Diffusion Kernel based Fast Adaptive Clustering of Single Cell RNA-seq Data

Abstract

Recently, with the advent of high throughput single-cell technologies, it has become possible to quantify the whole transcriptome of individual cells; however, it remains challenging to discover intrinsic rare cell-types from high throughput genes expression data. To overcome this challenge, various unsupervised clustering based approaches have been proposed such as GiniClust, SC3 and SIMLR clustering. These approaches identified appropriate rare cell types based on the ambiguous parametric settings and employed clustering algorithms are inefficient to discover meaningful clusters adaptively. However, the appropriate signal of interest can be observed along with the robust filtration, normalization, and transformation of raw count samples of single-cell data. Filtration, normalization, and transformation have become the essential primary procedure for down-stream analysis of single-cell data and to eliminate the risk of biological variation and technical noise. In this paper, we will evaluate the various methods to detect the rare cell-types from a large population and develop fast novel diffusion kernel based unsupervised framework (DKBUF) to identify rare cell types from single-cell RNA-seq data, more in an adaptive and attractive fashion. The DKBUF filters the non-stable genes, normalizes the genes, attractively detects subpopulations within single-cell datasets, and visualizes the discovered distinct subpopulations. Extensive experiments on single-cell datasets and comparisons with state-of-the-art methods validate the robustness of the DKBUF.