Single cell subtype sample generation method, system, equipment and medium

Abstract

The invention discloses a single cell subtype sample generation method, system, device and medium, and relates to the technical field of single cell subtype sample generation, the method comprises the following steps: obtaining a plurality of omics data; the omics data comprises a single cell genome, transcriptome sequencing expression data, clinical metadata and cell type annotation data; learning a meta-path graph of single-cell RNA sequencing data and clinical metadata through graph network data conversion, and learning representation of nodes and edges on the meta-path graph; a self-adaptive condition graph diffusion convolution model ACGDC model is constructed; inputting representations of nodes and edges into an adaptive condition graph diffusion convolution model through a graph network data conversion module, and sampling new single cell sample seeds in a Gaussian distribution space; decoding the single cell sample seeds to generate a new single cell subtype sample through a condition control sample generation method; according to the method, various high-quality single cell subtype samples can be efficiently generated.

Description

Technical Field

[0001] The present invention relates to the technical field of single cell sample generation, and in particular to a single cell subtype sample generation method, system, device and medium. Background Art

[0002] In recent years, deep generative algorithms have made significant progress in generating images, drug molecules, etc., and generative models on single-cell expression profiles have been applied in some applications.

[0003] At present, the application methods of generative models on single-cell expression profiles include methods based on generalized linear models, variational autoencoders (VAE), and multimodal single-cell data analysis frameworks using deep graph neural networks (GNNs), such as graph convolutional networks (GCNs), graph attention networks (GATs), graph autoencoders (GAEs), graph variational autoencoders (GVAEs), and graph diffuse convolutions (GDCs). These algorithms perform well in processing unstructured data.

[0004] However, although existing generative models perform well on specific datasets, they are clearly insufficient in their ability to generate new single-cell samples. They have difficulty capturing complex structures in single-cell data and are unable to process three or more omics data. At the same time, their performance is limited when processing multimodal and heterogeneous graph data, making it difficult to generate new samples that are highly similar to real samples. Summary of the invention

[0005] In view of the shortcomings of the prior art that it mainly focuses on single or two omics data and has difficulty in integrating three or more omics data, the present invention proposes a single-cell subtype sample generation method, system, device and medium, which solves the problems existing in the prior art by constructing an adaptive conditional graph diffusion convolution model (ACGDC) and utilizing network node attributes and neighborhood topological connections to reconstruct the relationship between the edges and attributes of the nodes.

[0006] A method for generating a single cell subtype sample comprises the following steps:

[0007] Acquire multiple omics data of a single cell sample; convert the multiple omics data into directed graph formats to obtain a node matrix and an adjacency matrix; the node matrix includes nodes formed by attributes of the multiple omics data, and the adjacency matrix includes edges between two nodes formed by omics data weights;

[0008] Constructing an adaptive conditional graph diffusion convolution (ACGDC) model; the ACGDC model includes a graph convolution layer and a noise diffusion layer;

[0009] The node matrix and the adjacency matrix are input into the ACGDC model, a new weight vector is assigned to the first-order neighbor edge of the node through the graph convolution layer, and the new weight vector is multiplied by the neighbor node vector of the node to obtain a new node representation; random noise is continuously added to the adjacency matrix containing the new node representation at a certain step interval in the noise diffusion layer, and the adjacency matrix is ​​converted into a high-noise matrix to obtain the trained ACGDC model;

[0010] The various omics data of the single-cell samples to be generated are input into the trained ACGDC model, and new single-cell subtype samples are generated by continuously denoising the Gaussian noise matrix.

[0011] Furthermore, the various omics data include single-cell RNA sequence data, single-cell genome, transcriptome sequencing expression data, clinical metadata and cell type annotation data.

[0012] Furthermore, after obtaining the multiple omics data, the data are preprocessed, and the preprocessing process includes the following steps:

[0013] Use the R package bitr to convert gene IDs for various omics data;

[0014] The enrichGO and enrichKEGG packages were used to annotate the converted omics data with the GO and KEGG databases;

[0015] The marker genes in the cell marker database are used to annotate the cell types of the omics data after the database annotation.

[0016] Furthermore, the method generates new single-cell subtype samples by continuously denoising the Gaussian noise matrix; specifically, it includes obtaining the label vector of the single-cell subtype, and generating new single-cell subtype samples by merging the denoised Gaussian noise matrix with the label vector of the single-cell subtype.

[0017] Furthermore, a cluster analysis method is used to evaluate the trained ACGDC model; specifically, the generated single-cell subtype samples are reduced in dimension by the T-SNE algorithm, the single-cell subtype samples after dimension reduction are clustered by K nearest neighbors, and the clustered single-cell subtype samples are evaluated and analyzed.

[0018] The present invention also includes a single cell subtype sample generation system, comprising:

[0019] An acquisition module is used to acquire multiple omics data of a single cell sample; the multiple omics data are converted into directed graph formats to obtain a node matrix and an adjacency matrix; the node matrix includes nodes formed by attributes of the multiple omics data, and the adjacency matrix includes edges between two nodes formed by omics data weights;

[0020] A model building module, used to build an adaptive conditional graph diffusion convolution (ACGDC) model; the AC GDC model includes a graph convolution layer and a noise diffusion layer;

[0021] A model training module is used to input the node matrix and the adjacency matrix into the ACGDC model, assign a new weight vector to the first-order neighbor edge of the node through the graph convolution layer, and multiply the new weight vector with the neighbor node vector of the node to obtain a new node representation; in the noise diffusion layer, random noise is continuously added to the adjacency matrix containing the new node representation at a certain step interval, and the adjacency matrix is ​​converted into a high-noise matrix to obtain the trained ACGDC model;

[0022] The generation module is used to input the multiple omics data of the single-cell samples to be generated into the trained ACGDC model, and generate new single-cell subtype samples by continuously denoising the Gaussian noise matrix.

[0023] The present invention also includes a computer device for generating single-cell subtype samples, including: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the single-cell subtype sample generation method when executing the computer program.

[0024] The present invention also includes a readable storage medium, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the single cell subtype sample generation method.

[0025] The present invention provides a method, system, device and medium for generating single cell subtype samples, which have the following beneficial effects:

[0026] The present invention integrates multiple omics data of single cells by converting them into directed graph formats respectively, and trains an adaptive conditional graph diffusion convolution model ACGDC constructed by the integrated data set, reconstructs the relationship between the edges and their attributes between nodes by using node attributes and neighborhood topological connections, and adaptively learns the different importance of each different type of edge when neighbors aggregate information by assigning new weight vectors to the first-order neighbor edges of the nodes, thereby solving the problem that different edge types represent different meanings, and finally converts the adjacency matrix into a Gaussian noise matrix. By continuously denoising the Gaussian noise matrix, the denoised matrix obtained is merged with the label vector of the single cell subtype, thereby realizing conditional binding of the cell subtype label with the single cell sampling sample, and generating a new single cell subtype sample. The present invention can efficiently integrate multiple omics data, thereby generating a variety of high-quality single cell subtype samples, which provides a new solution to the high cost of obtaining single cell samples and the sample diversity problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Schematic diagram of a process for generating a single cell subtype sample in an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0029] The present invention proposes a method for generating a single cell subtype sample, such as Figure 1 As shown, including:

[0030] S1. Obtain multiple omics data of single-cell samples; obtain single-cell RNA sequence data from the GTEx database, and collect other omics data at the same time, including single-cell genome, transcriptome sequencing expression data, clinical metadata, and cell type annotation data.

[0031] S2. Data preprocessing module: In this module, the collected data is preprocessed for subsequent analysis and construction of graph networks. The methods used include epigenetic annotation, cell type marker annotation, GO and KEGG database annotation, etc. First, the R package bitr is used to convert the gene ID, and then the enrichGO and enrichKEGG packages are used to annotate the GO and KEGG databases, and the marker genes in the cell marker database are used to annotate the cell types. These gene annotations are all to increase the regulatory relationships in the gene regulatory network. These relationships are the edges in the graph network, so as to obtain a higher quality graph network.

[0032] S3, graph network construction module, converts the various omics data into directed graph format, obtains node matrix and adjacency matrix; converts the preprocessed data into graph network format, generates node file and edge file. Wherein the node file includes nodes (nodes include genes, transcription genes, the above-mentioned annotation information meta-nodes, GO terms, KEGG terms, cell types, etc. are all nodes) and their attributes, and the node attributes include node name, node type, number of nodes, node id, etc. The edge file includes the edge relationship between nodes (edge ​​relationship is the generation of the above-mentioned respective annotation information, and the edge type includes hierarchical relationship, interaction relationship, composition relationship, sequence relationship, multimodal relationship) and edge attributes, and the edge attributes include the id of the edge, whether it exists, and the initial weight. Wherein the graph network construction module refers to the single cell genome, transcriptome sequencing expression data, clinical metadata and cell type annotation data in the format of a directed graph (edges formed by the relationship between nodes and their nodes, each node includes attributes such as node number, count value, node type, and edge includes edge type and edge weight) to obtain a node matrix and an adjacency matrix as the input of the model.

[0033] S4. Construct an adaptive conditional graph diffusion convolution ACGDC model, train the ACGDC model through the node matrix and adjacency matrix, and define learnable parameters for each relationship type to process heterogeneous graph data. Specifically, in the process of convolution and message propagation of input data, first perform feature convolution calculation on the node, then assign a new weight vector to the first-order neighbor edge of the node. After the convolution operation, multiply the vector with the neighbor node vector to generate a new node representation. Finally, perform full connection layer calculation on the node, as well as normalization and nonlinear activation operations, add noise to the adjacency matrix, thereby randomly changing the first-order neighbor relationship of the node; adaptively learn the different importance of each different type of edge when aggregating neighbor information through the added edge weight vector, thereby solving the problem that different edge types in heterogeneous graph networks represent different meanings.

[0034] S5. Use the conditional adaptive graph diffusion convolution model trained above to perform reverse sampling in the Gaussian distribution space to generate new single-cell sample seeds, decode the single-cell sample seeds obtained by reverse sampling, and generate new single-cell subtype samples. In the process of generating single-cell multi-omics data, ACGDC combines the diffusion model with the graph neural network method, and assigns a weight parameter to each edge type, thereby solving the problem that different edge types in heterogeneous graphs have different parameters. The representation of nodes and edges is used as the input of the adaptive conditional graph diffusion convolution model through the graph network conversion module, and new single-cell sample seeds are sampled in the Gaussian distribution space. Finally, new single-cell subtype samples are decoded through the conditional generation process.

[0035] Evaluation of the conditional adaptive graph diffusion convolution model: The performance of the trained adaptive conditional graph diffusion convolution model is evaluated, and the effectiveness, novelty, and diversity of the generated single-cell subtype samples are evaluated. The evaluation modules of the conditional control generated samples include: (1) cluster analysis (the generated single-cell samples are reduced to two dimensions through the T-SNE algorithm, and clustered and spatially distributed through K nearest neighbors); (2) gene expression profile analysis (i.e., gene expression profiles are used to find highly expressed genes, and the genes are ranked, and the ranked lists of highly expressed genes are compared by comparing the generated cell samples with the cell samples in the training set); (3) gene set enrichment analysis (after annotating the genes with pathways and biological functions, genes belonging to the same pathways or biological functions are grouped together, and the differences between the generated samples and the training samples are also compared); (4) similarity comparison with experimental samples (the generated samples and the training samples, that is, the samples obtained in the experiment, are directly calculated by gene expression profiles. The distance difference is the similarity, and the available generated samples are screened by comparing the similarity). The final output result is a comprehensive score of multiple evaluation criteria.

[0036] The present invention defines five edge types: hierarchical relationship, sequence relationship, composition relationship, interaction relationship, and multimodal relationship; adaptive graph diffusion convolution processing refers to message propagation and node feature aggregation of the input directed graph, feature extraction through the convolution operator, and continuous addition of noise to the adjacency matrix based on the diffusion theory, so that the adjacency matrix is ​​finally converted into a Gaussian noise matrix; in the conditional control generation module, first sampling is performed in the Gaussian distribution space to obtain a Gaussian noise matrix, and the Gaussian noise matrix is ​​continuously denoised through the trained model, and the denoised matrix obtained is merged with the label vector of the single cell subtype (the label is the vector after the specific type of the cell subtype is encoded), so as to realize conditional binding of the cell subtype label with the single cell sampling sample, and finally compare with the graph representation of the input single cell sample to calculate the loss, and realize parameter adjustment of the model through the gradient descent algorithm to reduce the loss size, and finally complete the entire training process, so as to realize the conditional control generation sample of the single cell subtype.

[0037] The present invention is specifically described with reference to a specific brain astrocyte sample data as an example, and includes the following steps:

[0038] S1. First, astrocyte single-cell gene and transcriptome sequencing data and clinical metadata were obtained from the GTEx database.

[0039] S2. Data cleaning and annotation, including gene ID conversion, gene / RNA classification, cell type annotation, pathway and function annotation, etc.

[0040] S3. Define the graph structure, including node and edge matrices, and construct a data set. The present invention improves the graph diffusion convolution method to adapt to heterogeneous graph data (i.e., three or more omics will have different types of edges, and this type of graph is called a heterogeneous graph), and introduces conditional sampling to control the directional generation process of samples.

[0041] S4. Use the processed data set to train the improved ACGDC model to learn the edge and attribute relationships between nodes. The specific improvements mainly include two additions to the GDC model: 1) The present invention defines five types of edges. During the message propagation process, when the node is updated by feature convolution, a new weight vector is assigned to the first-order neighbor edge. After the convolution operation, the vector is multiplied with the neighbor node vector to generate a new node representation. The traditional GDC model does not have this step, but directly averages the neighbor node vector, which is equivalent to treating different neighbor edges as equal weights (i.e., only suitable for homogeneous graphs), and then different edge types are of different importance to the node (the method of the present invention can process heterogeneous graphs), so the improvement of the present invention is more reasonable in theory; 2) Based on GDC, the present invention binds the label of the sample to the generation process in the reverse generation process to achieve the conditional generation of cell subtypes, thereby achieving more refined sample generation.

[0042] S5. Finally, use the trained model to generate new single-cell subtype samples through conditional control.

[0043] The ACGDC method proposed in the present invention utilizes a heterogeneous graph network to integrate a variety of omics data including single-cell genome, transcriptome sequencing expression data, clinical metadata, and annotation data such as cell types. This method can generate new single-cell subtype samples, which are highly similar to real samples in spatial distribution and bioinformatics characteristics, and also have certain novelty and diversity. Compared with known generation methods (such as VAE, GAN) and graph neural network methods (such as GDC and F2GNN), the ACGDC method proposed in the present invention shows superior performance in indicators such as model performance, average similarity and novelty of generated samples, and diversity. The generated single-cell subtype samples have been verified to be reliable through multi-dimensional bioinformatics analysis including sample distribution space, enrichment analysis (GO terms, KEGG terms), clustering, and cell subtype classification.

[0044] Based on the same inventive concept, the present invention also proposes a single cell subtype sample generation system, comprising:

[0045] An acquisition module is used to acquire multiple omics data of single-cell samples; the multiple omics data are converted into directed graph formats to obtain a node matrix and an adjacency matrix; the node matrix includes nodes composed of attributes of multiple omics data, and the adjacency matrix includes edges between two nodes composed of omics data weights.

[0046] The model building module is used to build an adaptive conditional graph diffusion convolution (ACGDC) model. The ACGDC model includes a graph convolution layer and a noise diffusion layer.

[0047] The model training module is used to input the node matrix and the adjacency matrix into the ACGDC model, assign a new weight vector to the first-order neighbor edge of the node through the graph convolution layer, multiply the new weight vector with the neighbor node vector of the node to obtain a new node representation; in the noise diffusion layer, random noise is continuously added to the adjacency matrix containing the new node representation at a certain step interval, and the adjacency matrix is ​​converted into a high-noise matrix to obtain the trained ACGDC model.

[0048] The generation module is used to input the multiple omics data of the single-cell samples to be generated into the trained ACGDC model, and generate new single-cell subtype samples by continuously denoising the Gaussian noise matrix.

[0049] Based on the same inventive concept, the present invention also proposes a computer device for generating single-cell subtype samples, comprising: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the single-cell subtype sample generation method when executing the computer program.

[0050] Based on the same inventive concept, the present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the single-cell subtype sample generation method.

[0051] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for generating a single cell subtype sample, characterized in that: The following steps are involved: Obtain multiple omics data from single-cell samples; Converting the plurality of omics data into directed graph formats respectively to obtain a node matrix and an adjacency matrix; the node matrix includes nodes formed by attributes of the plurality of omics data, and the adjacency matrix includes edges formed by weights of the omics data between two nodes; Constructing an adaptive conditional graph diffusion convolution (ACGDC) model; the ACGDC model includes a graph convolution layer and a noise diffusion layer; The node matrix and the adjacency matrix are input into the ACGDC model, a new weight vector is assigned to the first-order neighbor edge of the node through the graph convolution layer, and the new weight vector is multiplied by the neighbor node vector of the node to obtain a new node representation; random noise is continuously added to the adjacency matrix containing the new node representation at a certain step interval in the noise diffusion layer, and the adjacency matrix is ​​converted into a high-noise matrix to obtain the trained ACGDC model; The various omics data of the single-cell samples to be generated are input into the trained ACGDC model, and new single-cell subtype samples are generated by continuously denoising the Gaussian noise matrix.

2. A method for generating a single cell subtype sample according to claim 1, characterized in that: The various omics data include single-cell RNA-seq data, single-cell genome, transcriptome sequencing expression data, clinical metadata, and cell type annotation data.

3. A method for generating a single cell subtype sample according to claim 1, characterized in that: After obtaining the multiple omics data, the data are preprocessed, and the preprocessing process includes the following steps: Use the R package bitr to convert gene IDs for various omics data; The enrichGO and enrichKEGG packages were used to annotate the converted omics data with the GO and KEGG databases; The marker genes in the cell marker database are used to annotate the cell types of the omics data after the database annotation.

4. A method for generating a single cell subtype sample according to claim 1, characterized in that: The method generates new single-cell subtype samples by continuously denoising the Gaussian noise matrix, specifically including obtaining the label vector of the single-cell subtype, and generating new single-cell subtype samples by merging the denoised Gaussian noise matrix with the label vector of the single-cell subtype.

5. The method for generating a single cell subtype sample according to claim 1, characterized in that: The trained ACGDC model is evaluated by a cluster analysis method; specifically, the generated single-cell subtype samples are reduced in dimension by the T-SNE algorithm, the single-cell subtype samples after dimension reduction are clustered by K nearest neighbors, and the clustered single-cell subtype samples are evaluated and analyzed.

6. A single cell subtype sample generation system, characterized in that: include: The acquisition module is used to obtain multiple omics data of single-cell samples; Converting the plurality of omics data into directed graph formats respectively to obtain a node matrix and an adjacency matrix; the node matrix includes nodes formed by attributes of the plurality of omics data, and the adjacency matrix includes edges formed by weights of the omics data between two nodes; A model building module, used to build an adaptive conditional graph diffusion convolution (ACGDC) model; the AC GDC model includes a graph convolution layer and a noise diffusion layer; A model training module is used to input the node matrix and the adjacency matrix into the ACGDC model, assign a new weight vector to the first-order neighbor edge of the node through the graph convolution layer, and multiply the new weight vector with the neighbor node vector of the node to obtain a new node representation; in the noise diffusion layer, random noise is continuously added to the adjacency matrix containing the new node representation at a certain step interval, and the adjacency matrix is ​​converted into a high-noise matrix to obtain the trained ACGDC model; The generation module is used to input the multiple omics data of the single-cell samples to be generated into the trained ACGDC model, and generate new single-cell subtype samples by continuously denoising the Gaussian noise matrix.

7. A computer device for generating single cell subtype samples, characterized in that: include: A memory, a processor, and a computer program stored in the memory, wherein when the processor executes the computer program, the steps of the method for generating a single cell subtype sample according to any one of claims 1 to 5 are implemented.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the method for generating a single cell subtype sample according to any one of claims 1 to 5.

Images