A high-precision single-cell classification method and device based on artificial intelligence algorithm

Abstract

The present invention discloses a high-precision single-cell classification method and device based on an artificial intelligence algorithm, including: first, converting the original single-cell data into a csv format, and constructing an expression matrix to describe the expression level of genes under different conditions; then, preliminarily dividing the single-cell categories through dimensionality reduction and cluster analysis; further, combining PPI and GO information, using a graph neural network method to quantitatively characterize and classify genes; then, constructing a single-cell data embedding model, extracting features through an encoder and a decoder, and training the model through cross-validation and hyperparameter optimization; finally, using the hidden layer of the encoder in the model as an embedded expression matrix of a single cell, which includes original single-cell expression information, PPI, GO and other multi-dimensional data, and can achieve a variety of downstream tasks including single-cell accurate classification. The present invention helps to deeply understand gene expression patterns and provide data support for disease treatment and drug design.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, artificial intelligence and single-cell omics analysis, and in particular to a high-precision single-cell classification method and device based on an artificial intelligence algorithm. Background Art

[0002] Single-cell sequencing technology is a key technology in modern biological research. It can quantitatively measure gene expression levels, DNA methylation levels on the whole genome, chromatin open region accessibility, and single-cell spatial transcriptome expression levels. These data reveal the heterogeneity and complexity between cells, study their characteristics and functions, and gain a deep understanding of the sources and mechanisms of cellular heterogeneity. At the same time, single-cell classification can be applied to clinical diagnosis and treatment, providing a new perspective for understanding cell functions and disease mechanisms. In addition, single-cell classification can also be applied to transplantation medicine, regenerative medicine and other fields to provide more accurate and reliable information for clinical practice.

[0003] However, the analysis of single-cell data faces many challenges, such as the high dimensionality, large scale, missing values, batch effects, etc. of the data itself. At the same time, the classification of cell types is often nonlinear and cannot be simply divided by thresholds or boundaries. This requires more complex algorithms and models to handle this nonlinear relationship, which increases the difficulty of classification. Finally, due to the sensitivity of the technology, the data may contain a lot of noise. These problems limit the accuracy and reliability of single-cell classification.

[0004] The difficulty of this problem lies in finding an optimal nonlinear model that can simultaneously fit single-cell data of various characteristics and identify cell type subclasses and some rare cells. Several single-cell data analysis methods have been proposed, including Seurat launched by Satija Lab and totalVI algorithm developed by Nir Yosef et al. based on deep learning. The following is a brief introduction to these two single-cell multi-omics data analysis methods.

[0005] 1. Seurat: For each cell, Seurat first computes the set of k=20 nearest neighbors for each modality. Next, the molecular profiles of the transcriptomic neighbors are averaged (i.e., the molecular content of neighboring cells is predicted) and the average is compared to the original measurements. Seurat then uses the relative accuracy of these predictions to describe the relative information content of each cell. Modality weights are calculated using a Softmax transform.

[0006] 2. TotalVi: TotalVI learns a joint probabilistic representation of paired measurement data, which accounts for the unique noise and technical bias of each modality, as well as batch effects. For transcriptomes, totalVI uses a modeling strategy similar to variational autoencoders, and its results can be used for a variety of different analysis tasks.

[0007] Although there are several methods to analyze single-cell omics data, there are still some challenges in single-cell classification. To solve this problem, it is urgent to design appropriate statistical models and computational methods, and develop scalable and effective computational methods to reveal the biological laws of spatial patterns of cellular heterogeneity, biological states, cell types, cell development, and complex tissues. Summary of the invention

[0008] In response to the shortcomings of the prior art, the present invention provides a high-precision single-cell classification method and device based on an artificial intelligence algorithm. The method achieves high-precision classification of single cells by combining original single-cell characterization data, PPI, GO, graph neural network and single-cell data embedding model, effectively solving the challenges in the prior art and more comprehensively and systematically depicting the state and fate of cells. It is expected to have important application value in the fields of life sciences, medical disease diagnosis and treatment.

[0009] According to a first aspect of an embodiment of the present application, a high-precision single-cell classification method based on an artificial intelligence algorithm is provided, comprising the following steps:

[0010] (1) Preprocess the raw single-cell data and construct a single-cell expression matrix;

[0011] (2) The single-cell expression matrix is ​​divided into single-cell categories through dimensionality reduction and cluster analysis to obtain cell category labels;

[0012] (3) Combining protein interaction networks and gene ontology information, using graph neural network methods to quantitatively characterize and classify genes to obtain gene category labels;

[0013] (4) Combining the single-cell expression matrix, cell category labels, and gene category labels, a single-cell data embedding model is constructed. Features are extracted through the encoder and decoder, and a total loss function consisting of a cell category label loss function and a gene category label loss function is constructed to achieve a balance between cell and gene features. The model is trained through cross-validation and hyperparameter optimization.

[0014] (5) The hidden layer of the encoder in the trained model is used as the embedded expression matrix of the single cell. This layer includes multi-dimensional data of the original single cell expression information, protein interaction network, and gene ontology information; ultimately, the model is used to achieve high-precision classification of single cells.

[0015] Furthermore, in the step (1), the preprocessing includes standardization and global normalization processing; the single-cell expression matrix is ​​used to describe the expression level of genes in cells under different conditions.

[0016] Furthermore, the step (2) is specifically as follows: reducing the preprocessed single-cell data to a low-level dimension, and then analyzing the principal component distribution characteristics through scree plot observation, so as to preliminarily determine the total number of cell categories; using a clustering algorithm to perform clustering analysis on the single-cell data after dimensionality reduction, further determining the total number of cell categories, and obtaining cell category labels.

[0017] Furthermore, the step (3) includes the following sub-steps:

[0018] (3.1) Construction of protein interaction network graph data: Using protein names as nodes and interactions between proteins as edges, we constructed directed protein interaction network graph data:

[0019] (3.2) Calculation of the shortest gene pathway based on gene ontology: Parse the ontology file of gene ontology and extract the terms of biological process, cellular component and molecular function and the relationship between them; then, based on the hierarchical relationship in the ontology file, construct a tree structure of each part, i.e. a directed acyclic graph, where the nodes represent genes; then use the Dijkstra or Floyd algorithm to calculate the shortest path from the target gene to the root node in the directed acyclic graph, and concatenate the shortest paths as the gene ontology features of the target gene;

[0020] (3.3) Construction of graph data integrating protein interaction network and gene ontology information: The gene ontology features of the target gene are used as the node features of the protein interaction network graph data to obtain the primary gene graph network data;

[0021] (3.4) Unsupervised graph network node embedding and clustering: An unsupervised graph neural network method is used to implement node embedding on the primary gene graph network data to obtain a gene representation vector. A dimensionality reduction method is used to reduce the gene representation vector to a low-level dimension. A clustering algorithm is then used to cluster the reduced dimensionality gene representation vector to determine the total number of gene categories and the category to which each gene belongs.

[0022] Furthermore, in step (4), the encoder is used to extract and compress single-cell data features while removing irrelevant noise and redundant information; the encoder adopts a three-layer fully connected network, wherein the output dimension of the three-layer fully connected network decreases by 30%.

[0023] Furthermore, in step (4), the decoder is used to restore the encoded result to cells as samples and genes as features, while eliminating and weakening the noise of missing values, statistical errors, and experimental errors; the decoder adopts a three-layer fully connected network, wherein the output dimension of the three-layer fully connected network increases by 30%.

[0024] Furthermore, in step (4), the expression of the total loss function Loss_total is:

[0025] ;

[0026] In the formula, represents the control parameter, Loss1 represents the cell category label loss function, and Loss2 represents the gene category label loss function;

[0027] The cell category label loss function adopts the cross entropy loss function, which is expressed as:

[0028] ;

[0029] In the formula, b_size represents the size of a batch during training, that is, the number of single cells; C cell represents the total number of cell categories, j represents the jth cell, i represents the i-th cell category, and p j,i represents the probability that the jth cell belongs to the i-th cell category, represents the true probability that the jth cell belongs to the i-th cell category;

[0030] The gene category label loss function adopts the cross entropy loss function, which is expressed as:

[0031] ;

[0032] Where gene_m represents the output dimension of the last fully connected network of the decoder, C gene represents the total gene category, k represents the kth gene, s represents the sth gene category, P k,s represents the probability that the kth gene belongs to the sth gene category, represents the true probability that the kth gene belongs to the sth gene category.

[0033] Furthermore, in step (4), the Optuna algorithm is used to adjust the hyperparameters of the model. After the hyperparameters of the model are fixed, the model is retrained and cross-validation is performed to evaluate the performance of the model.

[0034] A second aspect of an embodiment of the present invention provides a high-precision single-cell classification device based on an artificial intelligence algorithm, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned high-precision single-cell classification method based on the artificial intelligence algorithm.

[0035] A third aspect of an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned high-precision single-cell classification method based on an artificial intelligence algorithm.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1) It provides a new AI-based approach for high-precision single-cell classification and discovery of new cell types;

[0038] 2) Integrate prior biological attribute data PPI and GO based on single-cell expression data to improve the scientificity and credibility of single-cell classification results from both theoretical and practical aspects;

[0039] 3) Designing a single-cell data embedding model from both the cell and gene dimensions effectively reduces the noise and batch effects in single-cell data;

[0040] 4) It helps to study the heterogeneity and complexity of single cells, and is also of reference significance for promoting basic clinical research and clinical application research. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions or specific implementations of the present invention, the following briefly introduces the drawings required for use in the technical description and specific implementations of the present invention. Obviously, the drawings described below are some implementations of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 is a flow chart of the method of the present invention;

[0043] Figure 2 This is an example diagram of the original single-cell data in the present invention;

[0044] Figure 3 It is a distribution diagram of missing values ​​of original single cell data cell samples in the present invention;

[0045] Figure 4 A stacked bar chart of cell missing values ​​of the original single-cell data in the present invention;

[0046] Figure 5It is the distribution diagram of gene missing values ​​of the original single-cell data in the present invention;

[0047] Figure 6 Schematic diagram of the cell dimension reduction clustering result of the original single cell data in the present invention; wherein, Figure 6 (a) is a schematic diagram of UMPA dimensionality reduction results. Figure 6 (b) is a schematic diagram of the clustering result after dimensionality reduction;

[0048] Figure 7 It is a schematic diagram of PPI in the present invention;

[0049] Figure 8 Schematic diagram of the gene shortest path calculation process based on gene ontology in the present invention; wherein, Figure 8 (a) is a schematic diagram of the BP tree extracted from the Gene Ontology. Figure 8 (b) is a schematic diagram of the CC tree extracted from the Gene Ontology. Figure 8 (c) is a schematic diagram of the MF tree extracted from the Gene Ontology. Figure 8 (d) is a schematic diagram of the shortest path construction of a gene in the BP tree. Figure 8 (e) is a schematic diagram of the shortest path construction of a gene in the CC tree. Figure 8 (f) is a schematic diagram of the shortest path construction of a gene in the MF tree;

[0050] Fig. 9 It is a schematic diagram of combining the gene GO features and PPI graph data in the present invention;

[0051] Fig.10 The figure is a schematic diagram of the unsupervised graph network in the present invention to realize the quantitative characterization and clustering process of genes; wherein, Fig.10 (a) is a schematic diagram of the gene quantification characterization process based on GO features and unsupervised graph networks and the dimensionality reduction clustering after characterization. Fig.10 (b) is a schematic diagram of the actual data dimensionality reduction result. Fig.10 (c) is a schematic diagram of the clustering results after dimensionality reduction of actual data;

[0052] Fig.11 A schematic diagram of a single-cell data embedding model combining single-cell normalized characterization data, PPI, and GO information in the present invention;

[0053] Fig.12 A schematic diagram of the model training and verification with cross-validation and the model hyperparameter search process of the single-cell data embedding model in the present invention;

[0054] Fig.13 A schematic diagram of the loss and accuracy results of the single-cell data embedding model with cross-validation in the present invention;

[0055] Fig.14 This is a schematic diagram of dimensionality reduction and clustering results of the encoder output after being characterized by the single-cell data embedding model in the present invention;

[0056] Fig.15 It is a schematic diagram of the structure of the device of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the present invention more clear, the present invention is further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0058] Example data description: In order to verify the actual effect of this solution, the "pbmc3k" classic single-cell sequencing data set derived from peripheral blood mononuclear cells (PBMC) was used for testing. It can be directly obtained through the official website of 10X Genomics (https: / / support.10xgenomics.com / single-cell-gene-expression / datasets / 1.1.0 / pbmc3k). The pbmc3k data contains 2700 cells and 13714 genes. The original data contains three files: barcodes.tsv (cell barcode information), features.tsv (quantitative gene identifiers) and matrix.mtx (gene expression values). After preprocessing, it can be converted into a matrix with rows and columns representing genes and cells respectively.

[0059] See also Figure 1 The present invention provides a high-precision single-cell classification method based on an artificial intelligence algorithm, comprising the following steps:

[0060] 1. Single-cell raw data analysis.

[0061] First, the original single-cell data stored in text form is read and converted into csv format for subsequent analysis and processing to construct a single-cell expression matrix.

[0062] The expression matrix is ​​a two-dimensional statistical table that describes the expression levels of genes in cells under different conditions, such as Figure 2 As shown, Figure 2Each row in the first column represents a gene name, and each column in the first row represents a cell code. This matrix usually consists of one row for one gene, and each column for one sample (such as a cell or a tissue). In the matrix, each cell contains the expression level of a gene in a sample. These data can be a direct measurement of gene expression activity or an inference of gene expression status. The expression matrix can be used as a comprehensive evaluation indicator of the expression pattern of genes in a specific environment or state in an organism. It can help biologists and doctors understand a large number of gene expression patterns in order to more effectively detect whether the expression of a specific gene is affected by a specific condition or environment. In addition, the expression matrix can also be used to understand the function of a specific gene in different organisms, and to compare the similarities and differences of gene expression in different organisms. In practical applications, the analysis method of the gene expression matrix can help to gain a deeper insight into the dynamic characteristics of the gene expression information network, as well as more effectively detect cancer antigen epitopes, identify drug resistance markers of tumor cells, and study a variety of biological processes such as the regulation of genetic expression, the regulation of DNA methylation, the regulation of gene mutations, and the role of the immune system. At the same time, research based on expression matrices can also be used to design new drugs or gene therapies, as well as to refine the treatment and prevention of various diseases.

[0063] First, we analyze missing values ​​from a cellular perspective, that is, we count the proportion of 0 values ​​in the expression matrix of each cell sample (since the data set has been cleaned, we do not count NA values). Figure 3 and Figure 4 As shown, the highest gene data missing ratio of cells is 98.45%, the lowest is 75.21%, and the average missing ratio is 93.8%. The missing data is extremely serious. Most genes in each cell are missing data and the missing data ratio of each cell is basically the same.

[0064] Similarly, missing values ​​are analyzed from a genetic perspective, such as Figure 5 As shown, the average missing percentage is as high as 93.83%.

[0065] In single-cell transcriptome sequencing analysis, some genes cannot be detected due to low expression or insufficient sequencing depth. This situation where the gene expression data detected is untrue due to the technology is called the dropout effect. At the same time, some genes are indeed not expressed in the cell, which is called true zero value. In the sequencing results, there are often a mixture of zero values ​​caused by the dropout effect and true zero values. Multiple sequencing can reduce the proportion of missing values ​​to a certain extent. This widespread data missing not only reveals the significant heterogeneity of the single-cell expression matrix, but also introduces a high level of signal noise, affecting the integrity of the data and the accuracy of subsequent analysis.

[0066] Specifically, the heterogeneity of single-cell datasets is caused by a variety of factors, including but not limited to the complexity of the cell's environment, the intrinsic biological diversity of the cell population, and the high genetic and phenotypic variability of specific cell types (such as cancer cells). These factors work together to cause gene expression data at the single-cell level to present a high degree of variability and uncertainty. Therefore, in order to accurately parse and interpret single-cell transcriptome data, some common data processing techniques and statistical methods must be used to correct the dropout phenomenon, reduce the impact of noise, and reveal the true biological differences between cells. Including quality control (QC), standardization, data correction, feature selection, dimensionality reduction, low-rank matrix filling, deep learning algorithms and other data recovery methods to estimate and fill in missing expression values, thereby improving the integrity of the data and the reliability of the analysis results.

[0067] Given the high proportion of missing values ​​in the original single-cell data, the data set is first standardized, which helps correct technical variation and makes the data between different samples or batches comparable. Then the cells are globally normalized, which can handle zero values ​​and reduce the variance in the data. The standardized data is saved and the single-cell expression matrix is ​​constructed as the basis for subsequent processing.

[0068] 2. Preliminary classification of single-cell types.

[0069] For the original single-cell data after standardization and global normalization, cells are used as samples for dimensionality reduction analysis. The methods that can be used include PCA, t-SNE and UMAP, etc. The purpose is to convert high-dimensional data into low-dimensional space for data visualization and further analysis. Figure 6 , Figure 6 (a) Using UMAP as an example, the standardized single-cell data is reduced to 2D, and the main component distribution characteristics are analyzed through scree plot observation to preliminarily determine the total number of cell categories C. cell Then, clustering algorithms are used to perform cluster analysis on the single-cell data after dimensionality reduction. The available methods include K-Means, Mean Shift, Hierarchical Clustering, DBSCAN, etc., to further determine the total number of cell categories C. cell , and get the cell category label. In practice, when performing cluster analysis, the total number of categories C cell It must be determined carefully. For example, when using K-Means for cluster analysis, it is recommended to use the Elbow Method to determine C cell The value of Figure 6 (b) C cell = 11. So far, a rough category label has been determined for the single cell.

[0070] 3. Quantitative characterization and classification of genes based on PPI, GO and graph neural networks.

[0071] 3.1 Construction of protein-protein interaction networks (PPI) graph data.

[0072] Protein interaction network (PPI) refers to two or more proteins interacting physically in cells to form complexes, which participate in regulating biological signal transmission, gene expression, metabolic processes and other biological functions. Commonly used PPI databases include STRING, bioGRID, GeNets, etc. The list of differentially expressed genes or proteins obtained based on experimental data (such as RNA-seq, proteome analysis) is searched in the STRING database (gene or protein name list) to obtain possible potential interactions between target proteins. Then, the directed PPI graph data is constructed using protein names as nodes and interactions between proteins as edges. G PPI ,like Figure 7 shown.

[0073] 3.2 Calculation of the shortest gene pathway based on Gene Ontology (GO).

[0074] Note that the PPI graph above G PPI There is only interaction data between proteins in the data, that is, only edge information but no node information. To solve this problem, the Gene Ontology (GO) related information is incorporated into G PPI GO data is a systematic classification framework used to describe the functions of genes and their products, including three main aspects: biological process (BP), cellular component (CC) and molecular function (MF).

[0075] First, download the GO ontology file from the GO official website (https: / / www.geneontology.org / ), usually in .obo format. Then use bioinformatics tools or programming languages ​​(such as Python) to parse the ontology file, extract the terms (terms) of the three parts of BP, CC, and MF, and the relationship between them. Then, based on the hierarchical relationship in the ontology file, construct a tree structure for each part, that is, a directed acyclic graph (DAG). The nodes in the graph represent genes (which can be translated into corresponding proteins), and the node names are represented by numbers, which can be converted from the original GO information. Then use the Dijkstra or Floyd algorithm to calculate the shortest path from the target gene to the root node. For example Figure 8 As shown, Figure 8 (a) is a schematic diagram of the BP tree extracted from the Gene Ontology. Figure 8 (b) is a schematic diagram of the CC tree extracted from the Gene Ontology. Figure 8 (c) is a schematic diagram of the MF tree extracted from the Gene Ontology. Figure 8 (d) is a schematic diagram of the shortest path construction of a gene in the BP tree. Figure 8 (e) is a schematic diagram of the shortest path construction of a gene in the CC tree. Figure 8 (f) is a schematic diagram of the shortest path construction of a gene in the MF tree; taking the PRCC target gene as an example, its shortest paths in the actual BP tree, the actual CC tree, and the actual MF tree are: [29,18,53], [32,76,12,39], and [95,35,62,77,41], respectively. The shortest paths of the gene in the three trees are spliced ​​together as the GO feature of the gene, namely:

[0076] GO PRCC =[29,18,53,32,76,12,39,95,35,62,77,41]

[0077] 3.3 Construction of graph data integrating PPI and GO information.

[0078] The gene GO features constructed above are used as node features of the PPI graph data to form a more complete graph data, such as Fig. 9 shown.

[0079] 3.4 Unsupervised graph network node embedding and clustering.

[0080] Previously, we obtained primary gene graph network data combining PPI and GO information. It is difficult to cluster genes directly on this network. In order to further integrate PPI and GO data on the graph data and realize gene clustering, unsupervised graph neural network methods (GAT, GCN, DeepWalk, Node2Vec) are used to implement node embedding on the primary gene graph network data, so that each gene is represented as a numerical vector with a length of 512 or 1024. On this basis, dimensionality reduction methods (PCA, UMAP, t-SNE) are used to reduce the gene representation vector to 2 dimensions or other suitable low-level dimensions; then clustering algorithms (K-Means, Mean Shift, Hierarchical Clustering, DBSCAN) are used to cluster the reduced dimensionality gene representation vector to determine the total number of gene categories C gene , and the category to which each gene belongs, that is, the gene category label is obtained. The whole process is as follows Fig.10 As shown, Fig.10 (a) is a schematic diagram of the gene quantification characterization process based on GO features and unsupervised graph networks and the dimensionality reduction clustering after characterization. Fig.10 (b) is a schematic diagram of the actual data dimensionality reduction result. Fig.10 (c) is a schematic diagram of the clustering results after dimensionality reduction of the actual data; gene =16.

[0081] 4. Construction of single-cell data embedding model.

[0082] The previous steps obtained the standardized single-cell expression matrix, the cell category labels obtained based on the clustering of the single-cell expression matrix, and the gene category labels obtained based on the PPI combined with GO information. Fig.11 The next step is to design a neural network model to combine these three types of data, so as to achieve the embedded expression of single-cell data, that is, to embed the standardized single-cell expression vector into a numerical vector of the same length. The embedded vector theoretically contains the original single-cell expression information, PPI information and GO information, which is more reasonable than using only the original single-cell expression information. The model structure is as follows: Fig.11 shown.

[0083] 4.1 Single-cell data encoder: After reading the standardized single-cell expression matrix, the encoder is responsible for extracting and compressing the single-cell data features, while removing irrelevant noise and redundant information. Due to the sparsity of the single-cell expression matrix, the encoder will play a key role in this network. The encoder uses a three-layer fully connected network. The output dimension of the three-layer fully connected network decreases by 30%, and the output dimension of the last layer is set to 512, that is, the final output dimension of the encoder is: h_shape=(b_size,h_size), where b_size represents the size of a batch during training, that is, the number of single cells, and h_size represents the encoder output dimension. Since the output of the three-layer network decreases, h_size is smaller than gene_m, and this paper sets it to h_size=512. In addition, each layer of the network x uses a ReLU activation layer to implement nonlinear mapping, as shown in formula (1), and each layer is set to dropout=10% to prevent overfitting.

[0084] (1)

[0085] The role of the encoder here is actually to regard a cell as a sample and a gene as a feature of the sample. After the training is completed, the key features of all genes are integrated and each cell is represented by h_size values.

[0086] 4.2 Single Cell Classification Network: The output of the encoder is essentially a cell sample after feature extraction and compression. It is input into the cell classification network to classify the cells. The cell classification labels are exactly the cell labels calculated in the previous chapter. The total number of categories is C cell The classification network also uses a three-layer fully connected network, and each layer is activated by the ReLU function. The output dimension of the third layer is: out_shape=(b_size,C cell ), and then the Softmax function (as shown in formula (2)) is used to convert the output value into a category probability. In a batch of cells, for the jth (j∈[1,b_size]) cell belongs to the ith (i∈[1,C cell ]) The probability of the type is p j,i , where x j,i That is, it represents the value of the i-th rank of the j-th cell (i.e., cell category i) in the output of the third layer:

[0087] (2)

[0088] The cell category label loss function adopts the cross entropy loss function, as shown in formula (3), where Represents the true probability that the jth cell belongs to the i-th class:

[0089] (3)

[0090] 4.3 Decoder: The encoder extracts, reorganizes and compresses gene features using cells as samples, while the purpose of the decoder here is to restore the encoded results to cells as samples and genes as features. The difference is that the decoded gene features are theoretically more reasonable than the gene features before encoding, and to a certain extent, the noise such as missing values, statistical errors, and experimental errors are eliminated and weakened. The decoder also uses a three-layer fully connected network. The output dimension of the three-layer fully connected network increases by 30%, and the output dimension of the last layer is set to gene_m, that is, the final output dimension of the decoder is: out_shape=(b_size,gene_m). In addition, each layer of the network uses a ReLU activation layer to implement nonlinear mapping, and each layer is set to dropout=10% to prevent overfitting.

[0091] 4.4 Transposition: The decoder output is transposed in order to use genes as samples and cells as features, so as to facilitate the subsequent classification of genes. The output dimension after transposition is: out_shape=(gene_m,b_size).

[0092] 4.5 Gene classification network: The first half of the model uses cells as samples and genes as features for analysis; the second half of the model uses genes as samples and cells as features for analysis. The output after decoding and transposition is for this purpose. Since the previous article has specifically clustered genes, that is, each gene is assigned a label, the total gene category is C gene Therefore, the final output dimension of the gene classification network here must be: out_shape=(gene_m,C gene ). The gene classification network still uses a three-layer fully connected network. The input layer and output layer have been determined before. The output of the middle layer is set to 1.2 times the sum of the input and output dimensions. Similarly, each layer of the network uses a ReLU activation layer to implement nonlinear mapping, and each layer is set to dropout = 10% to prevent overfitting. After the output, the Softmax function (as shown in formula (4)) is used to convert the output value into a category probability. For the kth (k∈[1,gene_m]) gene belonging to the sth (s∈[1,C gene ]) The probability of the type is P k,s , where y k,s That is, it represents the value of the sth rank of the kth gene (i.e., gene category s) in the output of the third layer:

[0093] (4)

[0094] The gene category label loss function adopts the cross entropy loss function, as shown in formula (5): represents the true probability that the kth gene belongs to the sth class:

[0095] (5)

[0096] 4.6 Definition of total loss function: According to the previous definition, the model has two branches for cells and genes, corresponding to the cell category label loss function Loss1 and the gene category label loss function Loss2 respectively. In order to integrate the two, it is also necessary to construct a total loss function Loss_total, which is defined as shown in formula (6):

[0097] (6)

[0098] The total loss function balances the two branches of cells and genes, where the control parameter , in practice it is usually chosen between 0.6 and 0.8.

[0099] 4.7 Model training: The present invention adopts the six-fold cross-validation method to make full use of limited data resources and improve the generalization ability of the model. The data set is evenly divided into six parts, and five of them are selected as training data each time in a rotating manner, and the remaining one is used as validation data. Cross-validation is a statistical method that reduces the risk of model overfitting and improves the accuracy of model evaluation by dividing the data set into multiple subsets and then training and validating on these subsets in turn.

[0100] At the same time, in order to improve the classification accuracy and training efficiency of the model, the present invention adopts the Optuna algorithm to adjust the hyperparameters of the model. Optuna is a Bayesian optimization tool based on the TPE (Tree-structured Parzen Estimator) algorithm, which can intelligently select the next set of experimental parameters and accelerate the hyperparameter search process. After the hyperparameters are optimized, two methods can be selected to determine the final model. The first is to directly use the optimized hyperparameters and the corresponding model as the final model. The second is to use the optimized hyperparameters to retrain the model on the entire data set to obtain a more robust model performance. Finally, the final hyperparameters are used to retrain the model. Fig.12 shown.

[0101] After fixing the model’s hyperparameters, the model was retrained and cross-validation was implemented to further evaluate the model’s performance. Fig.13The results of cross-validation are shown, including loss and accuracy, which is the ratio of the number of samples correctly predicted by the model to the total number of samples. In addition, the performance changes of the model in each cross-validation are shown to evaluate the stability of the model. For example, it is possible to observe the fluctuation of the accuracy of the model on different training-test splits, and whether the model is overfitting or underfitting a specific data subset. Finally, the most stable and best performing model is selected as the final model.

[0102] 5. Clustering (classification) of single-cell embedding data.

[0103] The entire process from data preprocessing to model building and complete training has been described in detail above. The ultimate goal of this solution is to achieve more accurate single-cell classification, so it is also necessary to extract the last layer of encoder weights from the trained model as the final embedding vector of the single cell, that is, Fig.11 Mark the part with the shape "☆". The reason for choosing this layer (hidden layer) as the final single-cell embedding expression matrix is ​​mainly because this layer is at the hub of the entire model. It is preceded by the original data and connected to the two dimensions of cells and genes. After the training, there is reason to believe that this layer encompasses the core information (i.e., original single-cell expression information, protein interaction networks, gene ontology information and other multi-dimensional data); and then the single-cell data embedding model is used to achieve high-precision classification of single cells. In the example of this scheme, 2700 single cells represented by 13714 genes are converted into vectors composed of 512 numerical features. In order to further understand the intrinsic characteristics of these representations, the representation results were first subjected to PCA and UMAP analysis, and then clustered separately. The results are shown below. Fig.14 As shown in the results, it can be seen that the single-cell classification results after processing by this scheme are very different from those obtained by directly using a single raw data for classification, especially the UMAP results. The boundaries between different cell clusters are clear, which provides the possibility for high-precision cell classification and even the discovery of new cell types.

[0104] See also Fig.15 An embodiment of the present invention provides a high-precision single-cell classification device based on an artificial intelligence algorithm, comprising one or more processors for implementing a high-precision single-cell classification method based on an artificial intelligence algorithm in the above-mentioned embodiment.

[0105] The embodiments of the high-precision single-cell classification device based on artificial intelligence algorithms of the present invention can be applied to any device with data processing capabilities, and the device with data processing capabilities can be a device or apparatus such as a computer. The device embodiment can be implemented through software, or through hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capabilities in which it is located reading the corresponding computer program instructions in the non-volatile memory into the internal memory for execution. From a hardware perspective, if Fig.15 As shown, it is a hardware structure diagram of any device with data processing capability where the high-precision single-cell classification device based on artificial intelligence algorithm of the present invention is located. Fig.15 In addition to the processor, memory, network interface, and non-volatile memory shown, any device with data processing capabilities in which the apparatus in the embodiments is located may also include other hardware, generally based on the actual functions of the device with data processing capabilities, which will not be described in detail.

[0106] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0107] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0108] Correspondingly, an embodiment of the present invention further provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the high-precision single-cell classification method based on the artificial intelligence algorithm in the above embodiment is implemented.

[0109] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capability and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0110] Those skilled in the art will readily appreciate other embodiments of the present application after considering the description and practicing the contents disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary technical means in the art that are not disclosed in the present application. The description and examples are intended to be exemplary only.

[0111] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0112] The above embodiments are only used to illustrate the design ideas and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, any equivalent changes or modifications made based on the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

Claims

1. A high-precision single-cell classification method based on artificial intelligence algorithm, characterized in that: The following steps are involved: (1) Preprocess the raw single-cell data and construct a single-cell expression matrix; (2) The single-cell expression matrix is ​​divided into single-cell categories through dimensionality reduction and cluster analysis to obtain cell category labels; (3) Combining protein interaction networks and gene ontology information, using graph neural network methods to quantitatively characterize and classify genes to obtain gene category labels; (4) Combine the single-cell expression matrix, cell category label, and gene category label to build a single-cell data embedding model; extract features through the encoder and decoder, and construct a total loss function composed of the cell category label loss function and the gene category label loss function to achieve a balance between cell and gene features, and train the model through cross-validation and hyperparameter optimization; the expression of the total loss function Loss_total is: ； In the formula, represents the control parameter, Loss1 represents the cell category label loss function, and Loss2 represents the gene category label loss function; The cell category label loss function adopts the cross entropy loss function, which is expressed as: ； In the formula, b_size represents the size of a batch during training, that is, the number of single cells; C cell represents the total number of cell categories, j represents the jth cell, i represents the i-th cell category, and p j,i represents the probability that the jth cell belongs to the i-th cell category, represents the true probability that the jth cell belongs to the i-th cell category; The gene category label loss function adopts the cross entropy loss function, which is expressed as: ； Where gene_m represents the output dimension of the last fully connected network of the decoder, C gene represents the total gene category, k represents the kth gene, s represents the sth gene category, P k,s represents the probability that the kth gene belongs to the sth gene category, represents the true probability that the kth gene belongs to the sth gene category; (5) The hidden layer of the encoder in the trained model is used as the embedded expression matrix of the single cell. This layer includes multi-dimensional data of the original single cell expression information, protein interaction network, and gene ontology information; ultimately, the model is used to achieve high-precision classification of single cells.

2. The high-precision single-cell classification method based on artificial intelligence algorithm according to claim 1, characterized in that: In the step (1), the preprocessing includes standardization and global normalization processing; the single-cell expression matrix is ​​used to describe the expression level of genes in cells under different conditions.

3. The high-precision single-cell classification method based on artificial intelligence algorithm according to claim 1, characterized in that: The step (2) specifically comprises: reducing the preprocessed single-cell data to a low-level dimension, and then analyzing the principal component distribution characteristics through scree plot observation, so as to preliminarily determine the total number of cell categories; using a clustering algorithm to perform clustering analysis on the single-cell data after dimensionality reduction, further determining the total number of cell categories, and obtaining cell category labels.

4. The high-precision single-cell classification method based on artificial intelligence algorithm according to claim 1, characterized in that: The step (3) includes the following sub-steps: (3.1) Construction of protein interaction network graph data: Using protein names as nodes and interactions between proteins as edges, we constructed directed protein interaction network graph data: (3.2) Calculation of the shortest gene pathway based on gene ontology: Parse the ontology file of gene ontology and extract the terms of biological process, cellular component and molecular function and the relationship between them; then, based on the hierarchical relationship in the ontology file, construct a tree structure of each part, i.e. a directed acyclic graph, where the nodes represent genes; then use the Dijkstra or Floyd algorithm to calculate the shortest path from the target gene to the root node in the directed acyclic graph, and concatenate the shortest paths as the gene ontology features of the target gene; (3.3) Construction of graph data integrating protein interaction network and gene ontology information: The gene ontology features of the target gene are used as the node features of the protein interaction network graph data to obtain the primary gene graph network data; (3.4) Unsupervised graph network node embedding and clustering: An unsupervised graph neural network method is used to implement node embedding on the primary gene graph network data to obtain a gene representation vector. A dimensionality reduction method is used to reduce the gene representation vector to a low-level dimension. A clustering algorithm is then used to cluster the reduced dimensionality gene representation vector to determine the total number of gene categories and the category to which each gene belongs.

5. The high-precision single-cell classification method based on artificial intelligence algorithm according to claim 1, characterized in that: In the step (4), the encoder is used to extract and compress single-cell data features while removing irrelevant noise and redundant information; the encoder adopts a three-layer fully connected network, wherein the output dimension of the three-layer fully connected network decreases by 30%.

6. The high-precision single-cell classification method based on artificial intelligence algorithm according to claim 1, characterized in that: In step (4), the decoder is used to restore the encoded result to cells as samples and genes as features, while eliminating and weakening the noise of missing values, statistical errors, and experimental errors; the decoder adopts a three-layer fully connected network, wherein the output dimension of the three-layer fully connected network increases by 30%.

7. The high-precision single-cell classification method based on artificial intelligence algorithm according to claim 1, characterized in that: In step (4), the Optuna algorithm is used to adjust the hyperparameters of the model. After the hyperparameters of the model are fixed, the model is retrained and cross-validation is performed to evaluate the performance of the model.

8. A high-precision single-cell classification device based on an artificial intelligence algorithm, comprising a memory and a processor, characterized in that: The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the high-precision single-cell classification method based on artificial intelligence algorithm as described in any one of claims 1 to 7 above.

9. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the high-precision single-cell classification method based on an artificial intelligence algorithm as described in any one of claims 1 to 7 is implemented.

Images