Abstract
Single-cell RNA-seq data analysis generally requires quality control, normalization, highly variable genes screening, dimensionality reduction and clustering. Among these processes, downstream analysis including dimensionality reduction and clustering are sensitive to the selection of highly variable genes. Though increasing number of tools for selecting the highly variable genes have been developed, an evaluation of their performances and a general strategy are lack. Here, we compare the performance of nine commonly used methods for screening variable genes by using single-cell RNA-seq data from hematopoietic stem/progenitor cells and mature blood cells, and find that SCHS outperforms other methods regarding to reproducibility and accuracy. However, this method prefers the selection of highly expressed genes. We further propose a new strategy SIEVE (SIngle-cEll Variable gEnes) by multiple rounds of random sampling, therefore minimizing the stochastic noise and identifying a robust set of variable genes. Moreover, SIEVE recovers lowly expressed genes as variable genes and substantially improves the accuracy of single cell classification, especially for the methods with lower reproducibility. The SIEVE software is freely available at https://github.com/YinanZhang522/SIEVE.