Single-cell RNA Sequencing Data Reconstruction Based On Compressed Sensing

Single-cell RNA sequencing(scRNA-seq)technology is a weapon for exploring the relationship between genes and phenotypes,and it has become one of the most important technologies in biomedical research,However,there are lots of factors that can bring about noise in scRNA-seq data,such as Dropout phenomenon.These noises may hinder the downstream analysis of single-cell sequencing,such as cell clustering.The researchers propose a variety of calculation methods to fill or complete single-cell sequencing data to lessen data noise.Single-cell RNA sequencing data will be extremely sparse due to the Dropout phenomenon etc.Based on the above issues,this paper proposes a computational framework based on compressed sensing to solve the quality problem of single-cell RNA sequencing data.Compressed Sensing(CS)can reconstruct the original data with a high probability utilizing a few observed data when data is sparse.This thesis mainly studies the filling method based on Compressed Sensing(CS)for single-cell RNA sequencing data,and makes the following two contributions:The first one is to propose a reconstruction model based on Compressed Sensing(Compressed Sensing Recovery(CSR)),which can be used to fill or restore single-cell RNA sequencing data.Firstly,CSR uses the K-Singular Value Decomposition(K-SVD)algorithm to calculate the dictionary matrix of scRNA-seq data.The dictionary matrix contains the important feature information of the original data,and the observed data can be further sparse,which make the compressed sensing has a better result.Secondly,a proper observation matrix is designed to observe the original data set,and the Orthogonal Matching Pursuit(OMP)algorithm is used for reconstruction of data.Finally,the CSR model is compared with other filling models through experiments,and the results prove the effectiveness of CSR model.The second one is to put forward the Network Recover(NR)model,which makes use of the Fully Connected Neural Network(FCNN)to revise the Negative value and part of zero value of the data matrix after recovering.Because there is no negative value of mRNA,this thesis assumes that there is a nonlinear relationship between cells and genes.The mapping relationship between cells and genes is calculated by using the FCNN,and then the negative values and some zero values are reasonably corrected.Finally,the CSR model is used to fill the data.The experimental results show that the NR method can effectively improve the performance of CSR.To sum up,in order to solve the data noise problem of single-cell RNA sequencing data,this thesis builds a CSR data filling model based on the compressed sensing framework,and constructs a Network Recovery model to solve the negative value problem in the CSR model.The experiment shows the effectiveness of the model.