Twinborn clustering method and system for explaining single-cell RNA sequencing data

The single-cell RNA sequencing data is preprocessed and enhanced in character through twin clustering. Combined with twin information fusion network and denoising ZINB model, the problems of insufficient information exploration and low anti-noise ability between cells are solved, and efficient clustering and Large-scale data processing provides solid data support for downstream tasks in biology.

CN119068996AActive Publication Date: 2024-12-03COMP NETWORK INFORMATION CENT CHINESE ACADEMY OF SCI

Patent Information

Application Number
CN202410989937.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-12-03
Estimated Expiration
2044-07-23

AI Technical Summary

Technical Problem

Existing single-cell RNA sequencing data clustering methods have shortcomings in exploring intercellular information, anti-noise ability, and scalability of processing large data sets, resulting in low clustering accuracy and scalability.

Method used

Using twin clustering method, the K nearest neighbor map between cells is constructed by preprocessing and feature space enhancement of single-cell RNA sequencing data, and the information encoding and denoising ZINB model is used to encode and code, combined with self-supervised learning Strategies are clustered to enhance noise immunity and the ability to process large-scale data.

Benefits of technology

It achieves higher clustering accuracy and noise resistance, can effectively process large-scale data sets, provides a solid data foundation for downstream biology tasks, and improves data characterization capabilities and visualization effects.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068996A_ABST
    Figure CN119068996A_ABST
Patent Text Reader

Abstract

The invention discloses a twinborn clustering method and system for explaining single-cell RNA sequencing data, and belongs to the technical field of biological information. The method comprises the following steps: preprocessing single-cell RNA sequencing data to obtain a preprocessed gene expression matrix, and constructing an intercellular K nearest neighbor graph based on the preprocessed gene expression matrix; performing feature space enhancement on the preprocessed gene expression matrix to obtain a disturbance attribute matrix # imgabs0 #, and performing structure space enhancement on the K nearest neighbor graph between the cells to obtain an enhanced graph; and performing information fusion and encoding and decoding based on the disturbance attribute matrix and the enhancement graph to obtain a clustering result corresponding to the single-cell RNA sequencing data. The scRNA-seq data clustering method can effectively solve the problems that an existing scRNA-seq data clustering method is insufficient in inter-cell information exploration, low in data anti-noise capacity and poor in large data set expandability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and relates to a twin clustering method and system for interpreting single-cell RNA sequencing data. Background Art

[0002] Single-cell RNA sequencing (scRNA-seq) technology is a breakthrough in the field of bioinformatics. This technology can capture comprehensive genetic information within a single cell in parallel, providing higher-resolution data support for revealing cell heterogeneity and diversity. This is crucial for understanding complex biological systems and disease progression. In scRNA-seq data analysis, cluster analysis is crucial as its initial step. Grouping individual cells by gene expression patterns can reveal the complex characteristics of different cell populations, thereby providing deeper insights into their functions and interactions. In addition, cluster analysis of gene expression patterns of large-scale scRNA-seq data helps in wet experimental design in cell and gene-related studies, and promotes various biological downstream tasks, such as marker gene identification, cell type annotation, and developmental trajectory prediction.

[0003] Cluster analysis of scRNA-seq data has been a vibrant research field for decades. Due to the inherent characteristics of scRNA-seq data, such as high dimensionality, high sparsity, high noise, nonlinearity, and frequent missing events, the relationship between variables is complex and nonlinear. In recent years, the research focus has gradually shifted to applying deep learning frameworks to learn the representation of scRNA-seq data and classify samples into different clusters. Self-supervised learning methods are a typical method that reveals effective representations of scRNA-seq data by recreating the input data, such as DCA, scDeepCluster, and scNAME. However, these methods mainly focus on extracting features from single cells and ignore the key intercellular structural information required to accurately describe cell differences. Graph neural networks (GNNs) have been applied to analyze scRNA-seq data by making full use of graph information. They can effectively extract cell topology embeddings and achieve good performance, such as scGAE, scDSC, and scGNN.

[0004] Although GNN-based cell clustering methods have achieved encouraging results, they still face limitations in addressing the following challenges: (1) insufficient exploration of intercellular information; (2) insufficient data noise resistance; (3) low scalability to large datasets. These challenges have prompted researchers to explore a general scRNA-seq data clustering model for interpreting scRNA-seq data. By collecting basic knowledge from raw scRNA-seq data and extending it to various downstream tasks, we hope to make breakthrough progress in the above challenges. Summary of the invention

[0005] The purpose of the present invention is to provide a twin clustering method and system (hereinafter referred to as scSCF) for interpreting single-cell RNA sequencing data, which not only has higher clustering accuracy, but also has stronger noise resistance and the ability to process large-scale data, providing a solid data foundation for various biological downstream tasks. The present invention can effectively solve the problems faced by existing scRNA-seq data clustering methods, such as insufficient exploration of intercellular information, low data noise resistance, and poor scalability to large data sets.

[0006] To achieve the above-mentioned object of the invention, the technical solution of the present invention includes the following contents.

[0007] A twin clustering method for interpreting single-cell RNA sequencing data, the method comprising:

[0008] Preprocess the single-cell RNA sequencing data to obtain the preprocessed gene expression matrix and constructing a K nearest neighbor graph between cells based on the preprocessed gene expression matrix;

[0009] The preprocessed gene expression matrix is ​​enhanced in feature space to obtain a disturbance attribute matrix

[0010] Performing structural spatial enhancement on the K nearest neighbor graph between cells to obtain an enhanced graph; wherein the enhanced graph includes: the enhanced graph after edge removal And the enhanced image after diffusion operation A r Represents the enhanced image after edge removal The adjacency matrix, A d Represents the enhanced image after the diffusion operation The adjacency matrix of

[0011] Information fusion and encoding and decoding are performed based on the perturbation attribute matrix and the enhancement graph to obtain a clustering result corresponding to the single-cell RNA sequencing data.

[0012] Furthermore, the single-cell RNA sequencing data is preprocessed to obtain a preprocessed gene expression matrix, including:

[0013] Filtering the single-cell RNA sequencing data, the filtering comprising: filtering out genes expressed in less than one cell and cells in which the number of gene expressions is less than one;

[0014] The filtered data were normalized and transformed by logTPM to filter out genes with no counts in any cell;

[0015] Highly expressed genes were screened to construct the preprocessed gene expression matrix.

[0016] Furthermore, the K nearest neighbor graph between cells is structurally enhanced to obtain an enhanced graph after edge removal. include:

[0017] After calculating the sample pairwise cosine similarity matrix based on the preprocessed gene expression matrix, a mask matrix M is generated;

[0018] Compute the adjacency matrix with edges removed Where D is the degree matrix of the K nearest neighbor graph between cells, is the normalized adjacency matrix of the K nearest neighbor graph between the cells, and I represents the identity matrix;

[0019] Generate an enhanced image after edge removal

[0020] Furthermore, the K nearest neighbor graph between cells is structurally enhanced to obtain an enhanced graph after diffusion operation. include:

[0021] The normalized adjacency matrix is ​​transformed into Convert to adjacency matrix A d ;

[0022] Generate an enhanced image after diffusion operation

[0023] Furthermore, information fusion and encoding and decoding are performed based on the perturbation attribute matrix and the enhancement graph to obtain a clustering result corresponding to the single-cell RNA sequencing data, including:

[0024] Construct a twin information fusion network autoencoder and a FAE decoder based on the denoising ZINB model; wherein each branch of the twin information fusion network autoencoder is equipped with a feature autoencoder and a symmetry graph autoencoder;

[0025] The perturbation attribute matrix As the input of the feature autoencoder in each branch, the first feature representation matrix is ​​obtained respectively and the second eigenvalue matrix

[0026] The perturbation attribute matrix and the adjacency matrix A r As the input of the symmetric graph automatic decoder in one branch, the perturbation attribute matrix and the adjacency matrix A d As the input of the symmetric graph automatic decoder in another branch, the first symmetric graph representation matrix is ​​obtained The first adjacency matrix And the second symmetric graph representation matrix The second adjacency matrix

[0027] Fusion first feature representation matrix and the second eigenvalue matrix Get the feature representation matrix Z FAE ;

[0028] Fusion first symmetric graph representation matrix and the second symmetric graph characterization matrix Get the symmetric graph representation matrix Z SGAE ;

[0029] The feature representation matrix Z FAE and the symmetric graph representation matrix Z SGAE Perform weighted sum calculation to obtain the initial fusion representation matrix Z I ;

[0030] Based on the normalized adjacency matrix Merge the local structural information between data to enhance the initial fusion representation matrix Z I , to obtain the expression matrix Z L ; Among them, the normalized adjacency matrix A normalized adjacency matrix representing the K nearest neighbor graph between the cells;

[0031] The expression matrix Z is obtained by using the non-local relationship between data in the initial information fusion space through the autocorrelation learning mechanism. G ; Among them, the initial information fusion space is based on the perturbation attribute matrix Build;

[0032] For the expression matrix Z L and the expression matrix Z G Perform weighted sum calculation to obtain the fusion representation matrix Z;

[0033] Performing self-supervised clustering based on the optimal transmission strategy on the fusion representation matrix Z to obtain a clustering result corresponding to the single-cell RNA sequencing data;

[0034] The distribution of the fusion representation matrix Z is extracted and visualized to obtain a visual display of the clustering result.

[0035] Furthermore, the process of training the twin information fusion network autoencoder, the FAE decoder based on the denoising ZINB model, and the self-supervised clustering based on the optimal transmission strategy includes:

[0036] Based on the perturbation attribute matrix Reconstructing the weighted attribute matrix Reconstructing the weighted attribute matrix Normalized adjacency matrix And reconstruct the adjacency matrix Constructing a reconstruction loss function based on feature autoencoders and symmetric graph autoencoders Wherein, the reconstructed weighted attribute matrix is the output of the feature autoencoder based on the fusion representation matrix Z, and the reconstructed weighted attribute matrix and the reconstructed adjacency matrix It is the output of the symmetric graph autoencoder based on the fusion representation matrix Z;

[0037] Based on the two-layer mutual correlation reduction mechanism, a cell-level mutual correlation reduction loss function is constructed. and gene-level cross-correlation reduction loss function Among them, the cell-level mutual correlation information reduces the loss function Includes: Feature mutual correlation reduction Cell-level mutual correlation reduction loss function in the channel The cell-level mutual correlation reduction loss function in the channel is reduced by the mutual correlation reduction of the graph The gene-level cross-correlation reduces the loss function Includes: Feature mutual correlation reduction Gene-level mutual correlation reduction loss function in the channel The loss function for reducing the gene-level correlation in the channel and the graph correlation

[0038] Based on the fusion representation matrix Z and the normalized adjacency matrix Constructing the propagation regularization loss function

[0039] Integrate cell-level cross-correlation to reduce loss function Gene-level cross-correlation reduction loss function Reconstruction loss function and propagation regularization loss function Get the loss function of the twin information fusion network automatic encoder

[0040] By assuming that the single-cell RNA sequencing data follows the ZINB distribution, the overall negative likelihood loss of the FAE decoder based on the denoising ZINB model is constructed

[0041] Constructing a self-supervised clustering loss function based on the optimal transmission strategy

[0042] Based on the loss function The overall negative likelihood loss And the clustering loss function Complete the training of the twin information fusion network autoencoder and the FAE decoder based on the denoising ZINB model.

[0043] Furthermore, the cell-level mutual correlation information reduction loss function in the feature mutual correlation information reduction channel is constructed based on the double-layer mutual correlation information reduction mechanism. and gene-level cross-correlation reduction loss function include:

[0044] By calculating the first eigenvalue matrix The i-th dimension feature and the second feature representation matrix The cosine similarity between the j-th dimension features in To construct the matrix S cell ;

[0045] Constructing feature mutual correlation reduction channels to reduce the cell-level mutual correlation reduction loss function Where N represents the potential embedding dimension and I represents the identity matrix;

[0046] Use the Readout function to represent the first feature matrix and the second eigenvalue matrix Projection is performed to obtain the first gene-level feature representation matrix and the second gene-level feature representation matrix

[0047] By calculating the first gene-level feature representation matrix The i-th dimension feature and the second gene-level feature representation matrix The cosine similarity between the j-th dimension features in To construct the matrix S gene ;

[0048] Constructing gene-level mutual correlation reduction loss function in feature mutual correlation reduction channel Among them, d is the dimension of the Readout function, Represents the identity matrix.

[0049] Furthermore, the double-layer mutual correlation information reduction mechanism is used to construct a cell-level mutual correlation information reduction loss function in the graph mutual correlation information reduction channel. and gene-level cross-correlation reduction loss function include:

[0050] Characterize the matrix by computing the first symmetry graph The i-th dimension feature and the second symmetric graph representation matrix The cosine similarity between the j-th dimension features in To construct the matrix T cell ;

[0051] Constructing graph mutual correlation to reduce cell-level mutual correlation in channels reduces loss function Where N represents the potential embedding dimension and I represents the identity matrix;

[0052] Use the Readout function to represent the matrix of the first symmetric graph and the second symmetric graph characterization matrix Projection is performed to obtain the first gene-level graph representation matrix and the second gene-level graph representation matrix

[0053] By calculating the first gene-level graph representation matrix The i-th dimension feature and the second gene-level graph representation matrix The cosine similarity between the j-th dimension features in To construct the matrix T gene ;

[0054] Constructing gene-level mutual correlation reduction loss function in feature mutual correlation reduction channel Among them, d is the dimension of the Readout function, Represents the identity matrix.

[0055] Furthermore, the overall negative likelihood loss of the FAE decoder based on the denoising ZINB model is in, represents the preprocessed gene expression matrix, π ij is the probability of a loss event, with a mean value of μ ij and the dispersion θ ij denote the negative binomial components, respectively.

[0056] Furthermore, the self-supervised clustering loss function is constructed according to the optimal transmission strategy. include:

[0057] Get the clustering result distribution matrix Q;

[0058] Combining the clustering result distribution matrix Q, a function of constructing a transportation plan matrix P using optimal transport;

[0059] Using the Sinkhorn distance and solving the function of the transportation plan matrix P through entropy constraints, the optimal transportation plan matrix is ​​obtained.

[0060] Based on the clustering result distribution matrix Q and the optimal transmission plan matrix Constructing a self-supervised clustering loss function in, is the optimal transmission plan matrix The elements in q ij is the element in the clustering result distribution matrix Q.

[0061] A twin clustering system for interpreting single-cell RNA sequencing data, the system comprising:

[0062] Preprocessing module, used to preprocess single-cell RNA sequencing data to obtain the preprocessed gene expression matrix and constructing a K nearest neighbor graph between cells based on the preprocessed gene expression matrix;

[0063] A data enhancement module is used to enhance the feature space of the preprocessed gene expression matrix to obtain a disturbance attribute matrix Performing structural spatial enhancement on the K nearest neighbor graph between cells to obtain an enhanced graph; wherein the enhanced graph includes: the enhanced graph after edge removal And the enhanced image after diffusion operation A r Represents the enhanced image after edge removal The adjacency matrix, A d Represents the enhanced image after the diffusion operation The adjacency matrix of

[0064] The twin information fusion module is used to perform information fusion and encoding and decoding based on the perturbation attribute matrix and the enhancement graph to obtain the clustering results corresponding to the single-cell RNA sequencing data.

[0065] Compared with the prior art, the present invention provides a general clustering framework for interpreting scRNA-seq data based on Siamese graph autoencoder, which can effectively solve the challenges faced by existing clustering methods. It not only has higher clustering accuracy, but also has stronger noise resistance and the ability to process large-scale data, providing a solid data foundation for various biological downstream tasks. The present invention can effectively solve the problems faced by existing scRNA-seq data clustering methods, such as insufficient exploration of intercellular information, low data noise resistance, and poor scalability to large data sets. It has the following advantages in detail:

[0066] (1) The present invention has faster training speed and more accurate clustering results;

[0067] (2) The present invention has stronger noise resistance and the ability to process large-scale data sets;

[0068] (3) In the process of clustering scRNA-seq data, the present invention can effectively learn the data representation matrix and has powerful data representation capabilities;

[0069] (4) The invention can have good data visualization capabilities;

[0070] (5) The representation matrix learned by the present invention can support downstream tasks in a variety of biological fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Diagram of the framework of the twin clustering method for interpreting single-cell RNA sequencing data.

[0072] Figure 2 Overall flowchart of the twin clustering approach for interpreting single-cell RNA sequencing data.

[0073] Figure 3 Visualization results of the present invention and four competing methods on human liver cells. DETAILED DESCRIPTION

[0074] The following will be combined with the accompanying drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0075] The twin clustering method of the present invention for interpreting single-cell RNA sequencing data, such as Figure 1 and Figure 2 As shown, it includes three steps:

[0076] Step 1: Data standardization preprocessing. The original scRAN-seq data is preprocessed by the SCANPY package in Python: First, the present invention filters out genes that are expressed in less than 1 cell, and filters out cells with less than 1 gene expression; Next, the present invention standardizes the data and transforms the data through logTPM to filter out genes that are not counted in any cell; Then, the gene expression matrix of the top 1000 highly expressed genes is screened out. And the K nearest neighbor graph between cells is constructed based on the gene expression matrix obtained by preprocessing. Finally, the preprocessed gene expression matrix and the K nearest neighbor graph between cells are used as the input of the core algorithm part of scSCF.

[0077] Step 2: Core algorithm part. This part consists of three key components: (1) Data enhancement module: enrich the data basis of representation learning through data enhancement in feature space and structure space respectively; (2) Twin information fusion module: first, reduce the correlation between cells and genes in the two weight-sharing channels respectively. Secondly, integrate the complex information inherent in cells through the dynamic fusion mechanism of cross-modal information; (3) Self-supervised learning module: construct a self-supervised clustering loss function based on the optimal transmission strategy; (4) Joint optimization: jointly optimize the cell representation space through the self-supervised module based on the clustering loss of optimal transmission, ZINB loss and the loss of the twin information fusion module.

[0078] Step 3. Visualization of clustering results and model evaluation.

[0079] The present invention realizes efficient clustering of scRNA-seq data through the above-mentioned technical means, solves the scalability and noise resistance problems faced by existing methods when processing large-scale data, and provides a reliable data foundation for downstream biological tasks.

[0080] Specifically, the contents of each implementation step of the present invention are as follows.

[0081] 1. Data standardization preprocessing:

[0082] The original scRNA-seq count matrix was obtained and preprocessed by calling the SCANPY package in Python. First, the present invention filtered out genes expressed in less than 1 cell and cells with less than 1 gene expression; next, the data was standardized and converted by logTPM to filter out genes with no counts in any cell; then, the expression matrix of the top 1000 highly expressed genes was screened to obtain the final preprocessed gene expression matrix. x ij(1≤i≤N,1≤j≤D) represents the expression of the jth gene in the i-th cell; finally, the K-nearest neighbor (KNN) graph between cells is constructed based on the preprocessed gene expression matrix. Where V = {v1,v2,…v n} and E represent the node set and the edge set between nodes respectively. Each node is a separate cell, and the edge represents the similarity between each cell and other cells.

[0083] Construction method: The Pearson correlation coefficient is used to construct a KNN graph, which is characterized by the preprocessed gene expression matrix X and the adjacency matrix A = (a ij ) N×N , here when (v i ,v j )∈E, a ij =1, otherwise a ij = 0. The corresponding degree matrix is and By D -1 (A+I) Normalize the adjacency matrix A to get in is the identity matrix.

[0084] Assumptions is the preprocessed gene expression matrix The representation matrix, c=[c1,c2,…,c N ] is the cluster assignment. This method will learn the representation matrix Z and the cluster assignment c simultaneously.

[0085] 2. Data enhancement module:

[0086] Without affecting the information conveyed by the original scRNA-seq data, the present invention creates novel and reasonable data, which helps to enhance the robustness and generalization ability of the model. The data enhancement module is divided into feature space enhancement and structure space enhancement.

[0087] 2.1 Feature space enhancement: Feature space enhancement is also called feature perturbation. It perturbs the feature expression of nodes on the preprocessed gene expression matrix, for example, by adding Gaussian noise to obtain the perturbation attribute matrix. As shown in Formula 1.

[0088]

[0089] Among them, ⊙ represents the Hadamard product, and N represents the Gaussian distribution. The random noise matrix, is the preprocessed gene expression matrix.

[0090] 2.2 Structural Space Augmentation: Structural space augmentation operates on the graph structure by removing edges or using diffusion matrices to generate a global view.

[0091] (1) Remove edges: To achieve edge removal, a mask matrix is ​​generated based on the sample pairwise cosine similarity matrix calculated in the latent space. The 10% of edges with the lowest similarity values ​​will be manually removed. Finally, it is normalized to obtain the adjacency matrix after removing the edges. The calculation process is remember This is the enhanced image after edge removal.

[0092] (2) Graph Diffusion: The present invention uses Personalized PageRank (PPR) to normalize the adjacency matrix Convert to graph diffusion adjacency matrix A d α is the transmission probability, allowing the message to propagate through higher-order neighbor nodes. In this invention, α is set to 0.1. Diffusion adjacency matrix A d will be calculated as remember Enhanced image after diffusion operation.

[0093] 3. Twin information fusion module

[0094] The present invention introduces a Siamese Information Fusion Module (SIFM), which includes an autoencoder, a feature decoder based on a denoising ZINB model, a dual mutual correlation information reduction mechanism and a cross-modal dynamic information fusion mechanism, aiming to fully explore inter-cellular information and learn cross-feature discriminative sample representations.

[0095] 3.1 Autoencoder:

[0096] As shown in the framework diagram of the scSCF algorithm, the encoder part of the present invention adopts a twin network structure. Through parameter sharing, each branch is equipped with a feature autoencoder (FAE) and a symmetric graph autoencoder (SGAE) to extract attribute and structural information respectively. In the decoder part, the present invention uses a clustering-oriented latent embedding Z and a normalized adjacency matrix As input, the inputs of the two sub-networks are reconstructed through the decoders of FAE and SGAE respectively. Unlike traditional autoencoders, the decoder of FAE has three output layers, which uses a ZINB-based distribution to model highly sparse and over-dispersed scRNA-seq data. In addition, SGAE uses a symmetric architecture to reconstruct the weighted attribute matrix and adjacency matrix. This approach enhances the ability of SGAE to capture the structural and attribute characteristics of the graph, thereby achieving more accurate and comprehensive information reconstruction and significantly enhancing the generalization ability of the network.

[0097] Feature Autoencoder:

[0098] A feature autoencoder is proposed to learn the discriminative representation matrix of scRNA-seq data by stacking three layers of dense networks in the encoder and decoder. As input, reconstruct the weighted attribute matrix The present invention expresses the feature encoder and decoder functions as:

[0099]

[0100] Where b and b′ represent the bias vectors of the encoder and decoder respectively, W FAE and W′ FAE are the weights of the layer learned from network training, and f enc (·) and f dec (·) denote the encoder and decoder respectively.

[0101] Symmetric Graph Autoencoder:

[0102] Symmetric graph autoencoder learning from perturbed attribute matrices and the adjacency matrix A with edges removed r (or the diffusion adjacency matrix A d ) to learn the representation matrix containing the topological information between cells. The network needs to reconstruct the weighted attribute matrix during the decoding process and the adjacency matrix Furthermore, the formula for one layer in the encoder and decoder is:

[0103]

[0104] in and denote the learnable parameters of the lth encoder layer and the hth decoder layer, respectively. σ is a nonlinear activation function, such as ReLU or Tanh.

[0105] In order to minimize the reconstruction loss function on FAE and SGAE, the present invention aims to minimize the hybrid loss, as shown in Formula 4:

[0106]

[0107] Here, η is a predefined hyperparameter set to 0.1. is the joint mean square error reconstruction loss of the node attributes and graph structure described by minimizing To optimize the objective function.

[0108] 3.2 FAE decoder based on denoising ZINB model:

[0109] The ZINB distribution can simulate highly sparse and overdispersed count data. In order to explain the characteristics of scRNA-seq data, the present invention assumes that the data follows the ZINB distribution, that is:

[0110]

[0111] in, Represents the preprocessed gene expression matrix, which is the feature matrix of the input encoder; π ij is the probability of a missing event (i.e., the probability that a non-zero gene expression value is observed to be 0), with a mean value of μ ij and the dispersion θ ij denote the negative binomial component, δ0(·) is the Dirac function (when X ij =0,δ0(X ij )=1, otherwise 0).

[0112] The decoder has three output layers, which are used to estimate the probability estimates of parameters π, μ, and θ respectively. Mean estimate and dispersion estimates, respectively:

[0113]

[0114] W π , W μ , W θ is the corresponding weight, the size factor s i is the ratio of the total number of cells to the median. Finally, the present invention minimizes the overall negative likelihood loss of the feature decoder based on the ZINB model, i.e.

[0115]

[0116] 3.3 Dual inter-correlation information reduction mechanism:

[0117] The learning process of the two-layer cross-correlation information reduction mechanism includes two channels, namely, feature cross-correlation information reduction (FCR) and graph cross-correlation information reduction (GCR). Except for the input data, the basic operations of each channel are the same, including cell-level cross-correlation information reduction (cell-CR) and gene-level cross-correlation information reduction (gene-CR), aiming to minimize the deviation of the cross-correlation matrix from the ideal unit matrix and reduce the redundant information between nodes in the latent space, thereby improving the performance of the scRNA-seq data clustering algorithm with the help of structural information.

[0118] Next, we will take the feature cross-correlation information reduction (FCR) channel as an example:

[0119] (1) Reduction of cell-level correlation information:

[0120] The cell-CR learning process consists of two steps. First, given the feature encoder learned feature representation matrix under two views and The present invention firstly Calculate the cosine similarity of the cell dimension, It represents the cosine similarity between the i-th dimension feature in the first view and the j-th dimension feature in the second view.

[0121] After that, this method makes the cell-CR matrix S cell Approximately equal to the identity matrix Taking the cell-CR objective loss function as the FCR channel, the formula is as follows:

[0122]

[0123] where N is the potential embedding dimension (also the number of samples). By ensuring that S cellThe diagonal elements of are approximately equal to 1, while the off-diagonal elements are approximately equal to 0. The present invention can ensure the embedding alignment of each cell in the two views and minimize the embedding consistency of different cells in different views. This helps the present method reduce redundant information in the latent space, thereby learning a more discriminative representation.

[0124] (2) Reduction of gene-level correlation information:

[0125] In addition to constructing a meaningful representation matrix by reducing the cellular correlation between two views, the present invention also refines the information correlation from the gene dimension. Specifically, gene-CR is performed in three steps.

[0126] In the first step, this method uses the Readout function The cell representation matrix under two views and Projection to gene-level representation matrix and The calculation formula is:

[0127]

[0128] In the second step, the present invention is again carried out by Calculate the gene-CR matrix S gene Each element in It represents the feature similarity between the i-th dimension feature in one view and the j-th dimension feature in another view.

[0129] In the third step, according to the target loss function target equation of the gene-CR of the FCR channel listed in formula 10, the present invention makes the gene-CR matrix S gene Equal to the identity matrix

[0130]

[0131] Where d is the potential embedding dimension. In Formula 10, Emphasizes bringing similar features closer together, Emphasis sets different features apart.

[0132] Subsequently, similar to the operation of the FCR channel, the present invention characterizes the matrix based on the symmetric graph and symmetric graph representation matrix Get the two target loss functions of the GCR channel and

[0133] The present invention linearly combines the objective functions of FCR and GCR to obtain and

[0134] 3.4 Cross-modal dynamic information fusion mechanism:

[0135] In order to filter redundant features and retain more discriminative features of scRNA-seq data, the present invention introduces a cross-modal dynamic information fusion strategy, which enables the scSCF network to learn more discriminative representations to avoid collapse, thereby improving clustering performance.

[0136] First, the present invention linearly combines the learned representation matrices of FAE and SGAE to obtain the initial fusion representation matrix As shown below:

[0137]

[0138] Wherein β is a learnable coefficient, which is initialized to 0.5 in the present invention, and selectively determines the importance of two information sources according to the characteristics of their respective data sets.

[0139] The present invention applies operations similar to graph convolution (i.e., message passing) to process the combined information and enhances the initial fusion representation matrix Z by incorporating local structural information from the data. I ,Right now

[0140] Next, the present invention introduces an autocorrelation learning mechanism to utilize the non-local relationship between samples in the initial information fusion space. Specifically, the present invention first uses Calculate the normalized autocorrelation matrix S. Then, use S as coefficients to reassemble Z by taking into account the global correlation between samples L , that is, Z G =SZ L .

[0141] Finally, the present invention adopts skip connections to facilitate the smooth transmission of information in the fusion mechanism, so that the clustering-oriented latent representation is obtained. Can be used for clustering:

[0142] Z=γZ G +Z L (Formula 12)

[0143] Wherein γ is a scale parameter, which is initialized to 0 in the present invention, and its weight is learned when training the network.

[0144] In this way, redundant features can be filtered while retaining more discriminative features in the latent space, so that scSCF can learn meaningful representations and avoid collapse, thereby improving clustering performance.

[0145] 3.5 Propagation Regularization

[0146] The present invention introduces propagation regularization to alleviate the over-smoothing problem during network training. The formula is as follows:

[0147]

[0148] where JSD(·) represents the Jensen-Shannon divergence. The network can capture long-range dependencies even with shallow network architectures, thus reducing over-smoothing as the propagated information goes deeper within the framework.

[0149] In summary, the target of the SIFM module is calculated by the following formula:

[0150]

[0151] where δ is a trade-off parameter set to 0.001.

[0152] 4. Self-supervision module

[0153] In order to better adapt to the complex structure of scRNA-seq data, especially the two characteristics of high dimensionality and high sparsity, the present invention adopts an optimal transmission strategy to deal with this complex structure. It uses the Student's t distribution as the kernel to measure the embedding point h i and cluster center c j The similarity between .

[0154]

[0155] where h i =f(x i )∈Z corresponds to characterization, ε = 1 is the degree of freedom of the Student t distribution, q ij is the probability of assigning sample i to cluster j (i.e., soft assignment). ij ] is defined as the distribution of all sample assignments.

[0156] After obtaining the clustering result distribution Q, the goal of learning is to optimize the data representation by learning a high-confidence assignment. Most traditional methods use a clustering guided loss function to force the generated sample representation to have minimal distortion relative to the pre-learned cluster centers. This approach aims to minimize the difference between the data representation and the cluster centers, thereby improving the cohesion of the clusters. However, in order to prevent the emergence of a degenerate solution that assigns all data points to a single (arbitrary) label, the present invention introduces a constraint that aligns the label distribution with the mixing ratio. This approach ensures improved clustering accuracy and a balanced contribution of each data point to the loss calculation. Therefore, the present invention constructs the target probability matrix P by solving the following optimal transmission:

[0157]

[0158] Here, the present invention regards the target distribution P as the transportation planning matrix in the optimal transportation theory and regards -logQ as the related cost matrix. T 1 N =N π , where π represents the proportion of points in each cluster, which can be estimated by the intermediate clustering results. This step successfully adds the constraint that the cluster distribution must be consistent with the mixing ratio. Considering the computational burden of direct optimization, the present invention uses the Sinkhorn distance and performs fast optimization through entropy constraints. The optimization process with entropy constrained Lagrange multipliers is as follows:

[0159]

[0160] Where H is the entropy function, and λ is the smoothing parameter that controls the cluster balance, which is set to 5. At this point, the uniqueness of P is guaranteed, and it can be solved by fixed-point iteration using Formula 17:

[0161]

[0162] Where t is the number of iterations, and in each iteration u (t) =1 N / (Q λ v (t-1) ) and v (t) =N π / (Q λ u (t) ) starts with v (0) =1 N After fixed-point iteration, the present invention can obtain the optimal transmission plan matrix During the training process, the present invention fixes and make Q Keep it consistent, so the clustering loss function is expressed as:

[0163]

[0164] The goal of the learning process is to minimize the clustering loss.

[0165] 5. Joint Optimization

[0166] The overall optimization objective of the proposed method consists of three parts: SIFM loss, ZINB loss and clustering loss:

[0167]

[0168] Wherein ρ and σ are adjustment parameters for balancing different losses. In the present invention, ρ is initialized to 10 and σ is initialized to 2.

[0169] The training process of scSCF includes three steps: (i) pre-training FAE and SGAE for at least 50 epochs respectively by minimizing the reconstruction loss function; (ii) integrating the two sub-networks into a unified framework and training them for another 100 epochs; (iii) training the entire network for at least 50 epochs until convergence under the guidance of Formula 19. Since the number of cell clusters k in practical applications is unknown, the present invention uses the K-means method on the clustering-oriented latent representation Z to determine the optimal value of k.

[0170] 6. Model Evaluation

[0171] In order to evaluate the effectiveness of the method, the present invention adopts three evaluation indicators widely used in performance evaluation of unsupervised learning scenarios: Accuracy (ACC), Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI). Assume that U is the cell type with a true label and V is the clustering result predicted by the model. The larger the value of these evaluation indicators, the higher the consistency between the two clustering assignments U and V, and the better the clustering effect of the method.

[0172] ACC is defined as measuring the best match between two cluster assignments U and V. If there is a data point i, l i is the true label, u i is the allocation of the clustering algorithm, then ACC is defined as:

[0173]

[0174] Here, m covers all possible one-to-one mappings between U and V, and the best mapping with respect to n cells can be found efficiently using the Hungarian algorithm.

[0175] NMI is a function that measures the consistency between the predicted labels and the true labels of n cells. Specifically,

[0176]

[0177] ARI is a function used to evaluate the similarity of two assignments. Given a set X of size n and two clustering results U = {u1,u2,…,u r},V={v1,v2,…,v s}, and n ij =|U i ∩V j |,i=1,2,…r,j=1,2,…,s, The calculation formula of ARI is as follows:

[0178]

[0179] From the above, it can be seen that the present invention provides a universal clustering framework for interpreting scRNA-seq data, which not only has higher clustering accuracy, but also has stronger noise resistance and the ability to process large-scale data, providing a solid data foundation for various biological downstream tasks.

[0180] 7. Visualization of clustering results

[0181] In order to obtain an intuitive biological interpretation of the clustering results, the present invention uses t-SNE to extract and visualize the distribution of the learned cluster embedding Z in two-dimensional space. Each point on the visualization represents a cell, and each color represents a predicted cell type.

[0182] In summary, the present invention provides a universal clustering framework for interpreting scRNA-seq data, realizes efficient clustering of scRNA-seq data, solves the scalability and noise resistance problems faced by existing methods when processing large-scale data, and provides a reliable data foundation for downstream biological tasks.

[0183] Below, a specific experiment is used to illustrate the twin clustering framework provided by the present invention for interpreting single-cell RNA sequencing data.

[0184] 1. Dataset Overview

[0185] To evaluate the performance of scSCF, seven real scRNA-seq datasets of known cell types downloaded from the GEO database were selected for clustering. Table 1 summarizes the details of the seven real datasets. These datasets are derived from different sequencing methods. Among the seven datasets, four are from human samples and three are from mouse samples. These datasets cover a wide range of cell types, including retina, lung, liver, kidney, pancreas, and peripheral blood mononuclear cells.

[0186] Table 1 Summary of scRNA-seq datasets

[0187] Dataset Sequencing methods Sample size Number of genes Number of cell clusters Shekhar mouse retina cells Drop-seq 27499 13166 19 Macosko mouse retina cells Drop-seq 14653 11422 39 QS mouse lung cells Smart-seq2 1676 23341 11 CITE_CMBC 10X genomics 8617 2000 15 Human liver cells 10X genomics 8444 4999 11 Human kidney cells 10X genomics 5685 25125 11 Human pancreas cells CEL-seq2 2122 19046 9

[0188] 2. Experimental environment and parameter settings

[0189] In the experiment, the present invention implemented scSCF in Python 3.7 based on PyTorch. The various parameter settings of the present invention are shown in the detailed steps.

[0190] 3. Evaluation indicators

[0191] In the experiment, the present invention uses three indicators, namely Accuracy (ACC), Normalized Mutual Information (NMI) and Adjusted Rand Index (ARI), to evaluate the performance.

[0192] 4. Experimental results analysis

[0193] (1) Here, the model performance of scSCF is compared with five competing methods (scDeepCluster, scNAME, scDSC, scGNN, and scCDCG), and experiments are conducted on seven real datasets. The experimental results are as follows:

[0194] Table 2 Clustering performance of all models on ten datasets (average value, scSCF experimental results are bolded)

[0195]

[0196]

[0197] According to the results, the present invention can be observed:

[0198] 1) The proposed method scSCF outperforms competing methods in all three indicators of all datasets. Different from previous methods, scSCF learns more discriminative scRNA-seq data representations through the reduction of double-layer information cross-correlation at the cellular level and gene level, as well as cross-modal dynamic information fusion, thereby improving the clustering effect.

[0199] 2) Combined with Table 1, taking Shekhar mouse retina cells and Macosko mouse retina cells as examples, scSCF performs well on large-scale datasets and achieves significant performance improvements. scSCF effectively utilizes the inherent rich information in scRNA-seq data and alleviates the over-smoothing problem, especially on large-scale datasets.

[0200] 3) The present invention significantly improves the existing technology, showing excellent representation learning ability on scRNA-seq data, while successfully solving the problem of representation collapse. Specifically, by thoroughly extracting high-order structural information between cells and integrating gene expression properties, scSCF effectively learns discriminative features in scRNA-seq data. The method of the present invention enhances the interpretability of data and provides strong support for revealing complex biological relationships between cells, thereby promoting the further development of biological research.

[0201] (2) In order to obtain an intuitive biological explanation of the clustering results, the present invention uses t-SNE to extract and visualize the distribution of the cluster embedding Z learned in two-dimensional space. Each point in the visualization represents a cell, and each color represents a predicted cell type. Figure 3 It can be seen that scSCF can effectively distinguish different cell types regardless of cluster size, forming clear boundaries between cell populations. In contrast, other methods cannot clearly demarcate clusters. These visualizations confirm that scSCF can detect various cell types and cluster similar cells together while separating different cell types, thereby improving the accuracy and biological interpretability of cluster analysis in scRNA-seq data.

[0202] The above implementation is only used to illustrate the technical solution of the present invention rather than to limit it. Ordinary technicians in this field can modify or replace the technical solution of the present invention with equivalents without departing from the scope of the present invention. The protection scope of the present invention shall be based on the claims.

Claims

1. A twin clustering method for interpreting single-cell RNA sequencing data, characterized in that The method comprises: Preprocess the single-cell RNA sequencing data to obtain the preprocessed gene expression matrix and constructing a K nearest neighbor graph between cells based on the preprocessed gene expression matrix; The preprocessed gene expression matrix is ​​enhanced in feature space to obtain a disturbance attribute matrix Performing structural spatial enhancement on the K nearest neighbor graph between cells to obtain an enhanced graph; wherein the enhanced graph includes: the enhanced graph after edge removal And the enhanced image after diffusion operation A r Represents the enhanced image after edge removal The adjacency matrix, A d Represents the enhanced image after the diffusion operation The adjacency matrix of Information fusion and encoding and decoding are performed based on the perturbation attribute matrix and the enhancement graph to obtain a clustering result corresponding to the single-cell RNA sequencing data.

2. The method according to claim 1, characterized in that The single-cell RNA sequencing data is preprocessed to obtain a preprocessed gene expression matrix, including: Filtering the single-cell RNA sequencing data, the filtering comprising: filtering out genes expressed in less than one cell and cells in which the number of gene expressions is less than one; The filtered data were normalized and transformed by logTPM to filter out genes with no counts in any cell; Highly expressed genes were screened to construct the preprocessed gene expression matrix.

3. The method according to claim 1, characterized in that The K nearest neighbor graph between cells is structurally enhanced to obtain an enhanced graph after edge removal. include: After calculating the sample pairwise cosine similarity matrix based on the preprocessed gene expression matrix, a mask matrix M is generated; Compute the adjacency matrix with edges removed Where D is the degree matrix of the K nearest neighbor graph between cells, is the normalized adjacency matrix of the K nearest neighbor graph between the cells, and I represents the identity matrix; Generate an enhanced image after edge removal 4. The method according to claim 1, characterized in that: The K nearest neighbor graph between the cells is structurally enhanced to obtain an enhanced graph after diffusion operation. include: The normalized adjacency matrix is ​​transformed into Convert to adjacency matrix A d ; Generate an enhanced image after diffusion operation 5. The method according to claim 1, characterized in that Information fusion and encoding and decoding are performed based on the perturbation attribute matrix and the enhancement graph to obtain a clustering result corresponding to the single-cell RNA sequencing data, including: Construct a twin information fusion network autoencoder and a FAE decoder based on the denoising ZINB model; wherein each branch of the twin information fusion network autoencoder is equipped with a feature autoencoder and a symmetric graph autoencoder; the perturbation attribute matrix As the input of the feature autoencoder in each branch, the first feature representation matrix is ​​obtained respectively and the second eigenvalue matrix The perturbation attribute matrix and the adjacency matrix A r As the input of the symmetric graph automatic decoder in one branch, the perturbation attribute matrix and the adjacency matrix A d As the input of the symmetric graph automatic decoder in another branch, the first symmetric graph representation matrix is ​​obtained The first adjacency matrix And the second symmetric graph representation matrix The second adjacency matrix Fusion first feature representation matrix and the second eigenvalue matrix Get the feature representation matrix Z FA E; Fusion first symmetric graph representation matrix and the second symmetric graph characterization matrix Get the symmetric graph representation matrix Z SGAE ; The feature representation matrix Z FAE and the symmetric graph representation matrix Z SGAE Perform weighted sum calculation to obtain the initial fusion representation matrix Z I ; Based on the normalized adjacency matrix Merge the local structural information between data to enhance the initial fusion representation matrix Z I , to obtain the expression matrix Z L ; Among them, the normalized adjacency matrix A normalized adjacency matrix representing the K nearest neighbor graph between the cells; The expression matrix Z is obtained by using the non-local relationship between data in the initial information fusion space through the autocorrelation learning mechanism. G ; Among them, the initial information fusion space is based on the perturbation attribute matrix Build; For the expression matrix Z L and the expression matrix Z G Perform weighted sum calculation to obtain the fusion representation matrix Z; Performing self-supervised clustering based on the optimal transmission strategy on the fusion representation matrix Z to obtain a clustering result corresponding to the single-cell RNA sequencing data; The distribution of the fusion representation matrix Z is extracted and visualized to obtain a visual display of the clustering result.

6. The method according to claim 5, characterized in that The process of training the twin information fusion network autoencoder, the FAE decoder based on the denoising ZINB model, and the self-supervised clustering based on the optimal transmission strategy includes: Based on the perturbation attribute matrix Reconstructing the weighted attribute matrix Reconstructing the weighted attribute matrix Normalized adjacency matrix And reconstruct the adjacency matrix Constructing a reconstruction loss function based on feature autoencoders and symmetric graph autoencoders Wherein, the reconstructed weighted attribute matrix is the output of the feature autoencoder based on the fusion representation matrix Z, and the reconstructed weighted attribute matrix and the reconstructed adjacency matrix It is the output of the symmetric graph autoencoder based on the fusion representation matrix Z; Based on the two-layer mutual correlation reduction mechanism, a cell-level mutual correlation reduction loss function is constructed. and gene-level cross-correlation reduction loss function Among them, the cell-level mutual correlation information reduces the loss function Includes: Feature mutual correlation reduction Cell-level mutual correlation reduction loss function in the channel The cell-level mutual correlation reduction loss function in the channel is reduced by the mutual correlation reduction of the graph The gene-level cross-correlation reduces the loss function Includes: Feature mutual correlation reduction Gene-level mutual correlation reduction loss function in the channel The loss function for reducing the gene-level mutual correlation in the channel and the graph mutual correlation Based on the fusion representation matrix Z and the normalized adjacency matrix Constructing the propagation regularization loss function Integrate cell-level cross-correlation to reduce loss function Gene-level cross-correlation reduction loss function Reconstruction loss function and propagation regularization loss function Get the loss function of the twin information fusion network automatic encoder By assuming that the single-cell RNA sequencing data follows the ZINB distribution, the overall negative likelihood loss of the FAE decoder based on the denoising ZINB model is constructed Constructing a self-supervised clustering loss function based on the optimal transmission strategy Based on the loss function The overall negative likelihood loss And the clustering loss function Complete the training of the twin information fusion network autoencoder and the FAE decoder based on the denoising ZINB model.

7. The method according to claim 6, characterized in that The double-layer mutual correlation reduction mechanism is used to construct a cell-level mutual correlation reduction loss function in the feature mutual correlation reduction channel. and gene-level correlation information to reduce loss function include: By calculating the first eigenvalue matrix The i-th dimension feature and the second feature representation matrix The cosine similarity between the j-th dimension features in To construct the matrix S cell ; Constructing feature mutual correlation reduction channels to reduce the cell-level mutual correlation reduction loss function Where N represents the potential embedding dimension and I represents the identity matrix; Use the Readout function to represent the first feature matrix and the second eigenvalue matrix Projection is performed to obtain the first gene-level feature representation matrix and the second gene-level feature representation matrix By calculating the first gene-level feature representation matrix The i-th dimension feature and the second gene-level feature representation matrix The cosine similarity between the j-th dimension features in To construct the matrix Sgene; Constructing gene-level mutual correlation reduction loss function in feature mutual correlation reduction channel Among them, d is the dimension of the Readout function, Represents the identity matrix.

8. The method according to claim 6, characterized in that The double-layer mutual correlation reduction mechanism is used to construct a cell-level mutual correlation reduction loss function in the graph mutual correlation reduction channel. and gene-level cross-correlation reduction loss function include: Characterize the matrix by computing the first symmetry graph The i-th dimension feature and the second symmetric graph representation matrix The cosine similarity between the j-th dimension features in To construct the matrix T cell ; Constructing graph mutual correlation to reduce cell-level mutual correlation in channels reduces loss function Where N represents the potential embedding dimension and I represents the identity matrix; Use the Readout function to represent the matrix of the first symmetric graph and the second symmetric graph characterization matrix Projection is performed to obtain the first gene-level graph representation matrix and the second gene-level graph representation matrix By calculating the first gene-level graph representation matrix The i-th dimension feature and the second gene-level graph representation matrix The cosine similarity between the j-th dimension features in To construct the matrix T gene ; Constructing gene-level mutual correlation reduction loss function in feature mutual correlation reduction channel Among them, d is the dimension of the Readout function, Represents the identity matrix.

9. The method according to claim 5, characterized in that The overall negative likelihood loss of the FAE decoder based on the denoising ZINB model in, represents the preprocessed gene expression matrix, π ij is the probability of a loss event, with a mean value of μ ij and the dispersion θ ij denote the negative binomial component, respectively; The self-supervised clustering loss function is constructed according to the optimal transmission strategy include: Get the clustering result distribution matrix Q; Combining the clustering result distribution matrix Q, a function of constructing a transportation plan matrix P using optimal transport; Using the Sinkhorn distance and solving the function of the transportation plan matrix P through entropy constraints, the optimal transportation plan matrix is ​​obtained. Based on the clustering result distribution matrix Q and the optimal transmission plan matrix Constructing a self-supervised clustering loss function in, is the optimal transmission plan matrix The elements in q ij is the element in the clustering result distribution matrix Q.

10. A twin clustering system for interpreting single-cell RNA sequencing data, characterized in that The system comprises: Preprocessing module, used to preprocess single-cell RNA sequencing data to obtain the preprocessed gene expression matrix and constructing a K nearest neighbor graph between cells based on the preprocessed gene expression matrix; A data enhancement module is used to enhance the feature space of the preprocessed gene expression matrix to obtain a disturbance attribute matrix Performing structural spatial enhancement on the K nearest neighbor graph between cells to obtain an enhanced graph; wherein the enhanced graph includes: the enhanced graph after edge removal And the enhanced image after diffusion operation A r Represents the enhanced image after edge removal The adjacency matrix, A d Represents the enhanced image after the diffusion operation The adjacency matrix of The twin information fusion module is used to perform information fusion and encoding and decoding based on the perturbation attribute matrix and the enhancement graph to obtain the clustering results corresponding to the single-cell RNA sequencing data.

Citation Information

Patent Citations

  • Single-cell sequencing data dimension reduction method fusing gene ontology and neural network

    CN111564183A

  • Bladder cancer depleted NKT cell subset, characteristic gene and application thereof

    CN111733244A

  • Single-cell RNA-seq data clustering method based on dual self-supervision

    CN114022693A

  • Single cell dimension reduction method and device based on graph adaptive aggregation auto-encoder

    CN115472228A

Cited By

  • Training method and device for single-cell drug response prediction model

    CN119889457A

  • A training method and device of a single-cell drug response prediction model

    CN119889457B

  • Action recognition method and system based on human skeleton structure

    CN121861729A