Quantitative assessment of single-cell RNA-sequencing methods

Abstract

Interest in single-cell whole-transcriptome analysis is growing rapidly, especially for profiling rare or heterogeneous populations of cells. We compared commercially available single-cell RNA amplification methods with both microliter and nanoliter volumes, using sequence from bulk total RNA and multiplexed quantitative PCR as benchmarks to systematically evaluate the sensitivity and accuracy of various single-cell RNA-seq approaches. We show that single-cell RNA-seq can be used to perform accurate quantitative transcriptome measurement in individual cells with a relatively small number of sequencing reads and that sequencing large numbers of single cells can recapitulate bulk transcriptome complexity.

Figure 1

Initial validation of single-cell RNA-seq methods. (a) Schematic of the experimental strategy. (b) Reproducibility, as evaluated by the percentage of genes detected in pairs of replicate samples out of the mean total number of genes detected in this pair of samples. (c) Sensitivity, as evaluated by overlap between genes detected by single-cell and bulk RNA-seq measurement. Bulk values listed exclude the overlap values. Percentages are calculated as the number of genes detected in both relative to the number of genes detected in the bulk measurement.

Figure 2

Correlation between single-cell RNA-seq and single-cell multiplexed qPCR for each sample preparation method. Correlation coefficients were computed from log₂- transformed values. A linear regression line (color) and the y = x line (black, dotted) are also shown in each panel. Shading represents the 95% confidence interval for each regression line. For RNA-seq data, FPKM values for each of gene of interest were normalized to the median FPKM for each cell and log₂ transformed. For qPCR data, threshold cycle (Ct) values for each gene of interest were normalized to the median Ct value for each cell (ΔCt), which also represents a fold change over the median expression. ΔCt values are already in log₂ space and were directly plotted. Error bars, standard error (n values as in Fig. 1a).

Figure 3

Comparison of gene expression distributions for 40 genes between samples prepared in microliter and nanoliter volumes. (a) Frequency distribution of expression values from single-cell qPCR shown as a violin plot for each gene. Expression (vertical axis) is the log₂-transformed fold change over median gene expression level for each sample. Width of the violin indicates frequency at that expression level. (b) Frequency distribution of expression from single-cell RNA-seq. Violin plots are presented as in a.

Figure 4

Merging many single-cell transcriptomes quantitatively recreates the bulk measurement. (a) Correlation between the merged single cells (“ensemble”) and the bulk RNA-seq measurement of gene expression. The ensemble was created by computationally pooling all the raw reads obtained from the 96 single-cell transcriptomes generated using the C1 system and then sampling 30 million reads randomly. The bulk and ensemble libraries were depth matched before alignment was performed. For each gene, the log₂-transformed median FPKM values from the ensemble and bulk were plotted against each other. (b) Variation in gene expression as a function of gene expression level across sample replicates for each preparation method. Variation (vertical axis) is the median absolute deviation of the FPKM divided by the median FPKM (MAD/M; see Online Methods for the equation). For each gene, the MAD/M is plotted against the log₂-transformed median FPKM value for that gene in order to visualize how variation changes with overall transcript abundance. Replicates for single-cell methods are biological replicates, whereas replicates for the bulk and ensemble are technical replicates, as each sample represents a subsampling of a pooled sample.

Figure 5

Saturation curves for the different sample preparation methods. Each point on the curve was generated by randomly selecting a number of raw reads from each sample library and then using the same alignment pipeline to call genes with mean FPKM >1. Each point represents four replicate subsamplings. Error bars, standard error.