Graph neural network construction method and apparatus, and electronic device and storage medium

Abstract

The present application relates to the technical field of bioinformatics. Provided are a graph neural network construction method and apparatus, and an electronic device and a storage medium. The method comprises: coding single-cell data in a single-cell data set into a low-dimensional latent space, so as to obtain a first feature vector in which a key feature and / or type-specific information of the single-cell data are / is kept; acquiring a constructed gene regulation network, and coding the gene regulation network to obtain a low-dimensional gene node; extracting a local topological feature from the low-dimensional gene node while keeping the global network structure of the gene regulation network, so as to obtain a second feature vector, wherein the second feature vector is used for reflecting interactions among all genes in the single-cell data set; and constructing a graph neural network on the basis of the first feature vector and the second feature vector, wherein the graph neural network is used for automatically performing labeling on the single-cell data. The present application solves the problem in the relevant art of cell labeling being inaccurate, time-consuming and labor-intensive, and being limited by a batch effect.

Description

Graph neural network construction method, device, electronic device and storage medium Technical Field

[0001] The present application relates to the field of bioinformatics, and more specifically, to a method, device, electronic device, and storage medium for constructing a graph neural network. Background Art

[0002] Cell annotation is a key step in processing single-cell data. Accurate cell type annotation is the basis for cellular communication, type-specific expression patterns, differentiation trajectories, and other downstream analyses.

[0003] Current cell annotation is based on single-cell transcriptome sequencing (scRNA-seq). Single-cell transcriptome sequencing (scRNA-seq) quantifies gene expression profiles at high resolution, enabling a deeper understanding of the molecular basis of cellular heterogeneity and tissue phenotypic differences. This allows us to identify various cell types at single-cell resolution and analyze differences in cell expression profiles. However, technical issues such as inaccurate cell annotation, time-consuming and labor-intensive processes, and batch effects still exist.

[0004] As can be seen from the above, the existing technology has technical problems such as inaccurate cell labeling, time-consuming and labor-intensive, and limited by batch effects.

[0005] Summary of the Invention

[0006] This application provides a graph neural network construction method, device, electronic device, and storage medium that can solve the problems of inaccurate cell labeling, time-consuming and labor-intensive, and limited by batch effects in related technologies. The technical solution is as follows:

[0007] According to one aspect of the present application, a method for constructing a graph neural network includes: encoding single-cell data in a single-cell dataset into a low-dimensional latent space to obtain a first eigenvector that retains key features and / or type-specific information of the single-cell data; obtaining a constructed gene regulatory network and encoding the gene regulatory network into low-dimensional gene nodes; while retaining the global network structure of the gene regulatory network, extracting local topological features of the low-dimensional gene nodes to obtain a second eigenvector; the second eigenvector is used to reflect the interactions between all genes in the single-cell dataset; based on the first eigenvector and the second eigenvector, constructing a graph neural network; the graph neural network is used to automatically annotate the single-cell data.

[0008] According to one aspect of the present application, a graph neural network construction device includes: a single-cell data encoding module, which is used to encode single-cell data in a single-cell dataset into a low-dimensional latent space to obtain a first eigenvector that retains the key features and / or type-specific information of the single-cell data; a gene regulatory network encoding module, which is used to obtain the constructed gene regulatory network and encode the gene regulatory network into low-dimensional gene nodes; a local topological feature extraction module, which is used to extract local topological features of the low-dimensional gene nodes while retaining the global network structure of the gene regulatory network to obtain a second eigenvector; the second eigenvector is used to reflect the interactions between all genes in the single-cell dataset; a graph neural network construction module, which is used to construct a graph neural network based on the first eigenvector and the second eigenvector; the graph neural network is used to automatically annotate single-cell data.

[0009] In an exemplary embodiment, the graph neural network construction module includes: a gene node initial feature acquisition unit, used to use the second feature vector as the initial feature of the gene node in the graph neural network; a cell node initial feature acquisition unit, used to use the first feature vector as the initial feature of the cell node in the graph neural network; an edge construction unit, used to construct the edge of the graph neural network between the gene node and the cell node; and a graph neural network acquisition unit, used to obtain the graph neural network from the edges between the gene nodes and the cell nodes.

[0010] In an exemplary embodiment, the edges of the graph neural network include at least one of the following: gene regulation edges connecting gene nodes; self-feedback edges pointing from gene nodes or cell nodes to themselves; and gene expression edges connecting between gene nodes and cell nodes.

[0011] In an exemplary embodiment, the local topological feature extraction module includes: a second feature vector acquisition unit, used to determine the graph adjacency matrix of the gene regulatory network; based on the graph adjacency matrix, the connected low-dimensional gene nodes are made to have similar embeddings through an autoencoder to capture the local topological features of the low-dimensional gene nodes, and by making the low-dimensional gene nodes with similar neighbor structures have similar embeddings, the second feature vector is obtained while maintaining the global network structure; the second feature vector includes information about the global network structure and the local topological features of each of the low-dimensional gene nodes.

[0012] In an exemplary embodiment, the gene regulatory network encoding module includes: a gene relationship collection unit, used to collect directional gene relationships; a gene regulatory relationship library calling unit, used to store the collected gene relationships in a gene regulatory relationship library, and obtain the gene regulatory network by calling the gene regulatory relationship library.

[0013] In an exemplary embodiment, the graph neural network construction device further includes: a graph neural network training module for training the graph neural network.

[0014] In an exemplary embodiment, the graph neural network training module includes: a reference data set acquisition unit, used to acquire a reference data set; the reference data set includes single-cell data carrying a label; the label indicates the cell type corresponding to the single-cell data; a prediction result acquisition unit, used to train the graph neural network based on the single-cell data in the reference data set to obtain a prediction result; the prediction result indicates the cell type corresponding to the single-cell data; a loss function calculation unit, used to calculate the loss function based on the prediction result and the label, and use the loss function to optimize the parameters of the graph neural network; a repetition unit, used to repeat the above steps based on other single-cell data in the reference data set until the loss function meets a specific condition, and the training of the graph neural network is completed.

[0015] In an exemplary embodiment, the graph neural network construction device also includes: a single-cell data acquisition module for acquiring newly sequenced single-cell data; a probability prediction module for inputting the single-cell data into the graph neural network, so that the cell nodes of the last layer of the graph neural network predict the probability that the single-cell data belongs to different cell types through a linear classifier; and a labeling result acquisition module for obtaining the labeling results of the single-cell data based on the predicted probability.

[0016] According to one aspect of the present application, an electronic device includes at least one processor and at least one memory, wherein the memory stores computer-readable instructions; the computer-readable instructions are executed by one or more of the processors, so that the electronic device implements the graph neural network construction method as described above.

[0017] According to one aspect of the present application, a storage medium stores computer-readable instructions thereon, and the computer-readable instructions are executed by one or more processors to implement the graph neural network construction method as described above.

[0018] According to one aspect of the present application, a computer program product includes computer-readable instructions, which are stored in a storage medium. One or more processors of an electronic device read the computer-readable instructions from the storage medium, load and execute the computer-readable instructions, so that the electronic device implements the graph neural network construction method as described above.

[0019] The beneficial effects of the technical solution provided by this application are:

[0020] In the above technical scheme, the first eigenvector and the second eigenvector are obtained according to the single cell data and the gene regulatory network in the single cell data set, and a graph neural network is constructed based on the first eigenvector and the second eigenvector. The resulting graph neural network can be used to automatically mark the newly sequenced single cell data. When constructing the graph neural network, large-scale gene regulatory relationships (gene regulatory networks) are integrated with single cell RNA sequencing data (single cell data), which can not only reflect the information flow between genes and cells, but also reflect the influence between genes and genes caused by gene regulatory relationships, that is, introduce gene regulatory relationships, consider the influence between genes, so that the graph neural network constructed can more accurately perform cell annotation, whether it is to perform cell annotation within the same data set, or between different data sets, tissues, sequencing platforms and even species. The present invention can achieve higher single cell data annotation accuracy, that is, overcome batch effect; in addition, the present invention carries a constructed gene regulatory network, which saves the time of manually collecting relevant information from various databases, is also more conducive to the user's operation, and saves time and effort. The present invention effectively solves the problems of inaccurate cell annotation, time-consuming and labor-intensive, and being limited to batch effect in the related art. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for describing the embodiments of the present application. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without inventive efforts.

[0022] FIG1 is a schematic diagram of an implementation environment according to the present application;

[0023] FIG2 is a flowchart of a method for constructing a graph neural network according to an exemplary embodiment;

[0024] FIG3 is a flow chart of an embodiment following step 260 in the embodiment corresponding to FIG2 ;

[0025] FIG4 is a flow chart of step 260 in one embodiment of the embodiment corresponding to FIG2 ;

[0026] FIG5 is a flow chart of step 240 in one embodiment of the embodiment corresponding to FIG2 ;

[0027] FIG6 is a flowchart of step 220 “obtaining a constructed gene regulatory network” in the embodiment corresponding to FIG2 in one embodiment;

[0028] FIG7 is a flow chart of an embodiment following step 260 in the embodiment corresponding to FIG2 ;

[0029] FIG8 is a schematic diagram of a specific implementation of a graph neural network construction method in an application scenario;

[0030] FIG9 is a schematic diagram showing a performance comparison between the present invention (scHGR) and other annotation tools within a dataset;

[0031] FIG10 is a schematic diagram of the performance evaluation of the present invention (scHGR) and other annotation tools on cross-sequencing platform datasets;

[0032] FIG11 is a schematic diagram of the performance evaluation of the present invention (scHGR) and other annotation tools on a cross-species dataset;

[0033] FIG12 is a structural block diagram of a graph neural network construction device according to an exemplary embodiment;

[0034] Fig. 13 is a structural block diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0035] The following describes in detail embodiments of the present application, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and are not to be construed as limiting the present application.

[0036] It will be understood by those skilled in the art that, unless expressly stated otherwise, the singular forms "a", "an", "said" and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" used in the specification of the present disclosure refers to the presence of the features, integers, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or groups thereof. It should be understood that when we refer to an element as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. In addition, "connected" or "coupled" as used herein may include wireless connections or wireless couplings. The term "and / or" used herein includes all or any units and all combinations of one or more associated listed items.

[0037] The following is an introduction and explanation of several terms involved in this application:

[0038] Single-cell transcriptome sequencing, also known as single-cell RNA sequencing (scRNA-seq), is a high-throughput gene sequencing technology used to study gene expression patterns in individual cells. It can help us understand differences between different cell types, changes in gene expression during cell development and differentiation, and information about cell function and metabolic pathways. Single-cell transcriptome sequencing can generate RNA sequencing data for individual cells.

[0039] The main steps of single-cell RNA sequencing technology are: step 1, extract a part of cells from the tissue; step 2, separate this group of extracted cells into individual cells; step 3, dissolve the cells to obtain the genetic material in each cell; step 4, synthesize complementary DNA (cDNA) from the RNA in a single cell through the action of reverse transcriptase; step 5, further culture the synthesized cDNA through in vitro transcription (IVT) or polymerase chain reaction (PCR) technology to obtain a large amount of cDNA (i.e., amplification and expansion); step 6, prepare a gene sequencing library and sequence the amplified cDNA to obtain gene expression values ​​(sequencing data).

[0040] A gene expression profile refers to the activity levels of all genes in an organism during the transcription and translation processes under specific conditions. By analyzing gene expression profiles, we can understand the activity of different genes in different tissues, at different developmental stages, or under different environmental conditions, thereby revealing gene functions and regulatory mechanisms. Gene expression profiles can be determined using RNA sequencing data: by sequencing the RNA of a single cell, we can determine which genes in that cell are transcribed into RNA, thereby obtaining the gene expression profile of that cell.

[0041] DeepWalk is an unsupervised learning method based on random walks for learning node embeddings. DeepWalk generates node sequences by performing random walks in a graph and converts these node sequences into embedding vectors using the Word2Vec algorithm. DeepWalk can learn node embeddings without requiring prior knowledge and can be applied to large-scale graph data.

[0042] Node2Vec is a supervised learning method based on random walks for learning node embeddings. Node2Vec treats node embedding as a classification problem and uses deep learning models to learn node embeddings. Node2Vec can capture the structural information between nodes and can be applied to large-scale graph data.

[0043] A hybrid graph neural network (HGNN) is a hybrid model that combines a graph neural network (GNN) with other neural network models. GNNs are a type of neural network model specifically designed for processing graph-structured data. They can learn feature representations for nodes and edges and capture information in the graph structure through message passing and aggregation. However, GNNs can face challenges when processing large-scale graph data, such as high computational complexity and limited information dissemination.

[0044] As mentioned above, the existing technology has technical problems such as inaccurate cell labeling, time-consuming and labor-intensive, and limited by batch effects.

[0045] One cell annotation approach relies solely on clustering of newly sequenced single-cell data. However, different algorithms in this approach can lead to different cell subpopulations, which can lead to biases in subsequent research. Furthermore, both collecting validated cell type-specific marker genes from databases and literature, and further annotating clusters with the help of experts, are time-consuming and labor-intensive.

[0046] Another cell labeling solution relies on automatic labeling technology of cell atlases. With the rapid accumulation of single-cell omics data, large-scale cell atlases are easy to obtain. Based on large-scale cell atlases, artificial intelligence models can be trained to establish mapping relationships from gene expression profiles to specific cell types, and obtain models with the ability to judge cell types based on gene expression profiles. This can automatically perform single-cell level labeling in a reliable and repeatable manner, but this technology does not take into account the information propagation between genes and the gene regulatory relationship.

[0047] From the above, we can see that the relevant technology still has defects such as inaccurate cell labeling, time-consuming and labor-intensive, and limited by batch effects.

[0048] To this end, the graph neural network construction method provided in this application can construct a graph neural network that can accurately and automatically annotate single-cell data. Accordingly, the graph neural network construction method is suitable for a graph neural network construction device, which can be deployed on an electronic device. The electronic device can be a computer device configured with a von Neumann architecture, for example, the computer device includes a desktop computer, a laptop computer, a server, etc.; the electronic device can also refer to a portable and mobile electronic device, for example, the electronic device includes a smartphone, a tablet computer, etc.

[0049] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0050] FIG1 is a schematic diagram of an implementation environment involved in a graph neural network construction method, which includes a terminal 110 and a server 130 .

[0051] Specifically, the terminal 110 may be an electronic device such as a desktop computer, a laptop computer, a tablet computer, a smart phone, etc., which is not limited here.

[0052] The server side 130 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, as well as big data and artificial intelligence platforms.

[0053] The server 130 establishes a communication connection with the terminal 110 in advance through wired or wireless means.

[0054] In one application scenario, a single-cell dataset is input into the terminal 110, and the terminal 110 uploads the single-cell dataset to the server 130, so that the server 130 constructs a graph neural network based on the single-cell dataset and the gene regulatory network. After the server 130 constructs and trains the graph neural network, the user can upload the newly sequenced single-cell data to be annotated to the server 130 through the terminal 110, obtain the annotation results through the graph neural network of the server 130, and transmit them to the terminal 110 for feedback to the user.

[0055] Please refer to Figure 2. An embodiment of the present application provides a graph neural network construction method, which is applicable to an electronic device. The electronic device can be the server 130 in the implementation environment shown in Figure 1, or the terminal 110 in the implementation environment shown in Figure 1.

[0056] In the following method embodiments, for ease of description, the execution subject of each step of the method is taken as an electronic device as an example for illustration, but this does not constitute a specific limitation.

[0057] As shown in FIG2 , the method may include the following steps:

[0058] Step 200: Encode the single-cell data in the single-cell dataset into a low-dimensional latent space to obtain a first feature vector that retains key features and / or type-specific information of the single-cell data.

[0059] Single-cell data can be RNA sequencing data or gene expression profiles, which are not specifically limited here.

[0060] Cell data has the characteristics of high dimension and sparseness, so the cell data can be preprocessed first, for example, by performing dimensionality reduction on the cell data.

[0061] In one possible implementation, the dimensionality reduction technique PCA (Principal Component Analysis) is used to encode the cell data into a low-dimensional latent space to obtain the first eigenvector.

[0062] Step 220: Obtain the constructed gene regulatory network and encode the gene regulatory network into low-dimensional gene nodes.

[0063] A gene regulatory network is a network with genes as nodes, which describes the interactions and regulatory relationships between genes. Each gene is regarded as a node in the network, and the interactions and regulatory relationships between genes are represented by edges connecting the nodes.

[0064] In one possible implementation, the gene regulatory network is constructed based on public databases and platforms.

[0065] Gene regulatory networks are complex and intricate. To better understand and analyze them and reduce their complexity, they can be encoded as low-dimensional gene nodes. By encoding gene regulatory networks as low-dimensional gene nodes, each gene can be represented as a vector with lower dimensionality, reducing computational and storage complexity and making visualization and analysis easier.

[0066] One possible implementation involves using graph embedding technology during the encoding process. Graph embedding can map complex graph structures into a low-dimensional vector space, preserving the relationships and similarities between nodes. By applying graph embedding, gene nodes in a gene regulatory network can be encoded as low-dimensional vectors while preserving their positions and relationships within the global network.

[0067] Step 240 : extracting local topological features of the low-dimensional gene nodes while retaining the global network structure of the gene regulatory network to obtain a second feature vector.

[0068] The second eigenvector is used to reflect the interactions between all genes in the single-cell dataset.

[0069] During the encoding process, we hope to preserve the global network structure, that is, the interaction relationship between genes. At the same time, we also hope to extract the local topological features of each gene to reflect the interaction between genes.

[0070] By extracting local topological features for each gene, we can understand each gene's neighbor nodes and interactions in the network. Local topological features can reflect the interactions between genes and help us understand the mechanisms of gene regulation. By integrating the interactions between all genes, we can more comprehensively analyze and understand the interactions between genes in the entire single-cell dataset (sequencing dataset).

[0071] In one possible implementation, the second eigenvector is obtained using a graph embedding technology, which may be SDNE (Structural Deep Network Embedding), DeepWalk, Node2Vec, etc., which is not specifically limited here.

[0072] Take SDNE, for example. SDNE is an unsupervised learning method based on deep learning that embeds nodes in a graph into a low-dimensional space. SDNE uses an autoencoder, a neural network that encodes input data into a low-dimensional vector and then decodes it back into the original data, to learn node embeddings. SDNE uses an autoencoder to learn node embeddings. It takes nodes as input and attempts to reconstruct them, embedding them into a low-dimensional space. SDNE can capture structural information between nodes and can be applied to large-scale graph data.

[0073] Step 260: construct a graph neural network based on the first eigenvector and the second eigenvector.

[0074] In one possible implementation, each cell and gene is used as a node in a graph neural network, and the first eigenvector and the second eigenvector are used as features of the node.

[0075] The constructed graph neural network extracts the nonlinear correlation between single-cell data and cell types, thereby enabling automatic annotation of newly sequenced single-cell data. As shown in FIG3 , after step 260 , the method may further include the following steps:

[0076] Step 300: Acquire newly sequenced single-cell data.

[0077] Single-cell data can be single-cell RNA sequencing data or gene expression profiles. Single-cell data can be obtained through single-cell RNA sequencing (scRNA-seq).

[0078] In step 320 , the single-cell data is input into the graph neural network, so that the cell nodes of the last layer of the graph neural network pass through a linear classifier to predict the probability that the single-cell data belongs to different cell types.

[0079] In one possible implementation, single-cell data is encoded into vectors and input into a graph neural network, which then processes and transmits the input vectors through a multi-layer neural network model.

[0080] Step 340: Obtain labeling results for the single-cell data based on the predicted probability.

[0081] In one possible implementation, the cell type with the highest probability is used as the annotation result of the single-cell data.

[0082] By the above process, the first eigenvector and the second eigenvector are obtained according to the single cell data and the gene regulatory network in the single cell data set, and a graph neural network is constructed based on the first eigenvector and the second eigenvector. The graph neural network obtained can be used for automatically marking the single cell data of the new sequencing. When constructing a graph neural network, large-scale gene regulatory relationships (gene regulatory networks) are integrated with single cell RNA sequencing data (single cell data), which can not only reflect the information flow between genes and cells, but also reflect the influence between genes and genes caused by gene regulatory relationships, i.e., introduce gene regulatory relationships, consider the influence between genes, so that the graph neural network constructed can more accurately perform cell annotation, whether it is to perform cell annotation within the same data set, or between different data sets, tissues, sequencing platforms and even species. The present invention can achieve higher single cell data annotation accuracy, i.e., overcome batch effect; in addition, the present invention carries the constructed gene regulatory network, saves the time of manually collecting relevant information from various databases, is also more conducive to the operation of the user, and saves time and effort. The present invention effectively solves the problem that the cell annotations present in the related art are inaccurate, time-consuming and labor-intensive, and are limited to batch effect.

[0083] Referring to FIG. 4 , in an exemplary embodiment, step 260 may include the following steps:

[0084] Step 400: Use the second eigenvector as the initial feature of the gene node in the graph neural network.

[0085] When constructing a graph neural network, each cell and gene is treated as a node in the graph neural network, and the characteristics of the nodes of the graph neural network are determined.

[0086] Step 420: Use the first eigenvector as the initial feature of the cell node in the graph neural network.

[0087] Step 440: construct edges of the graph neural network between the gene nodes and the cell nodes.

[0088] The edges of graph neural networks are used to determine the connection relationship between nodes. The construction of edges of graph neural networks can be based on attribute similarity, spatial proximity, common attributes, etc. between nodes.

[0089] In a possible implementation, if the expression level of a gene in a cell is higher than 0, there is a bidirectional edge between the gene and the cell node, that is, a gene expression edge connecting the gene node and the cell node.

[0090] In one possible implementation, if one gene regulates another gene, there is a unidirectional edge between the two genes, that is, a gene regulation edge connecting the gene nodes.

[0091] In one possible implementation, each node has an edge pointing to itself, that is, a self-feedback edge from a gene node or a cell node pointing to itself.

[0092] Step 460: Obtain a graph neural network based on the edges between gene nodes and cell nodes.

[0093] In the above embodiment, the second eigenvector reflecting the interactions between all genes in the single-cell dataset is used to construct a graph neural network. Compared with the prior art that ignores the influence between genes, this application introduces gene regulatory relationships into single-cell data annotation, which can improve the accuracy of cell annotation.

[0094] Referring to FIG. 5 , in an exemplary embodiment, step 240 may include the following steps:

[0095] Step 500: Determine the graph adjacency matrix of the gene regulatory network.

[0096] A graph adjacency matrix is ​​a data structure that represents a graph. It is a two-dimensional matrix in which rows and columns represent vertices in the graph, and the elements in the matrix represent the connection relationships between vertices.

[0097] The graph adjacency matrix of the gene regulatory network is used to represent the connection relationship between each node in the gene regulatory network.

[0098] In step 520, based on the graph adjacency matrix, the connected low-dimensional gene nodes are made to have similar embeddings through the autoencoder to capture the local topological features of the low-dimensional gene nodes, and the second eigenvector is obtained while maintaining the global network structure by making the low-dimensional gene nodes with similar neighbor structures have similar embeddings.

[0099] The second eigenvector includes information about the global network structure and the local topological features of each low-dimensional gene node.

[0100] As mentioned above, the graph adjacency matrix of the gene regulatory network is used to represent the connection relationship between the nodes of the gene regulatory network. Therefore, based on the graph adjacency matrix of the gene regulatory network, the connected low-dimensional gene nodes and the low-dimensional gene nodes with similar neighbor structures can be determined.

[0101] Under the action of the above embodiment, the gene regulatory network is reduced to a continuous vector so that it can serve as the initial feature of the graph neural network, thereby integrating the complex relationships between genes into the gene nodes, facilitating the information propagation between cell nodes and gene nodes in the graph neural network, and making the single-cell data annotation of the graph neural network more accurate.

[0102] Referring to FIG. 6 , in an exemplary embodiment, step 220 of “obtaining a constructed gene regulatory network” may include the following steps:

[0103] Step 600: Collect directional gene relationships.

[0104] Gene relationship refers to the interaction and influence between genes. Different relationships can exist between genes, including interaction, complementary effect, antagonism, etc., which does not constitute a specific limitation here.

[0105] In a possible implementation, the source of gene relationships includes public gene regulation information databases and platforms, such as GREDB, BioGRID, TRRUST, and RegNetwork, etc., which are not specifically limited here.

[0106] Among them, TRRUST is a manually annotated database that effectively extracts regulatory relationships from more than 20 million published articles, involving human and mouse data.

[0107] RegNetwork was developed based on 25 databases, integrating regulatory relationships within the database and potential relationships inferred based on transcription factor binding sites.

[0108] BioGRID is a biomedical interactive repository that includes focused low-throughput studies and large high-throughput datasets, spanning 80,848 papers and multiple species.

[0109] GREDB is a publicly available, manually curated biological database that preserves regulatory relationships in human gene expression.

[0110] Step 620: Store the collected gene relationships in a gene regulatory relationship library, and obtain a gene regulatory network by calling the gene regulatory relationship library.

[0111] The gene regulatory relationship library covers a variety of gene regulatory relationships and is the basic data for building and training graph neural networks.

[0112] In the above embodiments, the present invention constructs a large-scale gene regulation relationship library, which saves the time of manually collecting relevant information from various databases, saves time and labor, and is easy to use.

[0113] Referring to FIG. 7 , in an exemplary embodiment, after step 260 , the method may further include training the graph neural network. The training process includes the following steps:

[0114] Step 700: Acquire a reference dataset.

[0115] The reference dataset includes single-cell data with labels, where the labels indicate the cell type corresponding to the single-cell data. The single-cell data refers to expression data (such as RNA sequencing data or gene expression profiles).

[0116] With the rapid accumulation of single-cell omics data, large amounts of detailed labeling datasets (reference datasets) are readily available.

[0117] Step 720: Train a graph neural network based on the single-cell data in the reference dataset to obtain a prediction result.

[0118] The prediction results indicate the cell type to which the single-cell data corresponds.

[0119] Step 740: Calculate the loss function based on the prediction results and labels, and use the loss function to optimize the parameters of the graph neural network.

[0120] In one possible implementation, the parameters are optimized by minimizing the gap between the predictions and the labels (a loss function).

[0121] Step 760: Repeat the above steps based on other single-cell data in the reference dataset until the loss function meets specific conditions, and the training of the graph neural network is completed.

[0122] The specific condition may be that the loss function reaches a specific range, or that the loss function converges, which is not specifically limited here.

[0123] Through the above-mentioned embodiments, the graph neural network is trained to establish the relationship between the gene expression level and cell type in each cell.

[0124] Figure 8 is a schematic diagram of a specific implementation of a graph neural network construction method in an application scenario.

[0125] In this application scenario, A represents input. In Figure 8, the input includes a reference dataset and a query dataset. The reference dataset contains annotated gene expression profiles (single-cell data), and the query dataset includes newly sequenced gene expression profiles (unlabeled).

[0126] For each reference dataset, cell populations with fewer than 10 cells were first removed to ensure adequate training of the hybrid graph neural network. Different strategies were employed for gene selection depending on the data distribution. All genes were first sorted by dispersion, which is the quotient of the mean and variance. If the reference and query data distributions were similar, the top 1000 genes were selected and the raw reads were used as input for the technique. Otherwise, the top 5000 genes were retained, and the gene reads were log-normalized.

[0127] B is a gene regulatory relationship database, covering gene relationship records for humans and mice, and encompassing multiple gene regulatory information databases, including GREDB, BioGRID, TRRUST, and RegNetwork. It facilitates cell recognition by integrating signal propagation between genes. In this application scenario, only gene regulatory data with experimental evidence are collected.

[0128] C, the gene encoder, uses a graph embedding algorithm to generate initial embeddings for gene nodes using relevant gene information obtained from the gene regulatory relationship library B. Using the classic graph embedding technique SDNE, gene encoder C encodes the complex gene regulatory network into low-dimensional gene nodes. While preserving the global network structure, it extracts the local topological features of each gene to reflect the interactions between all genes in the sequencing dataset.

[0129] D is the cell encoder, which encodes single-cell data (gene expression profiles) into low-dimensional embeddings of cell nodes. Using the nonlinear dimensionality reduction technique PCA, the cell encoder D encodes the high-dimensional and sparse original gene expression profiles into a low-dimensional latent space, preserving key features and type-specific information from the original data.

[0130] E is a hybrid graph neural network. In this application scenario, each cell and gene is used as a node in the hybrid graph neural network, and the initial features are low-dimensional vectors constructed by the cell encoder D and the gene encoder C, respectively. The hybrid graph neural network E includes six types of edges, such as "intracellular gene expression", "gene regulation" and "self-feedback". Figure 8 shows five types of node relationships, such as a, b, c, d, and e. The hybrid graph neural network E extracts the nonlinear correlation between expression patterns (gene expression profiles) and cell types, and updates the node embeddings so that cell nodes of the same type are similar and cell nodes of different types are significantly different. The hybrid graph neural network E optimizes parameters and completes training by minimizing the gap between the predicted results and the reference labels.

[0131] F is the output of the hybrid graph neural network E, including automatic cell annotation, differentially expressed gene analysis (DEG analysis) based on propagation scores, revealing new cell subtypes with biological significance (Subtype revealing), and gene ontology enrichment (GO enrichment) for major diseases.

[0132] In this application scenario, a single-cell data automatic annotation tool based on a deep learning framework is provided. The tool uses existing large-scale single-cell data (single-cell atlas) as a reference, and explores the relationship between gene expression levels and cell categories in each cell through graph embedding technology, dimensionality reduction technology, and hybrid graph neural network technology. The cell type of the reference data set is migrated to the newly sequenced query data set to achieve the function of automatic annotation. Moreover, even if the reference data set and the query data set come from different batches, different sequencing technologies, different tissues, or even species, through the hybrid graph neural network, the tool can still achieve the cell type transfer between them, that is, overcome the batch effect between large-span data sets, and the tool can achieve accurate annotation of single-cell data across tissues, sequencing platforms, and even species. While annotating cell types, the present invention can screen key genes and perform downstream work such as differential gene analysis.

[0133] The present invention is verified by the following experiments.

[0134] As shown in FIG9 , it is a schematic diagram showing the performance comparison of scHGR (the present invention) and other annotation techniques within a dataset.

[0135] The annotation performance of the present invention (scHGR) was compared with that of other annotation technologies (scHPL, scGAC, scDeepSort, and scVI). Experimental validation was performed using a human peripheral blood mononuclear cell (PBMC-FACS) dataset and a mouse brain (AMB) dataset. To avoid experimental evaluation bias, a five-fold cross-validation strategy was employed for each dataset. Figure 9 shows the average values ​​of the evaluation metrics. As shown in Figure 9, multiple metrics (Accuracy, Precision, Recall, F1-Score, and MCC) were used to comprehensively quantify the effectiveness of different annotation technologies. For PBMC-FACS, all five metrics reached the same conclusion: scHGR (the present invention) outperformed other state-of-the-art technologies. In particular, its accuracy was 6% higher than the second-best candidate, demonstrating scHGR's strength in avoiding false positive samples. scVI ranked second only to scHGR because its generation module mitigates the inherent limitation of high sparsity in raw sequencing data.

[0136] In the AMB, scHGR (present invention) achieved near-perfect predictions, with mean precision and Matthews correlation coefficient (MCC) reaching 99%. Clearly, scHGR (present invention) achieved precision, recall, and F1-score 20% higher than other methods, demonstrating its superiority in overcoming false positives and annotating large datasets. Furthermore, the error band at the top of the histogram indicates that scHGR (present invention) achieved the lowest error, demonstrating its robustness and balanced performance.

[0137] As shown in FIG10 , it shows a schematic diagram of the performance evaluation of the present invention (scHGR) and other annotation tools on a cross-sequencing platform dataset.

[0138] This includes heatmaps of F1-score and accuracy derived from cross-technology experiments, with rows representing annotation tools and columns representing datasets from different sequencing platforms. Boxplots summarize the maximum, minimum, median, and quantile accuracy of different annotation tools, as well as the F1-score. The median value is indicated above the corresponding box.

[0139] Among them, the effectiveness of the present invention in annotating data across sequencing platforms was measured on datasets from six different sequencing platforms. The experimental results on the cross-sequencing platform datasets showed significant differences: PBMC1-SM2 achieved almost perfect annotation (scHGR: F1-Score = 94%), but the PBMC1-SW annotation results were generally poor (except scHGR, all methods had F1-Scores of 60%). Except for scHPL, the accuracy of each method did not differ significantly, all exceeding 70%. In contrast, scHGR's F1-Score was significantly higher than that of other technologies (scHPL: 26%, scGAC: 58%, scDeepSort: 63%, scVI: 67%, scHGR: 80%), demonstrating that the present application (scHGR) has cross-technology identifiability (that is, it overcomes batch effects between large-span datasets and can achieve accurate annotation of single-cell data across sequencing platforms). The gene regulation considered by scHGR is not constrained by the sequencing platform, which is also the key reason for its high robustness.

[0140] In order to prove that the present application scheme can achieve accurate annotation of single-cell data across species, the inventors used the human dataset as a reference to annotate the mouse dataset, and used the mouse dataset as a reference to annotate the human dataset.

[0141] As shown in Figure 11, it shows a schematic diagram of the performance evaluation of different annotation tools on a cross-species dataset. The bar graph shows the accuracy of all annotation tools in two cases, and the specific numerical values ​​are marked above the corresponding bar graphs.

[0142] Among them, the accuracy of the present invention and existing technologies in automatically annotating cross-species single-cell sequencing data was compared. Overall, scDeepSort and scVI showed unstable performance, with accuracy dropping by more than 20%, indicating that their effectiveness depends largely on the quality and scale of the reference dataset. In contrast, the reason why scHGR remains stable and effective is that the hybrid graph neural network forms a cross-species information flow through homologous genes. For human-mouse, scGAC, scDeepSort, scVI and scHGR perform significantly better than scHPL, among which scHGR slightly outperforms other technologies. However, in the mouse-human scenario, due to the insufficient number of reference cells, the prediction results are often unsatisfactory. In this case of insufficient training, only scHGR maintains a high degree of reliability.

[0143] Combining Figures 9, 10, and 11, it can be seen that the method of the present application (scHGR) achieves better results than other annotation tools, whether within a single dataset or across sequencing platform datasets and species datasets, and effectively solves the problems of inaccurate cell annotation, time-consuming and labor-intensive, and limited by batch effects in related technologies.

[0144] The following are embodiments of the apparatus of the present application, which can be used to execute the graph neural network construction method involved in this application. For details not disclosed in the apparatus embodiments of the present application, please refer to the method embodiments of the graph neural network construction method involved in this application.

[0145] Please refer to Figure 12. An embodiment of the present application provides a graph neural network construction device 900, including but not limited to: a cell data encoding module 910, a gene regulatory network encoding module 930, a local topological feature extraction module 950 and a graph neural network construction module 970.

[0146] The cell data encoding module 910 is configured to encode the single-cell data in the single-cell data set into a low-dimensional latent space to obtain a first feature vector that retains key features and / or type-specific information of the single-cell data.

[0147] The gene regulatory network encoding module 930 is used to obtain the constructed gene regulatory network and encode the gene regulatory network into low-dimensional gene nodes.

[0148] The local topological feature extraction module 950 is used to extract the local topological features of the low-dimensional gene nodes while retaining the global network structure of the gene regulatory network to obtain a second eigenvector; the second eigenvector is used to reflect the interactions between all genes in the single-cell dataset.

[0149] A graph neural network construction module 970 is used to construct a graph neural network based on the first eigenvector and the second eigenvector; the graph neural network is used to automatically annotate single-cell data.

[0150] It should be noted that the graph neural network construction device provided in the above embodiment only uses the division of the above-mentioned functional modules as an example when constructing the graph neural network. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the graph neural network construction device will be divided into different functional modules to complete all or part of the functions described above.

[0151] In addition, the graph neural network construction device and the graph neural network construction method provided in the above embodiments belong to the same concept, and the specific way in which each module performs operations has been described in detail in the method embodiments and will not be repeated here.

[0152] Please refer to FIG. 13 . An electronic device 4000 is provided in an embodiment of the present application. The electronic device 4000 may include a desktop computer, a laptop computer, a server, etc.

[0153] In FIG. 13 , the electronic device 4000 includes at least one processor 4001 and at least one memory 4003 .

[0154] Data exchange between the processor 4001 and the memory 4003 can be achieved through at least one communication bus 4002. The communication bus 4002 may include a path for transmitting data between the processor 4001 and the memory 4003. The communication bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. The communication bus 4002 can be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one thick line is used in FIG13 , but this does not mean that there is only one bus or one type of bus.

[0155] Optionally, the electronic device 4000 may further include a transceiver 4004, which may be used for data exchange between the electronic device and other electronic devices, such as data transmission and / or data reception. It should be noted that in actual applications, the number of transceivers 4004 is not limited to one, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of the present application.

[0156] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 4001 may also be a combination that implements computing functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, and the like.

[0157] The memory 4003 can be a ROM (Read Only Memory) or other types of static storage devices that can store static information and instructions, a RAM (Random Access Memory) or other types of dynamic storage devices that can store information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), a CD-ROM (Compact Disc Read Only Memory) or other optical disk storage, optical disc storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), a magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program instructions or codes in the form of instructions or data structures and can be accessed by the electronic device 400, but is not limited to these.

[0158] Computer-readable instructions are stored in the memory 4003 , and the processor 4001 can read the computer-readable instructions stored in the memory 4003 through the communication bus 4002 .

[0159] The computer-readable instructions are executed by one or more processors 4001 to implement the graph neural network construction method in the above-mentioned embodiments.

[0160] In addition, an embodiment of the present application provides a storage medium on which computer-readable instructions are stored, and the computer-readable instructions are executed by one or more processors to implement the graph neural network construction method as described above.

[0161] In an embodiment of the present application, a computer program product is provided. The computer program product includes computer-readable instructions. The computer-readable instructions are stored in a storage medium. One or more processors of an electronic device read the computer-readable instructions from the storage medium, load and execute the computer-readable instructions, so that the electronic device implements the graph neural network construction method as described above.

[0162] Compared with related technologies, this application can produce the following beneficial effects:

[0163] 1. Obtain the first eigenvector and the second eigenvector according to the single-cell data and gene regulatory network in the single-cell data set, construct a graph neural network based on the first eigenvector and the second eigenvector, and the resulting graph neural network can be used to automatically mark the newly sequenced single-cell data. When constructing a graph neural network, large-scale gene regulatory relationships (gene regulatory networks) are integrated with single-cell RNA sequencing data (single-cell data), which can not only reflect the information flow between genes and cells, but also reflect the influence between genes and genes caused by gene regulatory relationships, i.e., introduce gene regulatory relationships, take into account the influence between genes, so that the constructed graph neural network can more accurately perform cell annotation, whether it is to perform cell annotation within the same data set, or between different data sets, tissues, sequencing platforms, or even species. The present invention can achieve higher single-cell data annotation accuracy, i.e., overcome batch effect; in addition, the present invention carries a constructed gene regulatory network, which saves the time of manually collecting relevant information from various databases, is also more conducive to the user's operation, and saves time and effort. The present invention effectively solves the problems of inaccurate cell annotation, time-consuming and labor-intensive, and being limited to batch effect in the related art.

[0164] 2. Provides a single-cell data automatic annotation tool based on a deep learning framework. This tool uses existing large-scale single-cell data (single-cell atlases) as a reference. Through graph embedding, dimensionality reduction, and hybrid graph neural network technologies, it explores the relationship between gene expression levels and cell types in each cell. It then migrates cell types from the reference dataset to the newly sequenced query dataset, achieving automatic annotation. Furthermore, by overcoming batch effects between large-scale datasets, the hybrid graph neural network can accurately annotate single-cell data across tissues, sequencing platforms, and even species.

[0165] 3. Gene expression patterns are highly correlated with cell differentiation and biological functions. The present application introduces gene regulation relationships, a key factor affecting gene expression levels, into the task of cell identity annotation. Specifically, directed edges of gene regulation are added to the graph neural network to achieve information exchange between cells and genes, gene to gene, cell to cell, and between genes themselves and cells themselves, thereby promoting the accuracy of single-cell data annotation. Whether cell annotation is performed within a dataset or between different datasets, tissues, sequencing platforms, or even species, the present invention can achieve higher accuracy.

[0166] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a portion of the sub-steps or stages of other steps.

[0167] The above description is only part of the implementation methods of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for constructing a graph neural network, characterized in that: The method comprises: Encoding the single-cell data in the single-cell data set into a low-dimensional latent space to obtain a first feature vector that retains key features and / or type-specific information of the single-cell data; Acquire the constructed gene regulatory network, and encode the gene regulatory network into low-dimensional gene nodes; While retaining the global network structure of the gene regulatory network, extracting local topological features of the low-dimensional gene nodes to obtain a second feature vector; the second feature vector is used to reflect the interaction between all genes in the single-cell dataset; A graph neural network is constructed based on the first eigenvector and the second eigenvector; the graph neural network is used to automatically annotate single-cell data.

2. The method according to claim 1, characterized in that The constructing a graph neural network based on the first feature vector and the second feature vector includes: Using the second feature vector as the initial feature of the gene node in the graph neural network; Using the first feature vector as the initial feature of the cell node in the graph neural network; Constructing edges of the graph neural network between the gene nodes and the cell nodes; The graph neural network is obtained from the edges between gene nodes and cell nodes.

3. The method according to claim 2, characterized in that The edges of the graph neural network include at least one of the following: Gene regulatory edges connecting gene nodes; The self-feedback edge of a gene node or cell node pointing to itself; Gene expression edges connecting gene nodes and cell nodes.

4. The method according to claim 1, characterized in that The step of extracting local topological features of the low-dimensional gene nodes to obtain a second feature vector while retaining the global network structure of the gene regulatory network comprises: determining a graph adjacency matrix of the gene regulatory network; Based on the graph adjacency matrix, the connected low-dimensional gene nodes are made to have similar embeddings through an autoencoder to capture the local topological features of the low-dimensional gene nodes, and the low-dimensional gene nodes with similar neighbor structures are made to have similar embeddings while maintaining the global network structure. The second feature vector is obtained; the second feature vector includes information about the global network structure and local topological features of each of the low-dimensional gene nodes.

5. The method according to claim 1, characterized in that The step of obtaining the constructed gene regulatory network comprises: Collect directional genetic relationships; The collected gene relationships are stored in a gene regulation relationship library, and the gene regulation network is obtained by calling the gene regulation relationship library.

6. The method according to claim 1, characterized in that After constructing the graph neural network based on the first feature vector and the second feature vector, the method further includes: Training the hybrid graph neural network; The training process includes: Acquire a reference data set; the reference data set includes single-cell data carrying a label; the label indicates the cell type corresponding to the single-cell data; Training the graph neural network based on the single cell data in the reference data set to obtain a prediction result; the prediction result indicates the cell type corresponding to the single cell data; Calculating a loss function according to the prediction result and the label, and optimizing the parameters of the graph neural network using the loss function; The above steps are repeated based on other single-cell data in the reference data set until the loss function meets specific conditions, and the training of the graph neural network is completed.

7. The method according to any one of claims 1 to 6, characterized in that: After constructing the graph neural network based on the first feature vector and the second feature vector, the method further includes: Obtain newly sequenced single-cell data; Inputting the single cell data into the graph neural network so that the cell nodes of the last layer of the graph neural network predict the probability that the single cell data belongs to different cell types through a linear classifier; Based on the predicted probability, the annotation result of the single cell data is obtained.

8. A graph neural network construction device, characterized in that: include: A single-cell data encoding module, used to encode the single-cell data in the single-cell data set into a low-dimensional latent space to obtain a first feature vector that retains key features and / or type-specific information of the single-cell data; A gene regulatory network encoding module, used to obtain the constructed gene regulatory network and encode the gene regulatory network into low-dimensional gene nodes; A local topological feature extraction module, used to extract local topological features of the low-dimensional gene nodes while retaining the global network structure of the gene regulatory network to obtain a second feature vector; the second feature vector is used to reflect the interaction between all genes in the single-cell dataset; A graph neural network construction module is used to construct a graph neural network based on the first eigenvector and the second eigenvector; the graph neural network is used to automatically annotate single-cell data.

9. An electronic device, characterized in that: include: at least one processor and at least one memory, wherein: The memory has computer-readable instructions stored thereon; The computer-readable instructions are executed by one or more of the processors, so that the electronic device implements the graph neural network construction method as described in any one of claims 1 to 7.

10. A storage medium having computer-readable instructions stored thereon, characterized in that: The computer-readable instructions are executed by one or more processors to implement the graph neural network construction method as described in any one of claims 1 to 7.

Images