Dissecting within-embryo and between-embryo variation

We asked whether there are gene products that can reproducibly mark inter-blastomere differences within every embryo. We decomposed the cell-to-cell variation (sum of squares, SS) of transcript abundance of each gene into between-embryo variation (SSbe) and within-embryo variation (SSwe; SS = SSbe + SSwe) (Fig. 2A), and ordered genes by the proportion of SSwe (SSwe/SS) (Fig. 2B). A nontrivial number of 3467 and 7008 genes exhibited larger SSwe than SSbe at the 2-cell and the 4-cell stages, respectively. At the 2-cell stage, the top ranked genes (largest SSwe/SS) reproducibly presented a large FPKM (>1000, upper quartile normalized) in one blastomere but near zero FPKM in the other blastomere in eight or nine out of the 10 embryos (Fig. 2B). At the 4-cell stage, 590 genes exhibited ninefold or greater SSwe than SSbe (FDR ≤ 0.08) (Fig. 2C). These genes showed three to six orders of magnitude of expression differences between the blastomeres of the same embryo, but exhibited relatively stable expression levels across the embryos, opening the possibility of observing reproducible inter-blastomere differences.

Open in a separate window

Figure 2.

Variance decomposition and unsupervised clustering of single cells. (A) The total variation (SS) can be decomposed into the sum of within-embryo variation (SSwe) and between-embryo variation (SSbe). (B,C) Genes (y-axis) are rank-ordered from the largest to the smallest by SSwe/SS (x-axis). Expanded views of the genes with the largest and the smallest SSwe/SS are shown beneath, followed by the expression levels of selected genes. (E1–E10) Embryos 1 to 10. (B1, B2) Blastomeres 1 and 2. (FPKM) Upper quartile normalized FPKM. (D,E, left) Clustering of the blastomeres using all mRNAs (upper dendrogram) and high CV genes (lower dendrogram). (Ei-Bj) The jth blastomere of embryo i. (Right) A scatter plot of standard deviation (y-axis) vs. mean (x-axis) of all genes. The CV filter selected for the genes above the line.

Inter-blastomere differences dominate embryo differences and noise

How does the genome-wide signal strength of inter-blastomere differences compare to that of between-embryo differences and to the noise in gene expression? We did a series of clustering analyses of the blastomeres, starting from using all detected genes (coefficient of variation [CV] ≥ 0) and gradually restricting to using the genes with large overall cell-to-cell variation (CV > 0.25, 0.35, …, 0.85). Because CV represented the total variation rather than the within-embryo variation, this criterion would not favor any specific clustering outcome. However, the higher the CV filter, the more blastomeres of the same embryo were separated into different clusters (Supplemental Fig. S3). At CV > 0.65, the 288 genes left clustered the 2-cell stage blastomeres into two distinct groups, composed of nine (blue, Fig. 2D) and 11 cells (red), respectively. The nine cells of the blue cluster came from nine distinct embryos, (P-value = 6 × 10⁻⁵, hypergeometric test). Similarly, at CV > 0.70, every 4-cell embryo contributed blastomeres to the two furthest clusters (Fig. 2E). This unsupervised analysis suggests that the signal of within-embryo difference is strong as compared to between-embryo differences and random noise. Essentially most genes or gene sets with large total variation can distinguish sister blastomeres.

Bimodality of gene expression in sister blastomeres

It remains unclear whether any mammalian developmental processes or stem cells have bimodal gene expression (Shalek et al. 2013). A total of 138 and 205 genes showed bimodal expression patterns at the 2-cell and the 4-cell stages, respectively (FDR < 0.05, Dip Test of Unimodality [Hartigan and Hartigan 1985]). We then asked whether any genes were specifically bimodal among the sister blastomeres of the same embryo. After removing between-embryo variation, we found 12 and 13 genes with within-embryo bimodality at the 2-cell and the 4-cell stages, respectively (FDR < 0.05, Dip Test) (Fig. 3A,B; Supplemental Tables S1, S2). The majority of these genes had near zero FPKM in some blastomeres but 10³ − 10⁶ FPKM (upper quartile normalized) in the sister blastomeres in nearly every embryo. The reproducible bimodal gene expression between the sister blastomeres is incompatible with the equivalence hypothesis.

Open in a separate window

Figure 3.

Bimodality of gene expression in sister blastomeres. (A,B) Expression levels and their histograms. (E1–E10) Embryos 1 to 10. The two (four) blastomeres of each embryo are randomly assigned to two (four) colors. (FPKM) Upper quartile normalized FPKM. (C) Anticorrelation of 2-cell bimodality genes Ap2s1 and Rsph3a. Juxtaposition (upper) and inter-blastomere difference (lower) of expression levels in the same embryos. (D) Correlated genes at the 4-cell stage. Cells from the same embryo have the same color. (r) Pearson correlation. (**) FDR < 0.01.

We investigated whether the bimodality of one gene was predictive of another. At the 2-cell stage, the bimodal mRNAs of Ap2s1 and Rsph3a showed opposite directions in their inter-blastomere differences in nine out of 10 embryos (FDR < 6.7 × 10⁻⁴) (Fig. 3C). At the 4-cell stage, two pairs of bimodal genes, Camk1d-Acox3 and Gpr124-Csf2rb2, exhibited strong correlations (FDR < 0.01) (Fig. 3D). The correlation and anticorrelation of bimodal gene expression are unexpected from random fluctuations of gene expression.

We further interrogated whether the bimodal distribution of transcript abundance is reflected by protein abundance. We analyzed the protein expression of bimodal genes Ap2s1 and Zfp688 at the 2- and 4-cell stages, as well as Acox3 and Gadd45a at the 4-cell stage by immunocytochemistry. The immunofluorescence signals of alpha tubulin (control), AP2S1, ZFP688, and ACOX3 were comparable between the blastomeres (see Data Access). However, GADD45A exhibited visual and statistical differences in fluorescence intensities between the blastomeres in the majority (five to eight, depending on the method of assessment) of the eight 4-cell embryos scanned (Fig. 4; Supplemental Fig. S4; P-value ≤ 0.1 in Supplemental Table S3; see also Data Access). Moreover, each pair of blastomeres originating from the same mother cell, as identified by the spindle apparatus, exhibited similar immunofluorescence signals. In six cases, differences in protein expression were observed between the two pairs of blastomeres defined by the first cleavage. Thus, the distribution of GADD45A proteins resembled the bimodal distribution of Gadd45a mRNAs.

Open in a separate window

Figure 4.

GADD45A protein expression levels in two 4-cell mouse embryos. For each embryo (upper or lower panels), the confocal images were captured on the same z stack from an immunocytochemistry assay of GADD45A (Alexa-548, red) and alpha tubulin (Alexa-488, green). Arrows point to cells with greater expression of GADD45A. (DNA) DAPI fluorescence. (DIC) Differential interference contrast image. Scale bar, 50 μm.

Causes of bimodal gene expression

At least three plausible causes could produce transcriptome asymmetry in 2-cell embryos, namely, embryonic genome activation (EGA), polarized cell division, and RNA degradation. We compared the SMART-seq data of the bimodal genes across five mature oocytes (RNA-seq data from Ramskold et al. 2012; Xue et al. 2013), zygotes, and the 2-cell embryos. The mRNAs of Fzd2 and Zfp688 were not detected in oocytes and most of the zygotes, but were abundant (FPKM ≥ 1000, upper quartile normalized) in at least one blastomere in every 2-cell embryo (Fig. 5A; Supplemental Fig. S5). Thus, the Fzd2 and Zfp688 mRNAs were EGA transcripts (P-value = 3.5 × 10⁻⁴, hypergeometric test). On the other hand, the mRNAs of Ap2s1 and Rsph3a (Fig. 5B; Supplemental Fig. S5) were abundant in every zygote, whereas they had zero or near zero FPKM in precisely one blastomere in eight out of the 10 2-cell embryos, suggesting these mRNAs were differentially depleted between the sister blastomeres (maximum P-value = 3.5 × 10⁻⁴, hypergeometric test).

Open in a separate window

Figure 5.

Coverage plots of single-cell RNA-seq reads. Each row represents a cell. The gene models are at the bottom of the plots. (A) Fzd2, a candidate bimodality gene due to EGA. (B) Ap2s1, a candidate bimodality gene due to blastomere-specific RNA depletion.

Coexpression networks as further evidence for molecular asymmetry

We checked whether within-embryo differences could produce coexpression networks. After removing between-embryo variation, we clustered all the detected genes. Large coexpression modules emerged from both the 2-cell and the 4-cell stages (red blocks, Fig. 6A,B, left panels). To assess the statistical significance of these coexpression modules, we compared the distribution of pairwise correlations of genes with permutation-derived distributions. All the large coexpression modules disappeared from the permutation data (Fig. 6A,B, middle panels), and the numbers of strong (anti)correlations were greatly reduced (P-value < 2 × 10⁻²², Kolmogorov test) (Fig. 6A,B, right panels). The coexpression modules were enriched with gene ontology (GO) terms including “cell fate commitment” (Fig. 6C), and corresponded to known protein–protein or regulatory interactions (P-value ≤ 10⁻¹⁰, test by resampling) (Supplemental Fig. S6). The formation of coexpression modules provides further evidence favoring the asymmetric hypothesis.

Open in a separate window

Figure 6.

Coexpression networks resulted from inter-blastomere differences. (A,B) Correlation heatmaps of all genes after removing between-embryo variation (left) indicates the existence of coexpression networks (red blocks). The networks disappeared after permuting the cell labels (middle), which essentially produces data from the equivalence hypothesis. There were considerably more correlated and anticorrelated gene pairs in the real data than in the permuted data (right). (C) A subset of gene pairs with strong correlations at both 2-cell and 4-cell stages. This subset contained the genes with GO annotations “positive regulation of developmental processes” or “cell fate commitment.” (D) Consistency of gene expression differences at 4-cell and blastocyst stages. Gene pairs are categorized by their correlation (r) at the 4-cell stage (x-axis). A large r suggests that a gene pair shares this same direction of inter-blastomere differences. Odds ratio (OR) was used to represent the chances of these gene pairs exhibiting the same direction of expression differences between ICM and TE. OR had a monotone increase with respect to 4-cell r. (***) P-value < 10⁻³. (0.45, 0.55] is equivalent to 0.45 < r ≤ 0.55.

Probable mechanisms

Several probable mechanisms are compatible with the asymmetric hypothesis, including polarized cell division and differential gene regulation by signaling (Morris et al. 2012), TFs, epigenetics, post-transcriptional processing and transport, and degradation (for reviews, see Zernicka-Goetz 2004; Vogel 2005; Zernicka-Goetz et al. 2009; Takaoka and Hamada 2012; Saiz and Plusa 2013). Supportive data for these mechanisms have become evident for the 8–16 cell stage and onward, but not as clear for the earlier stages (Torres-Padilla et al. 2007; Plachta et al. 2011). The majority of genes or gene pairs implicated to be associated with 2-cell or 4-cell transcriptome asymmetry in this analysis are signaling molecules, TFs, epigenetic modifiers, and RNA processors (Supplemental Tables S5–S7).

WNT signaling

WNT signaling is overwhelmingly represented by 2-cell bimodality genes (Supplemental Table S4), starting with its ligand biogenesis enzyme Mgat5 (Matthews et al. 2009), receptor frizzled homolog 2 (Drosophila) (Fzd2), frizzled-binding modulator myocilin (Myoc) (Kwon et al. 2009), to its ligand endocytosis factor Ap2s1 (Blitzer and Nusse 2006), TF-binding Rsph3a (Grimsby et al. 2004), and to direct transcriptional targets Mgat5 (Sengupta et al. 2010) (potentially forming a feed forward loop) and TF gene Zeb1 (Sanchez-Tillo et al. 2011). Besides, both the Wnt inducing TF gene Tbx3 (also known as Brachury) and the canonical TF gene of the WNT pathway Tcf3 showed up in correlated gene pairs (Fig. 6C; Supplemental Table S6). WNT antagonist Apc strongly correlated with DNA methyltransferase Dnmt1 in both 2-cell and 4-cell stages (Supplemental Fig. S5). More signaling pathways were implicated in the 4-cell bimodality genes and within-embryo correlated gene pairs (Fig. 6C; Supplemental Tables S6, S7), including FGF receptor substrate Frs2, G protein-coupled receptor Gpr124, and WNT agonist receptor Lgr4.

EGA is implicated to the creation of asymmetry

Fifteen percent of the mRNAs with large SSwe (SSwe > 4 × SSbe) were not detectable in mature oocytes (Ramskold et al. 2012; Xue et al. 2013), and thus belonged to EGA transcripts. A subset of 2-cell bimodal genes was transcribed by the embryo genome in one and only one blastomere in 2-cell embryos (Fig. 5A; Supplemental Fig. S5A). This reproducible pattern cannot be explained by transcription noise alone. It is tempting to hypothesize that WNT-related TFs and epigenetic modifiers (Supplemental Tables S4–S7) regulate blastomere-specific EGA.

Blastomere-specific RNA depletion could account for another subset of bimodal mRNAs, especially those abundant in zygotes (Fig. 5B; Supplemental Fig. S5B). This depletion pattern could be explained by polarized cell division (Johnson and Ziomek 1981; Fleming and Johnson 1988) or cell-specific RNA degradation (Giraldez et al. 2006; Lykke-Andersen et al. 2008), even though no evidence of either mechanism has been found prior to the 8-cell stage. Surprisingly, at least four out of the 12 correlated gene pairs (Supplemental Table S7) are involved in regulation of cell polarity (Supplemental Table S8). These data suggest that, as early as the 2-4 cell stages, the blastomeres could have prepared differently for future asymmetric divisions. Consistent with this idea, bimodal Afap1 encodes a protein that interacts with atypical protein kinase C (aPKC) (Qian et al. 2002), which adopt polarized localization from the 8-cell stage onward and regulate cell fate decisions (Plusa et al. 2005a). In summary, these single-cell data from matched sister blastomeres argue against the equivalence hypothesis, and lead to several mechanistic insights into the proposed early asymmetry.

RNA library preparation and sequencing

The zygotes, blastomeres, and RNA extracted from ICM and TE were used for cDNA amplification using the SMARTer Universal Low Input RNA kit (Clontech). All amplifications were carried out in parallel with positive and no-template controls for quality assurance. For all samples, the total amplified cDNA was used for library preparation using Nextera XT DNA Sample Prep Kit (Illumina). In order to avoid batch effects, all the libraries for cells of the same embryo were prepared at the same time. The measurements of cDNA amplification library quality control were taken with Qubit (Invitrogen) and Bioanalyzer (Agilent). The libraries of cells composing one embryo were subsequently sequenced in the same lane of a flow cell to 100-base read length on HiSeq 2000 or HiSeq 2500 sequencers (Illumina).

Alignment of sequencing reads to the mouse genome

The 100-base long single-end reads were aligned to the mouse genome (mm10, Ensembl; Flicek et al. 2013), using the STAR aligner (Dobin et al. 2013), allowing for up to four mismatches. Only the reads aligned uniquely to one genomic location, and with alignment quality above 10 were retained for subsequent analyses. The duplicate reads were eliminated with Picard tools (http://picard.sourceforge.net).

Assessment of transcript coverage

The relative coverage of the reads along the entire length of the transcripts was used to assess the quality of single-cell RNA-seq. The RefSeq mRNA sequences, corresponding to mm10, were downloaded from the UCSC Genome Browser (Kuhn et al. 2012), indexed with Bowtie 2 v2.1.0 (Langmead and Salzberg 2012), and then aligned with the RNA-seq reads. The reads that aligned once to a sequence were retained for further filtering of duplicates with Picard tools. The coordinate of a read mapping (transcript and nucleotide) was used to calculate the position of the read relative to the total length of the transcript (0%–100%). The relative positions of all the reads were summarized into a cumulated density map.

Estimation of gene expression levels in single cells

The non-duplicated reads uniquely aligned to the mouse genome were used to estimate the gene expression levels (FPKM) by Cufflinks v2.0.2 (Trapnell et al. 2012) with the option of upper quartile normalization, which multiplied a scaling factor by the unnormalized FPKM (Dillies et al. 2013). Gene models were obtained from the Ensembl annotation (GRCm38.70). Requiring FPKM ≥ 1 in at least one blastomere of every 2-cell embryo, or at least two blastomeres of every 4-cell embryo, we obtained 8208 and 8054 mRNAs from the 2-cell and the 4-cell stages, respectively. Only the detected mRNAs were used in this analysis.

Building the web browser

The single-cell genome browser was built based on the portable codes of UCSC Genome Browser (Kent et al. 2002). We removed duplicated reads and converted the remaining aligned reads of each blastomere into a bigWig track. We normalized the read intensity in every track by the number of non-duplicated uniquely aligned reads of that library. We grouped the tracks by each embryo and ordered them by developmental stages. Genome annotation and sample information were provided alongside data tracks. This browser currently hosts 49 single cell tracks, four ICM tracks, three TE tracks, as well as two previously published sperm (Gan et al. 2013) and five previously published mature oocyte data sets (Ramskold et al. 2012; Xue et al. 2013).

Removing between-embryo variation

Between-embryo variation was removed in a subset of the analyses (Figs. 3, ​,5).5). Let be the log₁₀(FPKM + 1) of the gene in embryo blastomere . The following transformation is applied to every gene: , where is the mean of in embryo , and is the relative expression level after removing between-embryo variation.

Testing the deviations from unimodal distributions

The Dip Test (Hartigan and Hartigan 1985) was used to assess the deviation of data from unimodal distributions. The null hypothesis was the expression levels followed by a unimodal distribution and the alternative was a multimodal distribution. FDR was calculated subsequently (Pollard et al. 2005).

Permutation test of correlation

One hundred permutations were performed after removing the between-embryo variation. The cell labels of each gene were independently swapped in each permutation. The distribution of correlations was derived from sampling 336,815 and 324,133 gene pairs in each permutation from the 2-cell and the 4-cell stages, respectively. These sampled gene pairs added up to the same number of gene pairs in real data from the 100 permutations.

Association of coexpression in 4-cell embryos and in ICM/TE

Pairwise correlation was calculated for 4-cell data, after removing between-embryo variation. An association test was devised to assess whether the highly correlated gene pairs at the 4-cell stage tend to share the same direction of differential expression between ICM and TE. Gene pairs were binned by their 4-cell correlation, and then the numbers of gene pairs in each bin that shared the same direction of differential expression between ICM and TE were counted. These counts were used to compute odds ratio (OR), for which significance was assessed by either χ² or Fisher’s exact test.

Immunocytochemistry assay (IC)

Four genes were selected for further validation of RNA-seq results. Antibodies for the proteins are as follows: AP2S1 (Abcam, ab128950), ZFP688 (Abcam, ab72605), ACOX3 (Sigma, HPA035840), and GADD45A (Sigma, G3548). Embryos were collected at 2- or 4-cell stages and the zona pelucida was removed as described above. Embryos were fixed in methanol at −20°C overnight, washed in PBS (supplemented with 0.1% BSA), and incubated for 5 min in permeabilization solution (PBS, 0.1% BSA, 0.1% Triton X-100). Embryos were then incubated for 1 h in blocking solution (PBS, 0.1% BSA, 10% bovine fetal serum), following incubation with primary antibody (1:100 dilution in blocking solution) for 2 h at room temperature. After PBS (0.1% BSA) washes, indirect detection was possible by immersing the embryos in blocking solution with anti-rabbit antibody conjugated with Alexa Fluor-546 Goat Anti-Rabbit IgG (1:500 dilution, {"type":"entrez-nucleotide","attrs":{"text":"A11010","term_id":"492391","term_text":"A11010"}}A11010, Invitrogen). Following washes with PBS (0.1% BSA), embryos were immersed in blocking solution containing Mouse anti-Tubulin-Alexa-488 (1:25 dilution, 32-2588, Invitrogen) for 1 h. Embryos were washed in PBS (0.1% BSA) supplemented with NucBlue Fixed Cell Stain ReadyProbes (Invitrogen), and preserved with Prolong Gold (Invitrogen). Embryos were scanned with a Fluoview FV1000 confocal microscope (Olympus), with the aid of a 40× oil objective, excitation wavelength 405/473/559, and emission wavelength 461/520/572 (DAPI, Alexa-488 and Alexa-545, respectively). Scanning was performed at consecutive z-planes separated by 1 µm.

This work was supported by NIH DP2OD007417 and the March of Dimes Foundation. We thank Drs. Terence Hwa and Kun Zhang for useful suggestions.