Transcriptome-wide Cell-to-Cell Variation Is Similar across the Three Culture Conditions

An advantage of the single-cell approach is that we can study the distribution of expression levels across the population, thereby capturing cell-to-cell variability in gene expression (Figure 2A). To compare global levels of gene expression heterogeneity between the three different culture conditions, we used the coefficient of variation (CV) of normalized read counts (Figure S2). However, the CV of a gene depends strongly on its mean expression level and length, making it difficult to interpret differences between conditions. To account for the confounding factor of expression level, we therefore developed a measure of cell-to-cell variation by calculating the distance between the squared CV of each gene and a running median (Figures S2E and S2F). This is derived from the scatterplot of the mean normalized read counts versus the squared CV values, as in (Newman et al., 2006). We refer to this expression-level normalized measure of gene expression heterogeneity as distance to the median (DM) (refer to Supplemental Experimental Procedures for details).

Given the heterogeneous morphology of mESCs cultured in serum (Marks et al., 2012; Toyooka et al., 2008), as well as the heterogeneous expression of pluripotency factors (Canham et al., 2010; Hayashi et al., 2008; Kalmar et al., 2009; Singh et al., 2007), it was surprising that transcriptome-wide DM values are not significantly different across the three culture conditions (p = 0.625 by the Freidman rank sum test) (Figures S2B–S2D).

This prompted us to ask whether the levels of heterogeneity for genes belonging to individual functional categories are also consistent between conditions. We first performed gene set enrichment analysis for each culture condition to test whether genes belonging to Gene Ontology (GO) terms are enriched among those genes with extreme DM values. We observed that genes involved in translation, ribosome, RNA binding, structural molecule activity, and mRNA processing have a lower level of gene expression heterogeneity for all conditions (Figure S3D). In contrast, genes involved in plasma membrane, metal ion binding, lysosome, and integral component of membrane exhibit higher variation than expected by chance in all three conditions (p < 10⁻⁴). To gain more insight into how gene expression heterogeneity for functional categories differs between culture conditions, we compared the DM values of genes in pairs of culture conditions for each GO term (excluding 2i replicates containing 2C-like cells; for discussion of 2C-like cells, see below) (Figures 2B–2D, Figures S3A and S3B). We found that 712 GO terms (out of a total of 19,107 terms) exhibit a significant difference in levels of gene expression heterogeneity in at least one pairwise comparison (p < 0.01). For example, the expression of genes involved in “organ development” (p = 3.3 × 10⁻⁴) and “cell adhesion” (p = 4.8 × 10⁻⁴) is more heterogeneous in serum than in the inhibitory conditions (2i and a2i): these terms contain many pluripotency factors.

In contrast, genes involved in “cell cycle” (p = 5.4 × 10⁻³) and “nuclear division” (p = 5.9 × 10⁻⁶) have higher levels of gene expression heterogeneity in 2i compared to serum (Figures 2B–2D, Figures S3A and S3B). When we included 2i replicates containing 2C-like cells, we observed a similar trend (Figure S3C). When clustering cells based on cell-cycle genes only, we found that 2i cells separate into two groups: one with high expression of G2 and M genes and a second with lower expression of these genes (Figures 2E and 2F). Cells in serum and 2i also show different doubling kinetics with a rapid initial growth rate in 2i (24 hr). At the time of harvest, however (48 hr after plating), the doubling time of cells in 2i is 25 hr and in serum it is 11 hr, indicating that cells grown in 2i cycle more slowly, probably due to a longer G1 phase (refer to Supplemental Experimental Procedures).

As an independent validation, we performed the same analysis using data published previously (Grün et al., 2014). Consistent with our observations, global levels of gene expression heterogeneity between cells grown in 2i and in serum were comparable, while GO categories for development and differentiation were more heterogeneous in serum than in 2i, and cell-cycle genes were more heterogeneous in 2i than in serum (Figures S3E–S3G, Supplemental Experimental Procedures, Table S2).

Three Subpopulations Can Be Delineated in Serum-Grown mESCs

Genes with heterogeneous expression, especially those with clear bimodal expression (Figure 2A), may indicate the existence of underlying subpopulations. Indeed, hierarchical clustering of subsets of known pluripotency genes and differentiation markers reveals that serum-grown cells split into three distinct groups (Figure 3A). Similar to others, we found heterogeneous expression of Nanog (Faddah et al., 2013; Kalmar et al., 2009; MacArthur et al., 2012; Singh et al., 2007), Esrrb (van den Berg et al., 2008), and Zfp42 (Toyooka et al., 2008) in serum, as well as heterogeneous expression of Nr0b1 and Utf1.

One subpopulation consists of 39 cells (15%) that express higher levels of markers of differentiation, for example Fos or Hes1, and high levels of cytoskeletal genes such as keratins (Krt8 and Krt18), actins (Acta1 and Acta2), and annexins (Anxa1, Anxa2, and Anxa3). At the same time, these 39 cells have low levels or no expression of transcription factors involved in the maintenance of pluripotency (e.g., Nanog, Sox2, and Oct4) (Figures 3B and 3C), suggesting that these cells have exited pluripotency and committed to differentiation. A second group consists of 42 cells (17%) with somewhat lower expression levels of some pluripotency genes, such as Dppa3 and Nanog, and some expression of differentiation genes, yet high expression of Oct4 and Sox2. These cells may correspond to a previously described “differentiation permissive” set (Chambers et al., 2007; Islam et al., 2014; Kalmar et al., 2009). The largest group, which consists of 169 cells (68%), expresses the highest levels of pluripotency factors and exhibits very low expression of keratins or actins (Figures 3B and 3C).

We observe that the 39-cell and 42-cell populations, which have begun to move forward on the differentiation pathway, have heterogeneous expression of cell-cycle genes (Figure 3D). A shift in the distribution of the expression of G2/M genes, such as Cks2 or Cdc20, toward lower levels suggests that there are relatively more G1/S cells in these two groups as well. We inferred that more differentiated cells have a relatively longer G1 phase, as we sample more cells in G1 from this subpopulation relative to more pluripotent cells. This indicates that the 39-cell and 42-cell subsets that we identified proliferate more slowly than Nanog-high ground state pluripotent cells (Figure 3D). Moreover, we performed principal component analysis (PCA) of our data together with cells from an mESC-to-NPC (neural progenitor cell) differentiation time course (Bibel et al., 2007). We observed that cells belonging to the differentiating subpopulation overlap with cells that are differentiating toward NPCs (Figure S4). This strongly supports our earlier hypothesis that these cells are indeed progressing down a differentiation pathway (Figure 3A).

Ground State mESCs Cultured in Different Media Have Non-Overlapping Transcriptomes

Kalmar et al. (2009) suggested that mESCs grown in 2i are similar, or potentially identical, to the Nanog-high mESC subpopulation cultured in serum. To investigate whether ground state mESCs in serum (i.e. the cells we identified as most pluripotent within serum-only media) have a similar transcriptome to 2i or a2i mESCs, we clustered each population based on their global expression profiles. PCA (Figure 4A) demonstrates three separate clusters, revealing that ground state mESCs grown under different culture conditions in fact have distinct transcriptome identities. This is consistent with observations comparing bulk RNA-sequencing of Rex1-high (Zfp42-high) cells in serum and cells in 2i (Marks et al., 2012).

The Spearman correlation coefficient of mean gene expression levels between cells grown in the two inhibitory conditions is 0.95; between 2i and serum, it’s 0.88; and between a2i and serum, it’s 0.91. While these results suggest that 2i and a2i cells are more similar to each other than to serum-cultured cells, differences do exist between the two populations. Differences between 2i and a2i arise from the use of different inhibitors (Shimizu et al., 2012). Inhibition of Mek1/2 results in dephosphorylation of Erk1/2, while inhibition of Src does not have this effect, as we show by western blotting (Figure S1K).

To examine what differences in gene expression between the culture conditions explain the separation into distinct clusters, we performed GO enrichment analysis. We found that genes involved in development and differentiation, MAPK signaling, and basic metabolism are responsible for the separation (Figures 4B and 4C). To identify specific genes, we used DESeq, where each cell was considered a replicate of its culture condition, to test for significant differences in expression (Anders and Huber, 2010) as described in the Supplemental Experimental Procedures. There is a substantial amount of technical noise in single-cell data, so we considered only genes that are expressed on average above 50 normalized counts, as the technical bias is most pronounced for lowly expressed genes (Brennecke et al., 2013). This results in 4,587 differentially expressed genes between serum-grown cells and 2i-grown cells, 3,056 between serum and alternative 2i, and 2,061 genes between the two inhibitory conditions (the list of DE genes is available at http://www.ebi.ac.uk/teichmann-srv/espresso; Figure S5).

The two most enriched GO categories in genes differentially expressed between serum and 2i are in utero embryonic development (GO:0001701) and positive regulation of transcription from RNA polymerase II promoter (GO:0045944) (Table S1). Many of the transcription factors in the latter are key genes involved in pluripotency, such as Nanog, Oct4, Klf4, and Sox2. The differences between the two inhibitory conditions are smaller, and key terms are related to cell cycle, metabolism, and translation. Importantly, Oct4, Sox2, and Klf4 are not differentially expressed between 2i and a2i. Additionally, while Nanog is significantly differentially expressed, the expression level difference between 2i and a2i is smaller than between 2i and serum (log fold change = −0.71, adj p < 10⁻⁶ and log fold change = 2.4, adj p < 10⁻⁹⁸, respectively). We observed the same pattern of differential gene expression in the bulk RNA-sequencing experiments (Table S1).

We hypothesized that differences between 2i and a2i, which are related to cell cycle and metabolism, may originate from different proportions of G1/S to G2/M phase cells in each condition. Indeed, using pre-defined cell cycle marker genes, we found that roughly 60% of cells in 2i are in G2/M and only 35% of cells are in G2/M in a2i. We therefore split the cells in each culture condition into G2/M and G1/S subgroups and compared G2/M cells from 2i with G2/M cells from a2i, and we did similarly for the G1/S subgroups. Subsequently, we considered the intersection of genes that were differentially expressed in each comparison and performed GO enrichment and KEGG pathway enrichment analyses. Overall, there are 97 genes with higher expression in 2i and 449 genes with higher expression in a2i. The genes that are upregulated in a2i are involved in RNA processing and transport, translation, and basic metabolism (Figure S6A).

The fact that, even after accounting for cell cycle, differentially expressed genes relate to basic cellular processes led us to explore whether cells cultured in a2i have more mRNA than cells cultured in 2i. To do this we exploited an external spike in molecules that we added to one batch of the cells (2i2, a2i2, and serum 3). The same number of molecules was added to each cell lysate, meaning that the ratio of all reads mapped to the spike ins to all reads mapped to exons can be considered as a proxy for cellular mRNA content (Ding et al., 2015; Stegle et al., 2015). Confirming the reliability of our approach, when we divided 2i and a2i cells into G1/S and G2/M subpopulations and compared their mRNA content, we found that cells in G2/M have significantly more mRNA than cells in G1/S (Figure S6B). Importantly, we observed that cells in 2i contain significantly fewer mRNA molecules than cells in serum and a2i (Wilcoxon test p < 10⁻¹⁵ for both comparisons), which supports the differential expression of genes involved in basic cellular processes.

mESC Transcriptomes Are Similar, but Not Identical, to Those of Blastocyst Cells

It has been suggested that the pluripotent state of 2i cells resembles the cell state of early epiblast cells in the blastocyst (Boroviak et al., 2014; Nichols and Smith, 2009). The recent availability of single cell RNA-sequencing data from different stages of mouse embryonic development allowed us to assess the relationship of in vitro ESCs and in vivo blastomeres (Deng et al., 2014). Our cells were prepared with a very similar protocol, so we used PCA to overlay our data with the published embryonic time course. As expected, mESCs are most similar to the blastocyst stage cells from which they were derived, but they do not overlap (Figures 4D and 4E). The difference between in vivo blastocyst cells and cells cultured in 2i may originate from differences in the mouse strains and/or sequencing protocols, as well as transcriptome changes resulting from in vitro adaptation. mESCs grown in the inhibitory conditions are the most similar to the in vivo blastocyst cells, while serum cultured cells are somewhat more distant (Figures 4D and 4E), which has also been shown previously using cell ensembles (Boroviak et al., 2014).

The dispersion of mESCs in each culture condition is smaller than the dispersion between cells in the blastocyst. This may be explained by noting that mESCs are derived by clonal expansion and cultured in homogeneous conditions relative to the complexity of cellular niches within the embryo. Moreover, blastocyst cells were obtained from several embryos, thus adding an additional factor that may increase heterogeneity. We quantified global transcriptome noise using the DM measure to compare the heterogeneity of blastocyst cells from three stages (early, mid, and late) versus mESCs cultured in 2i. In all comparisons, blastocyst cells are significantly more heterogeneous than the cultured cells (p < 10⁻⁴ by Wilcoxon signed rank test).

Identification and Characterization of 2C-like Cells in 2i Medium

To find 2C-like cells in our samples, we examined the expression profile of genes shown previously to have at least 10-fold enrichment in 2C-like marker genes relative to the remaining mESCs (Macfarlan et al., 2012). Hierarchical clustering suggested the presence of ten 2C-like cells in 2i, and none in the a2i or serum culture conditions (2C-like cells may still be present in a2i and serum, but at a very low rate) (Figure 5A).

Globally, the transcriptomes of 2C cells are altered, and only about 50% of reads on average map to exons, compared to 60% in the remaining population in 2i (Figure 5B). Additionally, we observed substantial MERVL expression in 2C-like cells and no expression in the remaining cells (Figure 5C). Subsequently, we calculated the mean expression level of genes identified by Macfarlan et al. (2012) as differentially expressed in 2C-like cells and observed a similar pattern in our data (Figure 5D). Interestingly, we also observe that 2C-like cells have more upregulated genes than downregulated genes (Figure 5E).

It should be noted that globally, 2C-like cells are more similar to 2i cells and blastocyst cells than to cells from the two-cell stage of the in vivo embryo. 2C-like cells cluster together with 2i cells (Figure 5A), and there are only 294 differentially expressed genes between 2C-like cells and the remaining 2i cells (examples in Figure 5F). In comparison, we find 1,700 genes between 2C-like cells and blastocyst and 1,779 between 2C-like cells and two-cell stage cells (for differential expression results see http://www.ebi.ac.uk/teichmann-srv/espresso). In terms of expression of Nanog, Oct4, Sox2, and Myc, 2C-like cells are also similar to 2i cells in comparison to the two-cell and blastocyst stages of the embryo (Figure 5G).

Transcriptional Regulatory Interactions in mESCs Revealed by Gene-to-Gene Correlations

Above, we mined our high-throughput single cell RNA-sequencing data from the perspective of comparing in vitro and in vivo pluripotent cell populations. We next examined its potential as a rich resource for analyzing correlations in gene expression across culture conditions as a strategy to identify candidate regulators of pluripotency. This allows us to develop hypotheses about the transcriptional regulatory networks that regulate pluripotency in mESCs, which is known to be highly interconnected and complex (Boyer et al., 2005; Kim et al., 2008; Loh et al., 2006).

We found that in serum-cultured mESCs, Nanog expression correlates positively with transcription factors (Esrrb, Klf4, Oct4/Pou5f1, Sox2, and Zfp42), genes involved in DNA methylation (Dnmt3a, Tet1, and Tet2), and other genes such as nuclear receptor Nr0b1 and histone lysine acetyltransferase Kat6b. Nanog is negatively correlated with differentiation regulators including transcription factors Gata3 and Klf7 (Figure 6). These findings concur with known interactions in the pluripotency regulatory network, where Nanog regulates Esrrb (Boyer et al., 2005), Zfp42 (Shi et al., 2006), and Klf4 (Zhang et al., 2010). Beyond confirming known interactions, we identified correlations between characterized pluripotency genes and candidate components of the pluripotency transcriptional regulatory network.

Of the candidate genes we selected seven genes for validation: Ptma, Zfp640, Zfp710, Dpy30, Set, Etv5, and Kat6b. First, using ChIP-seq and ChIP-ChIP data from the ESCAPE database, we found that the promoters of six of the candidate genes are bound by core pluripotency genes (Figure 7A) (Xu et al., 2013). To provide insight into the functional role of these genes, we downregulated their expression using a CRISPR/dCas9 repressor that targeted their promoters (Figure 7B) (Gao et al., 2014) before examining changes in their transcriptomes using bulk RNA-sequencing.

We narrowed down our analysis to four cases that showed significant repression of the targeted gene (Figure 7C) and performed differential expression analysis between samples and control gRNA using DESeq. After correcting for multiple hypothesis testing, we found significantly differentially expressed genes in two cases: Ptma and Zfp640 (Figure 7E). In the samples with repressed Ptma, we observed a decrease in the expression of pluripotency genes and an increase in the expression of genes associated with differentiation (Figure 7D, where pluripotency and differentiation genes are as in Figure 3). Zfp710 and Zfp640 show a similar but milder phenotype, while for Dpy30 there is no clear change in the expression of pluripotency genes (Figure S6C). The lack of effect of Dpy30 downregulation on pluripotency gene expression is consistent with a previous report (Jiang et al., 2011). Overall, these results suggest that Ptma and Zfp640, and potentially also Zfp710, are candidate genes involved in regulating the exit from pluripotency.

Cell Culture of mESCs

The G4 (C57BL/6Ncr x 129S6/SvEvTac) mouse hybrid (George et al., 2007) ESCs were obtained from Mount Sinai Hospital and were maintained on STO feeders in serum-containing media at 5% CO₂ and 37°C. They were sub-cloned, and a line with normal karyotype was selected for further analysis. The cells were split onto gelatinized plates (10 cm, Corning) and expanded in serum-containing media or chemically defined media (standard 2i or alternative 2i) for at least three passages. Cells were harvested by trypsinization (0.05% trypsin/EDTA, GIBCO) for 10 min, at which point they reached 70%–80% confluence for single-cell capture.

The three media are as follows:

• (1)

  Serum-containing media: Knockout DMEM (GIBCO), 1X penicillin-streptomycin-glutamine (GIBCO), 1X non-essential amino acids (GIBCO), 100 U/ml recombinant human leukemia inhibitory factor (Millipore), 15% fetal bovine serum (HyClone), 0.1mM β-mercaptoethanol (Sigma).

• (2)

  Standard 2i media: N2B27 basal media (NDiff 227, StemCells), 100 U/ml recombinant human LIF (Millipore), 1 μM PD0325901 (Stemgent), 3 μM CHIR99021 (Stemgent).

• (3)

  Alternative 2i media: N2B27 basal media (NDiff 227, StemCells), 100 U/ml recombinant human LIF (Millipore), 1 μM CGP77675 (Sigma), 3 μM CHIR99021 (Stemgent).

cDNA Library Preparation from Single Cells using the Fluidigm C1

For each culture condition, 4,000 cells were loaded on to a 10–17 μm Fluidigm C1 Single-Cell Auto Prep IFC, and cell capture was performed according to the manufacturer’s instructions. The capture efficiency was inspected using a microscope to remove samples from the analysis with more than one cell captured. Upon capture, reverse transcription and cDNA preamplification were performed in the 10–17 μm Fluidigm C1 Single-Cell Auto Prep IFC using the SMARTer PCR cDNA Synthesis Kit (Clontech) and the Advantage 2 PCR Kit (Ramsköld et al., 2012). cDNA was harvested and diluted to a range of 0.1–0.3 ng/μl and Nextera libraries were prepared using the Nextera XT DNA Sample Preparation Kit and the Nextera Index Kit (Illumina) following the instructions in the Fluidigm manual “Using the C1™ Single-Cell Auto Prep System to Generate mRNA from Single Cells and Libraries for Sequencing.” Libraries from one chip were pooled, and paired-end 100 bp sequencing was performed on four lanes of an Illumina HiSeq2000.

Bulk RNA-Sequencing

Bulk RNA-sequencing libraries were prepared and sequenced using the Wellcome Trust Sanger Institute sample preparation pipeline with Illumina’s TruSeq RNA Sample Preparation v2 Kit. RNA was extracted from 1–2 million cells using the QIAGEN RNA Purification Kit on a QiaCube robot. The quality of the RNA sample was checked using gel electrophoresis. For library preparation, poly-A RNA was purified from total RNA using oligo-dT magnetic pull-down. Subsequently, mRNA was fragmented using metal-ion catalyzed hydrolysis. The cDNA was synthesized using random hexamer priming, and end repair was performed to obtain blunt ends. A-tailing was done to enable subsequent ligation of Illumina paired-end sequencing adapters, and samples were multiplexed at this stage. The resulting library was amplified using 10 cycles of PCR, substituting the Kapa Hifi polymerase for the polymerase in the Illumina TruSeq Kit. Samples were diluted to 4nM, and 100 bp paired-end sequencing was carried out on an Illumina HiSeq2000. Sequencing Quality Control was performed by the Sanger sequencing facility.

Mapping Reads

Paired-end reads were mapped simultaneously to the Mus musculus genome (Ensembl version 38.73) using GSNAP (version gmap-2014-05-15_v2) using default parameters. Subsequently we counted reads for each gene with htseq-count and normalized them with size factors calculated from DESeq as reported previously (Brennecke et al., 2013). We also applied location and scale adjustments to the normalized read counts to remove technical variation among multiple batches (Supplemental Experimental Procedures).

Quality Control of Cells

To exclude poor quality cells from the downstream analysis, we removed cells according to the following criteria: (1) empty capture sites or capture sites with multiple cells or debris, as defined by visual inspection of the chip; (2) cells that had fewer than 500,000 reads mapped to exons; and (3) cells that had over 10% reads mapped to mitochondrial genes (refer to Supplemental Experimental Procedures for details).

Candidate Gene Expression Repression with CRISPR

45 guide RNAs targeting promoter regions of 7 candidate genes (Ptma, Set, Zfp640, Zfp710, Kat6b, Dpy30, and Etv5) were cloned into gRNA-mCherry plasmid (for a list of sequences, refer to Table S3). GFP-Oct4 reporter strain ESCs (Silva et al., 2008) were transfected with (1) Cas9 repressor-BFP, (2) transposase, and (3) a cocktail of gRNA plasmids (Gao et al., 2014) targeting the gene of interest in a 1:1 ratio using Lipofectamine2000 (Life Technologies). Subsequently, cells were cultured in medium containing 15% serum and LIF for 4 days before 10,000 mCherry and BFP-positive cells were sorted for each sample. RNA was extracted using QIAGEN RNeasy Mini Kit. The SmartSeq2 protocol was used for reverse transcription and amplification of cDNA (Picelli et al., 2014). Sequencing libraries were prepared using Nextera XT Kit according to the manufacturer’s guidelines, barcoded with Nextera XT Dual Index Kit, and sequenced on an Illumina HiSeq2500 in rapid mode.