Tumor Bulk sample RNA-seq sequencing data deconvolution method based on deep learning model

Abstract

The invention relates to a tumor Bulk sample RNA-seq sequencing data deconvolution method based on a deep learning model, and the method comprises the steps: integrating a plurality of scRNA-seq data sets containing tumor cells, carrying out the cell annotation, constructing an scRNA-seq reference data set, and generating a pseudo Bulk RNA-seq data set; carrying out standardization and gene screening on the two data sets, inputting the two data sets into two encoders, extracting deep potential features, and carrying out deconvolution on the extracted features to obtain a cell proportion and cell specific gene expression matrix; carrying out iteration on model parameters through an error back propagation algorithm and a gradient descent method, and finally storing a model weight with an optimal verification effect; according to the method, the deep learning is combined with the construction and annotation of the scRNA-seq reference data set, the efficient deconvolution of the data is realized, the pseudo Bulk data is generated by adopting random sampling for training, and the generalization ability and prediction precision of the model are improved.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and computer applications, and specifically relates to a deconvolution method for tumor bulk sample RNA-seq sequencing data based on a deep learning model. Background Art

[0002] In the fields of bioinformatics and tumor research, bulk RNA-seq and single-cell RNA-seq (scRNA-seq) are two important transcriptomics research technologies, each with unique advantages and application scenarios. Bulk RNA-seq provides transcriptomic snapshots for tissues or cell populations by measuring the average gene expression level of a cell population. This method has the advantages of low cost and high throughput, can reveal the average gene expression profile at the macro level, and is suitable for studying complex biological processes. In contrast, scRNA-seq analyzes gene expression characteristics at the level of individual cells, can identify rare cell types, reveal heterogeneity in cell populations, and discover cell states or transition states that may be missed in bulk analysis. However, scRNA-seq is costly, requires a large experimental and computational workload, and requires specialized computational analysis methods to handle its data complexity.

[0003] Existing bulk RNA-seq deconvolution tools have several significant flaws, including dependence on known reference genome datasets, neglect of heterogeneity and dynamic changes in cell types in samples, and deficiencies in processing technical errors and biological noise. In addition, there is currently no tool that can accurately deconvolve tumor bulk samples. For example, the Chinese patent application number is US20230049525A1, entitled "Method for Identifying Cell Type-Specific Gene Expression Levels by Deconvolving Bulk Gene Expression", which discloses a method for identifying gene expression levels in a specific cell type based on a large number of gene expression levels measured in a tissue sample containing multiple cell types. The Chinese patent application number is WO2021108556A1, entitled "Method for Identifying Cell Type-Specific Gene Expression Levels by Deconvolving Bulk Gene Expression", which discloses a method for identifying gene expression levels in a specific cell type based on a large number of gene expression levels measured in a tissue sample containing multiple cell types. The identification methods involved in the above two patents are representative, and neither can provide an accurate deconvolution tool specifically for tumor bulk samples. Summary of the invention

[0004] In order to solve the shortcomings of the current tumor bulk sample RNA-seq data deconvolution method, the present invention discloses a tumor bulk sample RNA-seq sequencing data deconvolution method based on a deep learning model, comprising the following steps:

[0005] S1. Integrate multiple scRNA-seq datasets containing tumor cells, perform cell annotation, and construct a large scRNA-seq reference dataset;

[0006] S2. Simulate the cellular composition of real Bulk RNA-seq to provide training and validation data for the model: Randomly sample from the annotated scRNA-seq reference dataset and combine them according to the specified cell ratio to generate a pseudo Bulk RNA-seq dataset;

[0007] S3. Data preprocessing: Standardize and filter genes of the scRNA-seq reference dataset and pseudo bulk RNA-seq dataset to ensure the consistency of input data in feature expression and gene set;

[0008] S4. The preprocessed scRNA-seq reference dataset and pseudo Bulk RNA-seq dataset are input into the reference set encoder and the Bulk encoder to extract the deep potential features representing the two datasets, and then the decoder is used to deconvolute the extracted features into cell proportions and cell-specific gene expression matrices;

[0009] S5. The model parameters are continuously updated and optimized through the error back propagation algorithm and gradient descent method: In each training iteration, the error between the model prediction and the actual pseudo-bulk data is evaluated through the mean absolute error and mean square error loss functions, the prediction error is gradually reduced, and finally the model weight with the best verification effect is saved.

[0010] Preferably, in S1, principal component analysis and UMAP visualization are performed on the scRNA-seq dataset, and the scVI model is used to remove batch effects on the entire scRNA-seq dataset;

[0011] The Leiden algorithm was used to perform cluster analysis on the scRNA-seq reference dataset and divide it into multiple clusters. The highly expressed marker genes in each cluster were displayed by dot plots to determine the specific cell types of each cluster. Finally, the scRNA-seq reference dataset was annotated into 10 different cell types, namely T cells, NK cells, B cells, plasma cells, myeloid cells, mast cells, fibroblasts, endothelial cells, epithelial cells and unidentified cells. Among them, unidentified cells will be discarded in subsequent analysis.

[0012] Preferably, further subgroup annotation is performed on some of the cell types: all T cells and NK cells are extracted from the scRNA-seq reference dataset, and the previous steps are repeated to further annotate T cells and NK cells as resting T cells, helper T cells, cytotoxic T cells, exhausted T cells, regulatory T cells and NK cells; similarly, subgroup annotation is performed on B cells, myeloid cells, fibroblasts and endothelial cells.

[0013] Preferably, the inferCNV algorithm is used to analyze the scRNA-seq reference dataset to identify malignant tumor cells; for epithelial cells, cells with higher CNV scores are defined as malignant tumor cells, thereby obtaining the final annotated scRNA-seq reference dataset for subsequent deep learning network model training.

[0014] Preferably, in S2, the annotated scRNA-seq reference dataset is randomly sampled to construct a pseudo BulkRNA-seq dataset, and the specific steps are as follows:

[0015] S2.1. Randomly select an integer between 1000 and 5000 as the total number of cells M of the bulk sample;

[0016] S2.2. Randomly select N numbers between 0 and 1, where N is the number of cell types finally annotated in the scRNA-seq reference dataset, and normalize these numbers so that their sum is 1, which is used as the proportion P of each cell type in the bulk sample;

[0017] S2.3. Calculate the number of cells T of each cell category based on the total number of cells M and the proportion of cell categories P; then randomly select Ti (i = 1, 2, 3 …, N) cells for each cell category from the scRNA-seq reference dataset;

[0018] S2.4. Calculate cell type-specific gene expression for randomly selected cells and obtain pseudo-bulk RNA-seq data by addition;

[0019] Through the above steps, pseudo Bulk RNA-seq datasets with sample sizes of a and b are constructed for training and validation sets, respectively, where a>b.

[0020] Preferably, in S3, the scRNA-seq reference dataset is standardized: the total reads of each cell are standardized to c and logarithmically transformed; then, the Seurat_v3 algorithm in scanpy is used to screen the highly variable genes, the top d highly variable genes are selected, and then the cell type classification addition operation is performed on the screened dataset to generate a feature matrix of cell categories and highly variable genes;

[0021] Next, based on these d highly variable genes, gene screening is performed on the pseudo-bulk RNA-seq data and the pseudo-cell-specific gene expression matrix to ensure that the genes in the pseudo-bulk sample are consistent with the highly variable genes in the scRNA-seq reference dataset; finally, the pseudo-bulk RNA-seq data are min-max normalized, and the pseudo-cell-specific gene expression matrix is ​​log1p transformed; where a, b, c, and d are generally integer multiples of 10,000.

[0022] Preferably, in S4, the preprocessed scRNA-seq feature matrix and pseudo Bulk RNA-seq data are respectively input into two fully connected layers with residual connection modules: a reference set encoder and a Bulk encoder, so as to perform deep low-dimensional feature learning and obtain latent features Zref and Zbulk; then, matrix multiplication and dot multiplication operations are performed on these two latent features to calculate the cell proportion feature Zprop and the cell-specific gene expression feature Zexpr; finally, Zprop and Zexpr are respectively input into the cell proportion decoder and the cell-specific gene expression decoder for decoding, so as to obtain the predicted cell proportion prop and the cell-specific gene expression matrix Expr.

[0023] Preferably, in S5, the entire deep network model is optimized by minimizing the error loss, and the optimization goal is:

[0024]

[0025] in:

[0026] is the known true cell proportion in the pseudo-bulk data;

[0027] is the known true cell-specific gene expression matrix in the pseudo-bulk data;

[0028] is the proportion of cells predicted by Prop Decoder;

[0029] is the cell-specific gene expression matrix predicted by Expr Decoder;

[0030] It is a hyperparameter used to adjust the weight between the two parts of the loss;

[0031] Represents the mean absolute error (Mean Absolute Error), which is used to evaluate the difference between the predicted cell proportion and the true proportion;

[0032] Represents the mean squared error, which is used to evaluate the difference between the predicted cell-specific gene expression matrix and the true value.

[0033] Preferably, when the training reaches a preset stop condition or the error converges to a lower level, the model weights with the best verification effect are saved for subsequent deconvolution analysis of real Bulk RNA-seq data.

[0034] Preferably, the model weights are applied by inputting the preprocessed bulk RNA-seq data into a pre-trained deep learning model, which predicts the proportion of each cell subpopulation in the bulk sample and the corresponding cell type-specific gene expression matrix.

[0035] The tumor bulk sample RNA-seq sequencing data deconvolution method based on the deep learning model of the present invention has at least the following advantages:

[0036] 1. This paper achieves efficient deconvolution of tumor bulk RNA-seq data by combining deep learning with the construction and annotation of a large-scale scRNA-seq reference dataset.

[0037] 2. Random sampling is used to generate pseudo Bulk data for training, and the residual connection module is combined with the optimized loss function to improve the generalization ability and prediction accuracy of the model.

[0038] 3. It has enhanced adaptability to diverse tumor samples and can be widely used in the study of various tumor types.

[0039] 4. It provides a powerful and convenient tool for in-depth analysis of the tumor microenvironment, helping to reveal tumor heterogeneity, cell-cell interactions, and key mechanisms of tumor occurrence and development.

[0040] 5. It can better analyze the response of the tumor microenvironment to treatment options such as immunotherapy and targeted therapy, and provide support for the development of precision medicine and individualized treatment strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0042] Figure 1 This is the UMAP visualization result of the scRNA-seq reference dataset before and after batch removal.

[0043] Figure 2It is the cell clustering diagram of the scRNA-seq reference dataset after Leiden clustering analysis and its corresponding cell type annotation diagram.

[0044] Figure 3 is a dot plot of marker gene expression in different cell clusters in the scRNA-seq reference dataset.

[0045] Figure 4 It is the final cell type annotation result of the scRNA-seq reference dataset.

[0046] Figure 5 It is a process of deconvolution of bulk sample RNA-seq sequencing data based on deep learning models. DETAILED DESCRIPTION

[0047] The present invention is further described in detail below through examples so that those skilled in the art can implement the invention with reference to the description.

[0048] It should be understood that the terms such as “having”, “including” and “comprising” used herein do not exclude the existence or addition of one or more other elements or combinations thereof.

[0049] The present invention is further described in conjunction with the accompanying drawings, comprising the following steps:

[0050] 1) Construction of a large scRNA-seq reference dataset

[0051] We obtained 6 lung cancer-related scRNA-seq datasets from the Curated Cancer Cell Atlas (3CA) database. These datasets are in raw count format and come from Bischoff et al. 2021, Chan et al. 2021, Kim et al. 2020, Laughney et al. 2020, Qian et al. 2020, and Xing et al. 2021. After quality control of these datasets, we finally obtained a scRNA-seq dataset containing 120 samples, 33,410 genes, and 378,115 cells. We performed principal component analysis (PCA) and UMAP visualization on this dataset, and the results are shown in the figure below. Figure 1 As shown on the left. Subsequently, the scVI model was used to remove the batch effect of the entire scRNA-seq dataset, and the results were as follows Figure 1 Shown right.

[0052] 2) Cell type annotation of large scRNA-seq reference datasets

[0053] First, we used the Leiden algorithm to perform cluster analysis on the scRNA-seq reference dataset and divided it into multiple clusters, such as Figure 2 As shown on the left. Then, through Figure 3 The dot plots shown show the highly expressed marker genes in each cluster, so as to determine the specific cell types of each cluster. Finally, we annotated the scRNA-seq reference dataset into 10 different cell types, namely T cells, NK cells (Natural Killer cells), B cells, Plasma cells, Myeloid cells, Mast cells, Fibroblast cells, Endothelial cells, Epithelial cells and Unidentified cells. Unidentified cells may be double cells or low-quality cells, so they will be discarded in subsequent analysis. The annotation results of each cell type are as follows Figure 2 Shown right.

[0054] Next, we further annotated some cell types into subpopulations. For example, we extracted all T cells and NK cells from the scRNA-seq reference dataset, repeated the previous steps, and further annotated T cells and NK cells into Quiescent T cells, Helper T cells, Cytotoxic T cells, Exhausted T cells, Regulatory T cells, and NK cells. Similarly, B cells, myeloid cells, fibroblasts, and endothelial cells were annotated as subpopulations; specifically, B cells were further annotated into naive B cells (Naive B), memory B cells (Memory B), and germinal center B cells (GC B); fibroblasts were further annotated into proinflammatory cancer-associated fibroblasts (iCAFs), stromal cancer-associated fibroblasts (mCAFs), proliferative cancer-associated fibroblasts (pCAFs), metabolic cancer-associated fibroblasts (meCAFs), peripheral nerve-like cancer-associated fibroblasts (pnCAFs), and antigen-presenting cancer-associated fibroblasts (apCAFs); endothelial cells were further annotated into lymphatic endothelial cells (Lymphatic EC) and vascular endothelial cells (Vascular EC); myeloid cells were further annotated into neutrophils (Neutrophil), CD14+ monocytes (CD14+Mono), CD16+ monocytes (CD16+ Mono), CD14+CD16+ monocytes (CD14+ CD16+ TAMs include mono), plasmacytoid dendritic cells (pDCs), classical dendritic cells type 1 (cDC1), classical dendritic cells type 2 (cDC2), interferon-associated tumor-associated macrophages (INF TAMs), inflammatory cytokine-enriched tumor-associated macrophages (Inflam TAMs), angiogenic tumor-associated macrophages (Angio TAMs), pro-proliferative tumor-associated macrophages (Prolif TAMs), lipid-associated tumor-associated macrophages (LA TAMs), and tissue-resident tumor-associated macrophages (TRM TAMs).

[0055] Finally, we used the inferCNV algorithm to analyze the scRNA-seq reference dataset to identify malignant tumor cells. For epithelial cells, cells with higher CNV scores were defined as malignant tumor cells, thus obtaining the final annotated scRNA-seq reference dataset for subsequent deep learning network model training. The annotation results of the scRNA-seq reference set are shown in Figure 2. Figure 4 shown.

[0056] 3) Construction of pseudo bulk RNA-seq dataset

[0057] Random sampling of the annotated scRNA-seq reference dataset was performed to construct a pseudo Bulk RNA-seq dataset. The specific steps are as follows:

[0058] i. Randomly select an integer between 1000 and 5000 as the total number of cells M in the Bulk sample

[0059] ii. Randomly select N numbers between 0 and 1 (N is the number of cell categories finally annotated in the scRNA-seq reference dataset), and normalize these numbers so that their sum is 1, which is used as the proportion P of each cell category in the bulk sample.

[0060] iii. Calculate the number of cells T in each cell category based on the total number of cells M and the proportion of cell categories P. Then, randomly select Ti (i = 1, 2, 3 …, N) cells for each cell category from the scRNA-seq reference dataset.

[0061] iv. Calculate cell type-specific gene expression for randomly selected cells and obtain pseudo-bulk RNA-seq data through addition calculation.

[0062] Through the above steps, we constructed pseudo Bulk RNA-seq datasets with sample sizes of 80,000 and 20,000 (or 100,000 and 10,000), which were used for training and validation sets, respectively.

[0063] 4) Data preprocessing

[0064] First, the scRNA-seq reference dataset was standardized. Specifically, the total number of reads per cell was standardized to 10,000 and logarithmically transformed (log1p). Next, the Seurat_v3 algorithm in scanpy was used to filter the highly variable genes and select the top 20,000 highly variable genes. Then, the cell type classification addition operation was performed on the filtered dataset to generate a feature matrix of cell classes and highly variable genes.

[0065] Next, based on these 20,000 highly variable genes, pseudo-bulk RNA-seq data and pseudo-cell-specific gene expression matrices were screened to ensure that the genes in the pseudo-bulk samples were consistent with the highly variable genes in the scRNA-seq reference dataset. Finally, the pseudo-bulk RNA-seq data were min-max normalized, and the pseudo-cell-specific gene expression matrix was log1p transformed.

[0066] 5) Training of deep network models:

[0067] like Figure 5As shown in Figure 1, the preprocessed scRNA-seq feature matrix and pseudo Bulk RNA-seq data are respectively input into two fully connected layers with residual connection modules: Reference Encoder and Bulk Encoder, to perform deep low-dimensional feature learning and obtain the potential feature Z ref and Z bulk Subsequently, matrix multiplication and dot multiplication are performed on these two latent features to calculate the cell ratio feature Z prop and cell-specific gene expression signature Z expr Finally, Z prop and Z expr The data are respectively input into the cell proportion decoder (Prop Decoder) and the cell-specific gene expression decoder (Expr Decoder) for decoding to obtain the predicted cell proportion prop and cell-specific gene expression matrix Expr.

[0068] The entire deep network model is optimized by minimizing the error loss, and the optimization goal is:

[0069]

[0070] in,

[0071] is the known true cell proportion in the pseudo-bulk data;

[0072] is the known true cell-specific gene expression matrix in the pseudo-bulk data;

[0073] is the proportion of cells predicted by Prop Decoder;

[0074] is the cell-specific gene expression matrix predicted by Expr Decoder;

[0075] It is a hyperparameter used to adjust the weight between the two parts of the loss;

[0076] Represents the mean absolute error (Mean Absolute Error), which is used to evaluate the difference between the predicted cell proportion and the true proportion;

[0077] Represents the mean squared error, which is used to evaluate the difference between the predicted cell-specific gene expression matrix and the true value.

[0078] The model parameters are continuously updated and optimized through the error backpropagation algorithm (Backpropagation) and the gradient descent method. In each training iteration, the model gradually reduces the gap between the predicted value and the true value, thereby improving the accuracy of the model in the deconvolution task. Finally, when the training reaches the preset stop condition or the error converges to a lower level, the model weights with the best verification effect are saved for subsequent deconvolution analysis of real Bulk RNA-seq data.

[0079] This embodiment takes the lung cancer sample GSM3827136 in the GEO database as an example, and provides a deconvolution method for RNA-seq sequencing data of tumor bulk samples based on deep learning, including the following steps:

[0080] Load the deep network model:

[0081] First, load the deep learning model pre-trained on the lung cancer scRNA-seq reference dataset and import the corresponding model weights. At the same time, obtain the 20,000 highly variable genes screened during the pre-training process.

[0082] Preprocessing Bulk RNA-seq Data:

[0083] Next, the raw count data of Bulk RNA-seq were screened using the 20,000 highly variable genes screened during pre-training, and zero values ​​were used to fill in the missing genes. Subsequently, the screened Bulk RNA-seq data were subjected to min-max normalization to ensure data standardization and consistency.

[0084] Predictions from the network model:

[0085] The preprocessed bulk RNA-seq data is input into the pre-trained deep learning model, which will predict the proportion of each cell subpopulation in the bulk sample and the corresponding cell type-specific gene expression matrix.

[0086] Notes:

[0087] Bulk RNA-seq: A technique that measures the average gene expression level of all cells in a tissue or cell population, providing macroscopic gene expression information but unable to distinguish the expression characteristics of individual cells.

[0088] scRNA-seq: Single-cell RNA sequencing technology analyzes gene expression at the single-cell level, reveals the characteristics and heterogeneity of different cell types, and identifies rare cell types.

[0089] Deep learning: Deep learning is a big data analysis method based on artificial neural networks that can automatically extract features, learn complex patterns, and make predictions through large amounts of data training. In bioinformatics, deep learning can be used to process and analyze large amounts of genomic data to achieve tasks such as deconvolution, classification, and feature extraction.

[0090] Tumor microenvironment: Tumor microenvironment refers to the complex environment surrounding tumor cells, including cancer cells, immune cells, fibroblasts, vascular cells, matrix components, and secreted molecules. These components interact with each other and affect tumor growth, metastasis, immune escape, and response to treatment.

[0091] Tumor Biology: Tumor biology is the science of studying the occurrence, development, metastasis and interaction of tumors with the host. It covers a wide range of areas from molecular mechanisms, gene regulation to cell behavior, helping to understand the basic mechanisms of tumors and develop new treatments.

[0092] Tumor heterogeneity: Tumor heterogeneity refers to the presence of multiple different cell types, gene mutations, expression patterns, and microenvironmental characteristics inside and outside the tumor. Heterogeneity causes the same tumor to show significant differences in different patients, and also causes the same tumor to have different evolutionary trajectories and treatment responses during progression.

[0093] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the embodiments shown and described herein.

Claims

1. A deconvolution method for bulk RNA-seq sequencing data of tumor samples based on a deep learning model, characterized in that: The following steps are involved: S1. Integrate multiple scRNA-seq datasets containing tumor cells, perform cell annotation, and construct a large scRNA-seq reference dataset; S2. Simulate the cellular composition of real Bulk RNA-seq to provide training and validation data for the model: Randomly sample from the annotated scRNA-seq reference dataset and combine them according to the specified cell ratio to generate a pseudo Bulk RNA-seq dataset; S3. Data preprocessing: Standardize and filter genes of the scRNA-seq reference dataset and pseudo bulk RNA-seq dataset to ensure the consistency of input data in feature expression and gene set; S4. The preprocessed scRNA-seq reference dataset and pseudo Bulk RNA-seq dataset are input into the reference set encoder and the Bulk encoder to extract the deep potential features representing the two datasets, and then the decoder is used to deconvolute the extracted features into cell proportions and cell-specific gene expression matrices; S5. The model parameters are continuously updated and optimized through the error back propagation algorithm and gradient descent method: In each training iteration, the error between the model prediction and the actual pseudo-bulk data is evaluated through the mean absolute error and mean square error loss functions, the prediction error is gradually reduced, and finally the model weight with the best verification effect is saved.

2. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 1, characterized in that: In S1, principal component analysis and UMAP visualization were performed on the scRNA-seq dataset, and the scVI model was used to remove batch effects on the entire scRNA-seq dataset; The Leiden algorithm was used to perform cluster analysis on the scRNA-seq reference dataset and divide it into multiple clusters. The highly expressed marker genes in each cluster were displayed by dot plots to determine the specific cell types of each cluster. Finally, the scRNA-seq reference dataset was annotated into 10 different cell types, namely T cells, NK cells, B cells, plasma cells, myeloid cells, mast cells, fibroblasts, endothelial cells, epithelial cells and unidentified cells. Among them, unidentified cells will be discarded in subsequent analysis.

3. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 2, characterized in that: Further subpopulation annotation was performed on some of the cell types: all T cells and NK cells were extracted from the scRNA-seq reference dataset, and the previous steps were repeated to further annotate T cells and NK cells into resting T cells, helper T cells, cytotoxic T cells, exhausted T cells, regulatory T cells and NK cells; similarly, subpopulation annotation was performed on B cells, myeloid cells, fibroblasts and endothelial cells.

4. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 3, characterized in that: The inferCNV algorithm was used to analyze the scRNA-seq reference dataset to identify malignant tumor cells; for epithelial cells, cells with higher CNV scores were defined as malignant tumor cells, thereby obtaining the final annotated scRNA-seq reference dataset for subsequent deep learning network model training.

5. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 1, characterized in that: In S2, the annotated scRNA-seq reference dataset is randomly sampled to construct a pseudo BulkRNA-seq dataset. The specific steps are as follows: S2.

1. Randomly select an integer between 1000 and 5000 as the total number of cells M of the bulk sample; S2.

2. Randomly select N numbers between 0 and 1, where N is the number of cell types finally annotated in the scRNA-seq reference dataset, and normalize these numbers so that their sum is 1, which is used as the proportion P of each cell type in the bulk sample; S2.

3. Calculate the number of cells T of each cell category based on the total number of cells M and the proportion of cell categories P; then randomly select Ti (i = 1, 2, 3 …, N) cells for each cell category from the scRNA-seq reference dataset; S2.

4. Calculate cell type-specific gene expression for randomly selected cells and obtain pseudo-bulk RNA-seq data by addition; Through the above steps, pseudo Bulk RNA-seq datasets with sample sizes of a and b are constructed for training and validation sets, respectively, where a>b.

6. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 5, characterized in that: In S3, the scRNA-seq reference dataset was standardized: the total number of reads of each cell was standardized to c and logarithmically transformed; then, the Seurat_v3 algorithm in scanpy was used to filter the highly variable genes, and the top d highly variable genes were selected. Then, the cell type classification addition operation was performed on the filtered dataset to generate a feature matrix of cell categories and highly variable genes; Next, based on these d highly variable genes, the pseudo-bulk RNA-seq data and pseudo-cell-specific gene expression matrix were screened to ensure that the genes in the pseudo-bulk samples were consistent with the highly variable genes in the scRNA-seq reference dataset; finally, the pseudo-bulk RNA-seq data were min-max normalized, and the pseudo-cell-specific gene expression matrix was log1p transformed; where a, b, c, and d are integer multiples of 10,000.

7. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 1, characterized in that: In S4, the preprocessed scRNA-seq feature matrix and pseudo Bulk RNA-seq data are respectively input into two fully connected layers with residual connection modules: the reference set encoder and the Bulk encoder, for deep low-dimensional feature learning to obtain latent features Zref and Zbulk; then, matrix multiplication and dot multiplication operations are performed on these two latent features to calculate the cell proportion feature Zprop and the cell-specific gene expression feature Zexpr; finally, Zprop and Zexpr are respectively input into the cell proportion decoder and the cell-specific gene expression decoder for decoding to obtain the predicted cell proportion prop and cell-specific gene expression matrix Expr.

8. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 1, characterized in that: In S5, the entire deep network model is optimized by minimizing the error loss, and the optimization goal is: ； in: is the known true cell proportion in the pseudo-bulk data; is the known true cell-specific gene expression matrix in the pseudo-bulk data; is the proportion of cells predicted by Prop Decoder; is the cell-specific gene expression matrix predicted by Expr Decoder; It is a hyperparameter used to adjust the weight between the two parts of the loss; Represents the mean absolute error (Mean Absolute Error), which is used to evaluate the difference between the predicted cell proportion and the true proportion; Represents the mean squared error, which is used to evaluate the difference between the predicted cell-specific gene expression matrix and the true value.

9. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 1, characterized in that: When the training reaches the preset stop condition or the error converges to a lower level, the model weights with the best verification effect are saved for subsequent deconvolution analysis of real Bulk RNA-seq data.

10. The method for deconvolution of tumor bulk sample RNA-seq sequencing data based on a deep learning model according to claim 1, characterized in that: For the model weights, the preprocessed bulk RNA-seq data is input into the pre-trained deep learning model, which will predict the proportion of each cell subpopulation in the bulk sample and the corresponding cell type-specific gene expression matrix.

Images