Deep Spectral Clustering of Single-Cell RNA-seq Data

In recent years, Single cell RNA sequencing (scRNA-Seq) has become widely popular in bioinformatics. Single cell RNA-seq clustering is critical for determining cell type heterogenesity at single cell level and aims to assign cells that have similar transcriptomes into the same group. Since single cell RNA sequencing data are very complex and high dimensional classical unsupervised clustering techniques may not present satisfactory biological clustering performance. In this study, we propose to use deep spectral clustering method on three publicly available scRNA datasets and compare the clustering performance of the obtained model with different classical clustering algorithms. By using Normalized Mutual Information (NMI) evaluation metric, results show that deep spectral clustering method provides accurate and improved clustering performance.

I. Introduction

Single cell RNA sequencing (scRNA) is the next generation technology to analyze cells for better understanding of genes transciptome structure. scRNA technologies have become increasingly popular in many bioinformatics studies thanks to its advantages [1]. It allows researchers describe transcriptome heterogenesity, enable to identify novel cell types and organization of cells, and help reveal diseases [2]. Clustering studies that take into account transcriptome similarities have shown successful results in the identification of cell types [3]–[4]. With clustering, cells belonging to the same cell type are grouped based on their gene expression patterns, and the gene expression patterns of the groups allow the separation of different cell types and functions. Transcriptome based cell type identification provides a consisted and unbiased approach that can be applied to any sample. Thus, it can provide comprehensive references for all cell types in an organism or tissue at various stages of development and assist in a deeper understanding of the biology and sources of disease [5].