Main

Human embryonic stem cells (hESCs) derived from the inner cell mass (ICM) of the developing blastocysts1 have the capacity to self-renew and differentiate into any of the three (ectoderm, mesoderm and endoderm) germ lineages, a feature defined as pluripotency. Intriguing data2,3,4,5,6,7,8,9 also suggest that hESCs, in contrast to mouse embryonic stem cells (mESCs), can be induced to differentiate into extra-embryonic lineages such as trophectoderm cells, a capacity normally restricted to totipotent stem cells. The mechanisms restricting epiblast cells from entering trophectoderm lineage in the developing blastocyst remain elusive.

Conventional hESCs resemble the post-implantation epiblast and their features are similar to murine epiblast stem cells10,11, emanating the name: primed hESCs12. Pre-implantation-like hESCs, referred to as ‘naïve’, recapitulate the gene expression, X chromosome re-activation and DNA methylation patterns observed in the pre-implantation ICM in vivo13,14,15,16,17,18. Genomic, transcriptomic, proteomic and epigenomic analyses have been conducted to understand the molecular mechanisms governing distinct pluripotent states19,20,21,22,23,24,25,26,27.

Promoter bivalency, the presence of both a ‘repressive’ histone post-translational modification H3K27me3 and ‘active’ mark H3K4me3, has been attributed an important role in regulating developmental genes28,29,30. However, not much attention has been directed towards exploring a possible role of this epigenetic signature in naïve pluripotency, owing partly to reports that polycomb repressive complex 2 (PRC2), responsible for H3K27 methylation, is required for primed but dispensable for naïve pluripotency31,32,33. Polycomb repressive complex 1 (PRC1), responsible for H2AK119 mono-ubiquitination (hereafter H2Aub), has been shown to be essential for self-renewal, pluripotency and ordered differentiation of mESCs and hESCs34,35,36,37,38,39,40, but the H2Aub landscape of naïve hESCs is unexplored.

In this Article, we use quantitative epigenome and transcriptome profiling of naïve and primed hESCs, and functionally characterize distinct roles for PRC2 in the two states through pharmacologic inhibition and genetic targeting. To understand the functional role of PRC2 in setting up the naïve epigenomic and transcriptional landscape, we depleted H3K27me3 by inhibiting PRC2 in naïve and primed state. We discover a key role of PRC2 in shielding naïve human pluripotency from trophectoderm and mesoderm differentiation.

Results

Diffuse gain and distinct H3K27me3 patterns in naïve cells

To elucidate the function of promoter bivalency in naïve and primed hESCs, we performed quantitative chromatin immunoprecipitation followed by sequencing (ChIP–seq) on H9 female hESCs, maintained in naïve (t2iLGö) or primed (E8) culture conditions, with and without EZH1/2 inhibitor (EZH2i, EPZ-643841) for 7 days (Fig. 1a). We quantitatively profiled three histone modifications associated with bivalent genes, H3K4me3, H3K27me3 and H2Aub with MINUTE-ChIP42 (Extended Data Fig. 1a)42. Phenotypically, EZH2i-treated naïve and primed hESCs maintained their expected morphology over 7 days (Extended Data Fig. 2a).

Fig. 1: Diffuse H3K27me3 and H2Aub cover the naïve pluripotent genome.
figure 1

a, Experimental design of MINUTE-ChIP experiment comparing hESCs in naïve and primed, untreated and treated (10 µM EZH2i) culture conditions. Three biological replicates of each condition were barcoded and combined into a single MINUTE-ChIP pool. See Extended Data Fig. 1a for a scheme of the MINUTE-ChIP workflow. b, Global levels of H3K27me3 as determined by MINUTE-ChIP input-normalized read counts (INRCs) in naïve or primed hESCs, cultured with or without EZH2 inhibitor. P values of pairwise comparisons (two-sided unpaired Student’s t-test) are given. See Extended Data Fig. 1b for corresponding analysis of H2Aub and H3K4me3. c, Histone H3K27me3, H3K4me3 and H2Aub levels by chromatin state (reads per genome coverage, RPGC). See Extended Data Fig. 1c for individual replicates. d, Genome-wide analysis of H3K27me3, H2Aub and H3K4me3 levels by 10 kb bins, comparing naïve and primed hESCs. e, Genome browser examples of genomic regions with differential occupancy of H3K27me3, H3K4me3 and/or H2Aub. Group-scaled histone modification signals and gain/loss tracks comparing naïve and primed signals are shown. f, Chromosome average enrichment of H3K4me3, H3K27me3 and H2Aub in naïve and primed hESCs. Box plot boxes show the 25th and 75th percentile with the median, and whiskers indicate 1.5 times the interquartile range. All individual data points are shown. g, Chromosome density plot comparing the X chromosome signals in naïve and primed hESCs to an autosome with similar size (chr7). See Extended Data Fig. 4c for a genome browser example of X chromosome region. h, Treemap showing a proportional representation of the total (integrated) amount of H3K27me3 by chromosome (area) and average density (colour intensity). i, Density plots of log2FC in promoter H3K27me3 levels (left) and RNA-seq output (middle) of genes grouped by chromosome, comparing naïve and primed hESCs. Density plots of log2FC in RNA-seq output of genes grouped by chromosome, comparing untreated and EZH2i-treated naïve hESCs (right). Median values by chromosome are given and indicated as vertical lines in the density plot. In cg, three combined biological replicates are shown.

Source data

H3K27me3 levels were substantially (~3.3-fold) higher in naïve than in primed hESCs (Fig. 1b), an unexpected observation as H3K27me3 has been reported to be lower in the naïve state17,43,44. However, our findings are consistent with quantitative mass spectrometry data comparing naïve and primed hESCs26, further confirmed by immunofluorescence and immunoblotting (Extended Data Fig. 2b–d). H2Aub signal was also significantly (~2.1-fold) higher in naïve cultures (Extended Data Fig. 1b), suggesting a concerted regulation of PRC1/PRC2 activity. Naïve and primed H3K27me3 levels were depleted by 97% and 92%, after 7-day treatment with EZH2i (Fig. 1b and Extended Data Fig. 2c,d). Immunofluorescence analysis established that H3K27me3 was largely lost between days 2 and 4 and further depleted at day 7 (Extended Data Fig. 2e,f).

We next performed genome-wide enrichment analysis on functionally annotated chromatin states (Fig. 1c and Extended Data Fig. 1c), as well as in 10 kb windows (Fig. 1d and Extended Data Fig. 3a). Despite no substantial change in global levels or at bivalent promoters, as previously defined in the primed state45 (Fig. 1b,c and Extended Data Figs. 1b and 3a–c), H3K4me3 was increased marginally at active promoters in the naïve state (Fig. 1c and Extended Data Fig. 3b).

Naïve cells had a higher basal level of H3K27me3 across non-polycomb chromatin states and most of the 10 kb bins (Fig. 1c,d and Extended Data Fig. 3a), a phenomenon previously described in naïve mESCs42,46. While more than 80% of promoters did not significantly change H3K27me3 status (Fig. 1c and Extended Data Fig. 3b,c), approximately 5% of promoters, including a number of HOX genes and other classical bivalent genes, showed reduced H3K27me3 in naïve hESCs (Fig. 1e and Extended Data Fig. 3b, c). Among these, naïve-specific genes such as KLF4 entirely lost H3K27me3 in the naïve state (Fig. 1e). On the other hand, more than 10% of promoters, for example CDH2, DUSP6 and OTX2, strongly gained H3K27me3 (Fig. 1d,e and Extended Data Fig. 3b,c).

Diffuse gain in H2A ubiquitination in naïve pluripotency

PRC1 is thought to act through H3K27me3-dependent and H3K27me3-independent mechanisms in mESCs47,48,49,50. While genome-wide H2Aub and H3K27me3 levels were reasonably correlated in primed and naïve hESCs, H2Aub did not mirror many of the gains and losses of H3K27me3 (Extended Data Fig. 3b–d). For example, levels of H2Aub at DUSP6 and OTX2 genes were similar in naïve and primed states (Fig. 1e).

Intriguingly, the H2Aub landscape was quantitatively maintained in the absence of H3K27me3 (Fig. 1c and Extended Data Fig. 1b). Less than 0.15% of 10 kb bins genome wide showed a significant reduction in H2Aub, including ~0.25% of annotated bivalent promoters45 (Extended Data Fig. 3e). For example, H2Aub levels were broadly reduced by 10–30% at HOX gene clusters (Fig. 1e), with maximal loss of ~50% at HOXC5, HOXB8 and HOXB9 promoters. Together, these results suggest that PRC1 recruitment and activity was largely independent of H3K27me3.

H3K27me3 accumulation on X chromosomes in naïve cells

Strikingly, H3K27me3 levels of the naïve X chromosomes were approximately twofold higher than naïve autosomes and approximately fivefold higher than the average signal on chromosomes in the primed state of our naïve female H9 hESCs (Fig. 1f,g and Extended Data Fig. 4a,b). Published data from naïve female hESCs showed a similar enrichment (Extended Data Fig. 4b). At the megabase scale, regions of very high H3K27me3 enrichment alternated with less enriched regions of active transcription along the X chromosomes (Fig. 1g and Extended Data Fig. 4c). Integrating the H3K27me3 signal across chromosomes, we estimated that the X chromosomes carry ~10% of all H3K27me3 marks, more than any autosome (Fig. 1h).

Despite the strong increase in H3K27me3, RNA-seq showed that overall transcriptional output from the X chromosomes relative to autosomes did not change between naïve and primed states, and the distribution of differentially expressed genes was similar to autosomes (Fig. 1i and Extended Data Fig. 4d). A phenomenon termed X chromosome dampening is thought to reduce dosage of the X chromosomes in the naïve state in a mechanism that is distinct from X inactivation and not well understood51. Hence, we wondered if H3K27me3 contributed to dosage compensation in naïve hESCs. However, depletion of H3K27me3 did not lead to a global upregulation of transcription from the X chromosomes in naïve hESCs (Fig. 1i and Extended Data Fig. 4e). Even though some genes, such as VGLL1, were marked with H3K27me3 in the naïve state and strongly derepressed upon EZH2i treatment, up- and downregulated genes were similar in number and showed a similar fold-change range as those of autosomes (Fig. 1i and Extended Data Fig. 4e,f). Thus, we conclude that H3K27me3 hypermethylation on the naïve X chromosome does not confer chromosome-scale dosage compensation.

Quantitative ChIP defines naïve-specific bivalent genes

Since prior definitions of bivalent gene sets were based on H3K27me3/H3K4me3 enrichment in primed hESCs only17,45, we defined bivalent promoters de novo from our datasets in five categories based on their H3K27me3/H3K4me3 enrichment in primed and/or naïve hESCs (for criteria, see Extended Data Fig. 5a). We called 533 H3K4me3-positive promoters with higher H3K27me3 in the primed state (primed bivalent), 1,551 H3K4me3-positive promoters with higher H3K27me3 in the naïve state (naïve bivalent), 3,403 promoters bivalent in both states (common bivalent) and 12,391 H3K4me3-positive promoters without H3K27me3 (H3K4me3 only) (Fig. 2a, Extended Data Fig. 5a and Supplementary Table 2).

Fig. 2: H3K27me3 is adaptive to gene expression changes between naïve and primed pluripotent states and contributes to repression of non-state-specific genes.
figure 2

a, De novo annotation of bivalent promoters based on DESeq2 analysis. Five promoter classes were defined (for criteria, see Extended Data Fig. 5a): primed bivalent (Pr » Ni), naïve bivalent (Ni » Pr), common bivalent, H3K4me3 only and H3K4me3 negative (not shown). Average H3K27me3, H2Aub and H3K4me3 profile plots (fragments per genome coverage, FPGC) in naïve and primed hESCs for each class and corresponding heat maps for the first three states are shown. Profiles for EZH2i-treated naïve and primed conditions are shown as dashed lines. b, Alluvial plot showing H3K27me3 and H3K4me3 gains and losses (DESeq2 adjusted P < 0.05, fold change >1.5 from three replicates) at bivalent promoters between naïve and primed hESCs. Select connections are annotated. c, Context-specific transcriptional response to global H3K27me3 depletion in different classes of bivalent genes. RNA-seq changes (DESeq2 log2FC from three replicates each) are plotted, comparing naïve and primed conditions as well as EZH2i treatment and the respective control (left, naïve only; middle, primed only; right, shared bivalent promoters). The distribution of fold changes of all genes is shown as violin plots, and class-specific genes are shown as jitter points. The class-specific group was compared with all genes using a two-sided unpaired Wilcoxon test and Cohen’s d. d, Density plot of fold changes of H2Aub levels following H3K27me3 depletion in hESCs. Only genes that were derepressed upon EZH2i treatment (DESeq2 adjusted P < 0.05, fold change >1.5 based on three replicates) were included in the analysis. Bivalent promoters (hence including promoters of the naïve-bivalent and shared class) are compared with H3K27me3-devoid promoters. For an analysis of individual classes, see Extended Data Fig. 5b. e, Heat map showing RNA-seq expression (log2-transformed TPM) of previously defined marker genes for naïve and, primed pluripotency53 in naïve and primed cultures (± EZH2i treatment), as well as the H3K4me3 and H3K27me3 levels (RPGC) at their respective promoter. RPGC from combined replicates are shown for H3K4me3 and H3K27me3, whereas the individual replicate TPM values are plotted for RNA-seq data. n.s., not significant.

Source data

H3K4me3 was co-enriched at bivalent promoters in both naïve and primed condition but tended to be higher in naïve state (Fig. 2a). Intersecting significant changes in H3K4me3 and H3K27me3 between naïve and primed hESCs, three interesting scenarios occurred commonly (Fig. 2b): the first set of genes, including SOX11 and SIX3, gained H3K4me3 and lost H3K27me3 in the primed state; a second set of genes, including KLF4, KLF5, BMI1, TBX3 and DNMT3L, gained H3K4me3 and lost H3K27me3 in the naïve state; a third set, including STS, CGB8 and XAGE2, had higher levels of both H3K27me3 and H3K4me3 in the naïve state (Fig. 2b).

H3K27me3 repression of bivalent genes in naïve pluripotency

We next interrogated our RNA-seq datasets to elucidate if H3K27me3 played a functional role in regulating bivalent promoters. Naïve-bivalent genes were indeed predominantly lower expressed (P = 5 × 10−103, Cohen effect size medium) in naïve state (Fig. 2c, left), and primed-bivalent genes were predominantly lower expressed (P = 3 × 10−49, Cohen effect size large) in primed state (Fig. 2c, middle).

As expected, treating naïve hESCs with EZH2i derepressed naïve-bivalent genes (P = 1 × 10−77, Cohen effect size medium), with no significant effect on the expression of primed-bivalent genes (Fig. 2c). Treating primed hESCs with EZH2i derepressed primed-specific genes (P = 4 × 10−75, Cohen effect size large) while naïve-bivalent genes showed a variable response without a significant trend towards derepression (Fig. 2c). EZH2i treatment significantly increased expression of shared bivalent genes in both states (P ≤ 1.36 × 10−149, Cohen effect size medium) (Fig. 2c, right).

H2Aub in hESCs is largely independent of H3K27me3

The canonical pathway of polycomb repression involves recruitment of canonical PRC1 complex via H3K27me3, following H2A ubiquitination, chromatin compaction and silencing. H2Aub followed similar class-specific trends between naïve and primed states as H3K27me3, albeit within a smaller dynamic range (Fig. 2a), and changed little following H3K27me3 depletion (Fig. 2a). We observed only a weak trend to lose H2Aub (median loss 15%) at derepressed genes (Fig. 2d and Extended Data Fig. 5b). Hence, our data suggest a weak anti-correlation of transcriptional activity and H2Aub at bivalent promoters but does not resolve if the modest reduction in H2Aub is a prerequisite or a consequence of gene activation.

H3K27me3 acts as a barrier between naïve and primed hESC

Among the genes switching bivalency status between naïve and primed state (Fig. 2a) were many known marker genes of naïve and primed pluripotency (see also examples above, KLF4, OTX2 and DUSP6 in Fig. 1e). Another key transcription factor uniquely expressed in naïve hESCs and recently implicated in the establishment of the naïve transcriptional landscape is TFAP2C52. The promoters of TFAP2C gene, as well as the related TFAP2A gene, were indeed highly enriched in H3K4me3 and devoid of H3K27me3 in the naïve state, but acquired a bivalent state with intermediate H3K4me3 levels and high H3K27me3 enrichment in the primed state (Extended Data Fig. 5c). Intriguingly, depleting H3K27me3 in the primed state basally activated TFAP2A/C transcription, albeit not to the same high level observed in the naïve state. Thus, depletion of H3K27me3 in primed cells removed PRC2-mediated repression but was not able to revert the TFAP2A/C genes to a full ‘on’ state, presumably because additional activating factors were not expressed in the primed state.

We observed a similar switch from H3K4me3-only promoter status in naïve to bivalent in primed for three additional transcription factors implicated in setting up the naïve pluripotency (KLF4, KLF5 and TBX3), and these also increased basal transcription in response to H3K27me3 depletion in the primed state (Fig. 2e). However, the majority of naïve markers, such as DNMT3L, TRIM60 and ZNF729, did not accumulate H3K27me3 in the primed state and were also not responsive to EZH2i treatment (Fig. 2e). Most markers of primed pluripotency53 had H3K4me3-only promoters in the primed state, while assuming a H3K27me3-high/H3K4me3-low bivalent promoter status in the naïve state (Fig. 2b,e). H3K27me3 depletion in the naïve state resulted in increased transcription of most of these genes (including OTX2, CD24 and DUSP6), albeit in no case reaching the levels of the primed state (Figs. 1e and 2e). In conclusion, PRC2 appears to establish an epigenetic barrier between the two states through adaptive repression of different gene sets in naïve and primed pluripotency.

PRC2 represses trophectoderm lineage in naïve cultures

We next wanted to examine whether bivalent promoters in general may be repressed by H3K27me3 in the naïve state. Differential expression analysis identified 1,894 upregulated and 766 downregulated genes upon EZH2i treatment (Fig. 3a–c). About half of the upregulated genes showed H3K27me3 enrichment in their promoter region (Fig. 3c). A strikingly different transcriptional response of primed cells to EZH2i inhibition (518 up- and 34 downregulated; Extended Data Fig. 5d) highlights the exquisitely context-specific function of PRC2 in repressing largely non-overlapping gene sets in closely related pluripotent states (Fig. 3c). As a number of known developmental genes, including IGF2, FRZB, EPAS1 and GATA2, were among the upregulated genes in EZH2i-treated naïve cells (Fig. 3a), we evaluated a comprehensive list of lineage marker genes derived from single-cell RNA (scRNA)-seq of the developing human embryo54 (Fig. 3d). EZH2i treatment of naïve cells reduced marker genes of ICM and epiblast while trophectoderm, amnion and mesoderm markers were upregulated (Fig. 3d). Primed hESCs treated with EZH2i did not show major perturbations for any of these lineages (Extended Data Fig. 5e).

Fig. 3: PRC2 inhibition derepresses a naïve-specific subset of bivalent genes.
figure 3

a, Volcano plot showing differentially expressed genes (DESeq2 FDR <5%, |log2FC| >1) between 7-day EZH2i-treated naïve hESCs and untreated hESCs. Bivalent gene classes as shown in Fig. 2a are indicated in red (primed bivalent), blue (shared bivalent) and teal (naïve bivalent). See Extended Data Fig. 5d for corresponding analysis in primed hESCs. b, Fractions of significant (FDR <5%) transcriptional changes in response to EZH2 inhibitor treatment of naïve and primed hESCs, among all or bivalent promoter classes. c, Intersection of genes derepressed after EZH2i treatment in naïve and primed hESCs. Each Venn intersection is further annotated by the H3K27me3 promoter class. d, Strip chart showing expression of a comprehensive set of marker genes defined from human embryo single-cell data comparing naïve hESCs and naïve hESCs treated with EZH2i. Markers are grouped into pre-lineage, ICM, epiblast, primitive endoderm (PrE), trophectoderm (TE), cytotrophoblast (CTB), syncytiotrophoblast (STB), extravillous trophoblast (EVT), amnion, extra-embryonic mesoderm (exMes), advanced mesoderm (AdvMes), emerging mesoderm (EmMes), nascent mesoderm (NasMes), axial mesoderm (AxMes) and primitive streak (PriS). Significant differences (DESeq2 FDR <5% from triplicates) are highlighted in blue (downregulated) or orange (upregulated).

Source data

Given the strong upregulation of extra-embryonic lineage markers, we hypothesized that PRC2 shields naïve hESCs from trophectoderm differentiation. Indeed, among the most highly upregulated marker genes were ENPEP, the gene for a surface protein (APA) that marks trophectoderm progenitor cells capable of forming syncytiotrophoblast;9,55 ABCG2, a plasma membrane transporter highly expressed in trophectoderm and human placenta;5,56 TP63, a p53 family member defining the cytotrophoblast stem cell compartment5,57. Well-known transcription factors specifying trophectoderm lineage, CDX258,59, GATA39,59,60, MSX261 and NR2F262, were also specifically upregulated in EZH2i-treated naïve hESCs (Fig. 4a and Extended Data Figs. 6a–c). Investigating their promoter status, we noticed that transcription factors GATA3, CDX2, MSX2, NR2F2, GATA2 and Wnt agonist FRZB and AKT agonist IGF2 all shared bivalent promoters (Fig. 4a and Extended Data Fig. 6a). ChIP–seq in trophoblast progenitors has shown that GATA3 binds to the promoters of many marker genes of trophectoderm and placental development, including KRT23, VGLL1, CGA and TP639. Intersecting with the published binding data9, we found 20% of upregulated genes in EZH2i-treated naïve cells to be bound by GATA3 (Extended Data Fig. 6b). Genes like ENPEP, KRT23 and ERP27 lacked H3K4me3 in the naïve state but featured GATA3 binding sites (Fig. 4a and Extended Data Fig. 6c), suggesting they may be activated downstream of GATA3. We sought to confirm GATA3 induction on the protein level, also including a second PRC2 inhibitor, EED226 (EEDi), which targets the EED subunit63 in addition to EZH2i. Both inhibitors effectively reduced H3K27me3 at 7 days of treatment (Fig. 4b,c). PRC2 inhibition induced heterogeneity within colonies, with a fraction of cells acquiring GATA3 protein expression (11.9% for EZH2i and 2.6% for EEDi), concomitant with loss of NANOG (Fig. 4d). Similar induction of GATA3+ cells (6.3%) following EZH2 inhibition was confirmed in an additional naïve hESC line, HS975 (Extended Data Fig. 7a). In contrast, primed H9 hESCs did not show a loss of pluripotency markers or induced GATA3 expression with EZH2 inhibition (Extended Data Fig. 5f).

Fig. 4: Loss of H3K27me3 in naïve hESCs activates trophectoderm gene expression programmes.
figure 4

a, Heat map showing expression (TPM) in naïve and primed hESCs (± EZH2i treatment), as well as H3K4me3, H3K27me3 and H2Aub promoter status of selected trophectoderm and placenta-specific genes. RPGC from combined replicates are used for H3K4me3 and H3K27me3, whereas the three individual replicate TPM values are plotted for RNA-seq data. GATA3 binding as determined by ChIP–seq peaks during trophectoderm differentiation9 is indicated. For genome browser examples, see Extended Data Fig. 6a,c. b, Immunofluorescence confocal microscopy images of naïve H9 hESC colonies without treatment or with EZH2i (EPZ-6438) or EEDi (EED226) treatment for 7 days, assessing H3K27me3 levels, expression of pluripotency marker NANOG and trophectoderm transcription factor GATA3. Merged images show NANOG in purple, H3K27me3 in red, GATA3 in green, and Hoechst in blue. Scale bars, 10 µm. Data shown represent three independent experiments for EZH2i and two independent experiments for EEDi. For analogous experiment in naïve HS975 hESCs, see Extended Data Fig. 7a. c,d, CellProfiler image analysis of the experiment described in b. Per-nucleus H3K27me3 and GATA3 immunofluorescence intensities of a total of 670 nuclei derived from one experiment (165 untreated, 235 EZH2i and 270 EEDi). Box plot boxes show the 25th and 75th percentile with the median, and whiskers indicate 1.5 times the interquartile range. P values are determined by two-sided unpaired t-test (H3K27me3 P < 2.2 × 10−16 (top), GATA3 P = 2.4 × 10−5 for EZH2i and P = 0.046 for EEDi treatment). The threshold for GATA3+ cells (dashed line) is defined as 1.5× mean of untreated cells. Scatter plots contrasting per-nucleus GATA3 with H3K27me3 or NANOG immunofluorescence intensities. e, Scheme of experiment for CRISPR/Cas9-mediated acute deletion of EED. f, Representative image of experiment described in e. Example shown represents two independent experiments. Scale bars, 20 µm. g, CellProfiler image quantification of the experiment described in e. Scatter plots contrasting per-nucleus GATA3 and H3K27me3 immunofluorescence intensities of untransfected cells, NT gRNA or EED-targeting gRNA (EED gRNA) transfected cells. In total, 4,312 nuclei from two independent experiments were analysed. For statistics by individual colonies, see Extended Data Fig. 7b.

Source data

As an orthogonal approach to pharmacologic inhibition, we genetically targeted PRC2 using clustered regularly interspaced short palindromic repeats (CRISPR)/Cas9 (Fig. 4e). We electroporated naïve hESCs with synthetic Cas9/guide RNA (gRNA) complex and plated for clonal growth over 7 days (Fig. 4e,f). Targeting with EED gRNAs, we achieved complete loss of EED protein in a large fraction of colonies (Fig. 4g and Extended Data Fig. 7b). While we observed few GATA3-expressing cells within untreated (1.4% GATA3+) and NT-gRNA colonies (1.4% GATA3+), most of the EED-depleted colonies showed GATA3+ cells (total 16.7% GATA3+) (Extended Data Fig. 7b). On the single-cell level, GATA3+ cells were, with few exceptions, low or negative for EED (Fig. 4g).

To better resolve the cellular heterogeneity of the treated naïve cultures, we performed scRNA-seq during EZH2i treatment in naïve and primed cultures. To capture subpopulations and assign their cellular identity, we integrated a comprehensive cellular reference (Fig. 5a and Extended Data Fig. 8a,b), which included data from in vitro cultured blastocysts23,64, a human stem cell-based post-implantation amniotic sac embryoid (PASE) model65 and an in vivo gastrulation-stage human embryo specimen66. While over 97% of untreated naïve stem cells clustered with the expected pre-implantation epiblast cells of the reference, a small fraction (1.8%) clustered within the trophectoderm reference, suggesting a low but detectable spontaneous differentiation towards trophectoderm lineage in naïve cultures (Fig. 5b). A smaller fraction (1%) clustered within the mesoderm reference (Fig. 5b). This is in agreement with another single-cell study identifying a rare ‘intermediate population’ in naïve cultures53, which we confirmed to express trophectoderm and mesoderm markers (Extended Data Fig. 8d). With EZH2 inhibition, the fraction of these differentiated cells progressively reached 9.1% trophectoderm-like cells (TLCs) and 11.7% mesoderm-like cells (MeLCs) after 7 days (Fig. 5b). The MeLCs clustered mainly with the definitive mesoderm but to a lesser degree also to extra-embryonic mesoderm reference of the CS7 gastrula cells (annotated as yolk sac mesoderm). EZH2i treatment of primed cells, on the other hand, produced very few differentiated cells and 99.96% of the cells remained within the epiblast-like cell (ELC) cluster (Fig. 5c). MeLCs generated by our EZH2i treatment strongly expressed markers that also specified the mesoderm lineage within the reference datasets, namely TMEM88, LIX1 and PMP22 (Fig. 5d). TLCs expressed well-known trophectoderm factors MSX2, GATA2 and CLDN4 (Fig. 5d). TLCs and MeLCs also shared some upregulated genes, including KRT19, transcription factor HAND1 and CD24, a known surface marker of primed hESCs25,67 (Fig. 5d). Mapping selected marker genes on our UMAP reference, TMEM88 and GATA2 showed unique expression in cells of the MeLC and TLC population, respectively, and were essentially absent in ELCs (Fig. 5e). The well-established transcription factor for primitive streak and mesodermal lineage TBXT was induced after 7-day EZH2i treatment in a subset of cells of the ELCs and MeLCs (Fig. 5e). GATA3 appeared sporadically in ELCs and gained expression in TLCs (Fig. 5e). Although over 75% of the cells remained in the ELC population, our analysis indicated that the majority of cells display a transcriptional shift in the UMAP space with induction of lineage markers following 4 and 7 days of EZH2i treatment (Fig. 5b). Hence, we performed differential gene expression analysis within the ELC population following EZH2i treatment, identifying 482 up- and 387 downregulated genes (Extended Data Fig. 8e,f and Supplementary Table 4). Among the significantly upregulated genes were transcription factors HAND1, CDX1, CDX2, TBXT, TBX3, TFAP2A, SOX9, MEIS2 and TWIST1, (Fig. 5f,g), many of which were classical bivalent in naïve and primed hESCs. They were not expressed in primed hESCs, and gained additional H3K4me3 and low basal expression in naïve hESCs (Extended Data Fig. 9a,b).

Fig. 5: Single-cell transcriptomic profiling of EZH2i-treated naïve and primed cells.
figure 5

a, UMAP projection reference annotation (TE: trophectoderm; Exe_Mes: extra-embryonic mesoderm). The colour of each data point represents the cell annotations retrieved for each publication. Light-grey-coloured data points represent cells in the multi-data integration that were not used to infer cell annotations. The shape of data points for the embryonic cells indicates the data source. b UMAP projection and cell identity annotations of single-cell transcriptome datasets obtained for EZH2i-treated naïve and primed cells at indicated timepoint, showing ELCs, TLCs, MeLCs, AMLCs and HLCs. Percentage of cells within major clusters are indicated in the maps. Bar chart (bottom) shows the percentage of cells per cluster at each time point. c, UMAP projection and cell identity annotations as in b for primed hESCs treated with EZH2i. In ac, UMAP was generated from an extensive reference dataset, including data from human embryonic and PASE model (Extended Data Fig. 8a,b and Methods). d, Top ten significantly upregulated genes in MeLCs or TLCs, compared with ELC population. Bubble plot shows proportion of expressing cells and average expression per gene and condition. e, Single-cell expression of select TLC and MeLC markers mapped onto UMAP. log2 expression of selected genes overlaid on the UMAP. f, Selected genes differentially expressed within ELC population comparing 7-day-EZH2i-treated samples and untreated samples. Bubble plot shows proportion of expressing cells and average expression per gene and condition. g, Single-cell log2 expression of select factors mapped onto UMAP.

Source data

An ‘activated’ naïve state precedes lineage branching

To further describe how naïve hESCs exit from pluripotency, we performed a trajectory analysis using 864 differentially expressed genes across the different naïve hESC-derived populations and EZH2i time course. The analysis yielded a bifurcated trajectory with three principal states (Extended Data Fig. 9c), on which naïve hESCs first transition towards a bifurcation point from which two independent cell fates, trophectoderm and mesoderm, are accessed (Fig. 6a). Mapping our reference annotation (Fig. 5a,b) onto the trajectory, we were able to resolve a lineage bifurcation point already within the ELC population (Extended Data Fig. 9d). We subclassified cells within the ELC population by their position on the trajectory and pseudotime axis as follows: we termed the starting population ‘ground state’ ELC (gELC) and the trajectory leading up to the bifurcation point ‘activated’ ELC (aELC). We termed the ELC branch towards MeLC mesoderm-activated ELC (MaELC), and the other trophectoderm-activated ELC (TaELC) (Fig. 6c). In total, 91.4% of untreated epiblast-like naïve hESCs resided in the ground state (Fig. 6c,d). Intriguingly, 2.5% and 6.1% of single cells within the untreated naïve population also populated the activated and lineage states, respectively (Fig. 6c,d). This differentiation accelerated over 2 and 4 days of EZH2i treatment, and by 7 days 66.2% ELC cells were activated and branched into the two lineages at approximately even proportions (Fig. 6c,d). Interestingly, the fraction of aELCs plateaued at 16–18% after 4 and 7 days of EZH2i treatment, indicating that single cells shift through this state only transiently towards a stable, lineage-committed end state.

Fig. 6: Trajectory inference and gene expression dynamics for EZH2i-treated naïve hESCs.
figure 6

a,b, Trajectory inference colour-coded by inferred pseudotime and cell types. c, Further subclassification of ELC cells based on cell state and pseudotime distribution: gELC, aELC, TaELC and MaELC. Ridge plot (bottom) shows distribution of single cells along the pseudotime axis for naïve hESCs untreated or treated for 2, 4 and 7 days with EZH2i. Two dashed lines on ridge plot represent the pseudotime for transition from gELC to aELC, and bifurcation into MaELC and TaELC, respectively. d, Cells from untreated and EZH2i-treated (2, 4 and 7 days) naïve hESCs mapped onto trajectory. e, Volcano plot showing differentially expressed genes (24 up and 27 down) comparing aELC and gELC population (combined data from 4-day and 7-day EZH2i treatment of naïve hESCs). f, Volcano plot showing differentially expressed genes comparing mesoderm (MaELC, 67 genes) and trophectoderm (TaELC, 27 genes) branching at bifurcation point in 7-day-EZH2i treated naïve hESCs. g, Expression dynamics (pseudotime) of selected genes during EZH2i treatment on naïve cells. The confidence interval (95%) is indicated by bandwidth. MeLC branch and TLC branch are indicated by pink and blue lines, respectively.

Source data

As the trajectory analysis allowed us to separate early (generation of aELC) and late (bifurcation into TaELC and MaELC) steps in the differentiation process, we wanted to further dissect the trajectory underlying activation and lineage choice. First, we performed differential expression analysis comparing aELC and gELC to understand the process of activation. We identified 24 upregulated and 27 downregulated genes in the aELC, including CDX1, HAND1, TBXT and WNT5B, all key genes involved in streak and mesoderm induction (Fig. 6e and Supplementary Table 4).

Differential gene expression analysis comparing TaELC and MaELC identified 67 mesoderm-enriched and 27 trophectoderm-enriched genes (Fig. 6f, false discovery rate (FDR) <5%; Supplementary Table 4). For bifurcation towards mesoderm, we identified secreted factors LEFTY1, LEFTY2, NODAL and WNT5B, implying involvement of several signalling pathways, together with transcription factors TBXT, ID1, MIXL1 and UTF1. Towards trophectoderm, we confirmed core trophectoderm transcription factor TFAP2C, and additionally found surface proteins CD24 and CDH1, as well as tight junction protein CLDN4, among the top differentially expressed genes (Fig. 6f,g and Supplementary Table 4).

Finally, we wondered if we could delineate the definitive exit from pluripotency within our trajectory analysis. In agreement with population-wide downregulation of pluripotency factors already observed in bulk RNA-seq (Fig. 3a), we observed a continuous decline of NANOG and SOX2, while POU5F1 continued to be highly expressed throughout aELC, steeply dropping only after the bifurcation point (Fig. 6g). Together, this characterizes the activated ELC population as a state of opposing pluripotency and differentiation cues, from which more than one lineage trajectory remains accessed.

Discussion

Capturing quantitative differences between epigenomic profiles, as achieved with MINUTE-ChIP, has allowed us to uncover a more complex pattern of gains and losses of H3K27me3 between naïve and primed states than previously appreciated14,16,17,44. Importantly, we found that bivalency was largely maintained in the naïve state and identified hundreds of promoters that gained bivalency uniquely in the naïve state. Surprisingly, despite the extensive overlap of target promoters between naïve and primed states, inhibiting PRC2 catalytic activity activated independent gene sets, characteristic for each state. Hence, H3K27me3 per se cannot be attributed a constitutive repressive function at bivalent promoters; instead, we view bivalency as a default state of many developmental genes that can optionally acquire a transcriptionally poised state (that is, a state of low transcription in which solely H3K27me3 impedes activation) in the presence of cell-type- or lineage-specific activators.

We identified a network of transcription factors that may act as such activators and appear to cooperatively mediate activation and subsequent lineage commitment. These include factors already expressed in naïve pluripotency (TFAP2C, LEFTY2 and NODAL), and transcription factors, including TBXT, CDX1, HAND1 and GATA3, that drive the transition towards an activated state and lineage commitment. We observed the activated state to be populated with low frequency in naïve hESC cultures, and the transition may be reversible. However, a committing step may be reached at the bifurcation of mesoderm and trophectoderm lineages. Given the widespread promoter bivalency in naïve pluripotent cells, it is plausible that PRC2 basally represses early transcription factors and their targets, hence effectively intercepting extrinsic stimulatory signals (Wnt and Smad signalling) and positive feedback between the transcription factors itself. Recent studies highlight how multi-lineage transcription factors, including GATA3, TFAP2A/C, HAND1, TBXT and CDX1, cooperate in driving exit from primed pluripotency9,60.

By inhibiting or genetically ablating PRC2, we demonstrated an important function for H3K27me3 in counteracting the action of potent lineage-specifying factors in the pluripotent state. We identified trophectoderm and mesoderm as the endpoints of a differentiation invoked in the absence of H3K27me3. We traced the trajectory towards these lineages, including a bifurcation point at which a choice is made towards one of the two possible fates. Our data demonstrate an intrinsic propensity of naïve hESC cultures to produce bona fide trophectoderm and mesoderm lineages through defined differentiation trajectories and assign a previously undescribed role to PRC2 in shielding naïve pluripotency.

Our findings contrast in several aspects with earlier reports that PRC2 is dispensable for maintenance of naïve pluripotency in 5iLAF and t2iLGö media31,32. Shan et al.31 have reported that NANOG, OCT4 and SOX2 levels are unaffected in naïve EZH2−/− hESCs with additional short hairpin RNA-mediated depletion of EZH1. However, the remaining EZH1 or H3K27me3 levels were not explicitly assessed, leaving the possibility that our EZH2i treatment and genetic targeting resulted in a greater depletion of H3K27me3 and, thus, a more pronounced phenotype. Of note, the EZH2 inhibitor EPZ-6438 also inhibits EZH1 with nanomolar IC5068, probably achieving complete inhibition of EZH1 and EZH2 at the 10 µM concentration used here (Figs. 1b and 2a), and further experiments with pharmacologic inhibition and genetic ablation of EED, the core subunit of PRC2, corroborate our conclusion that PRC2 is indispensable to prevent differentiation of naïve hESC cultures in t2iLGö. In another study published in this issue, Zijlmans et. al.69 recapitulate key findings regarding the rewiring of bivalent promoters in naïve hESCs, but did not observe a differentiation response to PRC2 inhibition in naïve hESCs cultured in PXGL medium. This suggests that culture in PXGL medium alleviates the requirement for PRC2 to maintain naïve pluripotency. To test this hypothesis, we performed a direct comparison of PXGL and t2iLGö cultures and observed 0.2% GATA3+ cells after EZH2i treatment in PXGL versus 8.3% in t2iLGö (Extended Data Fig. 7c). Providing further rationale for this differences in response, the Tankyrase/WNT pathway inhibitor XAV939 contained in PXGL medium has been shown to reduce basal expression of GATA3 and TBXT in naïve hESCs70.

Genetic ablation of PRC2 in primed hESCs results in meso-endoderm development rather than trophectoderm development31,33. It is important to note that the process of exiting primed pluripotency and meso-endoderm speciation was not a rapid response to depletion of PRC2 in primed hESCs, but emerged over 3–4 weeks of culture31,33. Therefore, the transcriptional rewiring and appearance of TLCs and MeLCs within 7 days of inhibiting PRC2 in naïve hESCs (Fig. 5a) is in striking contrast to the slow exit from pluripotency of primed hESCs. This is confirmed by our control experiments treating primed hESCs with EZH2i (Fig. 5c), which did not detect any differentiation in the 7-day window. Together with the observations that inhibition of PRC2 enhances directed differentiation towards trophectoderm69, our data suggest that the naïve state is uniquely poised towards trophectoderm and mesoderm lineages through a unique bivalent gene signature of key developmental regulators.

In summary, our quantitative comparison of epigenetic landscapes in naïve and primed hESCs reveals an extensive rewiring of promoter bivalency in which PRC2-mediated H3K27me3 constitutes an important epigenetic barrier that dynamically follows and reinforces lineage choices and developmental progression. It will be an exciting focus of future work to elucidate if and how PRC2 is involved in guiding the totipotency-to-pluripotency transition in the early human embryo.

Methods

Culture of hESCs

Primed female H9 hESCs (Wicell, WA09) p28(11) were thawed using Nutristem hESC XF medium (Biological Industries, 05–100–1 A) and cultured on tissue culture plates (Sarstedt, 83.3922) pre-coated with 10% LN521 (Biolamina, LN521-02). Passaging was done when confluency reached 70–80%. Naïve H9 p28(8) naïve p7 that had been previously converted to the naïve stem cell state using NaïveCult induction medium (Stemcell Technologies, 05580) were thawed onto plates with high-density (strain ICR) inactivated mouse embryonic fibroblasts (MEFs; Gibco, A24903) using NaïveCult expansion medium (t2iLGö; Stemcell Technologies, 05590) supplemented with 10 µM ROCKi (Merck, Y-27632). Twenty-four hours after thawing, medium was changed to fresh NaïveCult expansion medium. Passaging was done every 4–5 days with medium supplemented with 10 µM ROCKi. Naïve HS97571 p35(17) cells were cultured in 5iLAF medium17 on inactivated MEFs (Gibco, A24903) under ethical permit 2011/745:31/3 issued by the Swedish Ethical Review Authority. All hESCs were cultured at 37 °C, 5% O2 and 5% CO2.

EZH2i treatment of H9 naïve and primed hESCs

For EZH2i treatment, primed and naïve cells were grown in respective media, supplemented with 10 µM EZSolution EPZ-6438 (BioVision, 2428-5) for 7 days, with daily medium changes. All cells were passage on d3-4 counting from start of treatment. H9 naïve cells were also maintained in PXGL70 consisting of N2B27 basal medium supplemented with 1 μM PD0325901, 2 μM XAV939, 2 μM Gö6983 and 10 ng ml−1 human leukaemia inhibitory factor, with or without EZH2i for 7 days before being fixed with 4% paraformaldehyde (FF-Chemicals, FFCHFF22023000) and used for immunofluorescence staining. For sample collection, primed and naïve H9 were dissociated using TrypLE Select (Gibco, 12563011). Primed cells were washed once with phosphate-buffered saline (PBS; Sigma, D8537) and counted using Moxi Z mini automated cell counter (Orflo, MXZ001). Collected naïve cells were centrifuged at 300g for 4 min, resuspended in fresh NaïveCult expansion medium and kept on ice. Mouse feeder removal microbeads (Milteny Biotech, 130-095-531) were used according to the manufacturer’s protocol to reduce the amount of MEFs in the naïve cell samples. In brief, naïve cells were mixed with the microbeads and incubated at 4 °C for 15 min. Meanwhile, the columns were equilibrated with NaïveCult medium. The cell–microbead suspension was added to the column mounted onto a magnet, allowing unbound naïve cells to pass through into a collection tube. Columns were rinsed using NaïveCult medium, and the flow-through was collected into the same collection tube.

EEDi treatment of H9 naïve hESC

Naïve H9 hESCs were maintained in NaïveCult expansion medium supplemented with 10 µM EED226 (Cayman Chemical, CAYM22031-10) or 2.5 µM UNC1999 (Sigma, SML0778). The latter was also applied to naïve H9 hESCs maintained in PXGL medium. Treatments lasted for 7 days, and the medium was changed daily. All cells were passaged on d3-4 counting from the start of treatment. On day 7, cells were fixed using 4% paraformaldehyde (FF-Chemicals, FFCHFF22023000) for 10 min. Samples were left in DPBS−/− (Gibco, 14190144) at 4 °C until proceeding with the immunofluorescence protocol.

EED gRNA and NT gRNA targeting in H9 naïve cells

Naïve cells were cultured as detailed above. MEF cells were plated 24 h before electroporation. On the day of electroporation, MEF wells were rinsed with DPBS−/− (Gibco, 14190144) before 2 ml fresh NaïveCult expansion medium supplemented with 10 µM ROCKi was added to the wells. Naïve colonies were dissociated into a single-cell suspension using TrypLE select and counted using the small particle setting on Moxi Z mini automated cell counter. The cell solution was passed through a 40 µm cell strainer (VWR, 732-2757) before it was centrifuged at 300g for 4 min. Meanwhile, 0.5 µl TrueCut Cas9 Protein v2 (Invitrogen, A36496) was complexed with 1 µl 100 µM EED or non-targeting (NT) gRNA (both from Synthego; GeneKnockout kit V2, GKO-HS1-000-0-1.5n-0-0), respectively, for 10–15 min. Supernatant was removed and the cell pellet was washed in PBS−/− and centrifuged at 300g for 4 min. The cell pellet was resuspended in the appropriate Buffer R volume as described in the manufacturer’s protocol. Cells were distributed into the tubes containing the ribonucleoprotein mixtures and briefly vortexed. Neon electroporator (Invitrogen, MPK5000) settings were: 1,400 V, 20 ms and 1 pulse. Medium was changed 24 h post-transfection, and daily medium changes followed until 7 days post-transfection when cells were fixed using 4% paraformaldehyde (FF-Chemicals, FFCHFF22023000). In the first electroporation experiment untransfected cells were used as control, and in the second electroporation experiment NT gRNA electroporated cells were used as control.

ChIP–seq

Triplicate pellets of 1 × 106 cells were collected for all conditions, flash frozen and stored at −80 °C before use. Mouse feeder removal microbeads (Milteny Biotech, 130-095-531) were used according to the manufacturer’s protocol to reduce the amount of MEFs in the naïve cell samples. Samples were prepared for ChIP–seq following the MINUTE-ChIP protocol (https://doi.org/10.17504/protocols.io.8nkhvcw)42. Briefly, native cell pellets were lysed, MNase digested to mono- to tri-nucleosome fragments and ligated with double-stranded DNA adaptors (containing T7 promoter, 8 bp sample barcode and a 6 bp unique molecular identifier (UMI)) in a one-pot reaction. Barcoded samples were then pooled and aliquoted into individual ChIP reactions with Protein A/G magnetic beads (Bio-Rad, 161- 4013/23) coupled with the desired antibodies (5 μg each of H3K27me3 (Millipore, 07-449), H3K4me3 (Millipore, 04-745) and H2AUb (Cell Signaling, 8240 S)). Upon incubation for 4 h at 4 °C with rotation and washing steps, ChIP DNA was isolated and set up in sequential reactions of in vitro transcription, RNA 3′ adapter ligation, reverse transcription and PCR amplification to generate final libraries for each ChIP (Extended Data Fig. 1a). After quality assessment and concentration estimation, libraries were diluted to 4 nM, combined and sequenced on the Illumina NextSeq500 platform with paired-end settings.

Bulk RNA-seq

Per condition, 1 × 106 cells were collected, resuspended in Buffer RLT (Qiagen, 74106) and spiked in with 5 × 104 Drosophila cells per sample. Then, total RNA was extracted using the RNeasy Plus Mini Kit (Qiagen, 74136) according to the manufacturer’s protocol. Mouse feeder removal microbeads (Milteny Biotech, 130-095-531) were used according to the manufacturer’s protocol to reduce the amount of MEFs in the naïve cell samples. Purified RNA quantities were estimated using the Qubit RNA HS assay kit (Life Technologies, Q32852), and the samples were subsequently flash frozen. RNA-seq libraries were generated and sequenced through BGI service (www.bgi.com) for strand-specific RNA-seq with poly(A) selection (DNBseq Eukaryotic Transcriptome De novo Sequencing).

scRNA-seq

Sample preparation

H9 naïve cells were treated with EZH2i for 2, 4 or 7 days and collected along with untreated control cells. Primed H9 cells were treated with the same inhibitor for 7 days and collected along with untreated control cells. Mouse feeder removal microbeads (Milteny Biotech, 130-095-531) were used according to the manufacturer’s protocol to reduce the amount of MEFs in the naïve cell samples.

Cell Multiplexing Oligo labelling (Cellplex)

Samples EZH2i d2, d4 and untreated control naïve as well as EZH2i d7 and untreated control primed cells were multiplexed following the manufacturer’s protocol. In brief, cells were dissociated, treated with mouse feeder removal beads (naïve cells), counted and resuspended in Cell Multiplexing Oligo. After oligo incubation, cells were washed three times using DPBS−/− containing 1% BSA (Sigma Aldrich, A7284) and counted using the NucleoCounter NC-200 automated cell counter. At this point, cells were pooled accordingly: EZH2i d2, d4 and untreated naïve cells, and EZH2i d7 and untreated primed cells, respectively, and each pool was counted once more.

Chromium Next GEM Single Cell 3′ Reagent Kits v3.1 (dual index)

Complementary DNA preparation and library construction was done according to the manufacturer’s detailed protocol. Briefly, the appropriate number of cells, as defined by the protocol, were complexed with gel beads in emulsion and barcoded. After cleanup, cDNA was amplified and purified. For the samples that were multiplexed, the supernatant as well as the pellet were purified, whereas for the single samples only the pellet was purified. Samples were analysed using TapeStation High Sensitivity 5000 at the Bioinformatics and Expression Analysis core facility at Karolinska Institutet. Following the quality control results, the samples went through fragmentation, end repair and A-tailing followed by adaptor ligation and index PCR. The transferred supernatant from the multiplexed samples was indexed through a sample index PCR. Once samples were purified and eluted in Buffer EB (Qiagen, 19086), they were analysed using TapeStation High Sensitivity 1000 at the Bioinformatics and Expression Analysis facility. Finally, cDNA libraries were sequenced using Illumina Nextseq 2000 Platform 100 cycles P2 and Illumina Nextseq 550 High Output Kit v2.5 (150 cycles).

Immunofluorescence

Samples fixed with 4% paraformaldehyde as described above and permeabilized using 0.3% Triton X-100 (Sigma Aldrich, T9284-100ML) in PBS (Gibco, 14190144) for 10 min, after which three washes with PBS were carried out. The samples were blocked for 2 h using 0.1% Tween-20 (Sigma Aldrich, P9416-100ML) and 4% FBS (Thermo Fisher, 10082147) in PBS. Primary antibodies (GATA3 clone L50-823 (1:200; BD, 558686), H3K27me3 C36B11 (1:500; Cell Signaling Technologies, 9733 S), OCT4 (1:200; SantaCruz, sc-5279), SOX2 clone EP103 (1:3; Biogenex, AN833), NANOG (1:200; RnD, AF1997-SP) and EED (E4L6E) XP (1:200; Cell Signaling Technology, 85322)) were diluted in blocking solution and added to the samples, which were incubated at 4 °C overnight. Excess antibodies were washed away using blocking buffer. Secondary antibodies, donkey anti-mouse IgG (H + L) Alexa Fluor 555, donkey anti-rabbit IgG (H + L) Alexa Fluor 647, donkey anti-goat IgG (H + L) Alexa Fluor 647, donkey anti-mouse IgG (H + L) Alexa Fluor 488, donkey anti-rabbit IgG (H + L) Alexa Fluor 555 (all from Thermo Fisher; A-31570, A-31573, A-21447, A-21202 and A-31572, respectively), were diluted in blocking solution, added to the samples and then incubated for 2 h at room temperature. Again, excess antibodies were washed away and samples were incubated with Hoechst 33342 (Thermo Fisher, H3570), which was followed by another set of washes. Samples were mounted using DAKO fluorescent mounting medium (DAKO, S3023). Images were acquired using a Nikon Eclipse Ti spinning disk confocal microscope with a 20× air and 60× oil immersion objective, respectively, and Z-stacks were analysed using ImageJ.

Image analysis

Image analysis was performed using CellProfiler 4 software72. Briefly, cell nuclei were segmented on the basis of Hoechst staining (three-class global thresholding strategy with Otsu method, threshold smoothing scale of 1.3488) and the intensity of these objects/nuclei across all fluorescence channels was quantified. The image analysis pipeline and raw data are provided as supplementary information. The resulting raw data (mean immunofluorescence intensities) were analysed and figures generated using R. Statistics were plotted with ggpubr (v0.4.0) package. Cells were counted as GATA3 positive if their mean intensity was above 1.5 times the mean of the GATA3 intensity distribution in the matching control sample across all experiments.

Immunoblotting

For each growth condition, 1 × 106 cells were lysed in 100 μl of ice-cold radioimmunoprecipitation buffer (0.1% sodium deoxycholate, 0.1% SDS, 1% Triton X-100, 10 mM HEPES (pH 7.6), 1 mM EDTA, 5% glycerol and 140 mM NaCl) supplemented with Protease Inhibitor Cocktail (PIC, Roche) on ice for 10 min. Lysates were homogenized by sonication for eight to ten cycles at high power, 30 s on/off in a Bioruptor sonicator (Cosmo Bio). Samples were boiled at 95 °C for 5 min with 6× SDS sample buffer before loading onto 4–20% Tris–glycine gels (Bio-Rad). Resolved proteins were transferred to nitrocellulose membranes using the Trans-Blot Turbo system (Bio-Rad) according to the manufacturer’s instructions. Membranes were then blocked for 1 h in 1% casein prepared in Tris-buffered saline and 0.1% Tween-20 (TBS-T) before blotting with respective primary antibodies diluted in TBS-T, overnight at 4 °C. Blots were washed three times with TBS-T and incubated with secondary antibodies in the same buffer for 1 h at room temperature (protect from light). After three TBS-T washes, the membranes were imaged on a LI-COR Odyssey FC system. Quantitation of signal and analysis was performed using the LI-COR Image studio software. Primary antibodies included total H3 1:10,000 (Active Motif, 39763), H3K4me3 1:5,000 (Millipore, 04-745), H3K27me3 1:5,000 (Millipore, 07-449) and H2Aub (Cell Signaling, 8240 S). The secondary antibodies were IRDye 680RD anti-rabbit and IRDye 800CW anti-mouse (LI-COR) at 1:5,000 dilution.

MINUTE-ChIP analysis

Preparation of FASTQ files

Sequencing was performed using 50:8:34 cycles (read 1:index 1:read 2). Illumina bcl2fastq was used to demultiplex paired-end sequencing reads by 8 nt index 1 read (PCR barcode). NextSeq lanes were merged into single fastq files, creating the primary fastq files. Read 1 starts with 6 nt UMI and 8 nt barcode in the format NNNNNNABCDEFGH.

Primary analysis

MINUTE-ChIP multiplexed FASTQ files were processed using minute73, a data processing pipeline implemented in Snakemake74. To ensure reproducibility, a conda environment was set. Source code and configuration are available on GitHub (https://github.com/NBISweden/minute). Main steps performed are described below.

Adaptor removal

Read pairs matching parts of the adaptor sequence (SBS3 or T7 promoter) in either read 1 or read 2 were removed using cutadapt v3.275.

Demultiplexing and deduplication

Reads were demultiplexed using cutadapt (v3.2)75 allowing only one mismatch per barcode. Demultiplexed reads were written into sample-specific FASTQ files used for subsequent mapping and GEO submission.

Mapping

Sample-specific paired FASTQ files were mapped to the human genome (hg38) using bowtie2 (v2.3.5.1)76 with –fast parameter. Alignments were processed into sorted BAM files with samtools (v1.10)77. Pooled BAM files were generated from replicates using samtools.

Deduplication

Duplicate reads are marked using UMI-sensitive deduplication tool je-suite (v2.0.RC)78. Read pairs are marked as duplicates if their read1 (first-in-pair) sequences have the same UMI (allowing for one mismatch) and map to the same location in the genome. Blacklisted regions were then removed from BAM files using BEDTools (v2.29.2)79.

Generation of coverage tracks and quantitative scaling

Input coverage tracks with 1 bp resolution in BigWig format were generated from BAM files using deepTools (v3.5.0)80 bamCoverage and scaled to a reads per genome coverage (RPGC) of one (1× RPGC, also referred to as ‘1× normalization’) using hg38 genome size 3095978588. ChIP coverage tracks were generated from BAM files using deepTools (v3.5.0) bamCoverage. Quantitative scaling of the ChIP–seq tracks among conditions within each pool was based on their input-normalized mapped read count (INRC). INRC was calculated by dividing the number of unique hg38-mapped reads by the respective number of input reads: #mapped[ChIP]/#mapped[Input]. This essentially corrected for an uneven representation of barcodes in the input, and we previously demonstrated that the INRC is proportional to the amount of epitope present in each condition42. Untreated naïve hESC (pooled replicates) was chosen as the reference condition, which was scaled to 1× coverage (also termed RPGC). All other conditions were scaled relative to the reference using the ratio of INRCs multiplied by the scaling factor determined for 1× normalization of the reference: (#mapped[ChIP]/#mapped[Input])/(#mapped[ChIP_Reference]/#mapped[Input_Reference]) × scaling factor.

Quality control

FastQC was run on all FASTQ files to assess general sequencing quality.

Picard (v2.24.1) was used to determine insert size distribution, duplication rate and estimated library size. Mapping statistics were generated from BAM files using samtools (v1.10)77 idxstats and flagstat commands. Final reports with all the statistics generated throughout the pipeline execution are gathered with MultiQC81.

Downstream analysis and visualization

Total mapped read counts from BAM files were used to calculate relative global levels of histone modifications. Summary values for fixed sized bins or custom intervals were calculated from scaled BigWig files using wigglescout (v0.13.5).

(https://github.com/cnluzon/wigglescout), which aggregates data calculated using rtracklayer (v1.54.0)82. Differentially H3K27me3-enriched bins and gene transcription start site (TSS) were calculated using DESeq2 (v1.32.0)83 using fixed size factors since BigWig files were already scaled by the minute pipeline. Data handling was done using tidyverse (v1.3.1)84 suite. Figures were created using R (v4.1.0) ggplot2 (v3.3.5)85. Additionally, combined heat maps were created using heatmaply (v1.2.1)86 and extra statistics were plotted with ggpubr (v0.4.0) package when required. Figure 1i was made with ggridges (v0.5.3). Figure 1g was made with karyoploteR package (v1.18.0)87. Figure 2b was done with ggalluvial (v0.12.3)88. Genome track figures were made using Integrative Genomics Viewer89 image export function. Corresponding source code for data analysis and figures can be found in the GitHub repository companion to this publication (https://github.com/elsasserlab/hesc-epigenomics). The repository was built using workflowr (v1.6.2)90. Data analyses were rendered from R markdown notebooks, and results can be navigated at the corresponding website, generated using workflowr package (v1.6.2)90.

RNA-seq analysis

Primary analysis

Bulk RNA-seq data were analysed by RNA-seq pipeline (v2.0) available on nf-core91 (https://nf-co.re/rnaseq/2.0) with hg38 as reference and RefSeq gene annotation, using STAR92 as read aligner and RSEM93 to quantify read counts.

Downstream analysis and visualization

Read counts produced by RSEM were used as input for DESeq283 differential expression analysis with default parameters. log2 fold-change (log2FC) shrinkage apegm94 was used to filter low-read-count genes. Significance adjusted P value cut-off was set to P < 0.05 and a fold change of 2. In heat map RNA-seq figures, transcripts per million (TPM) values are shown as (log2(TPM) + 1) instead of raw counts. Figures were created using R (v4.1.0) ggplot2 (v3.3.5). Additionally, combined heat maps were created using heatmaply (v1.2.1) and extra statistics were plotted with ggpubr (v0.4.0) package when required.

Pre-processing scRNA-seq data

Raw scRNA-seq FASTQ files for EZH2i d2, d4 and untreated naïve cells, and EZH2i d7 and untreated primed cells, were aligned to the human GRCh38 reference genome (v.3.0.0, GRCh38, downloaded from the 10X Genomics website) using Cell Ranger version 6.1.1 with default settings for the ‘cellranger multi’ pipeline (10X Genomics). Reads from EZH2i d7 naïve were aligned on the same reference with the default setting for the ‘cellranger count’ pipeline (10X Genomics). To minimize differences associated with the sequencing platform used and data processing, published 10X scRNA-seq reads from the PASE model65 were reprocessed as the same as for EZH2i d7. Published Smart-Seq2 datasets were also mapped on the same reference using the same aligner with default settings, and only uniquely mapped reads were kept for gene expression quantification. Raw read counts were further estimated using rsem-calculate-expression from RSEM tool93 with the option of ‘–single-cell-prior’. To filter low-quality cells, a cut-off based on the number of expressed genes (nGene) and percentage of mitochondrial genes (percent.mito) was used. High-quality cells from the EZH2i-related dataset were required to have 1,000 < nGene < 6,000 and percent.mito less than 0.15. Cells from published Smart-Seq2 datasets were required to have at least 2,000 nGene and percent.mito less than 0.125. Cells from PASE model were filtered using the same cut-off reported in their paper65. Cells belonging to haemogenic endothelial progenitors and erythroblasts from the CS7 gastrula66 were excluded in the following analysis. After quality control and exclusion of mitochondrial genes, we focused on genes with one or more counts in at least five cells (assessed for each dataset separately). We then calculated the log-normalized counts using the deconvolution strategy implemented by the computeSumFactors function in R scran package (v1.14.6)95 followed by rescaled normalization using the multiBatchNorm function in the R batchelor package (v1.2.4)96 so that the size factors were comparable across batches. log-Normalized expression after rescaling was then used for the human dataset integration and differentially expressed gene detection.

Data integration, dimensionality reduction and clustering

As discussed in ref. 54, to avoid over-integration of the amnion with the trophectoderm cells, mutual nearest neighbours were utilized to integrate all datasets together (cells from naïve cells and primed cells were treated as separate datasets). In detail, this was done by performing a principal component analysis using the top 2,500 most variable genes selected by RunFastMNN function and then correcting the principal components according to their mutual nearest neighbours. We then selected the corrected top 30 principal components for downstream UMAP dimensional reduction and clustering analysis using the RunUMAP function and FindClusters function in the R Seurat (v3.1.4) package97. The clustering resolution was set to 0.4 when using the FindClusters function. Cluster identities for 14 individual clusters were assigned manually using co-localization of cell identities from all embryonic datasets as reference, as well as key marker gene and gene signature expression. In detail, cells from EZH2i treated and untreated naïve cells were reclassified as TLCs (C7 and C11), ELCs (C0, C2, C4, and C12), hypoblast-like cells (HLCs) (C9), amnion-like cells (AMLCs) (C3 and C5) and MeLC (C1 and C10), respectively (Extended Data Fig. 8b). The remaining EZH2i-treated and untreated naïve cells (16 cells) were assigned as undefined cells. Fifty and 41 cells annotated as ‘ICM’ and ‘PSA-EPI’ from Xiang et al. were excluded for UMAP visualization as they are misclassified54,98. Cells from pre-implantation TE, cytotrophoblast, syncytiotrophoblast and extravillous trophoblast were all labelled as ‘TE’, and cells from nascent mesoderm, emergent mesoderm, advanced mesoderm and axial mesoderm were merged as ‘mesoderm’ on UMAP visualization. A shiny app generated by shinyCell99, which includes the above integration, can be browsed at https://petropoulos-lanner-labs.clintec.ki.se.

Trajectory inference and pseudotime analysis

Single-cell trajectory analysis of EZH2i-treated and untreated naïve cells was utilized by R Monocle2 package (v.2.14.0) using DDRTree method with default parameters100. In detail, raw UMI count was fed as the input for Monocle2. After default normalization, 864 top differentially expressed genes between clusters with q value less than 1 × 10−20 were identified by differentialGeneTest function, followed by reconstructing the single-cell trajectory, estimating functional states of cells and calculating pseudotime by reduceDimension and orderCells function. On the basis of the assumption that 95% untreated naïve cells belong to ‘ground state’ for differentiating, cells with a pseudotime less than 5.949 were identified as the ground ELC. Activated ELC was defined as the ELC cells with pseudotime higher than 5.949, which were further divided into activated ELC, trophectoderm-activated ELC and mesoderm-activated ELC on the basis of their functional states.

Detection of significant differentially expressed genes using scRNA-seq data

Detection of differentially expressed genes from scRNA-seq data was performed using ‘MAST’ test101 implemented in FindMarkers function from Seurat package. Genes with an FDR less than 0.05, with a fold change more than 1.5 and expressed in more than 10% of the cells were considered differentially expressed. Pseudocount 0.1 was added to avoid dividing by 0.

Lineage marker-gene detection

Gene expression from E-MTAB-392923, GSE13644764 and E-MTAB-938866 were used to select lineage marker genes. According to published annotations (epiblast cells from E5-E7, E8-E14 and Carnegie stage 7 were grouped as Early Epi, Middle Epi and Late Epi, respectively), pairwise differential expression analysis between lineages was performed using the ‘roc’ test of FindMarkers function from R package Seurat97. The top 50 upregulated marker genes with average power of more than 0.3 conserved in all comparisons were selected.

Statistics and reproducibility

For each ChIP, data from three biological replicates were analysed, fulfilling the ENCODE Consortium recommendations of two biological replicates for ChIP–seq data analysis102. No statistical method was used to pre-determine sample size. No data were excluded from the analyses. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.