Single-cell expression profile data enhancement method based on deep generative adversarial network model

Through the deep generation adversarial network model, the problem of low accuracy in inference of missing values ​​of single-cell RNA sequencing data is solved, and the randomness and diversity of the data is improved, and it is suitable for complex biological processes.

CN116884489BActive Publication Date: 2024-11-15NORTHWESTERN POLYTECHNICAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310776539.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-06-29
Publication Date
2024-11-15
Estimated Expiration
2043-06-29

Smart Images

  • Figure CN116884489B_ABST
    Figure CN116884489B_ABST
Patent Text Reader

Abstract

The present invention discloses a single-cell expression spectrum data enhancement method based on a deep generative adversarial network model, firstly constructing an input matrix based on the scRNA-seq data to be filled; then constructing a deep generative adversarial network, the deep generative adversarial network includes three generative adversarial networks, each of which includes a generator and a discriminator, and the generator and the discriminator are composed of a fully connected layer, a batch normalization layer, a convolutional layer and an activation layer; next, the matrix of the input discriminator is preprocessed, and the constructed deep generative adversarial network is trained in stages using a data set; finally, the scRNA-seq data to be filled is inferred for missing values ​​using the trained deep generative adversarial network. The present invention greatly improves the randomness and diversity of the data, so that the inferred gene expression data is more consistent with the dependency between cells and genes in complex biological processes.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a single-cell expression profile data enhancement method. Background Art

[0002] With the continuous development of second-generation sequencing technology, single-cell RNA sequencing (scRNA-seq) has gradually become a powerful tool that enables researchers to explore the cellular heterogeneity of complex biological systems at the single-cell level, providing new insights into cellular processes and disease mechanisms. Unlike traditional bulk RNA sequencing (bulk RNA-seq), scRNA-seq can detect the gene expression profiles of individual cells in a sample, enabling unbiased identification of rare or previously unknown cell types in tissues or organisms, thereby better understanding cell heterogeneity and function. In addition, scRNA-seq technology can track the developmental relationship between heterogeneous cell states over time, enabling researchers to study the dynamic process of cell differentiation or response to stimulation, such as embryonic development or cancer progression. scRNA-seq technology has also shown great potential in precision medicine, including clinical drug selection and new drug development.

[0003] Although scRNA-seq can provide high-resolution data at the cellular level, it still faces many challenges. Unlike bulk RNA-seq, which measures the average gene expression level of a group of cells, scRNA-seq may cause a large number of missing gene expression values ​​due to technical limitations, such as mRNA degradation after cell lysis, low capture efficiency, unstable amplification, excessive dilution of cell libraries, and low sequencing depth. At the same time, low RNA content in individual cells or silent gene expression during development may also lead to the existence of missing values. Missing values ​​in scRNA-seq data may lead to deviations in cell type clustering and reduce statistical power in downstream analysis (such as differential gene analysis). Therefore, inferring missing values ​​in scRNA-seq data is crucial to ensure the reliability and robustness of downstream analysis.

[0004] In order to accurately infer scRNA-seq data, a variety of inference methods have been developed, which can be roughly divided into four categories: smoothing or diffusion-based methods, statistical model-based methods, low-rank matrix-based methods, and deep learning-based methods. Although these methods have been widely used, they still have some defects, such as: (1) A large number of methods use the similarity between cells or genes and simply use the average value as the inference value to fill in the scRNA-seq data. However, this type of method will cause the loss of randomness and diversity of cellular gene expression, and diversity and randomness are crucial in biological processes; (2) There are also a large number of methods that rely on simple assumptions about data distribution, such as zero-inflated negative binomial distribution, but this type of method ignores the complex gene expression mechanism of cells, which may lead to inaccurate gene expression estimates, especially in the presence of complex scRNA-seq data distribution. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides a single-cell expression spectrum data enhancement method based on a deep generative adversarial network model. First, an input matrix is ​​constructed based on the scRNA-seq data to be filled; then a deep generative adversarial network is constructed, and the deep generative adversarial network includes three generative adversarial networks, each of which includes a generator and a discriminator, and the generator and the discriminator are composed of a fully connected layer, a batch normalization layer, a convolutional layer, and an activation layer; next, the matrix of the input discriminator is preprocessed, and the constructed deep generative adversarial network is trained in stages using a data set; finally, the trained deep generative adversarial network is used to infer missing values ​​for the scRNA-seq data to be filled. The present invention greatly improves the randomness and diversity of the data, so that the inferred gene expression data is more consistent with the dependency between cells and genes in complex biological processes.

[0006] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0007] Step 1: Construct an input matrix based on the scRNA-seq data to be populated;

[0008] Step 2: construct a deep generative adversarial network, which includes three generative adversarial networks, each of which includes a generator and a discriminator, and the generator and the discriminator are both composed of a fully connected layer, a batch normalization layer, a convolutional layer, and an activation layer;

[0009] Step 3: Preprocess the matrix of the input discriminator and use the dataset to train the constructed deep generative adversarial network in stages;

[0010] Step 4: Use the trained deep generative adversarial network to infer missing values ​​for the scRNA-seq data to be filled.

[0011] Furthermore, the specific process of step 1 is as follows:

[0012] Step 1-1: Obtain the gene expression matrix and cell type annotation file of the scRNA-seq dataset. The rows of the gene expression matrix represent genes and the columns represent cells.

[0013] Step 1-2: Perform TPM normalization on the gene expression matrix and perform logarithmic transformation to obtain the normalized gene expression matrix;

[0014] Step 1-3: Set the part of the gene expression matrix whose expression value is greater than 0 to 1, and set the part whose expression value is equal to 0 to 0, thereby obtaining the mask matrix corresponding to the gene expression matrix;

[0015] Step 1-4: Encode the cell type annotation file and convert it into the form of 0, 1, 2... to obtain the cell label vector;

[0016] Step 1-5: Slice the normalized gene expression matrix by column and convert each column into a square matrix;

[0017] Step 1-6: Slice the mask matrix by columns and convert each column into a square matrix.

[0018] Furthermore, the conversion methods mentioned in steps 1-5 and 1-6 are specifically as follows:

[0019] Calculate the square root of the number of rows (genes), and round the result up to an integer. Calculate the difference between the square of the integer and the number of rows, and then add a number of rows equal to the difference to each column, and then set all the data in the added rows to 0.

[0020] Furthermore, the three generative adversarial networks in step 2 are generative adversarial networks for generative adversarial networks for generative adversarial networks for generative adversarial networks for generative matrixes, and padding networks;

[0021] The optimization objective function of the gene expression data generation network is:

[0022]

[0023] Among them G x represents the generator of the gene expression data generation adversarial network, D x represents the discriminator of the gene expression data generation adversarial network, represents the set of all 1-Lipschitz functions, V xrepresents the loss function of the generative adversarial network for gene expression data;

[0024] The optimization objective function of the mask matrix generation network is:

[0025]

[0026] Among them G m Denotes the generator of the mask matrix generative adversarial network, D m Denotes the discriminator of the mask matrix generative adversarial network, V m represents the loss function of the mask matrix generation adversarial network, and α represents the hyperparameter of network training;

[0027] The optimization objective function of the filling network is:

[0028]

[0029] Among them G imp represents the generator of the padding network, D imp represents the discriminator of the padding network, V imp Represents the loss function of the padding network.

[0030] Furthermore, the loss function It is expressed as:

[0031]

[0032] in represents the expected value, pdata(x) represents the distribution of real data, and p z (z) represents the distribution of noise, λ represents the hyperparameter of network training, Representation calculation For variables The 2-norm of the gradient of .

[0033] Furthermore, the variable The samples obtained by linear interpolation of real samples and generated samples are expressed as:

[0034]

[0035] where x real represents the real sample, x fake represents the generated sample, ∈ is a value randomly sampled from a uniform distribution.

[0036] Furthermore, the hyperparameters involved in step 2 are defined as follows:

[0037] Ⅰ. The value of the hyperparameter α is between 0 and 1;

[0038] Ⅱ. The value of the hyperparameter λ is between 1 and 10.

[0039] Furthermore, the preprocessing method in step 3 is to combine the gene expression matrix and the mask matrix in the following manner:

[0040]

[0041] Where x is the gene expression matrix of a single cell processed in step 1, m is its corresponding mask matrix, and ⊙ represents the Hadamard product, that is, the two matrices are multiplied element by element. is the complement matrix corresponding to the mask matrix, and τ is the network hyperparameter.

[0042] Furthermore, in step 3, the training is performed in stages as follows:

[0043] Step 3-1: In the first training stage, the preprocessed gene expression data and mask matrix are input into the network, and the noise vector and data label vector are respectively input into the gene expression data generator and the mask matrix generator to generate gene expression data and mask matrix respectively; the generated mask matrix and the real mask matrix are input into the mask matrix discriminator for judgment, and the generated gene expression data and mask matrix are first combined according to the data preprocessing method in step 3, and then combined with the preprocessed gene expression data Input into the gene expression data discriminator for discrimination; optimize the objective function in step 2 through continuous training;

[0044] Step 3-2: In the second training phase, first fix the parameters of the gene expression data generator obtained in the first training phase, input random noise and data label vector into the generator to generate real gene expression data; input the gene expression data processed in step 1, mask matrix, data label vector and random noise into the filling network to generate the filled gene expression data; the data generated by the filling network and the data generated by the gene expression data generator are input into the discriminator for discrimination. Optimize the objective function in step 2 through continuous training.

[0045] Furthermore, the value of τ in step 3 is between 0 and 1.

[0046] Furthermore, the learning rate of the generative adversarial network is set at 10 -5 ~10 -3 between.

[0047] Furthermore, the inference method in step 4 is:

[0048] Step 4-1: After the second stage of training in step 3 is completed, the trained filling network is obtained;

[0049] Step 4-2: Input the gene expression matrix, mask matrix, cell label vector and random noise obtained in step 1 into the filling network to generate the filled gene expression data;

[0050] Step 4-3: Convert the gene expression data of each cell from the square matrix back into a vector form and remove the extra rows added in step 1. Merge the gene expression data of all cells by column to obtain the inferred complete gene expression matrix.

[0051] The beneficial effects of the present invention are as follows:

[0052] (1) The single-cell expression profile data enhancement method based on the deep generative adversarial network model provided by the present invention uses a neural network to mine the complex relationships between cells or genes, rather than inferring missing values ​​through simple diffusion or smoothing, which greatly improves the randomness and diversity of the data, making the inferred gene expression data more consistent with the dependency relationship between cells and genes in complex biological processes.

[0053] (2) The single-cell expression profile data enhancement method based on the deep generative adversarial network model provided by the present invention eliminates the need for users to assume the distribution pattern of complex data in advance, and to a certain extent avoids the reduction in the accuracy of data missing value inference due to the constructed statistical model not conforming to the actual situation. It is suitable for gene expression data with complex distribution. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic diagram of the process of deep generative adversarial network of the present invention.

[0055] Figure 2 It is a schematic diagram of the structure of the gene expression data generator and the mask matrix generator of the present invention.

[0056] Figure 3 It is a structural schematic diagram of the filling network generator of the present invention.

[0057] Figure 4 A schematic diagram of the structure of the generative adversarial network discriminator of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0059] The purpose of the present invention is to solve the problem of low accuracy of scRNA-seq data inference in existing methods, and propose a single-cell expression profile data enhancement method, model and application based on a deep generative adversarial network model. The technical solution adopted by the present invention to solve the above technical problems is: a single-cell expression profile data enhancement method, model and application based on a deep generative adversarial network model, comprising the following steps:

[0060] ①Build an input matrix based on the scRNA-seq data to be filled;

[0061] ② Construct a deep generative adversarial network, which includes three generative adversarial networks. Each generative adversarial network includes a generator and a discriminator. The generator and the discriminator are composed of a fully connected layer, a batch normalization layer, a convolutional layer, and an activation layer.

[0062] ③ Preprocess the matrix of the input discriminator and use the dataset to train the constructed generative adversarial network in stages;

[0063] ④ Use the trained deep generative adversarial network to infer missing values ​​of the scRNA-seq data to be filled.

[0064] The specific process of step ① is as follows:

[0065] Ⅰ. Obtain the gene expression matrix and cell type annotation file of the scRNA-seq dataset. The rows of the gene expression matrix represent genes and the columns represent cells.

[0066] II. Perform TPM normalization on the gene expression matrix and perform logarithmic transformation to obtain the normalized gene expression matrix;

[0067] III. Set the part of the gene expression matrix whose expression value is greater than 0 to 1, and set the part whose expression value is equal to 0 to 0, thereby obtaining the mask matrix corresponding to the gene expression matrix;

[0068] IV. Encode the cell type annotation file and convert it into the form of 0, 1, 2...;

[0069] V. Slice the normalized gene expression matrix by column, and convert each column into a square matrix according to a certain method;

[0070] VI. Slice the mask matrix by columns as well, and convert each column into a square matrix using a certain method.

[0071] Furthermore, the conversion method mentioned in steps V and VI is to calculate the square root of the row (number of genes), and round the result up to an integer value. Calculate the difference between the square of the integer value and the number of rows, and then add a certain number of rows to each column, the number is the difference mentioned above. Then set all the data of the added rows to 0.

[0072] The three GAN networks in step ② are gene expression data generation adversarial network, mask matrix generation adversarial network and filling network respectively;

[0073] The optimization objective function of the gene expression data generation network is:

[0074]

[0075] Among them G x represents the generator of the gene expression data generation adversarial network, D x represents the discriminator of the gene expression data generation adversarial network, represents the set of all 1-Lipschitz functions, V x Represents the loss function of the generative adversarial network for gene expression data.

[0076] The optimization objective function of the mask matrix generation network is:

[0077]

[0078] Among them G m Denotes the generator of the mask matrix generative adversarial network, D m Denotes the discriminator of the mask matrix generative adversarial network, V m represents the loss function of the mask matrix generative adversarial network, and α represents the hyperparameter of network training.

[0079] The optimization objective function of the filling network is:

[0080]

[0081] Among them G imp represents the generator of the padding network, D imp represents the discriminator of the padding network, V imp Represents the loss function of the padding network.

[0082] Further loss function It can be expressed as:

[0083]

[0084] in represents the expected value, pdata(x) represents the distribution of real data, and p z (z) represents the distribution of noise, λ represents the hyperparameter of network training, Representation calculation For variables The 2-norm (Euclidean norm) of the gradient of . Further The samples obtained by linear interpolation of real samples and generated samples can be expressed as:

[0085]

[0086] where x real represents the real sample, x fakerepresents the generated sample, ∈ is a value randomly sampled from a uniform distribution.

[0087] Further, the hyperparameters involved in step ② are defined as follows:

[0088] Ⅰ. The value of the hyperparameter α is between 0 and 1.

[0089] Ⅱ. The value of the hyperparameter λ is between 1 and 10.

[0090] The preprocessing method in step ③ is to combine the gene expression matrix and the mask matrix in the following manner:

[0091]

[0092] Where x is the gene expression matrix of a single cell processed in step ①, m is its corresponding mask matrix, and ⊙ represents the Hadamard product, that is, the two matrices are multiplied element by element. is the complement matrix corresponding to the mask matrix, and τ is the network hyperparameter.

[0093] In step ③, the training is carried out in two stages:

[0094] Ⅰ. In the first training stage, the preprocessed gene expression data and mask matrix are input into the network, and the noise vector and data label vector are input into the gene expression data generator and mask matrix generator respectively to generate gene expression data and mask matrix respectively. The generated mask matrix and the real mask matrix are input into the mask matrix discriminator for judgment, and the generated gene expression data and mask matrix are first combined according to the data preprocessing method in step ③, and then combined with the preprocessed gene expression data Input into the gene expression data discriminator for discrimination. Optimize the objective function in step ② through continuous training.

[0095] Ⅱ. In the second training stage, the parameters of the gene expression data generator obtained in the first training stage are first fixed, and random noise and data label vector are input into the generator to generate "real" gene expression data. The gene expression data, mask matrix, data label vector and random noise processed in step ① are input into the filling network to generate the filled gene expression data. The data generated by the filling network and the data generated by the gene expression data generator are input into the discriminator for discrimination. The objective function in step ② is optimized through continuous training.

[0096] Furthermore, the value of τ in step ③ is between 0 and 1.

[0097] Furthermore, the learning rate of the above-mentioned generative adversarial network is set at 10 -5 ~10 -3 between.

[0098] The inference method in step ④ is:

[0099] Ⅰ. After the second stage of training in step ③, the trained filling network can be obtained;

[0100] II. Input the gene expression data, mask matrix, data label vector and random noise obtained in step ① into the filling network to generate the filled gene expression data;

[0101] III. Convert the gene expression data of each cell from the square matrix back into a vector form and remove the extra rows added in step ①. Merge the gene expression data of all cells by column to obtain the inferred complete gene expression matrix.

[0102] Specific implementation method 1. Combination Figure 1 The present implementation method is described as a single-cell expression profile data enhancement method, model and application based on a deep generative adversarial network model, and the method specifically comprises the following steps:

[0103] Step 1: construct an input matrix based on the scRNA-seq data to be filled;

[0104] Step 2: construct a deep generative adversarial network, which includes three generative adversarial networks, each of which includes a generator and a discriminator, wherein the generator and the discriminator are both composed of a fully connected layer, a batch normalization layer, a convolutional layer, and an activation layer;

[0105] Step 3: Preprocess the data input to the discriminator and use the data set to train the constructed generative adversarial network in stages;

[0106] Step 4: Use the trained deep generative adversarial network to infer missing values ​​for the scRNA-seq data to be filled.

[0107] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that the specific process of step 1 is as follows:

[0108] Step 11: Download 10x single-cell RNA gene expression data from the CellBench dataset as the dataset to be filled;

[0109] The CellBench dataset was created by reference 1 (Luyi Tian, ​​Xueyi Dong, Saskia Freytag, Kim-Anh Cao, Shian Su, Abolfazl Jalal Abadi, Daniela Amann-Zalcenstein, Tom S Weber, Azadeh Seidi, Jafar S Jabbari, et al. Benchmarking single cell rna-sequencing analysis pipelines using mixture control experiments. Nature methods, 16(6):479–487, 2019.) Published and shared.

[0110] Step 1 and 2: Read the acquired gene expression data and convert it into behavioral genes, list them as the gene expression matrix of the cells, and extract the cell type information in the column name as the cell label information.

[0111] Step 13: perform TPM normalization on the converted gene expression data, and perform logarithmic transformation on the normalized gene expression matrix;

[0112] Step 14: Normalize the gene expression matrix converted in step 13 according to the expression value in each column, and convert the expression values ​​of all genes to the range of [0,1]. The calculation formula is:

[0113]

[0114] in, is the expression value of a column of the gene expression matrix in step 1-3, is the expression value of the corresponding column of the normalized gene expression matrix, and max() is the value of The maximum expression value in .

[0115] Step 15: Calculate the number of rows (number of genes) of the expression matrix, take its square root and round up to get an integer result, calculate the difference between the square of the integer value and the number of rows, and then fill a certain number of rows into the entire expression matrix, the number of rows is the above difference, and all fill values ​​are set to 0;

[0116] Step 16: Divide the expression matrix filled in step 15 by columns, and convert each column into a matrix form, where the length and width of the matrix are the integer results calculated in step 15;

[0117] Step 17: Use the transformed matrix of each column in step 16, set the positions greater than 0 in the matrix to 1, and the positions equal to 0 in the matrix to 0, use the processed matrix as the mask matrix of the gene expression matrix, and retain the matrix before processing.

[0118] Step 18: Encode the cell label information obtained in steps 1 and 2 and convert it into a vector form of 0, 1, 2...

[0119] The obtained gene expression matrix, mask matrix, and cell label vector are used as input to the generative adversarial network.

[0120] The other steps and parameters are the same as those in the first embodiment.

[0121] Specific implementation method three: This implementation method is different from any one of specific implementation methods one to two in that the three generative adversarial networks in step two are generative adversarial networks for generative adversarial data, generative adversarial networks for masked matrices, and padding networks, respectively. The difference between the generative adversarial network for generative adversarial data and the generative adversarial network for masked matrices is that the last activation layer of the generators of the two generative adversarial networks is different. The last activation layer of the generator of the generative adversarial network for generative adversarial data is a Sigmoid function, while the last activation layer of the generator of the masked matrix generative adversarial network is a Sigmoid function with parameters, namely:

[0122]

[0123] Where λ is an additional parameter. When λ is smaller, the function σ λ The closer (x) is to 0, the closer it is to a step function, and the more it resembles a discrete value that is either 0 or 1.

[0124] The other steps and parameters are the same as those in Specific Implementation 1 to 2.

[0125] Specific implementation method four: This implementation method is different from specific implementation methods one to three in that the discriminator structures of the three generative adversarial networks in step two are the same, and the other steps and parameters are the same as those of specific implementation methods one to three.

[0126] Specific implementation five: This implementation differs from one of specific implementations one to four in that the generator of the filling network in the three generative adversarial methods in step two is a UNet network structure, and the other steps and parameters are the same as those in one of specific implementations one to four.

[0127] Specific implementation method 6: This implementation method is different from any one of the specific implementation methods 1 to 5 in that the optimization objective function of the gene expression data generation adversarial network in step 2 is:

[0128]

[0129] Among them G x represents the generator of the gene expression data generation adversarial network, D x represents the discriminator of the gene expression data generation adversarial network, represents the set of all 1-Lipschitz functions, V x Represents the loss function of the generative adversarial network for gene expression data.

[0130] The optimization objective function of the mask matrix generation adversarial network is:

[0131]

[0132] Among them G m Denotes the generator of the mask matrix generative adversarial network, D m Denotes the discriminator of the mask matrix generative adversarial network, V m represents the loss function of the mask matrix generative adversarial network, and α represents the hyperparameter of network training.

[0133] The optimization objective function of the filling network is:

[0134]

[0135] Among them G imp represents the generator of the padding network, D imp represents the discriminator of the padding network, V imp Represents the loss function of the padding network.

[0136] The loss function V(D,G) of the three networks is:

[0137]

[0138] in is the approximate value of the Wasserstein distance between the true distribution and the generated distribution, x is the single-cell RNA gene expression data obtained in step 1, z is the input noise, G(z) is the generated single-cell RNA gene expression data, pdata(x) is the true data distribution, p z (z) is the distribution of noise z, D() is the discriminator function, For expectations; is the gradient penalty term, λ is the gradient penalty coefficient, which is a hyperparameter. For For variables The gradient of , |·|2 is the second norm. The samples obtained by linear interpolation of real samples and generated samples can be expressed as:

[0139]

[0140] where x real represents the real sample, x fakerepresents the generated sample, ∈ is a value randomly sampled from a uniform distribution. The other steps and parameters are the same as those in the first to fifth embodiments.

[0141] Specific implementation method 7: This implementation method is different from one of specific implementation methods 1 to 6 in that the matrix input to the discriminator in step 3 is preprocessed by combining the gene expression matrix and the mask matrix in the following manner:

[0142]

[0143] Where x is the gene expression matrix of a single cell processed in step ①, m is its corresponding mask matrix, and ⊙ represents the Hadamard product, that is, the two matrices are multiplied element by element. is the complement matrix corresponding to the mask matrix, and τ is the network hyperparameter.

[0144] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0145] Specific implementation eight: This implementation differs from specific implementations one to seven to one in that the specific process of the three-stage training step is as follows:

[0146] Step 3-1: During the first stage of training, the processed data set in step 1 is input into the deep generative adversarial network constructed in step 2, and the noise vector and cell label vector are respectively input into the gene expression data generator and the mask matrix generator to generate gene expression data and mask matrix, and the generated mask matrix and the real mask matrix are input into the mask matrix discriminator for discrimination, and (real gene expression data, real mask matrix) and (generated gene expression data, generated mask matrix) are processed according to the data preprocessing method in step 3 and then input into the gene expression data discriminator for discrimination. The objective functions of the gene expression data generative adversarial network and the mask matrix generative adversarial network in step 2 are continuously optimized through training.

[0147] Step 32: During the second stage of training, the parameters of the gene expression data generator trained in step 31 are fixed, and the noise vector and cell label vector are input into it to generate "real gene expression data". The gene expression data, mask matrix, cell label vector and random noise processed in step 1 are input into the generator of the filling network constructed in step 2 to obtain the generated gene expression data. The "real gene expression data" and the generated gene expression data are input into the discriminator of the filling network for discrimination, and the objective function of the filling network in step 2 is continuously optimized through training.

[0148] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0149] Specific implementation method 9: This implementation method is different from specific implementation methods 1 to 8-1 in that the value of the hyperparameter α in step 2 is between 0 and 1. The value of the hyperparameter λ is between 1 and 10.

[0150] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0151] Specific implementation method ten: This implementation method is different from specific implementation methods one to nine to one in that the value of τ in the preprocessing method in step three is between 0 and 1.

[0152] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0153] Specific implementation eleven: This implementation is different from specific implementations one to ten in that the learning rate value of the deep generative adversarial network is set at 10 -5 ~10 -3 between.

[0154] The other steps and parameters are the same as those in Specific Implementation Examples 1 to 10.

Claims

1. A single-cell expression profile data enhancement method based on a deep generative adversarial network model, characterized in that: The steps include: Step 1: Construct an input matrix based on the scRNA-seq data to be populated; Step 1-1: Obtain the gene expression matrix and cell type annotation file of the scRNA-seq dataset. The rows of the gene expression matrix represent genes and the columns represent cells. Step 1-2: Perform TPM normalization on the gene expression matrix and perform logarithmic transformation to obtain the normalized gene expression matrix; Step 1-3: Set the part of the gene expression matrix whose expression value is greater than 0 to 1, and set the part whose expression value is equal to 0 to 0, thereby obtaining the mask matrix corresponding to the gene expression matrix; Step 1-4: Encode the cell type annotation file and convert it into the form of 0, 1, 2... to obtain the cell label vector; Step 1-5: Slice the normalized gene expression matrix by column and convert each column into a square matrix; Step 1-6: Slice the mask matrix by columns and convert each column into a square matrix; The method used to convert to square matrix form is: calculate the square root of the number of rows (genes), and round the result up to an integer value; calculate the difference between the square of the integer value and the number of rows, and then add a number of rows equal to the difference in each column, and then set all the data in the added rows to 0; Step 2: construct a deep generative adversarial network, which includes three generative adversarial networks, each of which includes a generator and a discriminator, and the generator and the discriminator are both composed of a fully connected layer, a batch normalization layer, a convolutional layer, and an activation layer; The three generative adversarial networks are generative adversarial networks for ... matrices, and padding networks; The optimization objective function of the gene expression data generation network is: Among them G x represents the generator of the gene expression data generation adversarial network, D x represents the discriminator of the gene expression data generation adversarial network, represents the set of all 1-Lipschitz functions, V x represents the loss function of the generative adversarial network for gene expression data; The optimization objective function of the mask matrix generation network is: Among them G m Denotes the generator of the mask matrix generative adversarial network, D m Denotes the discriminator of the mask matrix generative adversarial network, V m represents the loss function of the mask matrix generation adversarial network, and α represents the hyperparameter of network training; The optimization objective function of the filling network is: Among them G imp represents the generator of the padding network, D imp represents the discriminator of the padding network, V imp represents the loss function of the padding network; Step 3: Preprocess the matrix of the input discriminator and use the dataset to train the constructed deep generative adversarial network in stages; The preprocessing method is to combine the gene expression matrix and the mask matrix as follows: Where x is the gene expression matrix of a single cell processed in step 1, m is its corresponding mask matrix, and ⊙ represents the Hadamard product, that is, the two matrices are multiplied element by element. is the complement matrix corresponding to the mask matrix, τ is the network hyperparameter, and the value of τ is between 0 and 1; Step 3-1: In the first training stage, the preprocessed gene expression data and mask matrix are input into the network, and the noise vector and data label vector are respectively input into the gene expression data generator and the mask matrix generator to generate gene expression data and mask matrix respectively; the generated mask matrix and the real mask matrix are input into the mask matrix discriminator for judgment, and the generated gene expression data and mask matrix are first combined according to the data preprocessing method in step 3, and then combined with the preprocessed gene expression data Input into the gene expression data discriminator for discrimination; optimize the objective function in step 2 through continuous training; Step 3-2: In the second training stage, firstly, the parameters of the gene expression data generator obtained in the first training stage are fixed, and random noise and data label vector are input into the generator to generate real gene expression data; the gene expression matrix, mask matrix, data label vector and random noise processed in step 1 are input into the filling network to generate the filled gene expression data; the data generated by the filling network and the data generated by the gene expression data generator are input into the discriminator for discrimination, and the objective function in step 2 is optimized through continuous training; Step 4: Use the trained deep generative adversarial network to infer missing values ​​for the scRNA-seq data to be filled.

2. According to claim 1, a single-cell expression profile data enhancement method based on a deep generative adversarial network model is characterized in that: Loss Function It is expressed as: in represents the expected value, pdata(x) represents the distribution of real data, and p z (z) represents the distribution of noise, λ represents the hyperparameter of network training, Representation calculation For variables The 2-norm of the gradient of .

3. According to claim 2, a single-cell expression profile data enhancement method based on a deep generative adversarial network model is characterized in that: The variable The samples obtained by linear interpolation of real samples and generated samples are expressed as: where x real represents the real sample, x fake represents the generated sample, ∈ is a value randomly sampled from a uniform distribution.

4. The single-cell expression profile data enhancement method based on a deep generative adversarial network model according to claim 3, characterized in that: The hyperparameters involved in step 3 are defined as follows: Ⅰ. The value of the hyperparameter α is between 0 and 1; Ⅱ. The value of the hyperparameter λ is between 1 and 10.