A cell subpopulation label mapping method, system, device and storage medium

Abstract

The present invention discloses a method, system, device and storage medium for mapping cell subpopulation labels. The present invention applies self-supervised learning capabilities on large-scale unlabeled single-cell sequencing data sets, improves the versatility of the model, and can effectively overcome the batch effect in single-cell sequencing data. The present invention discretizes and encodes gene expression to reduce the impact of gene expression variation under different experimental conditions. The present invention strengthens the importance of transcription factors to ensure that the role of these key genes in distinguishing cells is fully considered. The present invention not only captures the information of a single gene, but also reflects the interaction between genes by combining gene representation embedding and gene expression embedding. At the same time, through a self-supervised learning strategy based on Linformer, the model can learn the basic patterns of gene expression and their interactions from a large amount of single-cell data, thereby effectively constructing a single-cell gene panorama.

Description

Technical Field

[0001] The present invention relates to the field of biological information processing technology, and in particular to a cell subpopulation label mapping method, system, device and storage medium. Background Art

[0002] As single-cell sequencing technology becomes more mature and popular, large-scale projects such as the Human Cell Atlas (HCA) continue to advance, providing an unprecedented single-cell research reference framework for life science research. Therefore, it is becoming increasingly important to make full use of existing single-cell research results and migrate existing reference materials to new samples (such as disease tissues) to characterize differences in the cell composition of new samples and detect new sample cell types, making single-cell-based patient assessment possible.

[0003] The data sets generated by single-cell sequencing technology are very large. Each sample may contain information on tens of thousands of cells, and each cell involves the transcription levels of thousands to tens of thousands of genes. In order to reduce the complexity of analysis and reveal subtle differences between cells, researchers usually use unsupervised clustering algorithms, such as t-SNE and UMAP, to divide cells into several cell subpopulations, label the cells with cell subpopulation labels, and then conduct subsequent analysis. Today, cell subpopulations play a core role in single-cell sequencing research, such as annotating cell subpopulations to determine cell types or discover new cell types, or using cell clustering to discover key pathogenic cell subpopulations associated with drug response, thereby using the cell subpopulations to predict patients' drug responses.

[0004] The widespread use of unsupervised clustering algorithms in the single-cell field has brought new challenges to the utilization of single-cell research results. Unsupervised learning is non-deterministic and non-parametric, and there are differences between single-cell sequencing data generated under different experimental conditions. Therefore, the core cell subpopulation labels of previous studies are difficult to reproduce directly in new datasets. In addition, the single-cell analysis conclusions based on these cell subpopulations are difficult to maintain consistency in different datasets, which makes it more difficult to use existing single-cell research to evaluate new samples or new patients.

[0005] Existing technologies usually map cell subpopulation labels in two ways. The first is a correlation-based method that uses algorithms such as cosine similarity and Spearman correlation to calculate the similarity between key cell subpopulations and new samples to be queried, thereby finding the cells corresponding to the new samples. However, due to the high-dimensional and sparse characteristics of single-cell data, this method may be affected by batch effects across platforms and experiments, resulting in low robustness when calculating the correlation between two sets of single-cell data. Another method is to use machine learning to map cell subpopulation labels. However, due to the large number of genes, it may cause the problem of dimensionality disaster in machine learning. Therefore, it is necessary to reduce the dimension of the data in advance, select some highly mutated genes (Highly Variable Gene, HVG) as feature genes, and input them into the machine learning model. Highly mutated genes are some of the genes with the most obvious differential expression of a key cell subpopulation relative to other cell subpopulations, and are believed to be able to roughly distinguish cell subpopulations. However, the method based on characteristic genes will bring two problems. First, how to select characteristic genes is a difficult point. Due to the influence of batch effects, characteristic genes are variable between different data sets. If there is no appropriate method to obtain the truly important characteristic gene set, it is difficult to ensure the uniformity between cell subpopulation labels. At the same time, the selection of characteristic genes may also introduce human bias, which greatly affects subsequent analysis. In addition, the method based on highly mutated genes ignores the overall picture of genes, loses many unselected genes, and cannot effectively utilize gene-gene interactions. The joint action of these unselected genes or the interaction between unselected genes and highly mutated genes may greatly affect the final mapping results. Summary of the invention

[0006] The purpose of the present invention is to provide a cell subpopulation label mapping method, system, device and storage medium in view of the deficiencies in the prior art.

[0007] According to the first aspect of this specification, a method for mapping cell subpopulation labels is provided, the method comprising:

[0008] Preprocess the single-cell data and construct a single-cell corpus;

[0009] The gene embedding of each gene in the cell is expressed as the sum of the gene representation embedding and the gene expression embedding, and the gene embedding matrix corresponding to the cell is obtained, which is:

[0010] The genes in the total gene vocabulary are encoded into multi-dimensional vectors through the gene2vec model as gene representation embedding;

[0011] The expression amounts of genes in the total gene vocabulary are converted into discretized gene expression codes, and the gene expression codes are mapped into multidimensional vectors as gene expression embeddings using an embedding layer;

[0012] Based on the masked language model, some genes with non-zero expression are randomly masked to obtain the masked gene embedding matrix, and the pre-training model is obtained based on the self-supervised learning strategy of Linformer;

[0013] A single-cell sequencing dataset with cell subpopulation labels is input into a pre-trained model, a cell representation is generated through the pre-trained model, and a classifier is used to assign cell subpopulation labels.

[0014] Furthermore, the preprocessing of the single cell data is specifically as follows:

[0015] Read the raw count matrix of single-cell sequencing and generate a sparse matrix that records the gene expression of single cells;

[0016] Perform quality correction to filter out cells that do not meet the requirements for gene feature quantity, sequencing depth, and mitochondrial count ratio;

[0017] The gene expression data were Log-normalized to obtain the normalized single-cell gene expression data.

[0018] Furthermore, the gene expression code is obtained specifically as follows:

[0019] The genes in the gene vocabulary are divided into transcription factor related genes and other genes Two parts: Calculate the gene expression of each cell The non-zero median of The ratio r of the non-zero medians of

[0020] Discretize the gene expression value, and then use label encoding to assign an integer to each interval to replace the original gene expression value;

[0021] cell The range of all gene expression levels is recorded as , divide it into K continuous intervals , If the gene belongs to And the gene expression level in If the gene expression is If the gene belongs to And the gene expression level is r times If the gene expression is ; If the gene expression level is 0, the gene expression code is 0.

[0022] Furthermore, the method randomly masks some genes with non-zero expression based on the mask language model, specifically: Among all the genes used, 10% of the genes with non-zero expression were randomly selected for masking. For each selected gene, the random number Uniformly distributed in the interval [0,1), if , then mask the gene representation data if , then the gene expression data are masked.

[0023] Furthermore, the pre-trained model adopts Linformer with linear complexity self-attention mechanism as the model encoder, first calculates the query, key and value according to the gene embedding matrix, then uses two linear projection matrices to transform the key and value respectively to complete the matrix dimension compression, and finally uses the converted key-value pairs to calculate the attention score.

[0024] Furthermore, the loss function of the pre-trained model is The expression is: , where B is the total number of cells and N is the total number of genes. is the number of genes masked, It's cells The corresponding gene The true embedding vector of is the predicted embedding vector, is an indicator variable. The corresponding gene 1 if masked, 0 otherwise.

[0025] According to a second aspect of the present specification, a cell subpopulation label mapping system is provided, the system is used to implement the cell subpopulation label mapping method as described in the first aspect, and the system comprises:

[0026] Single-cell preprocessing and corpus construction module, used to preprocess single-cell data and construct a single-cell corpus;

[0027] The single-cell embedding module is used to represent the gene embedding of each gene in the cell as the sum of the gene representation embedding and the gene expression embedding, and obtain the gene embedding matrix corresponding to the cell;

[0028] The self-supervised learning module based on Linformer is used to randomly mask some genes with non-zero expression based on the masked language model, obtain the masked gene embedding matrix, and obtain the pre-trained model based on the self-supervised learning strategy of Linformer;

[0029] The supervised cell subpopulation label mapping module is used to input a single-cell sequencing data set with cell subpopulation labels into a pre-trained model, generate cell representations through the pre-trained model, and use a classifier to assign cell subpopulation labels.

[0030] According to the third aspect of this specification, an electronic device is provided, 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 cell subpopulation label mapping method as described in the first aspect.

[0031] According to a fourth aspect of the present specification, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the cell subpopulation label mapping method as described in the first aspect is implemented.

[0032] According to a fifth aspect of the present specification, a computer program product is provided, comprising a computer program / instruction, which, when executed by a processor, implements the cell subpopulation label mapping method as described in the first aspect.

[0033] The beneficial effects of the present invention are:

[0034] 1. In response to the batch effect problem of single-cell data across platforms and experiments, the present invention applies the ability of self-supervised learning on large-scale unlabeled single-cell sequencing data sets, improves the versatility of the model, and can effectively overcome the batch effect in single-cell sequencing data. The present invention eliminates the differences between different cell sequencing depths through standardization. The present invention discretizes continuous gene expression levels and performs label encoding to reduce the impact of gene expression variation under different experimental conditions. The present invention strengthens the importance of transcription factors to ensure that the role of these key genes in distinguishing cells is fully considered. These measures together ensure that single-cell data from different platforms and experimental conditions can be compared and analyzed under unified standards.

[0035] 2. In order to solve the problem of being unable to use the single-cell gene panorama and the interaction between genes, the present invention combines gene representation embedding with gene expression embedding to not only capture the information of a single gene, but also reflect the interaction between genes. At the same time, through the self-supervised learning strategy based on Linformer, the model can learn the basic patterns of gene expression and their interactions from a large amount of single-cell data, thereby effectively constructing a single-cell gene panorama. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0037] Figure 1The flowchart of a cell subpopulation label mapping method is shown as an exemplary embodiment;

[0038] Figure 2 A flowchart for generating gene embeddings is shown as an exemplary embodiment;

[0039] Figure 3 A schematic diagram of the structure of a single-cell pre-training model shown as an exemplary embodiment;

[0040] Figure 4 It is a schematic diagram of the structure of a cell subpopulation label mapping system shown in an exemplary embodiment;

[0041] Figure 5 The figure is a schematic diagram of the structure of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0042] In order to better understand the technical solution of the present application, the embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0043] It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0044] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings.

[0045] In this application, RNA sequencing technology refers to the technology of analyzing the entire genome by observing gene expression. RNA is sequenced and analyzed using high-throughput sequencing technology, which reflects the expression level (expression amount) of RNA. RNA sequencing technology usually breaks all cells in a tissue sample and performs sequencing based on the broken RNA. The sequencing result is usually the average RNA expression amount of all cells in a tissue, and it is impossible to distinguish the heterogeneity between cells.

[0046] In this application, single-cell sequencing refers to isolating a single cell, amplifying and high-throughput sequencing the mRNA of a single cell, and obtaining the RNA expression level at the single cell level. Based on the similarity of the result spectrum, different cell types can be distinguished, and even new cell types can be revealed, which helps to deeply explore the heterogeneity and life activities of cells, and has a very important guiding significance for personalized medicine.

[0047] In this application, cell subpopulations refer to smaller cell populations divided according to specific molecular characteristics or biological functions in a larger cell population, usually obtained by clustering cells. In single-cell analysis, the identification of cell subpopulations is one of the important means to study cell heterogeneity. The discovery of cell subpopulations can help scientists better understand information on the functions, differentiation pathways, disease states, etc. of different cell types.

[0048] The present invention provides a method for mapping cell subpopulation labels. Figure 1 As shown in the figure, the method includes single-cell preprocessing and corpus construction, generation of gene embedding matrix, obtaining pre-trained model based on Linformer self-supervised learning strategy, supervised cell subpopulation label mapping, etc. The specific implementation process of each part is elaborated in detail below.

[0049] S1, Single cell preprocessing and corpus construction

[0050] Single-cell data from different platforms are preprocessed uniformly, and a single-cell corpus is constructed based on the preprocessed data.

[0051] The specific steps of single cell pretreatment are as follows:

[0052] 1.1 Data reading: Read the raw count matrix of single-cell sequencing and generate a sparse matrix that records the single-cell gene expression;

[0053] 1.2 Quality correction: filter out low-quality cells with a small number of gene features, low sequencing depth, and high mitochondrial count ratio; for example, the quality correction standard is set to 500-3500 gene features, the number of UMIs (unique molecular identifiers) per cell is less than 20,000, and the mitochondrial count ratio is less than 7.5%;

[0054] 1.3 Data normalization: In order to remove the impact of different sequencing depths on gene expression, the gene expression data was log-normalized to obtain the standardized single-cell gene expression data; Log normalization can be performed in the following ways but is not limited to: the count of each gene is divided by the total count of the cell, multiplied by 10,000, and then log1p conversion is performed.

[0055] The single-cell data of the single-cell corpus comes from large human single-cell databases such as cellxgene and the Human Cell Atlas (HCA), and the cell types cover common human cells such as peripheral blood mononuclear cells, immune cells, kidney cells, and fibroblasts. The single-cell raw sequencing data obtained from large human single-cell databases are preprocessed according to the above steps and constructed into a unified single-cell corpus for subsequent processing.

[0056] S2, Generate gene embedding matrix

[0057] Each gene is considered as the smallest unit of information, equivalent to a token in natural language generation (NLG), and a single cell is considered to be composed of the expression of multiple genes, similar to a sentence in natural language. This step maps each gene to a high-dimensional vector space and represents the gene embedding of each gene in the cell as the sum of the gene representation embedding and the gene expression embedding.

[0058] 2.1 Gene Representation Embedding :

[0059] Gene representation embedding is similar to word embedding in natural language, where interacting genes need to have similar gene representation embedding. The present invention first uses N (about 25,000) protein-coding genes and non-coding RNA genes recorded in the Human Genome Nomenclature Database (HGNC) to generate a gene vocabulary; then for all genes in the gene vocabulary , gene2vec model is used to transform genes Encoded as a d-dimensional vector , i.e., the gene representation embedding of the model; in this example, d=200.

[0060] 2.2 Gene expression embedding :

[0061] Since single-cell sequencing data has a batch effect problem across datasets, that is, the gene expression of single cells under different single-cell sequencing methods is significantly different and cannot be directly compared. In order to solve this problem, the present invention converts the absolute value of gene expression in all cells into a relative value.

[0062] Specifically, the genes in the gene vocabulary are divided into transcription factor-related genes and other genes Two parts, among which transcription factor related genes There are about 1,800 genes encoding transcription factors that control the expression of genes, which may have lower raw expression values ​​but have higher ability to distinguish cells.

[0063] First, calculate the transcription factor-related genes expressed in each cell The non-zero median( ), and other genes The non-zero median( ), calculate the ratio r=median( ) / median( ).

[0064] Subsequently, the gene expression is discretized, and label encoding is used to assign an integer to each interval, replacing the original gene expression value. For example, the original gene expression level is [ ], assuming that the cell The expression levels of all genes ranged from , the present invention divides it into K continuous intervals , , among which cells The corresponding gene The expression code for:

[0065]

[0066] Through the above encoding function, the present invention converts the original gene expression into a discretized gene expression code [ ]. Finally, an embedding layer is used to encode gene expression Mapped to a d-dimensional vector , as the gene expression embedding; in this embodiment, d=200.

[0067] 2.3 Embedding Matrix :

[0068] The present invention converts gene In cells Gene embedding , represented as the gene representation embedding in step 2.1 and the gene expression embedding in step 2.2 The sum of Each cell The gene embedding matrix of the N genes in the corresponding input model is recorded as , due to gene embedding is a d-dimensional vector, ,Right now Since the present invention uses the method of masking genes to achieve self-supervised learning, some genes will be masked. The specific masking scheme will be described in the next step. The basic process of generating gene embedding is as follows: Figure 2 shown.

[0069] S3, Linformer-based self-supervised learning

[0070] The present invention uses a self-supervised learning strategy based on Linformer to obtain a single-cell pre-training model. Since single-cell sequencing data is non-sequential data, the present invention is based on a masked language model (MLM) to randomly mask about 10% of the genes with non-zero expression, of which 40% are gene representation and 60% are gene expression. Specifically, for each cell , all genes used were First, randomly select 10% of the non-zero expression genes for masking, and set is the set of selected masked genes, for each selected gene , assuming the random number is uniformly distributed in the interval [0,1), then if , masking gene representation data, if , then the gene expression data is masked. The masked embedding matrix Elements The formula is:

[0071]

[0072] in, is the mask vector.

[0073] The attention-based Transformer model has become ubiquitous for various problems in natural language processing, including translation, text classification, question answering, etc. One of the main applications of Transformer is to use it as a pre-trained language model, i.e., the model is first pre-trained using a language modeling objective on a large corpus, and then fine-tuned on the target task using supervised data. Transformer is built on top of the Multi-Head Self-Attention (MHA) mechanism.

[0074] Due to the large number of genes in the total gene vocabulary (about 25,000), the original embedding matrix The dimension of the input model is larger than the input length of the traditional Transformer. Therefore, the present invention introduces the Linformer with a linear complexity self-attention mechanism to replace the traditional Transformer as the model encoder to enhance the model capacity.

[0075] First, according to the embedding matrix Calculation query ,key Sum ,in, is a learnable weight matrix. Then, the present invention uses two dimensions as The linear projection matrix and , convert the original key K and value V, compress it from N dimensions to c dimensions, and get ; Where c < N, in this embodiment, c = 256. Finally, the attention score is calculated using the converted key-value pairs:

[0076]

[0077] In terms of model design, the present invention constructs a model with six encoder layers and four self-attention heads in each layer. Each layer also includes Add & Norm (residual connection and layer normalization) and a feedforward layer. The model architecture is as follows Figure 3 At the same time, the present invention selects mean square error as the loss function, and the loss function The expression is as follows:

[0078]

[0079] Among them, B is the number of cells, N is the total number of genes, is the number of genes masked, It's cells The corresponding gene The true embedding vector of is the predicted embedding vector, is an indicator variable. The corresponding gene It is 1 when masked, otherwise it is 0. Through this natural language-like self-supervised learning strategy, the present invention can learn the basic pattern of single-cell gene expression on a large amount of unlabeled single-cell sequencing data without the need for a large amount of manually annotated data.

[0080] S4, Supervised cell subset label mapping

[0081] This step achieves cell subpopulation label mapping of single-cell sequencing data by fine-tuning the pre-trained model obtained in step 3. The main process is to use the single-cell sequencing dataset with cell subpopulation labels provided by the user as input, generate high-quality cell representations through the pre-trained model, and then use the classifier to assign labels.

[0082] Specifically, the pre-trained model outputs a d-dimensional feature vector corresponding to each gene, then extracts the cell representation through one-dimensional convolution, and then uses the cell representation to assign cell subpopulation labels through a multi-layer perceptron (MLP). The classifier optimizes parameters based on the training data, and the optimization goal is to use the cross entropy loss between the predicted cell subpopulation labels and the actual cell subpopulation labels. The calculation method is as follows:

[0083]

[0084] Where B is the number of cells, For cells The true cell subset label, For cells Predicted cell subset labels.

[0085] The present invention uses 80% of the cells in the fine-tuning data as the training set, and the remaining 20% ​​of the cells as the test set. The performance evaluation indicators of the fine-tuning results include AUC (area under the curve) ± standard deviation and F1 score. In this way, the final output is the precise label annotation of individual cells, which can be used for downstream biological analysis, such as cell type annotation.

[0086] On the other hand, an embodiment of the present invention provides a cell subpopulation label mapping system for implementing the above-mentioned cell subpopulation label mapping method, such as Figure 4 As shown, the system includes:

[0087] The single-cell preprocessing and corpus construction module is used to preprocess single-cell data and construct a single-cell corpus;

[0088] The single-cell embedding module is used to represent the gene embedding of each gene in the cell as the sum of the gene representation embedding and the gene expression embedding, and obtain the gene embedding matrix corresponding to the cell;

[0089] The self-supervised learning module based on Linformer is used to randomly mask some genes with non-zero expression based on the masked language model, obtain the masked gene embedding matrix, and obtain the pre-trained model based on the self-supervised learning strategy of Linformer;

[0090] The supervised cell subpopulation label mapping module is used to input a single-cell sequencing data set with cell subpopulation labels into a pre-trained model, generate cell representations through the pre-trained model, and use a classifier to assign cell subpopulation labels.

[0091] Regarding the system in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0092] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The system 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 present application. A person of ordinary skill in the art can understand and implement it without creative work.

[0093] Accordingly, the present application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the cell subpopulation label mapping method as described above. Figure 5 As shown, it is a hardware structure diagram of any device with data processing capability in which the cell subpopulation label mapping method provided by the embodiment of the present invention is located. Figure 5 In addition to the processor, memory and network interface shown, any device with data processing capability in which the apparatus in the embodiment is located may also include other hardware, generally based on the actual functions of the device with data processing capability, which will not be described in detail.

[0094] Accordingly, the present application also provides a computer-readable storage medium on which computer instructions are stored, and when the instructions are executed by the processor, the cell subpopulation label mapping method as described above is implemented. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), an SD card, a flash card (Flash Card), etc. equipped on the device. Furthermore, the computer-readable storage medium can also include both an internal storage unit of any device with data processing capabilities 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 capabilities, and can also be used to temporarily store data that has been output or is to be output.

[0095] 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.

[0096] It should be understood that the present application is not limited to the exact construction that has been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof.

[0097] The above is only a preferred embodiment of the present invention. Although the present invention has been disclosed as a preferred embodiment, it is not intended to limit the present invention. Any technician familiar with the art can make many possible changes and modifications to the technical solution of the present invention by using the above disclosed methods and technical contents without departing from the scope of the technical solution of the present invention, or modify it into an equivalent embodiment of equivalent changes. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of protection of the technical solution of the present invention.

Claims

1. A cell subpopulation label mapping method, characterized in that: The method comprises: Preprocess the single-cell data and construct a single-cell corpus; The gene embedding of each gene in the cell is expressed as the sum of the gene representation embedding and the gene expression embedding, and the gene embedding matrix corresponding to the cell is obtained, which is: The genes in the total gene vocabulary are encoded into multi-dimensional vectors through the gene2vec model as gene representation embedding; The expression of genes in the total gene vocabulary is converted into a discretized gene expression code, and the gene expression code is mapped into a multidimensional vector using an embedding layer as a gene expression embedding; the gene expression code is obtained by dividing the genes in the total gene vocabulary into transcription factor related genes G tf and other genes G o Two parts: Calculate the G of gene expression in each cell o The non-zero median and G tf The ratio of the non-zero median of ; discretize the gene expression, and then use label encoding to assign an integer to each interval to replace the original gene expression; the range of all gene expression values ​​of cell j is recorded as Divide it into K continuous intervals k={1,2,…,K}, if the gene belongs to G o And the gene expression level in If the gene belongs to G tf And the gene expression level is r times If the gene expression is within , the gene expression code is k; if the gene expression is 0, the gene expression code is 0; Based on the masked language model, some genes with non-zero expression are randomly masked to obtain the masked gene embedding matrix, and the pre-training model is obtained based on the self-supervised learning strategy of Linformer; A single-cell sequencing dataset with cell subpopulation labels is input into a pre-trained model, a cell representation is generated through the pre-trained model, and a classifier is used to assign cell subpopulation labels.

2. The method according to claim 1, characterized in that The preprocessing of the single cell data is specifically as follows: Read the raw count matrix of single-cell sequencing and generate a sparse matrix that records the gene expression of single cells; Perform quality correction to filter out cells that do not meet the requirements for gene feature quantity, sequencing depth, and mitochondrial count ratio; The gene expression data were Log-normalized to obtain the normalized single-cell gene expression data.

3. The method according to claim 1, characterized in that The method randomly masks some genes with non-zero expression based on the mask language model, specifically: among all the genes used by cell j, 10% of the genes with non-zero expression are randomly selected for masking. For each selected gene, assume that the random number r j Uniformly distributed in the interval [0,1), if 0≤r j <0.4, then the gene representation data is masked, if 0.4≤r j <1, the gene expression data are masked.

4. The method according to claim 1, characterized in that The pre-trained model uses Linformer with linear complexity self-attention mechanism as the model encoder. First, the query, key and value are calculated according to the gene embedding matrix. Then, two linear projection matrices are used to transform the key and value respectively to complete the matrix dimension compression. Finally, the attention score is calculated using the converted key-value pairs.

5. The method according to claim 1, characterized in that The expression of the loss function L of the pre-training model is: Where B is the total number of cells, N is the total number of genes, |M j | is the number of genes masked, y ij is the true embedding vector of gene gi corresponding to cell j, is the predicted embedding vector, m ij is an indicator variable, which is 1 when the gene gi corresponding to cell j is masked, and 0 otherwise.

6. A cell subpopulation label mapping system, characterized in that: For implementing the cell subpopulation label mapping method according to any one of claims 1 to 5, the system comprises: Single-cell preprocessing and corpus construction module, used to preprocess single-cell data and construct a single-cell corpus; The single-cell embedding module is used to represent the gene embedding of each gene in the cell as the sum of the gene representation embedding and the gene expression embedding, and obtain the gene embedding matrix corresponding to the cell; The self-supervised learning module based on Linformer is used to randomly mask some genes with non-zero expression based on the masked language model, obtain the masked gene embedding matrix, and obtain the pre-trained model based on the self-supervised learning strategy of Linformer; The supervised cell subpopulation label mapping module is used to input a single-cell sequencing data set with cell subpopulation labels into a pre-trained model, generate cell representations through the pre-trained model, and use a classifier to assign cell subpopulation labels.

7. An electronic device 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 cell subpopulation label mapping method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the cell subpopulation label mapping method according to any one of claims 1 to 5 is implemented.

9. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the cell subpopulation label mapping method according to any one of claims 1 to 5 is implemented.

Images