Deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder
Through the Gaussian mixed graph variational autoencoder, combined with the graph convolutional neural network and the Gaussian mixed model, the problem of single-cell clustering method in processing high-dimensional sparse data is solved, and the prediction of single-cell clustering and gene regulation network is realized. , improving the dimensionality reduction effect of cellular data and the ability to mine biological information.
Patent Information
- Application Number
- CN202210506799.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-11
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing single-cell clustering methods are difficult to effectively process high-dimensional sparse genome sequencing data, ignore the interaction relationship between genes or open regions, and cannot simultaneously realize single-cell clustering and predict cell class-specific gene regulatory networks.
Using a deep unsupervised single-cell clustering method based on Gaussian mixed graph variational autoencoder, the gene regulation network is initialized by inputting protein-protein interaction relationship and single-cell gene expression data, and using graph convolutional neural network to learn single-cell low Dimensional expression and cell class-specific gene regulation networks are combined with Gaussian mixed model for cell clustering and feature extraction.
Effectively reduce the problem of high-dimensional sparseness of cell data, extract the low-dimensional hidden layer characteristics that meet the Gaussian hybrid model, represent the patterns of different cell categories, and predict the gene regulation network through the decoder, realizing cell clustering and biological information Digging makes up for the lack of missing information.
Smart Images

Figure CN114783526B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of single-cell clustering, and in particular to a deep unsupervised single-cell clustering method based on a Gaussian mixture graph variational autoencoder. Background Art
[0002] Single-cell sequencing technology refers to a technology that performs high-throughput sequencing analysis of the genome, transcriptome, and epigenome at the level of a single cell. It can reveal the genetic structure and gene expression status of a single cell, reflect the heterogeneity between cells, and play an important role in the fields of tumors, developmental biology, microbiology, neuroscience, etc., and is becoming the focus of life science research. One of the first tasks to be solved in single-cell sequencing is to identify the cell types contained in the sample, that is, to complete cell clustering based on an unsupervised algorithm. Single-cell clustering can establish a comprehensive reference for all cell types at different developmental stages in an organism or tissue. In addition to providing a deeper understanding of basic biology, it can also serve as a reference for disease research. Since many downstream analyses are based on cell clustering, the results of single-cell clustering may have a significant impact on downstream applications. Therefore, how to obtain reliable single-cell clustering is one of the key challenges in the single-cell field.
[0003] There are some methods designed to process high-dimensional and sparse genomic sequencing data, especially single-cell RNA-seq (scRNA-seq) data (RNA-seq is a transcriptome sequencing technology, which is to sequence and analyze mRNA, smallRNA, NONcoding RNA, etc. or some of them using high-throughput sequencing technology to reflect their expression levels) and single-cell ATAC-seq (scATAC-seq) data (ATAC-seq is a transposase-accessible chromatin sequencing technology, which is an effective method to use Tn5 transposase to mark the open positions of chromatin in the whole genome with sequencing adapters). Because genomic sequencing data is very high-dimensional and sparse, the distances between cells become similar, and the differences in distances are often small, so it is very difficult to identify cell types. Feature selection or dimensionality reduction can reduce noise and speed up calculations. Dimensionality reduction techniques such as principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) are often used to visualize data and examine the distribution of input data. There are many types of single-cell clustering methods. The most popular clustering algorithm is k-means, which iteratively identifies k cluster centers and assigns each cell to the nearest center. However, the algorithm is greedy and is not guaranteed to find the global minimum. Another disadvantage is that it tends to determine clusters of equal size, which may cause rare cell types to be hidden in larger clusters. Another clustering algorithm widely used for scRNA-seq is hierarchical clustering, which sequentially combines individual cells into larger clusters or splits clusters into smaller groups. Its disadvantage is that the time and memory requirements grow at least quadratically with the number of data points, which means that the cost of using hierarchical clustering for large datasets will be very high.
[0004] Recent studies have developed several methods specifically for scATAC-seq data analysis: chromVAR, which evaluates a group of peaks that share the same motif or functional annotation; scABC, which weights cells according to sequencing depth and applies weighted K-medoid clustering to reduce the impact of missing values, then calculates the label for each cluster and assigns cells to the nearest label based on Spearman correlation. However, each method has noteworthy problems: chromVAR can only analyze peaks in groups and lacks judgment on individual peaks; scABC relies heavily on label samples with high sequencing depth, and for data with many missing values, especially for scATAC-seq data, the Spearman correlation coefficient may be defined incorrectly.
[0005] In summary, the input data of existing single-cell clustering methods are mostly cell gene expression or chromatin openness, ignoring the interaction between these genes or open regions, and only using vectors as model input. Secondly, because the interaction between genes or open regions is not considered, existing methods can only achieve single-cell clustering and cannot simultaneously predict cell-type-specific gene regulatory networks. Summary of the invention
[0006] The purpose of the present invention is to provide a deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder to address the technical defects in the prior art.
[0007] In a first aspect, the present invention provides a deep unsupervised single-cell clustering method based on a Gaussian mixture graph variational autoencoder, comprising the following steps:
[0008] S1. Input protein-protein interaction PPIs data and single-cell gene expression data X;
[0009] S2. Use PPIs to initialize the trainable gene regulatory network A;
[0010] S3. Initialize the cell cluster C of each cell using the K-means method;
[0011] S4. Let A and X pass through the graph encoder to obtain the hidden layer;
[0012] hidden=GCN(A,X),
[0013] S5. Based on the Gaussian mixture model GMM, the cell cluster C is obtained, and then the cell low-dimensional representation Z is obtained by sampling from the Gaussian mixture model GMM:
[0014] S5.1. Based on the assumption that the hidden layer obeys the Gaussian mixture distribution, update the parameters of the Gaussian mixture model GMM:
[0015] μ=GCN(hidden,X)
[0016] σ=GCN(hidden,X)
[0017]
[0018] Where N(μ k ,σ k ) indicates that the mean is μ k , the variance is Gaussian distribution, α k Indicates the probability that sample m obeys the Gaussian distribution;
[0019] S5.2. Obtain cell clusters C based on Gaussian mixture model GMM:
[0020] C = argmax k (α k )
[0021] S5.3. Sampling from the Gaussian mixture model GMM obtains the low-dimensional representation Z of the cell: first sample from a standard Gaussian distribution with mean 0 and variance 1, and then transform it by the standard deviation and mean:
[0022] Z=μ+N(0,1)*σ
[0023] S6. Based on the low-dimensional representation Z of the cell, the decoder Graph Decoder is used to predict the gene regulatory network A;
[0024] A=ReLU(Z T *Z)
[0025] S7. Calculate the loss function and back-propagate to update A and GCN. The loss function includes cross entropy loss CrossEntropy and KL divergence. The calculation formula is as follows:
[0026] loss = CrossEntropy-KL
[0027] S8. Repeat the above steps S2-S5 until convergence;
[0028] S9. Output gene regulatory network A, cell low-dimensional representation Z, and cell cluster C.
[0029] Among them, the steps of initializing the gene regulatory network A are as follows:
[0030] Genes are numbered numerically;
[0031] If there is an interaction relationship between the protein expressed by the i-th gene and the protein expressed by the j-th gene in PPIs, the i-th row and j-th column of the gene regulatory network A is initialized to 1, otherwise it is initialized to 0.
[0032] The present invention not only features gene expression data, but also adds the relationship between genes, and uses graph convolutional neural network GCN to simultaneously learn single cell low-dimensional expression, cell clustering and cell category-specific gene regulatory network. An undirected graph is used to represent each single cell, nodes represent genes, and edges represent gene-gene interactions. An improved graph variational autoencoder is used to learn cell representation. The model assumes that the hidden layer obeys the Gaussian mixture model GMM, based on which cells are clustered, and a trainable adjacency matrix is used to learn cell category-specific gene-gene interactions.
[0033] By comparing the clustered regulatory networks, we can discover the similarities and differences in the structures of gene regulatory networks under different background conditions, and then analyze their practical significance in biology.
[0034] In a second aspect, the present invention provides a deep unsupervised single-cell clustering method based on a Gaussian mixture graph variational autoencoder, comprising the following steps:
[0035] S1. Input regulatory element interaction HiChIP data, single-cell gene expression data or regulatory element openness data X;
[0036] S2. Initialize a trainable gene regulatory network A using HiChIP;
[0037] S3. Initialize the cell cluster C of each cell using the K-means method;
[0038] S4. Let A and X pass through the graph encoder to obtain the hidden layer;
[0039] hidden=GCN(A,X),
[0040] S5. Based on the Gaussian mixture model GMM, the cell cluster C is obtained, and then the cell low-dimensional representation Z is obtained by sampling from the Gaussian mixture model GMM:
[0041] S5.1. Based on the assumption that the hidden layer obeys the Gaussian mixture distribution, update the parameters of the Gaussian mixture model GMM:
[0042] μ=GCN(hidden,X)
[0043] σ=GCN(hidden,X)
[0044]
[0045] Where N(μ k ,σ k ) indicates that the mean is μ k , the variance is Gaussian distribution, α k Indicates the probability that sample m obeys the Gaussian distribution;
[0046] S5.2. Obtain cell clusters C based on Gaussian mixture model GMM:
[0047] C = argmax k (α k )
[0048] S5.3. Sampling from the Gaussian mixture model GMM obtains the low-dimensional representation Z of the cell: first sample from a standard Gaussian distribution with mean 0 and variance 1, and then transform it by the standard deviation and mean:
[0049] Z=μ+N(0,1)*σ
[0050] S6. Based on the low-dimensional representation Z of the cell, the decoder Graph Decoder is used to predict the gene regulatory network A;
[0051] A=ReLU(Z T *Z)
[0052] S7. Calculate the loss function and back-propagate to update A and GCN. The loss function includes cross entropy loss CrossEntropy and KL divergence. The calculation formula is as follows:
[0053] loss = CrossEntropy-KL
[0054] S8. Repeat the above steps S2-S5 until convergence;
[0055] S9. Output gene regulatory network A, cell low-dimensional representation Z, and cell cluster C.
[0056] Among them, the steps of initializing the gene regulatory network A are as follows:
[0057] Genes are numbered numerically;
[0058] If there is an interaction relationship between the i-th regulatory element and the j-th regulatory element in HiChIP, the i-th row and j-th column of the gene regulatory network A are initialized to 1, otherwise they are initialized to 0.
[0059] The present invention not only features the degree of chromatin openness, but also adds the relationship between regulatory elements, and uses the graph convolutional neural network GCN to simultaneously learn single-cell low-dimensional expression, cell clustering, and cell-class-specific gene regulatory networks. An undirected graph is used to represent each single cell, with nodes representing regulatory elements and edges representing regulatory element-regulatory element interactions. An improved graph variational autoencoder is used to learn cell representation. The model assumes that the hidden layer (H) obeys the Gaussian mixture model GMM, based on which cells are clustered, and a trainable adjacency matrix is used to learn cell-class-specific regulatory element-regulatory element interactions.
[0060] By comparing the clustered regulatory networks, we can discover the similarities and differences in the structures of gene regulatory networks under different background conditions, and then analyze their practical significance in biology.
[0061] The present invention analyzes cell data by extracting hidden features. The algorithm framework combines variational graph auto-encoders (VGAE) and Gaussian mixture models (GMM). The low-dimensional hidden features of the data are extracted by deep learning methods to cluster the cell data and mine biological information, effectively solving the problem of high-dimensional sparsity of cell data. The encoder (Graph Encoder) of the graph variational auto-encoder extracts low-dimensional hidden features that meet the Gaussian mixture model, so that different dimensions have low coupling, and the patterns of different cell categories are characterized respectively. Cells of the same category share the parameters of the Gaussian mixture model with each other, thereby achieving the effect of sharing information to make up for each other's missing information, and finally predicting the final gene regulatory network A through the decoder (Graph Decoder). BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 It is a schematic diagram of the processing flow of the deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder of the present invention. DETAILED DESCRIPTION
[0063] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0064] like Figure 1 As shown, in a first aspect of an embodiment of the present invention, a deep unsupervised single-cell clustering method based on a Gaussian mixture graph variational autoencoder (VGAE) is provided, which is a method for obtaining a gene regulatory network A, a cell low-dimensional representation Z, and a cell cluster C using a graph variational autoencoder based on a Gaussian mixture distribution, comprising the following steps:
[0065] Step 1: Input protein-protein interaction relationships PPIs and single-cell gene expression data X;
[0066] Step 2: Use PPIs to initialize the trainable gene regulatory network A. The specific steps are as follows:
[0067] Step 2.1: Digitally number the genes;
[0068] Step 2.2: If there is an interaction relationship between the protein expressed by the i-th gene and the protein expressed by the j-th gene in PPIs, the i-th row and j-th column of the gene regulatory network A are initialized to 1, otherwise they are initialized to 0;
[0069] Step 3: Initialize the cell cluster C of each cell using the K-means method;
[0070] Step 4: Let A and X pass through the graph encoder to get the hidden layer;
[0071] hidden=GCN(A,X)
[0072] Among them, GCN represents graph convolutional neural network;
[0073] Step 5: Obtain cell clustering and low-dimensional representation of cells, including the following steps:
[0074] Step 5.1: Based on the assumption that the hidden layer (H) obeys a Gaussian mixture distribution, update the parameters of the Gaussian mixture model GMM:
[0075] μ=GCN(hidden,X)
[0076] σ=GCN(hidden,X)
[0077]
[0078] Where N(μ k ,σ k ) indicates that the mean is μ k , the variance is Gaussian distribution, α k Indicates the probability that sample m obeys the Gaussian distribution;
[0079] Step 5.2: Obtain cell cluster C based on Gaussian mixture model GMM:
[0080] C = argmax k (α k )
[0081] Step 5.3: Sample from the Gaussian mixture model GMM to obtain the low-dimensional representation Z of the cell: first sample from a standard Gaussian distribution with mean 0 and variance 1, and then transform it by the standard deviation and mean:
[0082] Z=μ+N(0,1)*σ
[0083] Step 6: Use the decoder Graph Decoder to predict the gene regulatory network A;
[0084] A=ReLU(Z T *Z)
[0085] Step 7: Calculate the loss function (including cross entropy loss CrossEntropy and KL divergence), and back-propagate to update A and GCN;
[0086] loss = CrossEntropy-KL
[0087] Step 8: Repeat steps 2-5 above until convergence;
[0088] Step 9: Output gene regulatory network A, cell low-dimensional representation Z, and cell cluster C.
[0089] In the embodiment of the present invention, a graph variational auto-encoder (VGAE) is used. An n*n convolution kernel is used in the encoder process to replace the encoder process with the convolution process of the graph convolutional neural network (GCN), and the decoder process is replaced with the graph decoder process. Specifically, the change of the previous and next adjacency matrices is used as loss. The input of the graph variational auto-encoder VGAE is the adjacency matrix of the undirected graph (such as the gene regulatory network A) and the feature matrix of the nodes in the undirected graph (single-cell gene expression data X). The Gaussian distribution (μ, σ) of the low-dimensional representation of the node vector can be obtained through the graph convolutional neural network (GCN). 2 ), and then the decoder generates the adjacency matrix (A') of the graph to obtain the gene regulatory network.
[0090] like Figure 1 As shown, in a second aspect of an embodiment of the present invention, a deep unsupervised single-cell clustering method based on a Gaussian mixture graph variational autoencoder is provided, which is a method for obtaining a gene regulatory network A, a cell low-dimensional representation X, and a cell cluster C using a graph variational autoencoder based on a Gaussian mixture distribution, comprising the following steps:
[0091] Step 1: Input regulatory element interaction HiChIP data, single-cell gene expression data or regulatory element openness data X.
[0092] Step 2: Use HiChIP to initialize the trainable gene regulatory network A. The specific steps are as follows:
[0093] Step 2.1: Digitally number the genes;
[0094] Step 2.2: If there is an interaction relationship between the i-th regulatory element and the j-th regulatory element in HiChIP, the i-th row and j-th column of the gene regulatory network A are initialized to 1, otherwise they are initialized to 0;
[0095] Step 3: Initialize the cell cluster C of each cell using the K-means method;
[0096] Step 4: Let A and X pass through the graph encoder to get the hidden layer;
[0097] hidden=GCN(A,X)
[0098] Step 5: Obtain cell clustering and low-dimensional representation of cells, including the following steps:
[0099] Step 5.1: Based on the assumption that the hidden layer (H) obeys the Gaussian mixture distribution, update the parameters of the Gaussian mixture model GCN:
[0100] μ=GCN(hidden,X)
[0101] σ=GCN(hidden,X)
[0102]
[0103] Where N(μ k ,σ k ) indicates that the mean is μ k , the variance is Gaussian distribution, α k It represents the probability that sample m follows the Gaussian distribution.
[0104] Step 5.2: Obtain cell cluster C based on Gaussian mixture model GMM:
[0105] C = argmax k (α k )
[0106] Step 5.3: Sample from the Gaussian mixture model GMM to obtain the low-dimensional representation Z of the cell: first sample from a standard Gaussian distribution with mean 0 and variance 1, and then transform it by the standard deviation and mean:
[0107] Z=μ+N(0,1)*σ
[0108] Step 6: Use the decoder Graph Decoder to predict the gene regulatory network A;
[0109] A=ReLU(Z T *Z)
[0110] Step 7: Calculate the loss function (including cross entropy loss CrossEntropy and KL divergence), and back-propagate to update A and GCN;
[0111] loss = CrossEntropy-KL
[0112] Step 8: Repeat steps 2-5 above until convergence;
[0113] Step 9: Output gene regulatory network A, cell low-dimensional representation Z, and cell cluster C.
[0114] The present invention uses a learnable adjacency matrix to learn a cell type-specific gene regulatory network A, and uses a Gaussian mixture distribution in a variational graph autoencoder (VGAE) to obtain a low-dimensional representation Z of the cell and a cell cluster C. The present invention achieves three goals at one stroke by applying a Gaussian mixture distribution model to a variational graph autoencoder (VGAE) to cluster cells and reduce the dimensionality of cell representations in the process of constructing a gene regulatory network A; in addition, compared to the traditional method of clustering single cells by cell features, the present invention introduces the gene-gene / regulatory element-regulatory element interaction relationship, providing a basis for cell clustering from the perspective of a gene regulatory network.
[0115] The basic principles and main features of the present invention and the advantages of the present invention are shown and described above. It is obvious to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic features of the present invention.
[0116] Therefore, no matter from which point of view, the embodiments should be regarded as illustrative and non-restrictive, and the scope of the present invention is limited by the appended claims rather than the above description. Therefore, it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any figure signs in the claims should not be regarded as limiting the claims involved.
[0117] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder, characterized in that: The following steps are involved: S1. Input protein-protein interaction PPIs data and single-cell gene expression data X; S2. Use PPIs to initialize the trainable gene regulatory network A; S3. Initialize the cell cluster C of each cell using the K-means method; S4. Let A and X pass through the graph encoder to obtain the hidden layer; hidden = GCN(A, X), S5. Based on the Gaussian mixture model GMM, the cell cluster C is obtained, and then the cell low-dimensional representation Z is obtained by sampling from the Gaussian mixture model GMM: S5.
1. Based on the assumption that the hidden layer obeys the Gaussian mixture distribution, update the parameters of the Gaussian mixture model GMM: μ=GCN(hidden,X) σ=GCN(hidden,X) Where N(μ k , σ k ) indicates that the mean is μ k , the variance is Gaussian distribution, α k Indicates the probability that sample m obeys the Gaussian distribution; S5.
2. Obtain cell clusters C based on Gaussian mixture model GMM: C=argmax k (α k ) S5.
3. Sampling from the Gaussian mixture model GMM obtains the low-dimensional representation Z of the cell: first sample from a standard Gaussian distribution with mean 0 and variance 1, and then transform it by the standard deviation and mean: Z=μ+N(0,1)*σ S6. Based on the low-dimensional representation Z of the cell, the decoder Graph Decoder is used to predict the gene regulatory network A; A=ReLU(Z T *WITH) S7. Calculate the loss function and back-propagate to update A and GCN. The loss function includes cross entropy loss CrossEntropy and KL divergence. The calculation formula is as follows: loss = CrossEntropy-KL S8. Repeat the above steps S2-S5 until convergence; S9. Output gene regulatory network A, cell low-dimensional representation Z, and cell cluster C.
2. The deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder according to claim 1, characterized in that: The steps to initialize the gene regulatory network A are as follows: Genes are numbered numerically; If there is an interaction relationship between the protein expressed by the i-th gene and the protein expressed by the j-th gene in PPIs, the i-th row and j-th column of the gene regulatory network A are initialized to 1, otherwise they are initialized to 0.
3. A deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder, characterized in that: The following steps are involved: S1. Input regulatory element interaction HiChIP data, single-cell gene expression data or regulatory element openness data X; S2. Initialize a trainable gene regulatory network A using HiChIP; S3. Initialize the cell cluster C of each cell using the K-means method; S4. Let A and X pass through the graph encoder to obtain the hidden layer; hidden = GCN(A, X), S5. Based on the Gaussian mixture model GMM, the cell cluster C is obtained, and then the cell low-dimensional representation Z is obtained by sampling from the Gaussian mixture model GMM: S5.
1. Based on the assumption that the hidden layer obeys the Gaussian mixture distribution, update the parameters of the Gaussian mixture model GMM: μ=GCN(hidden,X) σ=GCN(hidden,X) Where N(μ k , σ k ) indicates that the mean is μ k , the variance is Gaussian distribution, α k Indicates the probability that sample m obeys the Gaussian distribution; S5.
2. Obtain cell clusters C based on Gaussian mixture model GMM: C=argmax k (α k ) S5.
3. Sampling from the Gaussian mixture model GMM obtains the low-dimensional representation Z of the cell: first sample from a standard Gaussian distribution with mean 0 and variance 1, and then transform it by the standard deviation and mean: Z=μ+N(0,1)*σ S6. Based on the low-dimensional representation Z of the cell, the decoder Graph Decoder is used to predict the gene regulatory network A; A=ReLU(Z T *WITH) S7. Calculate the loss function and back-propagate to update A and GCN. The loss function includes cross entropy loss CrossEntropy and KL divergence. The calculation formula is as follows: loss = CrossEntropy-KL S8. Repeat the above steps S2-S5 until convergence; S9. Output gene regulatory network A, cell low-dimensional representation Z, and cell cluster C.
4. The deep unsupervised single-cell clustering method based on Gaussian mixture graph variational autoencoder according to claim 3 is characterized in that: The steps to initialize the gene regulatory network A are as follows: Genes are numbered numerically; If there is an interaction relationship between the i-th regulatory element and the j-th regulatory element in HiChIP, the i-th row and j-th column of the gene regulatory network A are initialized to 1, otherwise they are initialized to 0.
