ScRNA-seq missing data prediction method based on bidirectional auto-encoder

Abstract

The invention discloses an scRNA-seq missing data prediction method based on a bidirectional auto-encoder. The scRNA-seq missing data prediction method comprises the following steps: 1, acquiring an scRNA-seq data matrix; 2, carrying out normalization on the scRNA-seq matrix; and 3, respectively learning similarity characteristics between cells and between genes by using a row auto-encoder and a column auto-encoder, and respectively reconstructing data by using the learned characteristics. And 4, introducing a cross point loss function to restrain a cross point of reconstruction data of the two auto-encoders, and realizing information sharing between cells and between genes by sharing weights. And 5, predicting scRNA-seq data by using features learned by the row auto-encoder and the column auto-encoder, and taking a mean value of the scRNA-seq data as a final predicted value. Meanwhile, non-zero values in the original scRNA-seq data are identified and are regarded as real expression values, and only zero values are filled. According to the method, the missing value can be well recovered, the clustering of the cell subpopulation is promoted, the identification of the marker gene is perfected, and the inference of the cell development trajectory is facilitated.

Description

Technical Field

[0001] The present invention relates to the field of single-cell RNA sequencing missing data prediction in bioinformatics, and specifically to a scRNA-seq missing data prediction method based on a bidirectional autoencoder. Background Art

[0002] Single-cell RNA sequencing (scRNA-seq) is a revolutionary technology that can characterize the transcriptome of individual cells on a genome-wide scale. This technology can explore the heterogeneity and dynamics of cells and provide cell type-specific gene expression signals and trajectory inferences, which are not possible with traditional bulk RNA-seq technologies. Currently, scRNA-seq has become a powerful tool for biomedical research and has been widely used in fields such as cancer biology, immunotherapy, and drug development. However, a major challenge facing scRNA-seq data analysis is the occurrence of missing events, that is, due to low RNA capture efficiency, some expressed transcripts are mistakenly recorded as zero. The presence of missing values ​​in gene expression data poses a considerable obstacle to downstream analysis of scRNA-seq datasets. Therefore, it is particularly important to develop effective methods to compensate for these missing values.

[0003] In recent years, a variety of interpolation methods have emerged, which can be mainly divided into four categories. The first category is to use smoothing or data diffusion to restore missing values. For example, MAGIC uses a Markov matrix constructed by cell affinity to diffuse data, propagating information between similar cells to interpolate missing values. However, accurately determining the diffusion time is a challenge faced by MAGIC, which may lead to overestimation of missing values. DrImpute uses a clustering method for estimation, averaging the gene expression values ​​of similar cells determined by clustering, and obtaining the final estimate by averaging multiple clustering estimates. However, multiple clustering makes the DrImpute estimation process time-consuming and requires a lot of memory.

[0004] The second category applies probability models to directly calculate missing values ​​in the data. For example, scImpute uses a gamma-Poisson mixture model to estimate the probability of missing each gene in each cell, and then selects genes with low probability of missing in similar cells to estimate missing data. BayNorm uses a binomial model based on the mRNA capture mechanism to infer priors through the empirical Bayesian method and estimates gene expression levels using cross-cell expression data. However, most of these methods rely on assumptions about the relationship between cells, which may limit their performance when fewer cell types are involved.

[0005] The third category uses matrix decomposition techniques to estimate missing data. ALRA uses the non-negative property of the expression matrix and its inherent correlation structure to estimate the zeros in the expression matrix using the singular value decomposition (SVD) technique. However, ALRA can usually only capture linear relationships in the original expression matrix.

[0006] The fourth category uses deep learning models for estimation. For example, deepImpute uses a divide-and-conquer strategy to train sub-neural networks, using dropout layers and loss functions to identify patterns in the data and estimate missing values. However, deepImpute focuses mainly on learning the similarity of genes and may not be able to fully capture biological information at the cellular level. GE-Impute uses Euclidean distance to construct a similarity network between primitive cells, and generates an embedding matrix for each cell by simulating a random walk of a fixed length. It then reconstructs the similarity network and estimates the missing data of each cell by averaging the expression values ​​of neighboring cells. However, GE-Impute relies on the average expression values ​​of neighboring cells and may not be able to fully capture all biological changes.

[0007] In view of this, it is of great significance to study the interpolation method of missing data in scRNA-seq. The present invention proposes a method for interpolating missing data in scRNA-seq based on a bidirectional autoencoder. Summary of the invention

[0008] The present invention proposes a scRNA-seq missing data prediction method based on a bidirectional autoencoder, the main steps of which are as follows:

[0009] Step 1: Obtain scRNA-seq matrix.

[0010] scRNA-seq was obtained from the NCBI GEO database and named Where X represents the scRNA-seq data, m and n represent the number of cells and genes, respectively.

[0011] Step 2: Normalize the scRNA-seq matrix.

[0012] In order to reduce the technical deviation introduced in the sequencing process, the present invention performs minimum and maximum normalization preprocessing on the data set. ij Subtract the minimum expression value of the gene in the cell Then divide it by the difference between the maximum gene expression value and the minimum gene expression value in the cell. Between 0 and 1, where i represents different cells (from 1 to m) and j represents different genes (from 1 to n).

[0013] Step 3: Use row autoencoders and column autoencoders to learn the similarity features between cells and genes respectively, and use the learned features to reconstruct the data respectively.

[0014] Autoencoders are an unsupervised learning paradigm in neural network architectures, and have shown wide applicability in feature dimension reduction and data interpolation tasks. To this end, the present invention uses row autoencoders to learn the similarity features between cells and uses column autoencoders to learn the similarity features between genes. By introducing the row autoencoder loss function Loss row And the column autoencoder loss function Loss col Minimize the error between the reconstructed data and the original data to achieve effective feature extraction by the autoencoder.

[0015] Step 4: Introduce the intersection loss function to constrain the intersection of the two autoencoders to reconstruct the data, and realize information sharing between cells and cells and genes by sharing weights.

[0016] Since the two autoencoders process the same data matrix, the present invention introduces a third loss function Loss cross , that is, the intersection loss function, constrains the reconstructed data output by the two autoencoders to ensure their consistency at the intersection.

[0017] Step 5: Use the features learned by the row autoencoder and column autoencoder to predict the scRNA-seq data respectively, and take the average of the two as the final prediction value. At the same time, identify the non-zero values ​​in the original scRNA-seq data, regard them as the true expression values, and only fill in the zero values.

[0018] The predictions from the row and column autoencoders are summed and averaged, and only zero values ​​in the original scRNA-seq are filled in. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0020] Figure 2 This is a comparison chart of the interpolation performance of the present invention and other interpolation methods.

[0021] Figure 3 This is a cluster diagram of cell subpopulations before and after interpolation of the Zeisel data set according to the present invention.

[0022] Figure 4 This is a marker gene identification diagram of the present invention before and after interpolation of the Zeisel data set.

[0023] Figure 5This is the cell development trajectory inference diagram of the present invention before and after interpolation of the Klein dataset. DETAILED DESCRIPTION

[0024] The present invention is a method for predicting missing data of scRNA-seq based on a bidirectional autoencoder. The present invention is further described in detail below in combination with specific implementation methods and simulation experiments. It should be understood by professionals in the field that these implementation methods are only used to explain the technical principles of the present invention and are not intended to limit the scope of evidence collection of the present invention.

[0025] like Figure 1 As shown, a scRNA-seq missing data prediction method based on a bidirectional autoencoder specifically includes the following steps:

[0026] Preferably, the step of obtaining the scRNA-seq association matrix in step 1 is specifically as follows:

[0027] Four publicly available real datasets, Zeisel (GSE60361), Romanov (GSE74672), Usoskin (GSE59739), and Kelin (GSE65525), were obtained from the NCBI GEO database. , where X represents the scRNA-seq data, m and n represent the number of cells and genes, respectively.

[0028] Preferably, the normalized scRNA-seq matrix described in step 2 is specifically:

[0029] Use the minimum and maximum normalization preprocessing method. ij Subtract the minimum expression value of the gene in the cell Then divide it by the difference between the maximum and minimum gene expression values ​​in that cell.

[0030]

[0031] in, represents the original gene expression matrix consisting of m cells and n genes, X ij represents the expression value of the jth gene in the ith cell of the original data matrix, represents the minimum gene expression value in the ith cell, represents the maximum gene expression value in the ith cell, Represents the normalized gene expression matrix.

[0032] Preferably, the row autoencoder and the column autoencoder described in step 3 are used to learn the similarity features between cells and between genes, respectively, and the learned features are used to reconstruct the data, specifically:

[0033] A single cell row is fed into a row-wise autoencoder, and a single gene column is fed into a column-wise autoencoder. Subsequently, the row-wise encoder and column-wise encoder map the inputs into their respective latent spaces.

[0034]

[0035] in represents the normalized expression value of cell i, represents the normalized expression value of gene j, and represents the parameters of the two fully connected layers of the row encoder, represents the parameters of the two fully connected layers of the column autoencoder, RELU represents the nonlinear activation function, H (row) and H (col) Represent the outputs of the row encoder and column encoder respectively.

[0036] Then, the row decoder and column decoder reconstruct the data respectively. Specifically, they convert the data (H (row) and H (col) ) maps from the latent space dimensions back to the input space dimensions.

[0037]

[0038] in, represents the parameters of the two fully connected layers of the row decoder, represents the parameters of the two fully connected layers of the column decoder, represents the reconstructed data of cell i by the row decoder, Represents the reconstructed data of gene j by the column decoder.

[0039] Preferably, the intersection loss function introduced in step 4 constrains the intersection of the data reconstructed by the two autoencoders, and realizes information sharing between cells and cells and between genes and genes by sharing weights, specifically:

[0040] Three different loss functions are defined for specific tasks of row autoencoders and column autoencoders (Loss row ,Loss col , and Loss cross ), Loss row and Loss col They are responsible for the reconstruction loss in the row autoencoder and column autoencoder respectively. At the same time, a third loss function Loss is introduced crossResponsible for calculating the loss of the two autoencoder output matrices at the intersection point, thereby ensuring the consistency and equality of the intersection points in the output matrices.

[0041]

[0042] in, Represents the normalized matrix X (n) The mask matrix of Represent the expression value of gene j in cell i reconstructed by row autoencoder and column autoencoder respectively.

[0043] Preferably, the features learned by the row autoencoder and the column autoencoder described in step 5 are used to predict the scRNA-seq data respectively, and the average of the two is taken as the final prediction value. Specifically:

[0044] The features learned by the row autoencoder and column autoencoder are used to predict the scRNA-seq data respectively, and the average of the two is taken as the final prediction value. At the same time, the non-zero values ​​in the original scRNA-seq data are identified and regarded as the true expression values, and only zero values ​​are filled.

[0045]

[0046] in, represents the final interpolated value of gene j in cell i.

[0047] The technical effects of the present invention are further illustrated by experimental verification as follows:

[0048] 1. Experimental conditions and contents:

[0049] The experiments of the present invention were all completed on a 2.50 GHz CPU, NVIDIA GeForce RTX 4060 and Windows 11 operating system.

[0050] 2. Experimental results analysis:

[0051] Evaluations on four real scRNA-seq datasets consistently show that the proposed imputation method exhibits superior performance compared to existing imputation methods, which can well recover missing values, promote clustering of cell subpopulations, improve the identification of marker genes, and facilitate the inference of cell developmental trajectories.

[0052] The experimental comparison results of the present invention are as follows Figure 2-5 shown.

[0053] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in the field, understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention, which should be included in the protection scope of the claims of the present invention.

Claims

1. A method for predicting missing data in scRNA-seq based on a bidirectional autoencoder, characterized in that: The following steps are involved: Step 1: Obtain scRNA-seq matrix; Step 2: Normalize the scRNA-seq matrix; Step 3: Use row autoencoders and column autoencoders to learn the similarity features between cells and genes respectively, and use the learned features to reconstruct the data respectively; Step 4: Introduce the intersection loss function to constrain the intersection of the two autoencoders to reconstruct the data, and realize information sharing between cells and cells and genes by sharing weights; Step 5: Use the features learned by the row autoencoder and column autoencoder to predict the scRNA-seq data respectively, and take the average of the two as the final prediction value; at the same time, identify the non-zero values ​​in the original scRNA-seq data, regard them as the true expression values, and only fill in the zero values.

2. The scRNA-seq missing data prediction method based on a bidirectional autoencoder according to claim 1, characterized in that: In step 1, specifically: scRNA-seq was obtained from the NCBI GEO database and named ,in represents scRNA-seq data, and represent the number of cells and genes, respectively.

3. The scRNA-seq missing data prediction method based on a bidirectional autoencoder according to claim 1, characterized in that: In step 2, specifically: In order to reduce the technical deviation introduced in the sequencing process, the present invention performs minimum and maximum normalization preprocessing on the data set; each expression value in the data set is Subtract the minimum expression value of the gene in the cell , and then divided by the difference between the maximum gene expression value and the minimum gene expression value in the cell; the normalized expression value Between 0 and 1, where Represents different cells (from 1 to ), j represents different genes (from 1 to ).

4. The scRNA-seq missing data prediction method based on a bidirectional autoencoder according to claim 1, characterized in that: In step 3, specifically: Autoencoders are an unsupervised learning paradigm in neural network architectures, and have shown extensive practicality in feature dimension reduction and data interpolation tasks. To this end, the present invention uses row autoencoders to learn similarity features between cells, and uses column autoencoders to learn similarity features between genes. By introducing the row autoencoder loss function And the column autoencoder loss function Minimize the error between the reconstructed data and the original data to achieve effective feature extraction by the autoencoder.

5. The scRNA-seq missing data prediction method based on a bidirectional autoencoder according to claim 1, characterized in that: In step 4, specifically: Since the two autoencoders process the same data matrix, the present invention introduces a third loss function , that is, the intersection loss function, constrains the reconstructed data output by the two autoencoders to ensure their consistency at the intersection.

6. The scRNA-seq missing data prediction method based on a bidirectional autoencoder according to claim 1, characterized in that: In step 5, specifically: The predictions from the row and column autoencoders are summed and averaged, and only zero values ​​in the original scRNA-seq are filled in.

Images