Research On Imputation And Differentiation Trajectory Inference Methods For Single-cell Transcriptome Sequencing Data

Single-cell RNA sequencing technology can obtain differences in gene expression of cells in specific microenvironments,which facilitates the study of functional differences.By analyzing the state changes of single-cell sequencing data presented in biological processes,we can better understand cell differentiation trajectories,developmental cycles,and other cellular processes.Inferring the pseudotime differentiation trajectory of cells based on single-cell transcriptome sequencing data has become an important task in studying different states transitions between cells.However,due to the high dimensionality,missing events,and high noise of data generated by current single-cell sequencing technology,how to improve the reliability of data through single-cell data imputation algorithms and improve the accuracy of trajectory inference algorithms is still a hot research topic.To address these issues,this article mainly focuses on the following two aspects of research:Firstly,this paper proposes a single-cell sequencing data imputation algorithm,VAAI,based on a weighted variational autoencoder to complete missing data in the sc RNA-seq process.Considering the high dimensionality and sparsity of sc RNA-seq data,this study introduces a variational autoencoder model based on self-attention mechanism.It combines the ideas of autoencoders and probabilistic graph models to simulate the Gaussian distribution of single-cell sequencing data and weight the latent variables sampled by self-attention mechanism,enabling dimensionality reduction,compression,reconstruction,and generation of high-dimensional data.VAAI mainly maps the input data to the probability distribution in a low-dimensional space through a multilayer perceptron,and samples from the probability distribution as latent variables.Through the self-attention mechanism,important features in the latent variables are weighted and represented,and a multilayer perceptron is used to map the weighted latent variables to the reconstructed data space.This paper uses the VAAI model to learn the dependency structure between high-dimensional gene features,predict sc RNA-seq data,and complete missing values to ensure the integrity of the sc RNA-seq data.The accuracy of the VAAI algorithm is validated on five real sc RNAseq datasets,and the reliability of the imputation results is verified by downstream analysis.Compared to other imputation algorithms,VAAI can more accurately recover missing gene expression values in the sc RNA-seq process.This passage describes the development of an algorithm called MARV,which uses RNA velocity to infer the pseudotime trajectory of cell differentiation.Current highdimensional RNA velocity trajectory inference algorithms are often unstable or inaccurate due to the high noise and sparsity of the data.To address this,the authors propose a low-dimensional feature representation algorithm based on a multi-head attention neural network to estimate RNA velocity.First,context information from different expression spaces,such as spliced mRNA and unspliced mRNA count matrices,is learned using multi-head attention mechanism and latent features are extracted through multiple perception layers.Then,gene expression matrix,spliced mRNA expression matrix,and unspliced mRNA expression matrix are stacked horizontally and projected onto a low-dimensional space using PCA algorithm.Finally,the above information features are fused using attention mechanism to reconstruct the gene expression matrix.The entire network architecture uses a neural network with multi-head attention mechanism to extract the low-dimensional feature representation for calculating RNA velocity and inferring the differentiation direction of cell clusters in the next moment.The authors demonstrate the accuracy and differential differentiation directions among different cell clusters of MARV algorithm by visualizing the cell differentiation trajectory results.The MARV algorithm is validated on four real single-cell sequencing datasets with cell differentiation labels and is compared to other competing algorithms,showing that MARV can infer cell differentiation trajectories more accurately on multiple metrics than other algorithms.