Sphinx: modeling transcriptional heterogeneity in single-cell RNA-Seq

The significance of single-cell transcription resides not only in the cumulative expression strength of the cell population but also in its heterogeneity. We propose a new model that improves the detection of changes in the transcriptional heterogeneity pattern of RNA-Seq data using two heterogeneity parameters: ‘burst proportion’ and ‘burst magnitude’, whose changes are validated using RNA-FISH. Transcriptional ‘co-bursting’ – governed by distinct mechanisms during myoblast proliferation and differentiation – is described here. Advances in single-cell RNA-Seq technology have promoted in-depth investigation of heterogeneous gene expression at individual cell resolution . Single-cell RNA-Seq data exhibit significantly greater variability (i.e., larger overdispersion) than bulk-cell RNA-Seq data. We examined the read counts in two bulk-cell RNA-Seq datasets 4 and two single-cell RNA-Seq datasets . We observed that the estimated overdispersion parameters of single-cell data were typically greater than those from bulk datasets by orders of magnitude (Fig. 1a). Substantial variability of single-cell RNA-Seq data is due to various biological and technical aspects, including transcriptional stochasticity, cellular heterogeneity, and technical noise, among others. Of these aspects, the first two cannot be investigated through bulk-cell technologies. Mammalian gene transcription can be classified into two schemes called constitutive expression and stochastic ‘bursty’ expression , which lead to distinct transcriptional kinetic patterns (Supplementary Fig. 1). In a previous study of mouse embryonic development, transcriptional bursting is believed to be the key factor that contributes to the rapid expression dynamics observed in single-cell RNA-Seq data. Besides gene bursting, differences in cellular subpopulations also give rise to additional variance beyond what is observed in bulk-cell RNASeq data. Cells that undergo cellular processes such as differentiation of myoblasts also show high variability in gene expression between individual cells. Technical variability is another factor that contributes to large overdispersion of single-cell RNA-Seq data. Unique variability in single-cell RNA-Seq has resulted in bimodal distribution of sequencing reads that is not observed in bulk-cell data. Thus, a gene’s expression is detected only in a sub-population of cells. Methods have been proposed to analyze single-cell RNA-Seq data. The Poisson-Beta model was previously developed to model all theoretical kinetics for ‘bursty’ gene expression. However, in the presence of massive variability, fitting of the Poisson-Beta model is compromised by excessive overdispersion in read counts (Supplementary Results R1). Kharchenko et al. proposed the SCDE method to model extreme data points in single-cell count data as drop-out peer-reviewed) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprint (which was not . http://dx.doi.org/10.1101/027870 doi: bioRxiv preprint first posted online Oct. 1, 2015;