoencoder. Considering that the Pearson correlation coefficient better reflects the correlation between two random variables within a consistency benchmark, this embodiment of the application uses the Pearson correlation coefficient to measure the similarity of cells.

[0175] (a) Based on the potential representation of the ZINB autoencoder, the inter-cell Pearson correlation coefficient matrix S is calculated as follows:

[0176]

[0177] in, represents the average expression value of the i-th cell.

[0178] (b) Next, create a matrix P of the same size as S and take the top k highest correlation coefficient values for each row to generate the preliminary adjacency matrix P. The formula is:

[0179]

[0180] Among them, if S ij Belong to the first k highest values of the i-th row in the similarity matrix S, and P ij Set to 1; otherwise, P ij Set to 0;

[0181] (c) Obtain the adjacency matrix A and generate the final weighted adjacency matrix through symmetry processing and Hadamard product The formula is as follows:

[0182]

[0183] Where ⊙ represents the Hadamard product. The node features X are given by the ZINB autoencoder and the structural information ā, thus defining the cell graph G.

[0184] 4. Residual graph attention autoencoder.

[0185] The residual graph attention autoencoder of the embodiment of the present application includes an encoder GE and two decoders: a node feature decoder D X and the adjacent decoder D A .

[0186] Regarding the encoder GE, the encoder GE consists of a graph attention layer and a graph convolution layer.

[0187] Considering that the edges between clusters may cause improper message transmission and lead to over-smoothing, the embodiment of the present application uses residual connections to alleviate this problem.

[0188] The graph attention layer dynamically assigns neighborhood weights through multi-head attention coefficients and connects the residual terms to retain the original features. The formula of the graph attention layer can be expressed as:

[0189]

[0190] Among them, the attention coefficient α i，j Calculated by dot product:

[0191]

[0192] Similar to GAT, multi-head attention is then used to improve performance and learning stability:

[0193]

[0194] Where ‖ is the concatenation operation of the C-head attention operation; x i is the eigenvector of node i, x i ′ is the new feature vector N(i) representing the set of all nodes of neighbor node i.

[0195] The graph convolution layer propagates information through the normalized adjacency matrix and outputs the node feature matrix Z. The formula of the graph convolution layer is as follows:

[0196]

[0197] in, represents the weighted adjacency matrix, yes X' represents the node feature matrix output from the graph attention layer, and Z is the new node feature matrix after convolution.

[0198] Regarding two decoders, a node feature decoder D X and a neighbor decoder D A .

[0199] In the decoder part, Dx consists of two fully connected layers with LeakyReLU activation function and slope 0.2:

[0200]

[0201] in, is the reconstructed node feature matrix, and z is the output of the ZINB encoder.

[0202] Another decoder DA of the graph autoencoder is defined as the dot product of embeddings:

[0203]

[0204] in is the reconstructed adjacency matrix of A.

[0205] The goal of learning a graph autoencoder is to minimize the reconstruction loss, where γ is a coefficient that controls the relative weight of the two losses:

[0206]

[0207] 5. Deep graph embedding clustering.

[0208] The clustering algorithm is defined by the KL divergence between the distributions P and O:

[0209]

[0210] Total loss of end-to-end training: L total =L ZINB +L r +L c .

[0211] where Q is the soft label distribution of the Student’s t distribution.

[0212] t-distribution and P are the derivatives of the target distribution Q. iu is the embedding point Z i The soft label measures the similarity between the self-point and the cluster center and embeds u using the Student t distribution u :

[0213]

[0214] Generate initial cluster centers using k-means.

[0215] p iu is the target distribution, which can be calculated as follows:

[0216]

[0217] Since the target distribution P is defined in terms of Q, the embedding learning process of Q is self-optimizing to approach the target distribution P.

[0218] 6. Results

[0219] 1) Clustering performance comparison.

[0220] About indicators: NMI, ARI, CA.

[0221] Comparison methods: scDeepCluster, SSCC, Seurat, scGNN, scGAC.

[0222] Results: scZGA improves NMI by an average of 20% and ARI by 18% on six datasets (see Table 2-3), with CA reaching up to 82.1% ( Figure 6 ).

[0223] Figure 6 , which shows a comparison chart of clustering performance of different clustering methods according to an embodiment of the present application.

[0224] As shown in Table 2, scZGA performs best in terms of NMI scores on all datasets. Table 3 and Figure 2 It is shown that scZGA also outperforms all other methods in terms of ARI and CA metrics on all datasets.

[0225] Figure 6 Unlike the NMI and ARI metrics, CA utilizes information about the number of clusters. Since SCGNN cannot specify the number of clusters, it is not included in the calculation of the CA score.

[0226]

[0227] Table 2 NMI scores

[0228]

[0229] Table 3 ARI scores

[0230] 2) Parameter sensitivity analysis.

[0231] Number of attention heads: 5 heads achieves the best performance (NMI increases by 5%).

[0232] GCN layer number: Increasing the number of layers has little impact on performance (fluctuation < 2%).

[0233] Figure 7 ] shows the NMI score model performance diagram for different numbers of attention heads according to an embodiment of the present application. Figure 8 FIG shows the NMI score model performance diagram under different GCN layer numbers according to an embodiment of the present application.

[0234] like Figure 7 and Figure 8 As shown in Figure 3, the NMI score model performance under different numbers of attention heads and GCN layers. It is obvious that the number of attention heads has a greater impact on model performance, while the number of GCN layers has a smaller impact.

[0235] 3) Ablation experiment.

[0236] Ablation test In SCZGA, we combined ZINB autoencoder, graph autoencoder, encoder, and deep clustering model. The results show:

[0237] No_ZINB (ZINB module removed): NMI decreased by 12%;

[0238] No_DC (deep clustering removed): CA drops by 15%.

[0239] Figure 9 and Figure 10 FIG. 4 shows NMI score diagrams of six samples in three ablation experiments according to an embodiment of the present application.

[0240] like Figure 9 、 Figure 10 As shown, scZGA outperforms all experimental variants in terms of NMI.

[0241] 4) Visualization Analysis

[0242] For scZGA, we used UMAP for visualization, and all the results were as follows: Figure 5 shown.

[0243] UMAP visualization: scZGA latent space shows closer intra-cluster distances and farther inter-cluster distances (see Figure 5 );

[0244] Silhouette coefficient: 30%-50% higher than PCA and t-SNE (see Table 4).

[0245] Figure 11 2 shows the visualization results of different methods according to the embodiments of the present application.

[0246] like Figure 11 The following figure shows the visualization results of different methods. From top to bottom, the datasets are 10X_PBMC, Zeisel, mouse_bladder, Human3, Human1, and human_kidney.

[0247]

[0248] Table 4 Silhouette coefficients of four methods

[0249] For cell visualization, e.g. Figure 10 As shown in Figure 3, we also observed that scZGA has closer intra-group distances and farther inter-group distances than the other three methods. In the principal component analysis (PCA) visualization, the overall distribution of data points is chaotic; for t-SNE and UMAP, data points within the same cluster are closer, but the distances between different groups are insufficient.

[0250] Finally, we conclude that scZGA, through the deep integration of ZINB distribution modeling, graph attention mechanisms, and self-optimizing clustering, systematically addresses the high sparsity of scRNA-seq data, modeling complex cell relationships, and embedding-cluster separation. Experiments demonstrate that it significantly outperforms existing methods in clustering accuracy, robustness, and interpretability.

[0251] A single-cell RNA sequencing (scRNA-seq) data clustering method (scZGA) based on ZINB distribution and graph attention network is adopted in this application, aiming to solve the data sparsity caused by high dropout rate in the existing technology, insufficient modeling of nonlinear relationships between cells, and embedding and clustering separation problems.

[0252] The scRNA-seq data clustering method based on ZINB distribution and graph attention in this application mainly includes three core modules:

[0253] ZINB Autoencoder: Models scRNA-seq data based on the zero-inflated negative binomial distribution and generates robust latent representations through a denoising autoencoder, accurately capturing sparsity, overdispersion, and dropout events in gene expression.

[0254] Residual Graph Attention Autoencoder: This uses the Pearson correlation coefficient to construct a cell relationship graph, combines it with a multi-head attention mechanism to dynamically learn neighborhood weights, and uses residual connections to preserve the original features, alleviating the over-smoothing problem of graph convolution.

[0255] Self-optimizing deep clustering model: By minimizing the KL divergence of target distribution and soft label distribution, it achieves end-to-end joint optimization of embedding learning and clustering.

[0256] This application provides a new model for scRNA-seq data clustering (scZGA), which provides a high-precision, robust and interpretable clustering solution for single-cell data analysis and solves the problems of the existing technology to a certain extent.

[0257] The clustering results show that this application implements sCZGA to consistently achieve higher scores on six real-world SCRNA-seq datasets by using evaluation metrics such as normalized mutual information (NMI) and adjusted Rand index (ARI).

[0258] Example 2

[0259] This embodiment provides a scRNA-seq data clustering system based on ZINB distribution and graph attention. For details not disclosed in the scRNA-seq data clustering system based on ZINB distribution and graph attention in this embodiment, please refer to the specific implementation content of the scRNA-seq data clustering method based on ZINB distribution and graph attention in other embodiments.

[0260] Figure 12 A structural schematic diagram of the scRNA-seq data clustering system based on ZINB distribution and graph attention according to an embodiment of the present application is shown in FIG.

[0261] like Figure 12As shown in Figure 1, the scRNA-seq data clustering system based on ZINB distribution and graph attention includes a preprocessing unit 10, a ZINB autoencoder unit 20, a cell relationship graph construction unit 30, a residual graph attention autoencoder unit 40, and an end-to-end learning and clustering unit 50.

[0262] Specifically:

[0263] A preprocessing unit 10 is used to preprocess the original scRNA-seq data matrix, retain highly variable genes and perform normalization;

[0264] A ZINB autoencoder unit 20 for modeling the preprocessed data matrix through a ZINB autoencoder, while capturing the sparsity, overdispersion, and dropout events of gene expression, and generating a robust latent representation through a denoising autoencoder;

[0265] a cell relationship graph construction unit 30, configured to calculate the Pearson correlation coefficient between cells using the potential representation Z, construct a cell relationship graph G based on the Pearson correlation coefficient, and obtain a weighted adjacency matrix ā;

[0266] The residual graph attention autoencoder unit 40 is used to input the potential representation and the weighted adjacency matrix ā into the residual graph attention autoencoder, dynamically learn the neighborhood weights through the multi-head attention mechanism, and use residual connections to retain the original features;

[0267] The end-to-end learning and clustering unit 50 is used to jointly optimize the output of the residual graph attention autoencoder using a self-optimizing deep clustering model, and realize end-to-end embedding learning and clustering by minimizing the KL divergence target distribution and soft label distribution.

[0268] A single-cell RNA sequencing (scRNA-seq) data clustering method (scZGA) based on ZINB distribution and graph attention network is adopted in this application, aiming to solve the data sparsity caused by high dropout rate in the existing technology, insufficient modeling of nonlinear relationships between cells, and embedding and clustering separation problems.

[0269] The clustering results show that this application implements sCZGA to consistently achieve higher scores on six real-world SCRNA-seq datasets by using evaluation metrics such as normalized mutual information (NMI) and adjusted Rand index (ARI).

[0270] Example 3

[0271] This embodiment provides an scRNA-seq data clustering device. For details not disclosed in the scRNA-seq data clustering device of this embodiment, please refer to the specific implementation content of the scRNA-seq data clustering method or system based on ZINB distribution and graph attention in other embodiments.

[0272] Figure 13 Schematic diagram of the structure of the scRNA-seq data clustering device 400 according to an embodiment of the present application is shown in FIG.

[0273] like Figure 13 As shown, the scRNA-seq data clustering device 400 includes: a storage unit 402: for storing executable instructions; and a processing unit 401: for connecting with the storage unit 402 to execute the executable instructions to complete the scRNA-seq data clustering method based on ZINB distribution and graph attention.

[0274] Those skilled in the art will understand that Figure 13 The diagram is merely an example of the scRNA-seq data clustering device 400 and does not constitute a limitation of the scRNA-seq data clustering device 400 . The device may include more or fewer components than shown in the diagram, or a combination of certain components, or different components. For example, the scRNA-seq data clustering device 400 may also include input and output devices, network access devices, buses, and the like.

[0275] The so-called processing unit 401 (Central Processing Unit, CPU) can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor, or the processing unit 401 can also be any conventional processor. The processing unit 401 is the control center of the scRNA-seq data clustering device 400, and uses various interfaces and lines to connect various parts of the entire scRNA-seq data clustering device 400.

[0276] The storage unit 402 can be used to store computer-readable instructions. The processing unit 401 implements the various functions of the scRNA-seq data clustering device 400 by running or executing the computer-readable instructions or modules stored in the storage unit 402 and accessing the data stored in the storage unit 402. The storage unit 402 can mainly include a program storage area and a data storage area. The program storage area can store an operating system and at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the scRNA-seq data clustering device 400. In addition, the storage unit 402 can include a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a flash card, at least one disk storage device, a flash memory device, a read-only memory (ROM), a random access memory (RAM), or other non-volatile / volatile storage devices.

[0277] If the modules integrated into the scRNA-seq data clustering device 400 are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the present invention can also implement all or part of the processes in the above-mentioned method embodiments by instructing the relevant hardware through computer-readable instructions. The computer-readable instructions can be stored in a computer-readable storage medium. When executed by a processor, the computer-readable instructions can implement the steps of each of the above-mentioned method embodiments.

[0278] Example 4

[0279] This embodiment provides a computer-readable storage medium having a computer program stored thereon; the computer program is executed by a processor to implement the scRNA-seq data clustering method based on ZINB distribution and graph attention in other embodiments.

[0280] Those skilled in the art will appreciate that the terms used in the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. As used in the present invention and the appended claims, the singular forms "a," "the," and "the" are intended to include plural forms, unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.

[0281] It should be understood that although the terms "first," "second," "third," etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are merely used to distinguish information of the same type from one another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention. Depending on the context, the term "if" as used herein may be interpreted as "when," "when," or "in response to determining."

[0282] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0283] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A scRNA-seq data clustering method based on ZINB distribution and graph attention, characterized by: The following steps are involved: Preprocess the original scRNA-seq data matrix, retain highly variable genes and perform normalization; The preprocessed data matrix is modeled by the ZINB autoencoder, which simultaneously captures the sparsity, overdispersion, and dropout events of gene expression, and a robust latent representation Z is generated by the denoising autoencoder. Calculating the Pearson correlation coefficient between cells using the potential representation Z, constructing a cell relationship graph G based on the Pearson correlation coefficient and obtaining a weighted adjacency matrix ā; Input the potential representation Z and the weighted adjacency matrix ā into the residual graph attention autoencoder, and dynamically learn the neighborhood weights through a multi-head attention mechanism; A self-optimizing deep clustering model is used to jointly optimize the output of the residual graph attention autoencoder, achieving end-to-end embedding learning and clustering by minimizing the KL divergence target distribution and soft label distribution.

2. The data clustering method according to claim 1, characterized in that: The ZINB autoencoder modeling includes: Construct an encoder network that maps the preprocessed data into a low-dimensional latent space through three fully connected layers with ReLU activation function; Construct a decoder network and predict the mean, dispersion and dropout probability of the ZINB distribution through independent fully connected layers. The formula is: ZINB(X raw |π,μ,θ)=πδ0(X raw )+(1-π)NB(X raw |m,i) Among them, X raw is the raw count matrix, μ, θ, and π represent the mean, dispersion, and dropout probability, respectively; Finally, the model is optimized by the negative log-likelihood loss function to fill in missing values and generate robust latent representations.

3. The data clustering method according to claim 1, characterized in that: The construction of the cell relationship diagram is specifically as follows: First, define a grid in the matrix data as G = (V, E), where V = {v i } i＝1,…n It consists of a set of nodes, E={e ij } is the set of edges between nodes, where e ij Represents the connection node v i ; to node v j Edge, X∈R n×m is the node feature matrix, where x i Indicates v i The eigenvector of (a) Based on the potential representation of the ZINB autoencoder, the inter-cell Pearson correlation coefficient matrix S is calculated as follows: in, represents the average expression value of the i-th cell; (b) Take the top k highest correlation coefficient values for each row to generate the preliminary adjacency matrix P, the formula is: Among them, if S ij Belong to the first k highest values of the i-th row in the similarity matrix S, and P ij Set to 1; otherwise, P ij Set to 0; (c) Obtain the adjacency matrix A and generate the final weighted adjacency matrix ā through symmetry processing and Hadamard product. The formula is as follows: Here, ⊙ represents the Hadamard product.

4. The data clustering method according to claim 1, characterized in that The residual graph attention autoencoder includes an encoder GE; the encoder GE includes a graph attention layer and a graph convolution layer, specifically: The graph attention layer dynamically allocates neighborhood weights through multi-head attention coefficients and connects the residual terms to retain the original features; The graph convolution layer propagates information through the normalized adjacency matrix and outputs the node feature matrix Z.

5. The data clustering method according to claim 4, characterized in that: The residual graph attention autoencoder includes a node feature decoder D X and a neighbor decoder D A , specifically: Node feature decoder D X Reconstruct the feature matrix through two fully connected layers The formula is: Adjacent decoder D A Reconstruct the adjacency matrix through the inner product, the formula is:

6. The data clustering method according to claim 1, characterized in that: The self-optimizing deep clustering model is implemented by the following steps: (a) Initialize the cluster center and perform k-means clustering on the latent representation Z of the graph autoencoder to generate the initial cluster center {u u }; (b) Calculate the soft label distribution and use Student’s t distribution to calculate the similarity between the embedded point and the cluster center: Among them, z i is the embedding point, u u is the initial cluster center; (c) Generate target distribution by normalizing soft labels and optimizing clustering confidence iu , the formula is: (c) End-to-end optimization, minimizing the KL divergence loss to optimize the embedding space.

7. The data clustering method according to claim 1, characterized in that: It also includes evaluation and visualization of the model, including: Calculate NMI, ARI and CA indicators to verify clustering performance; Use UMAP to reduce the dimensionality of the latent representation to 2D space, visualize the cluster structure and calculate the silhouette coefficient.

8. A scRNA-seq data clustering system based on ZINB distribution and graph attention, characterized by: It includes a preprocessing unit, a ZINB autoencoder unit, a cell relationship graph construction unit, a residual graph attention autoencoder unit, and an end-to-end learning and clustering unit; specifically: The preprocessing unit is used to preprocess the original scRNA-seq data matrix, retain highly variable genes and perform normalization; The ZINB autoencoder unit is used to model the preprocessed data matrix through the ZINB autoencoder, while capturing the sparsity, overdispersion and dropout events of gene expression, and generating a robust latent representation Z through the denoising autoencoder; A cell relationship graph construction unit, configured to calculate the Pearson correlation coefficient between cells using the potential representation Z, construct a cell relationship graph G based on the Pearson correlation coefficient, and obtain a weighted adjacency matrix ā; A residual graph attention autoencoder unit is used to input the potential representation Z and the weighted adjacency matrix ā into the residual graph attention autoencoder, dynamically learn neighborhood weights through a multi-head attention mechanism, and retain the original features using residual connections; An end-to-end learning and clustering unit is used to jointly optimize the output of the residual graph attention autoencoder using a self-optimizing deep clustering model, and achieve end-to-end embedding learning and clustering by minimizing the KL divergence target distribution and soft label distribution.

9. A scRNA-seq data clustering device based on ZINB distribution and graph attention, characterized in that include: a storage unit for storing executable instructions; as well as A processing unit, configured to be connected to the memory to execute executable instructions to complete the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon; the computer program is executed by a processor to implement the method according to any one of claims 1 to 7.

Images