Single-cell classification methods, devices, equipment and storage media

Abstract

This disclosure provides a single-cell classification method and apparatus. The method includes: acquiring gene expression data, including the expression level of genes in each of multiple single cells; constructing a cell relationship graph among the multiple single cells based on the gene expression data, where nodes represent single cells and edges represent relationships between single cells; extracting features from the cell relationship graph using a graph convolutional neural network to obtain a feature embedding representation corresponding to the cell relationship graph; and classifying the multiple single cells using a classification network based on the feature embedding representation to obtain cell types corresponding to each single cell. Thus, by utilizing a graph convolutional neural network and a cell relationship graph that displays higher-order relationships among multiple single cells, the accuracy of single-cell classification is effectively improved.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and in particular to a single-cell classification method, apparatus, device, and storage medium. Background Technology

[0002] In recent years, single-cell sequencing technology has developed rapidly and has been widely applied in the biomedical field. Single-cell sequencing refers to the sequencing of the genome and transcriptome of a single cell. The sequencing results can provide a more granular perspective for studying cellular heterogeneity and complex cellular gene regulatory networks. Among the many research tasks related to single cells, cell type annotation and classification play an important role in studying the ribonucleic acid (RNA) sequences of single cells, facilitating downstream task analysis and helping researchers understand the mechanisms of action of disease cells. With the increasing amount of single-cell data, how to use sequencing results to achieve cell classification is currently one of the key research focuses. Currently, one single-cell classification method involves analyzing the cell similarity between reference data and query data based on sequencing results in a single-cell classification tool, and then classifying cell types based on this similarity.

[0003] However, the above methods have low accuracy in classifying single cells. Summary of the Invention

[0004] This disclosure provides a single-cell classification method, apparatus, device, and storage medium to improve the classification accuracy of single-cell data.

[0005] In a first aspect, embodiments of this disclosure provide a single-cell classification method, including:

[0006] Acquire gene expression data, wherein the gene expression data includes the expression level of genes in each of multiple single cells;

[0007] Based on the gene expression data, a cell relationship graph is constructed among the multiple single cells, wherein nodes represent single cells and edges represent the relationships between single cells in the cell relationship graph.

[0008] The cell relationship graph is subjected to feature extraction using a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph;

[0009] Based on the feature embedding representation, the multiple single cells are classified by a classification network to obtain the cell types corresponding to the multiple single cells.

[0010] Secondly, embodiments of this disclosure provide a single-cell sorting device, comprising:

[0011] An acquisition unit is used to acquire gene expression data, wherein the gene expression data includes the expression level of genes in each of multiple single cells;

[0012] A construction unit is used to construct a cell relationship graph between the plurality of single cells based on the gene expression data, wherein in the cell relationship graph, nodes represent single cells and edges represent the relationships between single cells;

[0013] The feature processing unit is used to extract features from the cell relationship graph through a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph;

[0014] A classification unit is used to classify the plurality of single cells according to the feature embedding representation through a classification network to obtain the cell types corresponding to the plurality of single cells respectively.

[0015] Thirdly, embodiments of this disclosure provide an electronic device, including: at least one processor and a memory; the memory stores computer execution instructions; the at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the single-cell classification method as described in the first aspect above.

[0016] Fourthly, embodiments of this disclosure provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the single-cell classification method described in the first aspect above.

[0017] Fifthly, this disclosure provides a computer program product comprising computer execution instructions, which, when executed by a processor, implement the single-cell classification method described in the first aspect above.

[0018] The single-cell classification method, apparatus, device, and storage medium provided in this disclosure construct a cell relationship graph among multiple single cells based on gene expression data. This analyzes the relationships between multiple single cells at the gene level and accurately describes these relationships using the cell relationship graph. Compared to analyzing the similarity between multiple single cells, analyzing the relationships at the gene level can consider higher-order relationships between single cells. This type of relationship can be used to accurately classify single cells from different species, platforms, and other sources. Next, a graph convolutional neural network is used to extract features from the cell relationship graph, obtaining the corresponding feature embedding representation. The graph convolutional neural network improves the dimensionality reduction and noise reduction effect of the cell relationship graph. Based on the feature embedding representation corresponding to the cell relationship graph, a classification network is used to classify the multiple single cells, obtaining the cell types corresponding to each single cell. Therefore, from multiple aspects such as cell relationship graph construction and graph convolutional neural network feature extraction, the classification accuracy of single cells is effectively improved, especially for single cells from different species, platforms, and other sources. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 Schematic flowchart of the single-cell classification method provided in the embodiments of this disclosure Figure 1 。;

[0021] Figure 2 Schematic flowchart of the single-cell classification method provided in the embodiments of this disclosure Figure 2 ;

[0022] Figure 3 An example graph showing the gene enrichment scores of individual cells in a pathway database;

[0023] Figure 4 Schematic flowchart of the single-cell classification method provided in the embodiments of this disclosure Figure 3 ;

[0024] Figure 5 Example diagrams of the single-cell classification process, model training process, and corresponding model structure provided for embodiments of this disclosure;

[0025] Figure 6 A structural block diagram of the single-cell sorting device provided in the embodiments of this disclosure;

[0026] Figure 7 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0028] In related technologies, single cells can be classified in the following ways:

[0029] Method 1 uses single-cell analysis tools such as Seurat, SingleR, and scmap to analyze the similarity between single cells and classify them based on this similarity. However, Method 1 does not consider higher-order relationships between single cells, resulting in poor classification performance for single-cell data from different species and platforms, and insufficient ability to distinguish cell subpopulations.

[0030] Method two involves using neural networks to classify single cells. These neural networks are typically trained using supervised or semi-supervised methods. In supervised or semi-supervised methods, training requires information from the test set, and retraining is necessary for each new test set, resulting in high training costs and low convenience.

[0031] In addition to the drawbacks mentioned above, single-cell data obtained through sequencing often contains a lot of noise, such as dropout events (noise during the detection process leads to extremely low RNA capture rates, which in turn cause genes with moderate or high expression levels in one cell to be undetectable in another cell); single-cell data is also affected by batch effects due to differences in sequencing methods, sequencing time, and sequencing location. All of these factors contribute to the low accuracy of single-cell classification based on sequencing-derived single-cell data.

[0032] Based on the aforementioned shortcomings, this disclosure provides a single-cell classification method, apparatus, device, and storage medium to improve the accuracy of single-cell classification. In this disclosure, a cell relationship graph is constructed based on gene expression data, accurately representing higher-order relationships between single cells. A Graph Neural Network (GCN) is used to extract features from the cell relationship graph, obtaining the corresponding feature embedding. Finally, a classification network is used to classify multiple single cells based on the feature embedding. Thus, not only are higher-order relationships between single cells considered, but the relationships between cells are also used as graph input to the GCN, achieving full utilization of the relationships between single cells and effectively improving the accuracy of single-cell classification, especially improving the classification effect on single-cell data from different species and platforms, and enhancing the ability to distinguish cell subpopulations. Furthermore, the feature extraction of the cell relationship graph by the GCN also plays a denoising role, reducing the impact of dropout events and improving the accuracy of single-cell classification.

[0033] The technical solutions of this disclosure and how they solve the aforementioned technical problems will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this disclosure will now be described with reference to the accompanying drawings.

[0034] Figure 1 Schematic flowchart of the single-cell classification method provided in the embodiments of this disclosure Figure 1 .like Figure 1 As shown, this single-cell classification method includes:

[0035] S101. Obtain gene expression data, which includes the expression level of genes in each of multiple single cells.

[0036] Among them, a single cell refers to a single cell.

[0037] Gene expression data can include the expression levels of multiple genes within each single cell. Gene expression level refers to the level of gene expression, obtained through single-cell sequencing technology. This involves sequencing the transcriptome of a single cell to obtain sequencing reads, which are positively correlated with gene expression levels.

[0038] In this embodiment, gene expression data can be obtained from a database; or, gene expression data input by a user can be received; or, gene expression data sent by a gene sequencing device can be received.

[0039] S102. Based on gene expression data, construct a cell relationship graph between multiple single cells. In the cell relationship graph, nodes represent single cells and edges represent the relationships between single cells.

[0040] In this embodiment, during the construction of the cell relationship graph, the expression level of genes within a single cell is used as a characteristic of that single cell. Similarity analysis is performed on multiple single cells to obtain the similarity between each pair of single cells. Based on the similarity between each pair of single cells, adjacent single cells are identified. In the cell relationship graph, nodes corresponding to each single cell are constructed, and the nodes corresponding to adjacent single cells are connected to each other via edges. Finally, a cell relationship graph between multiple single cells is obtained.

[0041] S103 uses a graph convolutional neural network to extract features from the cell relationship graph, obtaining the feature embedding representation corresponding to the cell relationship graph.

[0042] The graph convolutional neural network includes an encoder and a decoder. The encoder can be used to encode and denoise the cell relationship graph to obtain the feature embedding representation corresponding to the cell relationship graph.

[0043] In this embodiment, the cell relationship graph is encoded by the encoder in the graph convolutional neural network. During the encoding process, the cell relationship graph is simultaneously denoised to obtain the feature embedding representation output by the encoder, which is the feature embedding representation corresponding to the cell relationship graph.

[0044] S104. Based on the feature embedding representation, multiple single cells are classified through a classification network to obtain the cell types corresponding to each single cell.

[0045] The classification network is a pre-trained neural network for cell classification. During training, the classification network can be trained in a supervised manner based on the feature embeddings corresponding to the cell relationship graphs among the multiple single cells used for training and the cell types corresponding to the multiple single cells used for training.

[0046] In this embodiment, the feature embedding representation can be input into a classification network. In the classification network, the feature embedding representation is processed by multiple network layers to classify multiple single cells and obtain the cell types corresponding to the multiple single cells.

[0047] In this embodiment, based on gene expression data, higher-order relationships between multiple single cells are analyzed to obtain a cell relationship graph. A graph convolutional neural network is then used to extract features from the cell relationship graph, resulting in corresponding feature embedding representations. Finally, a classification network is used to classify the multiple single cells based on these feature embedding representations. This approach not only analyzes higher-order relationships between multiple single cells but also utilizes these relationships for single-cell classification through a graph convolutional neural network. Furthermore, it effectively reduces noise generated during sequencing to some extent, thereby improving the classification accuracy of single cells from multiple perspectives.

[0048] In some embodiments, gene expression data can be a gene expression matrix. In this matrix, each element corresponds to the expression level of a gene. Thus, the gene expression matrix concisely and clearly represents the expression levels of multiple genes within each of multiple single cells.

[0049] Optionally, in the gene expression matrix, one row corresponds to one single cell and one column corresponds to one gene.

[0050] Optionally, in the gene expression matrix, one row corresponds to one gene and one column corresponds to one single cell.

[0051] In some embodiments, a possible implementation of S202 includes: using the expression level of genes within a single cell as a characteristic of the single cell, determining the neighbor nodes of each single cell in multiple single cells using a proximity algorithm, and constructing a cell relationship graph among multiple single cells based on the neighbor nodes of each single cell. The proximity algorithm can also be called the K-nearest neighbors (KNN) algorithm. Thus, the accuracy of the cell relationship graph can be improved to some extent through the KNN algorithm.

[0052] In some embodiments, considering that the KNN algorithm is used to construct cell relationship maps based on the relationships between cells while ignoring the relationships between genes, the improvement in the accuracy of cell relationship maps is limited. To address this issue, another possible implementation of S202 includes: determining the similarity between multiple single cells in the pathway database based on the pathway database and gene expression data; and constructing a cell relationship map between multiple single cells based on the similarity between the multiple single cells in the pathway database. The pathway database includes multiple pathways. Genes are often divided into different sets according to different biological functions. A pathway is a set of genes that regulate the same biological process. For example, glycolysis is a biological metabolic pathway in the human body used to break down glucose into energy, which involves the regulation of genes such as HK3, HK1, and HK2. Therefore, the pathway corresponding to this biological metabolic pathway can be represented as [HK3, HK1, HK2].

[0053] In this embodiment, the similarity between multiple single cells can be analyzed from a biological function perspective based on the expression levels of genes within each single cell in the pathway and gene expression data of the pathway database. Since the similarity or identical biological functions reflect the relationship between genes, analyzing the similarity between multiple single cells from a biological function perspective is a cell similarity analysis performed under the premise of fully considering the relationship between genes, effectively improving the accuracy of the similarity between multiple single cells. Furthermore, based on the similarity between multiple single cells in the pathway database, a cell relationship map between multiple single cells is constructed, further improving the accuracy of the cell relationship map between multiple single cells.

[0054] In some embodiments, there can be multiple cell relationship maps, and different cell relationship maps can correspond to different pathway databases. That is, the similarity between multiple single cells is analyzed based on different pathway databases to construct cell relationship maps between multiple single cells under different pathway databases. Thus, by using different pathway databases, cell relationship maps from multiple perspectives can be provided for single cell classification, thereby improving the accuracy of single cell classification.

[0055] Optionally, the multiple pathway databases can be KEGG, Reactome, WikiPath, and yan, respectively.

[0056] Figure 2 Schematic flowchart of the single-cell classification method provided in the embodiments of this disclosure Figure 2 In this embodiment of the disclosure, there are multiple cell relationship diagrams and multiple pathway databases, with different cell relationship diagrams corresponding to different pathway databases. For example... Figure 2 As shown, this single-cell classification method includes:

[0057] S201. Obtain gene expression data.

[0058] The implementation principle and technical effects of S201 can be referred to in the aforementioned embodiments, and will not be repeated here.

[0059] S202. For each pathway database in multiple pathway databases, determine the similarity between multiple single cells in each pathway database based on gene expression data.

[0060] Each pathway database contains multiple pathways, and different databases may include different pathways. Since a pathway is a collection of genes that regulate the same biological process—that is, a collection of genes with similar or related functions—using multiple pathway databases allows for the determination of the similarity between multiple single cells within each database. This provides an opportunity to analyze the similarity between multiple single cells from multiple functional perspectives, improving the accuracy of similarity assessments.

[0061] In this embodiment, for each pathway database, the similarity between multiple single cells in the pathway database can be determined from the perspective of cell function based on the expression levels of genes within each single cell in the gene expression data of multiple pathways in the pathway database. Thus, in the cell relationship diagram corresponding to each pathway database, single cells are clustered together according to different functions, providing different relational perspectives for single cell classification and effectively improving the accuracy of single cell classification.

[0062] In one possible implementation, such as Figure 2 As shown, S202 includes: S2021, performing gene enrichment analysis on gene expression data based on each pathway database to obtain the gene enrichment score for each single cell in each pathway database; S2022, calculating the similarity between multiple single cells in each pathway database based on the gene enrichment score of each single cell in each pathway database. Thus, gene enrichment analysis improves the accuracy of similarity calculations between multiple single cells in each pathway database.

[0063] Gene enrichment analysis refers to the analysis of the distribution trends of genes in single cells within pathway databases. For each pathway database, the gene enrichment score of a single cell within that database can include the association score between the single cell and various pathways in the database. To a certain extent, the gene enrichment score of a single cell within a pathway database can reflect the distribution trends of genes in the single cell within the pathway database.

[0064] In S2021, for each pathway database, gene enrichment analysis can be performed on each single cell based on multiple pathways and gene expression data in the pathway database, obtaining the gene enrichment score for each single cell under the pathway database. During the gene enrichment analysis, the number of genes within a single cell falling into a pathway can be determined. Based on the gene expression level and the number of genes within a single cell falling into a pathway, the association score between the single cell and each pathway is determined. The gene enrichment score of each single cell under the pathway database is obtained by combining the association scores between the single cell and each pathway. Thus, the gene enrichment score of each single cell under each pathway database can be obtained.

[0065] As an example, Figure 3 This is an example graph showing the gene enrichment scores of individual cells within a pathway database. (Example:) Figure 3 As shown, the gene enrichment scores of individual cells in a pathway database can be represented as a matrix, where each row corresponds to a single cell and each column corresponds to a pathway. The i-th row and j-th column of the matrix corresponds to the association score between the i-th single cell and the j-th pathway. For example, Figure 3 In, N c N represents the number of single cells. p S represents the number of pathways in a pathway database. 11 S 12 ...S1N P They are single-cell C1 and N, respectively. p One path The correlation score between them, S 21 S 22 ...S2N P These represent the association scores between single-cell C2 and pathway P1, ... Single-cell With N p The correlation score between each pathway.

[0066] One possible implementation of S2021 includes: sorting the gene expression levels in each single cell from high to low based on gene expression data; determining the coding value of each gene in the single cell according to the order of gene expression levels; constructing a two-dimensional curve for each single cell under each pathway database, using the coding value of the gene in the single cell as the x-axis coordinate and the number of genes in the single cell falling within each pathway database as the y-axis coordinate; and calculating the area under the curve (AUC) of each single cell based on the two-dimensional curve for each pathway database to obtain the gene enrichment score for each single cell under each pathway database. Thus, by constructing a two-dimensional curve and calculating the AUC within it, the accuracy of the gene enrichment score is improved.

[0067] In this implementation, gene expression levels in individual cells can be obtained from gene expression data. For each individual cell, genes are sorted according to their expression levels from highest to lowest. These genes are then encoded according to their order, resulting in corresponding coding values. For example, starting from 0, the coding value for the first gene is 0, the second gene is 1, and so on. By sorting genes by expression level from highest to lowest, enriched genes (i.e., highly expressed genes) in the cell can be identified. These enriched genes serve as valuable characteristics within individual cells. Subsequent analysis of these enriched genes, based on pathway databases, can then be performed from a cellular function perspective, analyzing the relationships between them and improving the accuracy of similarity analysis between individual cells.

[0068] Next, taking a pathway from a cell and pathway database as an example, after sorting and encoding, the encoded value of a gene within a single cell can be used as the gene's coordinate on the x-axis, and the number of genes within a single cell falling within a pathway can be used as the gene's coordinate on the y-axis, thus constructing a two-dimensional curve. In this two-dimensional curve, the top N percent of genes can be selected according to their sorting order, and the AUC corresponding to these top N percent of genes can be calculated, which is the association score between the single cell and the pathway. In this way, the association scores between a single cell and multiple pathways in the pathway database can be calculated. Combining these association scores, the gene enrichment score of a single cell within the pathway database can be obtained. Furthermore, the gene enrichment score of a single cell within each pathway database can be obtained.

[0069] In this case, N is, for example, 5, which means taking the top 5% of genes.

[0070] Optionally, when the gene expression data is a gene expression matrix, with each row of the gene expression matrix corresponding to a single cell and each column corresponding to a gene, the genes in the single cell are sorted according to their expression levels from highest to lowest. This can include: for each row of the gene expression matrix, sorting the genes in the corresponding cell according to the element values ​​in that row from highest to lowest. This improves the efficiency and accuracy of gene sorting.

[0071] As an example, suppose the gene expression matrix is ​​as follows:

[0072]

[0073] In this gene expression matrix, the first, second, and third rows correspond to single cells C1, C2, and C3, respectively, while the first, second, and third columns correspond to genes G1, G2, and G3, respectively. Sort the genes in single cell C1 according to their expression levels from highest to lowest, resulting in the sequence G2, G3, G1. Encoding these genes according to this sequence yields the corresponding coding values ​​of 0, 1, and 2 for G2, G3, and G1, respectively. On the X-axis, 0 represents G2, 1 represents G3, and 2 represents G1. Then, for each pathway 'a' in the pathway database, the number of genes falling into pathway 'a' on the x-axis is counted, and this number is used as the y-axis coordinate for that gene. This allows for the creation of a two-dimensional curve corresponding to the single cell, which correlates with the distribution of genes within pathway 'a' in that single cell. Next, in the two-dimensional curve graph, the top 5% of genes are selected, i.e., the curve segment containing the points on the x-axis that fall within the top 5%. Finally, the AUC of this curve segment is calculated as the association score between a single cell and pathway a. This process is repeated to calculate the association scores between many single cells and different pathways.

[0074] In S2022, for each pathway database, the similarity between multiple single cells in the same database can be calculated based on the gene enrichment score of each single cell within that database. This allows for the determination of the similarity between multiple single cells across different pathway databases. Furthermore, in calculating the similarity between pairs of single cells, the similarity between their respective gene enrichment scores can be calculated. Therefore, gene enrichment analysis improves the accuracy of similarity calculations.

[0075] One possible approach to calculating the similarity between the gene enrichment scores of each pair of individual cells is to calculate the Euclidean distance between them. This Euclidean distance represents the similarity between the gene enrichment scores of each pair of individual cells, and thus, the similarity between the individual cells themselves.

[0076] In this implementation, for each pathway database, the gene enrichment score of a single cell in the pathway database is obtained by combining the association scores between the single cell and the pathway in the pathway database. Therefore, the gene enrichment scores of multiple single cells can form a score matrix. In this matrix, each row corresponds to the gene enrichment score of a single cell, so the gene enrichment score of a single cell is a row vector. The Euclidean distance between the gene enrichment scores of each pair of single cells can be calculated by calculating the Euclidean distance between each pair of row vectors.

[0077] Furthermore, the Euclidean distance between any two row vectors can be expressed as: ||xy||2, where x and y represent the row vectors corresponding to different single cells.

[0078] S203. Based on the similarity between multiple single cells in each pathway database, construct cell relationship diagrams corresponding to each pathway database.

[0079] In this embodiment, for each pathway database: First, after calculating the similarity between multiple single cells under the pathway data, adjacent single cells can be identified based on the similarity between the multiple single cells; in the cell relationship graph, the nodes corresponding to the single cells are constructed, and the nodes corresponding to adjacent single cells are connected to each other through edges to obtain the cell relationship graph between multiple single cells under the pathway database. Based on each pathway database, a cell relationship graph can be constructed separately, ultimately obtaining the cell relationship graph corresponding to each pathway database.

[0080] S204. Using a graph convolutional neural network, feature extraction is performed on the cell relationship graph to obtain the feature embedding representation corresponding to the cell relationship graph.

[0081] In this implementation, when there are multiple cell relationship graphs, each cell relationship graph can be encoded separately using an encoder in a graph convolutional neural network. During the encoding process, the cell relationship graphs are simultaneously denoised to obtain the feature embedding representations corresponding to each cell relationship graph.

[0082] In one possible implementation, there are multiple graph convolutional neural networks (GNNs), with different GNNs corresponding to different cell relationship graphs. Since different cell relationship graphs correspond to different pathway databases, this can also be understood as different pathway databases corresponding to different GNNs. For each GNN, it is used to encode and denoise the cell relationship graph obtained based on the pathway database corresponding to that GNN. Thus, different GNNs are provided for different pathway databases (or different cell relationship graphs), improving the effectiveness of encoding and denoising the cell relationship graph.

[0083] Based on the fact that different cell relationship graphs correspond to different graph convolutional neural networks, the cell relationship graph can be encoded by the encoder in the graph convolutional neural network corresponding to the cell relationship graph. During the encoding process, the cell relationship graph is denoised to obtain the feature embedding representation corresponding to the cell relationship graph.

[0084] S205. Based on the feature embedding representation, classify multiple single cells using a classification network to obtain the cell types corresponding to each single cell.

[0085] In this embodiment, after obtaining the feature embedding representations corresponding to each cell relationship diagram, the feature embedding representations corresponding to each cell relationship diagram can be fused and input into the classification network, or the feature embedding representations corresponding to each cell relationship diagram can be input into the classification network for fusion. In the classification network, based on the fused feature embedding identifiers, multiple single cells are classified to obtain the cell types corresponding to the multiple single cells.

[0086] In this embodiment, based on multiple pathway databases, the similarity between multiple single cells is analyzed from multiple perspectives of cell function, constructing multiple cell relationship maps for multiple single cells. One pathway database can construct one cell relationship map. Then, a graph convolutional neural network is used to extract features from the multiple cell relationship maps, obtaining corresponding feature embedding representations. Using a classification network and the feature embedding representations corresponding to the multiple cell relationship maps, the classification of multiple single cells is achieved. Therefore, the similarity analysis takes into account higher-order relationships between cells and relationships between genes, and performs similarity analysis from multiple perspectives, improving the accuracy and diversity of the cell relationship maps. This provides cell relationship maps from different perspectives for single-cell classification, improving the accuracy of single-cell classification.

[0087] In some embodiments, feature extraction of a cell relationship graph using a graph convolutional neural network to obtain a feature embedding representation corresponding to the cell relationship graph may include: inputting the cell relationship graph and gene expression data into the encoder of the graph convolutional neural network, encoding and denoising the cell relationship graph based on the gene expression data in the encoder, and obtaining the feature embedding representation corresponding to the cell relationship graph. Thus, introducing gene expression data into the graph convolutional neural network improves the encoding and denoising effects of the cell relationship graph.

[0088] Furthermore, when there are multiple cell relationship graphs, there are also multiple graph convolutional neural networks (GNNs). Different cell relationship graphs correspond to different GNNs. The cell relationship graphs and gene expression data are input into the encoder of the GNN. In the encoder, the cell relationship graphs are encoded and denoised based on the gene expression data to obtain a feature embedding representation. This can include: inputting the cell relationship graphs and gene expression data into the encoder of the corresponding GNN; in the encoder, the cell relationship graphs are encoded and denoised based on the gene expression data to obtain a feature embedding representation of the cell relationship graphs. Thus, by using different GNNs to process different cell relationship graphs and introducing gene expression data into the GNNs, the encoding and denoising performance of the GNNs on cell relationship graphs is effectively improved through these two methods.

[0089] In this implementation, the cell relationship graph is in the form of an adjacency matrix. The cell relationship graph can be standardized. The standardized cell relationship graph and gene expression data are then input into the encoder contained in the graph convolutional neural network corresponding to the cell relationship graph for encoding and denoising, so as to obtain the output data of the encoder, which is the feature embedding representation corresponding to the cell relationship graph.

[0090] Furthermore, before inputting gene expression data into a graph convolutional neural network, preprocessing operations can be performed on the gene expression data to reduce noise in the gene expression data.

[0091] The preprocessing operations for gene expression data may include at least one of the following: feature selection and standardization. Thus, by using one or more of these operations, the preprocessing effect can be improved.

[0092] Feature selection involves identifying the top-ranked genes by expression level in the gene expression data, such as the top 2000 highly expressed genes. This reduces the number of genes with low expression levels, thus minimizing data redundancy. Standardization operations may include median normalization, which determines the average gene expression level per cell based on the gene expression data and then adjusts the expression levels of each gene within each cell according to this average.

[0093] Furthermore, if the preprocessing includes median standardization, the expression level of each gene within each single cell can be calculated using the following formula: X / sum(X)*μ. Here, μ is the average gene expression level per single cell, X is the total gene expression level of the cell, and sum() is the summation function.

[0094] In some embodiments, the encoder of a graph convolutional neural network can be represented as:

[0095]

[0096] in, A represents a cell relationship diagram. This represents a standardized cell relationship diagram, where X represents gene expression data, and W... 1 Here, σ represents the encoder's model parameters, σ represents the activation function, and H represents the encoder's output. It is a degree matrix. It can be represented as i represents the i-th row of the matrix, and j represents the j-th column of the matrix.

[0097] In some embodiments, the decoder in a graph convolutional neural network can be represented as:

[0098]

[0099] in, W is the output of the decoder. 2 These are the model parameters for the decoder.

[0100] The training of a graph convolutional neural network requires the use of its encoder and decoder. Similarly, the application of a graph convolutional neural network to process cell relationship graphs requires the use of its encoder.

[0101] In some embodiments, graph convolutional neural networks (GCNNs) are trained using self-supervised learning. Self-supervised learning is a type of unsupervised learning that extracts supervisory information from unsupervised data for model training. By training GCNNs with self-supervision, they become inductive models, meaning they are less dependent on the training data and can be applied to multiple test sets, rather than requiring retraining on the training set for each new test set. Therefore, self-supervised training reduces the training time of GCNNs, making them more convenient and time-efficient to use.

[0102] In one possible implementation, a training process of a graph convolutional neural network may include: acquiring training data, wherein the training data includes a cell relationship graph for training and gene expression data for training, the cell relationship graph for training being obtained based on the gene expression data for training; masking the expression levels of some genes in the gene expression data for training to obtain masked gene expression data; inputting the masked gene expression data and the cell relationship graph for training into an encoder included in the graph convolutional neural network, where the encoder encodes and denoises the cell relationship graph for training based on the masked gene expression data to obtain a feature embedding representation of the cell relationship graph for training; then, inputting the masked gene expression data and the feature embedding representation of the cell relationship graph for training into a decoder included in the graph convolutional neural network, where the decoder decodes the cell relationship graph for training based on the masked gene expression data to obtain decoder output data; determining the loss value of the graph convolutional neural network based on the decoder output data, the masked gene expression matrix, and the loss function; and adjusting the network parameters of the graph convolutional neural network based on the loss value of the graph convolutional neural network. Thus, self-supervised training of graph convolutional neural networks is achieved. During the training process, the Dropout event is simulated by occluding the cell relationship graph, which enables the graph convolutional neural network to learn the inherent rules in the cell relationship graph, reduce noise, and reduce Dropout events.

[0103] The loss function is used to determine the difference between the decoder's output data and the masked gene expression matrix. Furthermore, the loss function can be expressed as: is the output of the decoder, and X represents the masked gene expression matrix.

[0104] The data processing procedure for the encoder during training can be referred to the data processing procedure for the encoder in the application of the convolutional neural network in the previous embodiments, and will not be repeated here. The data processing procedure for the decoder during training can be referred to the description of the decoder in the previous embodiments, and will not be repeated here.

[0105] When there are multiple graph convolutional neural networks, the above process can be used to train multiple graph convolutional neural networks respectively. Different graph convolutional neural networks use different training data. For example, the cell relationship graph used for training the first graph convolutional neural network is constructed based on the first channel database, the cell relationship graph used for training the second graph convolutional neural network is constructed based on the second channel database, and so on.

[0106] Furthermore, before inputting the gene expression data used for training into the graph convolutional neural network, it is necessary to preprocess the gene expression data used for training. The preprocessing operation for the gene expression data used for training can refer to the process of preprocessing gene expression data in the previous embodiment, and will not be repeated here.

[0107] Figure 4 Schematic flowchart of the single-cell classification method provided in the embodiments of this disclosure Figure 3 ,like Figure 3 As shown, this single-cell classification method includes:

[0108] S401. Obtain gene expression data.

[0109] S402. For each pathway database in multiple pathway databases, determine the similarity between multiple single cells in each pathway database based on gene expression data.

[0110] S403 constructs cell relationship maps corresponding to each pathway database based on the similarity between multiple single cells in each pathway database.

[0111] S404. Using a graph convolutional neural network, feature extraction is performed on each cell relationship graph to obtain the feature embedding representation corresponding to each cell relationship graph.

[0112] The implementation principles and technical effects of S401 to S404 can be referred to in the aforementioned embodiments, and will not be repeated here.

[0113] In the case of multiple cell relationship diagrams, there are also multiple feature embedding representations, with different cell relationship diagrams corresponding to different feature embedding representations.

[0114] S405 uses a multi-view learning network to fuse multiple feature embedding representations to obtain the target embedding representation.

[0115] In this context, multiple feature embeddings are equivalent to multiple views. Multi-View Networks are used to find the corresponding latent space representations for multiple views in the latent space (also known as the subspace) and fuse these latent space representations.

[0116] The multi-view learning network can be found in the relevant content of Cross Partial Multi-View Networks (CPM-Nets), and will not be described in detail here.

[0117] In this embodiment, using the feature embedding representation as the view, a multi-view learning network is employed to search for the latent space representations corresponding to multiple feature embedding representations in the latent space. In the latent space, the latent space representations are distributed according to cell type, with latent space representations for the same cell type being more concentrated, while the latent space representations for different cell types differ significantly. The latent space representations are fused to obtain the target embedding representation. Therefore, by using multi-view learning, the effect of fusing multiple feature embedding representations is improved.

[0118] In one possible implementation, it can be assumed that multiple mapping functions exist, each corresponding to a different feature embedding representation. These mapping functions can also be understood as reconstruction functions, used to reconstruct the output of the multi-view learning network into the corresponding feature embedding representation. For example, the first mapping function reconstructs the output data of the multi-view learning network into the first feature embedding representation, the second mapping function reconstructs the output data into the second feature embedding representation, and so on. Based on this, the training process of the multi-view learning network can include: initializing a number of latent space representations based on the number of training samples, and randomly initializing the parameters of multiple mapping functions; determining the reconstruction error of the multi-view learning network based on the training samples, latent space representations, mapping functions, and reconstruction loss function; updating the parameters of the mapping functions based on the reconstruction error; determining the classification error of the multi-view learning network based on the training samples, their corresponding sample labels, latent space representations, and classification loss function; updating the latent space representation based on the reconstruction error and classification error; and iteratively executing the operations of determining the reconstruction error, updating the parameters of the mapping functions, determining the classification error, and updating the latent space representation until the latent space representation converges. Thus, by combining reconstruction error and training error, the accuracy of multi-view learning networks in finding corresponding representations in the latent space can be improved.

[0119] The training samples may include feature embeddings corresponding to the cell relationship graph used for training. A single training sample may include multiple feature embeddings; for example, if the number of pathway databases is four, a single training sample may include four feature embeddings. The sample labels corresponding to the training samples include the cell type corresponding to each single cell in the cell relationship graph used for training.

[0120] In this implementation, based on the number of training samples, a number of latent space representations are initialized. Assuming the number of samples is N, then the number of latent space representations is... h iLet be the i-th latent space representation. Multiple mapping functions are randomly initialized, with the number of mapping functions matching the number of pathway databases and the number of cell relationship maps corresponding to a training sample. The latent space representation is reconstructed using the mapping functions, resulting in a reconstructed representation. In the reconstruction loss function, the reconstruction error is determined based on the difference between the reconstructed representation and the training sample. The parameters of the mapping functions are updated based on the reconstruction error. In the classification loss function, the classification error is determined based on the sample labels and latent space representations corresponding to the training and training samples. The reconstruction error and classification error can be weighted and summed to obtain the total error. The latent space representation is updated based on the total error.

[0121] Optionally, the reconstruction loss function can be expressed as:

[0122]

[0123] Among them, S n Let N represent the training set. v Indicates the number of samples. h represents the v-th view feature (i.e., feature embedding representation) of the n-th training sample. n Let d represent the latent space features of the nth training sample. v Let θ represent the v-th mapping function. v f v The parameter, l r This indicates the reconstruction error.

[0124] Optionally, the classification loss function can be expressed as:

[0125]

[0126] Where y represents the label of the training sample (i.e., cell type), if y n =y, then Δ(y) n , y)=0, otherwise Δ(y) n F(h, h) = 1, where y represents the label set, and F(h, h) = 1. n ) = h T h, where h represents the latent space representation of a sample, and E represents the expectation. This formula shows that the purpose of this classification loss function is to maximize the similarity of latent space representations between cell types of the same type, while reducing the similarity of latent space representations between different cell types.

[0127] The total loss is obtained by weighted summation of the reconstruction error and classification error. This is essentially combining the reconstruction loss function and the classification loss function using a weighting balancing factor to arrive at the total loss function. The total loss function can be expressed as:

[0128]

[0129] Wherein, λ represents the weight balancing factor, which is used to balance reconstruction error and classification error, and improve the rationality of adjusting the total error of the latent space representation.

[0130] Optionally, the parameters of the mapping function can be updated using stochastic gradient descent.

[0131] The parameter update formula for the v-th mapping function is:

[0132]

[0133] Where, N c That is, the aforementioned N v .

[0134] Optionally, the latent space representation can be updated using stochastic gradient descent.

[0135] The parameter update formula for the nth latent space representation is:

[0136]

[0137] S406, The target embedding representation is input into the classification network, and the classification network classifies multiple single cells based on the target embedding representation to obtain the cell types corresponding to the multiple single cells.

[0138] In this embodiment, the target embedding representation is input into the classification network, and feature extraction is further performed on the target embedding representation in the classification network. Based on the extracted features, multiple single cells are classified to obtain the cell types corresponding to the multiple single cells.

[0139] In this embodiment, multiple cell relationship graphs are constructed for multiple single cells in gene expression data based on multiple pathway databases; feature embedding representations corresponding to each of the multiple cell relationship graphs are extracted using a graph convolutional neural network to achieve dimensionality reduction and noise reduction of the cell relationship graphs; the feature embedding representations corresponding to each of the multiple cell relationship graphs are fused using a multi-view learning network, and multiple single cells are classified based on the fused feature representations, providing beneficial and complementary view information for single cell classification from the perspective of cell function, thereby improving the accuracy of single cell classification.

[0140] In some embodiments, the classification network can be a fully connected neural network. The classification network may include multiple fully connected layers and one activation layer, wherein the activation layer may employ a normalized exponential function (softmax) layer to normalize the predicted probabilities. Thus, the classification network with this structure can classify single cells.

[0141] Optionally, during the training of the classification network, the cross-entropy function can be used as the loss function, and the gradient descent method can be used to update the parameters of the classification network to improve the classification accuracy of the classification network.

[0142] As an example, Figure 5 The diagram illustrates the single-cell classification process, model training process, and corresponding model structure as provided in the embodiments of this disclosure. Figure 5 As shown in the figure, Figure a illustrates the process of constructing a cell relationship graph, Figure b illustrates the model structure of a graph convolutional neural network, the data processing process of a graph convolutional neural network, and the training process of a graph convolutional neural network, Figure c illustrates the data processing process of a multi-view learning network, and Figure d illustrates the model structure of a classification network.

[0143] like Figure 5 As shown in Figure 'a', in the gene expression matrix, each row corresponds to a single cell, and each column corresponds to a gene, with the element value being the gene expression level. There are four pathway databases, such as KEGG, Reactome, WikiPath, and yan. In one of these databases, the genes in the first pathway include RPS11 and RPS18, the genes in the second pathway include ELMO2 and GPR98, the genes in the third pathway include CREB3L1, and the genes in the fourth pathway include PNMA1. This demonstrates that a single pathway database can include multiple pathways. Gene enrichment analysis is performed based on the gene expression matrix and pathway databases, using AUC to obtain gene enrichment scores for multiple single cells across the four pathway databases. These scores are presented in matrix form, with each row corresponding to a single cell and each column corresponding to a pathway, with the element value being the association score between intracellular genes and pathways. Within each pathway database, the similarity between multiple single cells can be determined based on their gene enrichment scores. Figure 5 (Not shown in the image), and then using the KNN algorithm based on similarity, cell relationship maps between multiple single cells are constructed. Since there are 4 pathway databases, 4 cell relationship maps can be obtained.

[0144] like Figure 5 As shown in Figure b, the graph convolutional neural network includes an encoder and a decoder. The encoder takes a gene expression matrix and a cell relationship map as input and outputs feature embeddings corresponding to the four cell relationship maps. In the graph convolutional neural network, the expression levels of some genes in the gene expression matrix can be masked. The masked gene expression matrix X and the cell relationship maps used as training samples are input into the encoder for encoding to obtain feature embeddings. The decoder then decodes these feature embeddings to obtain the decoder's output data. Through X and The difference between the parameters and the loss function are used to determine the loss value of the graph convolutional neural network, and the parameters of the graph convolutional neural network are adjusted according to the loss value.

[0145] Among them, such as Figure 5 As shown, the loss function of the graph convolutional neural network is:

[0146] like Figure 5 As shown in c, through multi-view learning, the corresponding target embedding representation can be found in the latent space for the feature embedding representation from b. The target embedding representation can be reconstructed back into its corresponding feature embedding representation using four reconstruction functions f1, f2, f3, and f4. In the latent space, latent representations of the same or similar cell types are grouped together, while latent representations of dissimilar cell types are distributed more widely. In d, the classification network can be a fully connected layer network. Figure 5 Taking two fully connected layers as an example, a softmax layer can also be included. Figure 5 (Not shown), and finally, multiple single cells are classified to obtain prediction labels, which are the predicted cell types.

[0147] Corresponding to the single-cell classification method in the above embodiments, Figure 6 This is a structural block diagram of a single-cell sorting device provided in an embodiment of this disclosure. For ease of explanation, only the parts relevant to the embodiments of this disclosure are shown. (Refer to...) Figure 6 The single-cell classification device includes: an acquisition unit 601, a construction unit 602, a feature processing unit 603, and a classification unit 604.

[0148] Acquisition unit 601 is used to acquire gene expression data, wherein the gene expression data includes the expression level of genes in each of multiple single cells;

[0149] The construction unit 602 is used to construct a cell relationship graph between multiple single cells based on gene expression data. In the cell relationship graph, nodes represent single cells and edges represent the relationships between single cells.

[0150] The feature processing unit 603 is used to extract features from the cell relationship graph through a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph;

[0151] The classification unit 604 is used to classify multiple single cells through a classification network based on the feature embedding representation, so as to obtain the cell types corresponding to the multiple single cells respectively.

[0152] In some embodiments, there are multiple cell relationship diagrams, and different cell relationship diagrams correspond to different pathway databases. The construction unit 602 is specifically used to: determine the similarity between multiple single cells in each pathway database based on gene expression data; and construct cell relationship diagrams corresponding to each pathway database based on the similarity between multiple single cells in each pathway database.

[0153] In some embodiments, the construction unit 602 is specifically used to: perform gene enrichment analysis on gene expression data based on each pathway database to obtain the gene enrichment score of each single cell in each pathway database; and calculate the similarity between multiple single cells in each pathway database based on the gene enrichment score of each single cell in each pathway database.

[0154] In some embodiments, the construction unit 602 is specifically used for: sorting the gene expression levels in each single cell in descending order based on gene expression data; determining the coding value corresponding to the gene in each single cell based on the order of gene expression levels; constructing a two-dimensional curve of the single cell under each pathway database using the coding value corresponding to the gene in the single cell as the coordinate of the gene on the x-axis and the number of genes in the single cell falling within each pathway database as the coordinate of the y-axis; and calculating the AUC of the single cell under each pathway database based on the two-dimensional curve of the single cell under each pathway database to obtain the gene enrichment score of the single cell under each pathway database.

[0155] In some embodiments, the feature processing unit 603 is specifically used to: input cell relationship graphs and gene expression data into the encoder of a graph convolutional neural network, and encode and denoise the cell relationship graphs based on gene expression data in the encoder to obtain feature embedding representations; wherein the graph convolutional neural network is trained through a self-supervised training method.

[0156] In some embodiments, there are multiple cell relationship graphs and multiple graph convolutional neural networks. Different cell relationship graphs correspond to different graph convolutional neural networks. The feature processing unit 603 is specifically used to: input the cell relationship graph and gene expression data into the encoder contained in the graph convolutional neural network corresponding to the cell relationship graph, and encode and denoise the cell relationship graph based on the gene expression data in the encoder to obtain the feature embedding representation of the cell relationship graph.

[0157] In some embodiments, the feature embeddings are multiple, and different cell relationship diagrams correspond to different feature embeddings. The classification unit 604 is specifically used to: fuse multiple feature embeddings through a multi-view learning network to obtain a target embedding representation; input the target embedding representation into a classification network, and classify multiple single cells based on the target embedding representation in the classification network to obtain cell types.

[0158] In some embodiments, a training process of a graph convolutional neural network may include: acquiring training data, wherein the training data includes a cell relationship graph used for training; occluding the cell relationship graph used for training to obtain an occluded cell relationship graph; encoding and denoising the occluded cell relationship graph using an encoder included in the graph convolutional neural network to obtain a feature embedding representation corresponding to the occluded cell relationship graph; decoding the feature embedding representation corresponding to the occluded cell relationship graph using a decoder included in the graph convolutional neural network to obtain a cell relationship reconstruction graph; determining the loss value of the graph convolutional neural network based on the occluded cell relationship graph, the cell relationship reconstruction graph, and a loss function; and adjusting the network parameters of the graph convolutional neural network based on the loss value of the graph convolutional neural network.

[0159] In some embodiments, the training process of a multi-view learning network includes: initializing a number of latent space representations based on the number of training samples, and randomly initializing the parameters of multiple mapping functions; determining the reconstruction error of the multi-view learning network based on the training samples, latent space representations, mapping functions, and reconstruction loss functions; updating the parameters of the mapping functions based on the reconstruction error of the multi-view learning network; determining the classification error of the multi-view learning network based on the training samples, the sample labels corresponding to the training samples, latent space representations, and classification loss functions; updating the latent space representations based on the reconstruction error and classification error; and iteratively executing the operations of determining the reconstruction error, updating the parameters of the mapping functions, determining the classification error, and updating the latent space representations until the latent space representations converge.

[0160] The training processes of graph convolutional neural networks and multi-view learning networks can be executed on single-cell classification devices or on model training devices.

[0161] The single-cell classification device provided in this embodiment can be used to execute the technical solutions of the above-described single-cell classification method embodiments. Its implementation principle and technical effects are similar, and will not be repeated here.

[0162] refer to Figure 7The diagram illustrates a structural schematic of an electronic device 700 suitable for implementing embodiments of the present disclosure. The electronic device 700 can be a terminal device or a server. The terminal device can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, personal digital assistants (PDAs), portable Android devices (PADs), portable media players (PMPs), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0163] like Figure 7 As shown, the electronic device 700 may include a processing unit (e.g., a central processing unit, a graphics processor, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the electronic device 700. The processing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0164] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows electronic device 700 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 An electronic device 700 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0165] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 709, or installed from storage device 708, or installed from ROM 702. When the computer program is executed by processing device 701, it performs the functions defined in the methods of embodiments of this disclosure.

[0166] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0167] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0168] The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods shown in the above embodiments.

[0169] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0170] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0171] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The names of the units are not necessarily limiting in certain circumstances; for example, an acquisition unit can also be described as "a unit that acquires the page image and page description text of a webpage to be detected".

[0172] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0173] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0174] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0175] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0176] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A single-cell classification method, characterized in that, include: Acquire gene expression data, wherein the gene expression data includes the expression level of genes in each of multiple single cells; Based on the gene expression data, a cell relationship graph is constructed among the multiple single cells. In the cell relationship graph, nodes represent single cells, and edges represent the relationships between single cells. There are multiple cell relationship graphs, and different cell relationship graphs correspond to different pathway databases. The pathway database includes multiple pathways, and each pathway is a set of genes that regulate the same biological process. The cell relationship graph is subjected to feature extraction using a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph. There are multiple graph convolutional neural networks, and different cell relationship graphs correspond to different graph convolutional neural networks. There are also multiple feature embedding representations, and different cell relationship graphs correspond to different feature embedding representations. Based on the feature embedding representation, the multiple single cells are classified by a classification network to obtain the cell types corresponding to the multiple single cells respectively; The step of constructing a cell relationship map among the multiple single cells based on the gene expression data includes: For each pathway database in multiple pathway databases, the similarity between the multiple single cells in each pathway database is determined based on the gene expression data. Based on the similarity between the multiple single cells in each pathway database, a cell relationship map corresponding to each pathway database is constructed. The step of extracting features from the cell relationship graph using a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph includes: The cell relationship graph and the gene expression data are input into the encoder contained in the graph convolutional neural network corresponding to the cell relationship graph. The encoder encodes and denoises the cell relationship graph based on the gene expression data to obtain the feature embedding representation of the cell relationship graph. The step of classifying the plurality of single cells using a classification network based on the feature embedding representation to obtain the cell types corresponding to the plurality of single cells includes: The target embedding representation is obtained by fusing multiple feature embedding representations through a multi-view learning network. The target embedding representation is input into the classification network, and the multiple single cells are classified in the classification network based on the target embedding representation to obtain the cell type.

2. The single-cell classification method according to claim 1, characterized in that, The determination of the similarity between multiple single cells within each pathway database, based on the gene expression data, includes: Based on the respective pathway databases, gene enrichment analysis is performed on the gene expression data to obtain the gene enrichment score of each single cell under each pathway database. Based on the gene enrichment score of each single cell in each pathway database, the similarity between the multiple single cells in each pathway database is calculated.

3. The single-cell classification method according to claim 2, characterized in that, Based on the respective pathway databases, gene enrichment analysis is performed on the gene expression data to obtain the gene enrichment score for each single cell under each pathway database, including: Based on the gene expression data, the gene expression levels in each single cell are sorted in descending order; For each single cell, the coding value of the gene in the single cell is determined according to the order of gene expression levels within the single cell. For each single cell, a two-dimensional curve of the single cell under each pathway in the respective pathway database is constructed, with the coding value corresponding to the gene in the single cell as the coordinate of the gene on the x-axis and the number of genes in the single cell falling within each pathway in the respective pathway database as the coordinate of the y-axis. For each single cell, the area under the curve (AUC) is calculated based on the two-dimensional curve of the single cell under each pathway database to obtain the gene enrichment score of the single cell under each pathway database.

4. The single-cell classification method according to any one of claims 1 to 3, characterized in that, The graph convolutional neural network is trained using a self-supervised training method.

5. A single-cell sorting device, characterized in that, include: An acquisition unit is used to acquire gene expression data, wherein the gene expression data includes the expression level of genes in each of multiple single cells; A construction unit is used to construct a cell relationship graph between the multiple single cells based on the gene expression data. In the cell relationship graph, nodes represent single cells, and edges represent the relationships between single cells. There are multiple cell relationship graphs, and different cell relationship graphs correspond to different pathway databases. The pathway database includes multiple pathways, and each pathway is a set of genes that regulate the same biological process. The feature processing unit is used to extract features from the cell relationship graph through a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph. There are multiple graph convolutional neural networks, and different cell relationship graphs correspond to different graph convolutional neural networks. There are multiple feature embedding representations, and different cell relationship graphs correspond to different feature embedding representations. A classification unit is used to classify the plurality of single cells according to the feature embedding representation through a classification network to obtain the cell types corresponding to the plurality of single cells respectively; When constructing a cell relationship map among the plurality of single cells based on the gene expression data, the construction unit includes: For each pathway database in multiple pathway databases, the similarity between the multiple single cells in each pathway database is determined based on the gene expression data. Based on the similarity between the multiple single cells in each pathway database, a cell relationship map corresponding to each pathway database is constructed. When the feature processing unit extracts features from the cell relationship graph using a graph convolutional neural network to obtain the feature embedding representation corresponding to the cell relationship graph, it includes: The cell relationship graph and the gene expression data are input into the encoder contained in the graph convolutional neural network corresponding to the cell relationship graph. The encoder encodes and denoises the cell relationship graph based on the gene expression data to obtain the feature embedding representation of the cell relationship graph. When the classification unit classifies the plurality of single cells according to the feature embedding representation and through a classification network to obtain the cell types corresponding to the plurality of single cells, it includes: The target embedding representation is obtained by fusing multiple feature embedding representations through a multi-view learning network. The target embedding representation is input into the classification network, and the multiple single cells are classified in the classification network based on the target embedding representation to obtain the cell type.

6. An electronic device, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the single-cell classification method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the single-cell classification method as described in any one of claims 1 to 4.

Images