SSRE: Cell Type Detection Based on Sparse Subspace Representation and Similarity Enhancement

Abstract

15Accurate identification of cell types from single-cell RNA sequencing (scRNA-seq) data plays 16 a critical role in a variety of scRNA-seq analysis studies. It corresponds to solving an 17 unsupervised clustering problem, in which the similarity measurement between cells in a high 18 dimensional space affects the result significantly. Although many approaches have been 19 proposed recently, the accuracy of cell type identification still needs to be improved. In this 20 study, we proposed a novel single-cell clustering framework based on similarity learning, called 21 SSRE. In SSRE, we model the relationships between cells based on subspace assumption and 22 generate a sparse representation of the cell-to-cell similarity, which retains the most similar 23 neighbors for each cell. Besides, we adopt classical pairwise similarities incorporated with a 24 gene selection and enhancement strategy to further improve the effectiveness of SSRE. For 25 performance evaluation, we applied SSRE in clustering, visualization, and other exploratory 26 data analysis processes on various scRNA-seq datasets. Experimental results show that SSRE 27 achieves superior performance in most cases compared to several state-of-the-art methods. 28 29 30 analysis [1,2]. Unlike bulk sequencing averaging the expression of mass cells, scRNA-seq 34 technique quantifies gene expression at the single cell resolution. Single cell techniques 35 promote a wide variety of biological topics such as cell heterogeneity, cell fate decisions and 36 disease pathogenesis [3-5]. Among all the applications, cell type identification plays a 37 fundamental role and its performance has a deep impact on downstream researches [6]. 38 However, identifying cell types from scRNA-seq data is still a challenging problem because of 39 the high noise rate and high dropouts, which cannot be addressed by traditional clustering 40 methods well [7]. Therefore, new efficient and reliable clustering methods for cell type 41 identification are urgent and meaningful. 42In recent studies, several novel clustering approaches for detecting cell types from scRNA-43 seq data have been proposed. Among these methods, cell types are mainly decided on the basis 44 of cell-to-cell similarity learned from scRNA-seq data. SIMLR [8] visualizes and clusters cells 45 using multi-kernel similarity learning [9] , which performs well on grouping cells. SNN-Cliq 46 [10] firstly constructs a distance matrix based on the Euclidean distance, and then introduces 47 the shared k-nearest-neighbors model to redefine the similarity. SNN-Cliq provides both the 48 estimation of cluster number and the clustering results by searching for quasi-cliques. Jiang et 49 al [11] proposed the differentiability correlation between pairs of cells instead of computing 50 primary (dis)similarity using the Pearson correlation or the Euclidean distance. RAFSIL [12] 51 divides genes into multiple clusters and concatenates the informative features from each gene 52 cluster after dimension reduction, and finally applies the ...