Figure 1. Modeling single-cell RNA-seq measurement as a mixture of two processes.
a. Types of cell-to-cell variability observed in single-cell RNA-seq measurements. A smoothed scatter plot compares gene expression estimates from two cells of the same type (MEF cells), illustrating prevalence of drop-out events, over-dispersion, and high-magnitude outliers.
b. Single-cell variability throws off standard RNA-seq analysis methods, with top differentially expressed genes influenced by difference in drop-out (Rnaseh2a) or outlier (Bmp4) events. The examples are taken from CuffDiff214 comparison of 10 ESC and 10 MEF cells, with triangles showing expression magnitudes observed in different cells, and whiskers spanning the range of observed expression magnitudes.
c. To identify a reliable set of genes for fitting model parameters, our approach initially uses cross-comparison of single-cell measurements (using cells of the same type, e.g. MEF), determining whether the transcript is likely to have been successfully amplified in both experiments (correlated component). The true expression magnitude of such genes is estimated as a median expression level across cells in which the gene appears in a correlated component.
d. Each single-cell measurement is modeled as a mixture of drop-out and successful amplification processes. The parameters of the distributions and the magnitude-dependent mixing of the two processes are determined based on the expected population expression averages of genes appearing in many correlated components (c.).
e. Drop-out rates vary between different cell types. The rate of transcript detection failures (drop-out events) depends on the average expression magnitude of a gene in the cell population, and varies among the cells. In Islam et al. dataset2, higher drop-out frequencies are observed for mouse ES cells compared to MEF cells.
f. Drop-out rates for 4, 8 and 16-cell embryo samples examined by Deng et al.12 using a recently-developed protocol also show systematic differences.