Efficient endoderm induction from human pluripotent stem cells by logically directing signals controlling lineage bifurcations

SUMMARY

Human pluripotent stem cell (hPSC) differentiation typically yields heterogeneous populations. Knowledge of signals controlling embryonic lineage bifurcations could efficiently yield desired cell-types through exclusion of alternate fates. Therefore we revisited signals driving induction and anterior-posterior patterning of definitive endoderm to generate a coherent roadmap for endoderm differentiation. With striking temporal dynamics, BMP and Wnt initially specified anterior primitive streak (progenitor to endoderm), yet 24 hours later suppressed endoderm and induced mesoderm. At lineage bifurcations, cross-repressive signals separated mutually-exclusive fates: TGFβ and BMP/MAPK respectively induced pancreas versus liver from endoderm by suppressing the alternate lineage. We systematically blockaded alternate fates throughout multiple consecutive bifurcations, thereby efficiently differentiating multiple hPSC lines exclusively into endoderm and its derivatives. Comprehensive transcriptional and chromatin mapping of highly-pure endodermal populations revealed that endodermal enhancers existed in a surprising diversity of “pre-enhancer” states before activation, reflecting establishment of a permissive chromatin landscape as a prelude to differentiation.

INTRODUCTION

At developmental junctures, multipotent progenitors choose between multiple fates (Graf and Enver, 2009; Loh and Lim, 2011). Extrinsic signals often instruct a particular fate while repressing alternate lineages. It is critical to decipher the extrinsic signals that direct such lineage segregations in order to efficiently differentiate human pluripotent stem cells (hPSC) into pure populations of desired cell-types in the absence of mutually-exclusive, unwanted lineages. However the precise lineage outcomes specified by these signals at particular bifurcations remain to be fully clarified, despite informative insights from in vivo genetic perturbations (Tam and Loebel, 2007; Zorn and Wells, 2009) and explant approaches (Bernardo et al., 2011; Deutsch et al., 2001). Pertinent issues include how alternate lineages are segregated at each branchpoint as well as the exact order and kinetics of dynamic signaling switches that drive successive cell-fate transitions (Wandzioch and Zaret, 2009).

The present work revisits signaling dynamics that drive induction and anterior-posterior patterning of the definitive endoderm (DE) germ layer and subsequent organ formation. DE is the embryonic precursor to organs including the thyroid, lungs, pancreas, liver and intestines (Švajger and Levak-Švajger, 1974). The pluripotent epiblast (E5.5 in mouse embryogenesis) differentiates into the anterior primitive streak (~E6.5), which generates DE (~E7.0-E7.5) (Lawson et al., 1991; Tam and Beddington, 1987). DE is then patterned along the anterior-posterior axis into distinct foregut, midgut and hindgut territories (~E8.5) and endoderm organ primordia arise from specific anteroposterior domains (~E9.5) (Zorn and Wells, 2009).

Various methods to differentiate hPSC towards DE employ animal serum, feeder co-culture or defined conditions (Cheng et al., 2012; D'Amour et al., 2005; Touboul et al., 2010) but typically yield a mixture of DE and other contaminating lineages, with induction efficiencies fluctuating between hPSC lines (Cohen and Melton, 2011; McKnight et al., 2010). Viewed from the perspective of lineage bifurcations, these mixed lineage outcomes might stem from incomplete exclusion of alternate fates at such junctures. Heterogeneous early DE populations harboring contaminating lineages complicate the subsequent generation of endodermal organ derivatives (McKnight et al., 2010).

In vertebrate embryos and during PSC differentiation, TGFβ/Nodal/Activin signaling is imperative for DE specification whereas BMP broadly induces mesodermal subtypes (e.g., Bernardo et al., 2011; D'Amour et al., 2005; Dunn et al., 2004). Yet TGFβ signaling (even with additional factors) is insufficient to specify homogeneous DE (quantified by Chetty et al., 2013). BMP, FGF, VEGF and Wnt have also been employed together with TGFβ signals to generate DE (Cheng et al., 2012; Green et al., 2011; Kroon et al., 2008; Nostro et al., 2011; Touboul et al., 2010). However these factors have also been implicated in mesoderm formation (Davis et al., 2008) and their precise involvement in DE induction remains to be clarified.

We have systematically elucidated how mutually-exclusive lineages are separated at 4 consecutive steps of endoderm development: PS induction; segregation of endoderm versus mesoderm germ layers; DE anterior-posterior patterning; and bifurcation of liver and pancreas. Accurately defining which signals instructed or repressed specific fates at each endodermal bifurcation enabled homogeneous hPSC differentiation down one path or the other. Knowledge of precise temporal signaling dynamics, combined with efficient differentiation throughout successive developmental steps, culminated in a single strategy to universally differentiate diverse hPSC lines into pure populations of endodermal lineages by excluding alternate lineages at each branchpoint. This altogether provides a coherent view of signaling logic underlying multiple steps of endoderm induction and patterning. This also furnishes the means to molecularly profile highly homogeneous endoderm populations, allowing us to comprehensively capture transcriptional and chromatin dynamics underlying endoderm specification.

RESULTS

BMP, FGF, TGFβ and Wnt initially establish the primitive streak and anteroposteriorly pattern it

This work was preceded by findings that Activin, in conjunction with FGF, BMP and a PI3K inhibitor (“AFBLy”) (Touboul et al., 2010) or together with animal serum (D'Amour et al., 2005), induced hESC towards DE. However we and others (Chetty et al., 2013) observed these methods still yielded mixed lineage outcomes, evident during the differentiation of 5 hESC lines (Fig. 1a, Fig. 2b-c, Fig. S1-3). For example, AFBLy (Touboul et al., 2010) concurrently generated mesoderm, upregulating skeletal, vascular and cardiac genes (P<10⁻⁸; Fig. 1a, Fig. S1a-d), whilst Activin and serum treatment (D'Amour et al., 2005) yielded a proportion of undifferentiated cells (Fig. 2b,c,f). Creation of impure early DE populations might explain the emergence of non-endoderm lineages after downstream differentiation (Kroon et al., 2008; Rezania et al., 2012).

Figure 1. Dynamic signaling switch for primitive streak and endoderm formation.

A) Microarray analysis of genes upregulated >2-fold during AFBLy treatment of H9 hESC (Touboul et al., 2010) and GO analysis

B) To test effects of increasing FGF2 (10-40 ng/mL), Wnt3a (15-100 ng/mL), CHIR99021 (50-1000 nM) or BMP4 (3-20 ng/mL) (panels i, ii, iii and iv, respectively) and respective inhibitors (100 nM PD173074, 2 μM IWP2, 150 ng/mL Dkk1 and 250 nM DM3189) on PS formation, H1 hESC were differentiated towards PS for 24 hours with indicated base combinations of Activin (100 ng/mL), FGF2 (20 ng/mL) and 10 μM LY294002 (“AFLy” or “ALy”) in conjunction with the indicated signaling perturbations, and qPCR was performed (day 1)

C) To test effects of increasing BMP, FGF or Wnt signaling (10 ng/mL BMP4, 3 μM CHIR and 5-20 ng/mL FGF2; panels i, ii and iii respectively) on DE vs. mesoderm emergence from PS, H1 hESC were initially differentiated with AFBLy towards PS for 24 hours and then subsequently differentiated with AFLy, AFBLy or ALy + 250 nM DM3189 (“ADLy”) for 48 subsequent hours with indicated signaling perturbations, and qPCR was performed (day 3)

D) HES3 MIXL1-GFP⁺ PS (Davis et al., 2008) induced by 100 ng/mL Activin, 2 μM CHIR and 50 nM PI-103 (ACP) by day 1 of differentiation was blocked by concomitant addition of BMP inhibitors (300 ng/mL noggin or 250 nM DM3189) (i); truth table (ii) and schematic (iii) of dynamic signaling during differentiation See also Fig. S1

Figure 2. Efficient definitive endoderm induction in defined conditions by SR1.

A) H1 hESC differentiated by ACP for 24 hours stained for BRACHYURY, FOXA2, EOMES and LHX1 (nuclear staining by DAPI), scale bar=100 μm for all subsequent figures (left); FACS shows >99% of HES3 hESC are MIXL1-GFP⁺ (Davis et al., 2008) after 24 hours of ACP treatment (right)

B) Microarray heatmap of independent triplicates; undifferentiated HES3 hESC (Day 0), ACP-induced APS (Day 1), SR1-induced DE (Day 3) or hESC differentiated by AFBLy (Touboul et al., 2010) or serum (D'Amour et al., 2005) for 3 days

C) FOXA2 and SOX17 staining of SR1-, serum- or AFBLy-treated H1 hESC after 3 days of differentiation (top); summary of CXCR4⁺PDGFRα⁻ DE percentages in hPSC (grey) or after SR1 differentiation (blue) from 7 hPSC lines, dots depict experimental replicates (bottom left); histogram summarizing CXCR4⁺PDGFRα⁻ DE percentages after various differentiation protocols, error bars depict standard deviation (bottom right)

D) FACS analysis of H9 SOX17-mCHERRY hESC; reporter expression before or after 2 days of SR1 differentiation

E) FACS analysis of CXCR4 and PDGFRα expression before or after SR1 differentiation from indicated hPSC lines

F) Single-cell qPCR heatmap of 80 single cells (H7 hESC, or those differentiated by SR1, AFBLy or serum for 2 days)

G) To test their neural competence, H1 hESC after 0-2 days of SR1 induction were transferred (“→”) into neuralizing media (“N”, 3 days) and neural gene expression was compared to SR1-induced DE (“day 3 DE”), see Supplementary Experimental Procedures See also Fig. S2, S3

Guided by prior in vivo and in vitro findings, we selectively perturbed developmental signals (>3,200 signaling conditions) at specific embryonic stages of hPSC differentiation in serum-free conditions and assessed resultant lineage outcomes by qPCR (yielding >16,000 datapoints, Fig. S1-4). These signaling perturbations revealed elements of the signaling logic underlying DE induction (Fig. 1-4).

Figure 4. Bifurcation of liver versus pancreas from posterior foregut.

A) To test effects of increasing amounts of (i-iii) BMP/TGFβ signaling or (iv) FGF/MAPK signaling on pancreas vs. liver induction, day 3 DE was differentiated with indicated conditions with (i-ii) 5-20 ng/mL Activin or (ii-iii) 5-10 ng/mL BMP4 and respective inhibitors (1 μM A8301, 250 nM DM3189, 100 nM PD173074, 500 nM PD0325901) where indicated. Abbreviations for base conditions: (i) RS = 2 μM RA + SANT1; (ii-iii) RS+PD = RS + PD0325901; (iv) DRK = DM3189 + RA + KAAD-Cyclopamine

B) Depictions of (i) dynamic signaling inputs, (ii) truth table and (iii) dichotomy of BMP and TGFβ signaling for liver versus pancreas induction

C) AFP immunostaining of day 7 early liver progenitors and quantification

D) Substrate luciferase assay for CYP3A4 metabolic activity (i) and staining for LDLR expression and LDL-DyLight 594 uptake (ii) in hESC-derived late hepatic progeny

E) CAG-GFP⁺ hESC differentiated into early hepatic progenitors or late hepatic progeny were transplanted (Chen et al., 2013) (top left); human albumin levels in mouse sera, each dot is an individual mouse (fractions of successfully engrafted mice indicated; top right); recipient whole-liver cross-section with different lobes and subfields indicated, scale bar=5mm (middle right); costaining for human albumin and GFP in four distinct hepatic lobes, fields numbered above (bottom) See also Fig. S4, S5

In vivo, DE arises from the primitive streak (PS, ~E6.5) (Levak-Švajger and Švajger, 1974). The anteriormost PS (APS) generates DE (~E7.0-E7.5) whereas posterior PS (PPS) forms mesoderm (Lawson et al., 1991; Tam and Beddington, 1987). Determinants of anterior versus posterior PS from hPSC remain to be elucidated.

We found both APS and PPS were combinatorially induced by BMP, FGF and Wnt on day 1 of hESC differentiation. These signals have been individually implicated in PS induction (Bernardo et al., 2011; Blauwkamp et al., 2012; Gadue et al., 2006) but their roles in PS patterning have not been dissected in detail. If either BMP, FGF or Wnt was inhibited, both APS and PPS formation failed (Fig. 1b), corroborating the lack of PS in BMP and Wnt pathway knockout mice (Beppu et al., 2000; Liu et al., 1999; Mishina et al., 1995). FGF signaling was equally permissive for both APS and PPS emergence and endogenous FGF was sufficient to drive either outcome (Fig. 1bi, Fig. S2a-c). However, exogenous Wnt (either Wnt3a or GSK3 inhibition [CHIR]) was necessary to maximize PS induction and Wnt broadly promoted both APS and PPS (Fig. 1bii-iii). Limited PS formation could occur without exogenous Wnt but was dependent on endogenous Wnt (Fig. 1bii). BMP levels arbitrated between APS and PPS: lower (endogenous) BMP levels elicited APS whereas higher BMP yielded PPS (Fig. 1biv, Fig. S2b). Nonetheless, the absolute necessity of BMP for MIXL1-GFP⁺ APS induction (Fig. 1di, P<0.025) was unexpected as BMP was typically thought to be posteriorizing (Bernardo et al., 2011). Therefore, FGF, Wnt and low BMP were essential for APS specification.

A dynamic switch in BMP and Wnt signaling induces primitive streak but subsequently suppresses definitive endoderm emergence

To further differentiate APS towards DE, prior studies used similar factors to induce both lineages over 3-5 days (Nostro et al., 2011; Touboul et al., 2010). Instead we found APS and DE were sequentially driven by diametrically opposite signals within 24 hours of differentiation. BMP and Wnt initially specified APS from hESC on day 1, but 24 hours later, BMP and Wnt induced mesoderm and reciprocally repressed DE formation from PS on days 2-3 (Fig. 1ci-ii). Interestingly, not only removing exogenous BMP but neutralizing endogenous BMP (using noggin or DM3189/LDN-193189) was critical to eliminate mesoderm and to reciprocally divert PS differentiation towards DE (Fig. 1ci). This was evinced by ~3000-fold downregulation of MESP1 and concurrent upregulation of SOX17, HHEX, FOXA1 and FOXA2 in 2 hESC lines (Fig. S1c-e). Given that prolonged BMP and Wnt were known to induce mesoderm (Bernardo et al., 2011; Gadue et al., 2006), our results altogether argue against prior sustained BMP treatment to induce DE (Cheng et al., 2012; Nostro et al., 2011; Touboul et al., 2010), which we show abrogated DE and instead specified mesoderm. Timed BMP inhibition also improved DE induction from mESC, although which developmental step(s) it benefited remained unclear (Sherwood et al., 2011). In summary, understanding the precise kinetics of BMP signaling was essential to thwart extraneous mesoderm production.

Similarly, endogenous Wnt/β-catenin signals directed PS towards mesoderm, such that inhibiting endogenous Wnt (using IWP2, Dkk1 or XAV939) on days 2-3 blocked mesoderm formation from 2 hESC lines (Fig. 1cii, Fig. S1g-h). However, individually inhibiting either BMP or Wnt was sufficient to abolish mesoderm indicating that inhibiting both was redundant (Fig. S1h). Thus we subsequently only inhibited BMP to derive DE from PS. Finally, our results contrast with prolonged Wnt treatment to induce DE (Sumi et al., 2008), which we show instead specified mesoderm from PS and blocked DE. Altogether, BMP and Wnt induced mesoderm from PS and suppressed endoderm; therefore their inhibition ablated mesoderm and diverted differentiation towards DE.

While BMP and Wnt specified mesoderm (Gertow et al., 2013), we found DE formation from PS was jointly driven by FGF (Fig. 1ciii) and TGFβ signaling (Bernardo et al., 2011; D'Amour et al., 2005). If FGF was inhibited, mesoderm formation was re-enabled even in the absence of pro-mesodermal BMP (Fig. 1ciii), showing FGF prevented illegitimate conversion of prospective DE to mesoderm. FGF is also essential for DE formation from mESC, yet paradoxically it was previously found that exogenous FGF was detrimental to DE induction (Hansson et al., 2009), which we did not observe (Fig. 1ciii).

In conclusion, these data uncovered a signaling cross-antagonism in which BMP and Wnt versus FGF and TGFβ respectively induced mesoderm versus endoderm from the PS and did so by cross-repressing the alternate fate (Fig. 1dii-iii). Furthermore, BMP and Wnt yielded dichotomous lineage outcomes depending on the developmental time of exposure: their effects became reversed within 24 hours (Fig. 1d, Fig. S1f).

Universal generation of highly-purified DE from diverse hPSC lines through sequential APS formation and mesoderm suppression

The above findings that APS and DE were sequentially specified by opposing signals, together with the necessity of BMP inhibition to eliminate mesoderm from the PS, motivated a serum-free monolayer approach (“SR1”) for DE induction. We first differentiated hPSC to APS in 24 hours (Fig. 2a) while excluding ectoderm by combining high Activin/TGFβ with CHIR (emulating Wnt/β-catenin signaling) and PI3K/mTORC inhibition (Fig. S2c-e), abbreviated “ACP”. This yielded a 99.3±0.1% MIXL1-GFP⁺ PS population (Davis et al., 2008) in which pan-PS TF BRACHYURY was coexpressed with APS-specific TFs EOMES, FOXA2 and LHX1 (Fig. 2a, Fig. S2h). 24 hours later, CHIR was withdrawn and APS was subsequently differentiated into DE for 48 hours by high Activin concomitant with BMP blockade (DM3189) to exclude mesoderm. Exogenous FGF was superfluous as endogenous FGF sufficed (Fig. 1ciii, Fig. S2a).

Sequential APS formation followed by DE induction universally yielded a 94.0±3.1% CXCR4⁺PDGFRα⁻ DE population from 9 diverse hESC (H1, H7, H9, HES2, HES3) and hiPSC (BJC1, BJC3, HUF1C4, HUF58C4) lines by day 3 of differentiation (Fig. 2b-e, Fig. S2i), overcoming line-to-line induction variability. SR1 abundantly elicited SOX17⁺FOXA2⁺ DE (Fig. 2c, Fig. S3c) and downregulated hPSC marker CD90 (Fig. S2j). hESC (94.0±3.1%) and hiPSC (94.0±3.4%) did not markedly differ in DE induction efficiencies (P>0.97, Fig. S3d). We further exploited a SOX17-mCHERRY knockin hESC reporter line to quantify differentiation efficiencies and found SR1 induced >90% SOX17-mCHERRY⁺ DE (Fig. 2d). SR1 generated definitive instead of extraembryonic endoderm (ExEn) as evinced by lack of PDGFRα and SOX7 (Fig. 2e, Fig. S3a).

We directly compared DE induction by SR1 against 2 prevailing protocols, AFBLy (Touboul et al., 2010) or Activin and serum treatment (D'Amour et al., 2005) across 5 hESC lines (Fig. S3a). SR1 differentiation exclusively yielded DE (SOX17, FOXA1, FOXA2, CER1, FZD8) from all 5 hESC lines with minimal mesoderm, neuroectoderm or ExEn (Fig. 2b, Fig. S3a). In contrast, the other DE protocols generated modest amounts of SOX17⁺FOXA2⁺ DE (Fig. 2b-c, Fig. S3c) and produced mixed lineage outcomes: AFBLy upregulated mesoderm TFs (FOXF1, HAND1, MSX1, ISL1) whereas pluripotency TF expression (OCT4, SOX2, NANOG) persisted after serum induction across all 5 lines (Fig. 2b, Fig. S3a). At a clonal level, both FACS quantification (Fig. 2c, Fig. S3b) and single-cell qPCR (Fig. 2f) confirmed SR1 yielded purer DE than either AFBLy or serum treatment: 20/20 of SR1-differentiated cells were FOXA2⁺ whereas few cells after AFBLy (1/20 cells) or serum induction (2/20 cells) highly expressed FOXA2 (Fig. 2f). Thus, even though all 3 differentiation protocols utilized high Activin, clearly Activin alone was insufficient to generate pure DE.

Finally, neural competence was relinquished within 24 hours of SR1 induction (Fig. 2g), showing mutually exclusive ectoderm potential was lost upon APS commitment.

Anteroposterior patterning of hESC-derived DE into mutually exclusive AFG, PFG and MHG domains by BMP, FGF, RA, TGFβ and Wnt signaling

After its initial specification in vivo, DE is patterned along the anteroposterior axis into distinct domains which are the regional antecedents to endodermal organs (Zorn and Wells, 2009). The anterior foregut (AFG) gives rise to lungs and thyroid, the posterior foregut (PFG) to pancreas and liver and the midgut/hindgut (MHG) to small and large intestines (Fig. 3a,b). Therefore, having induced mostly homogeneous DE from hPSC by day 3, we next attempted to anteroposteriorly pattern it into distinct AFG, PFG or MHG populations by 4 subsequent days of differentiation (Fig. 3a), based on increasing knowledge of signals controlling DE patterning in vivo (Zorn and Wells, 2009) and in vitro (e.g., Green et al., 2011; Sherwood et al., 2011; Spence et al., 2011).

Figure 3. Anteroposterior patterning of hESC-derived definitive endoderm.

A) Overview of anteroposterior patterning strategy

B) TF expression in anteroposteriorly patterned endoderm in vivo, see Table S1

C) To test effects of (i) increasing BMP4 (10-25 ng/mL) or (ii) increasing CHIR (3-6 μM) on MHG induction, day 3 DE was differentiated for 4 subsequent days with indicated base conditions together with designated signaling perturbations until day 7, with AFG and PFG controls indicated (subsumed by Fig. S4a); (i) FGF+CHIR = 100 ng/mL FGF2 + 3 μM CHIR; (ii) BF = 10 ng/mL BMP4 + 100 ng/mL FGF2

D) OTX2, FOXA2 and CDX2 staining of H1-derived day 7 AFG and MHG respectively with quantification

E) Microarray heatmap of HES3-derived AFG, PFG, and MHG populations on day 7 in independent triplicate

F) qPCR of day 7 AFG, PFG, and MHG populations from H7 and HES3 hESC lines; HOX genes boxed

G) To test their pancreatic or hepatic competence, day 3 DE was patterned into AFG or PFG for 1-2 days, and each was then subsequently differentiated towards pancreas or liver for 3 further days, see Supplementary Experimental Procedures See also Fig. S4

In vertebrate embryos, tailbud mesoderm expresses BMP4, FGF4/8 and WNT3A and is juxtaposed with posterior endoderm, suggesting these signals might posteriorly pattern the nearby MHG. In vitro, we found BMP markedly posteriorized DE (Fig. 3ci), inducing MHG TFs (e.g., CDX2, EVX1 and 5’ HOX genes) and mirroring zebrafish data (Tiso et al., 2002). Wnt (emulated by CHIR) was similarly posteriorizing (Fig. 3cii) and FGF could also partially posteriorize PFG into MHG (Fig. S4a), confirming prior work (Sherwood et al., 2011; Spence et al., 2011). BMP, FGF and Wnt all reciprocally suppressed anterior endoderm TF SOX2 (Fig. 3c, Fig. S4a). Hence we used a combination of BMP, CHIR and FGF to pattern day 3 DE into >99% CDX2⁺ MHG (Fig. 3d) while suppressing foregut (Fig. 3e) in serum-free conditions.

Conversely, inhibiting posteriorizing BMP signals broadly yielded anterior endoderm (foregut). Combining BMP inhibition with TGFβ inhibition (Green et al., 2011) yielded >98% OTX2⁺ AFG pharyngeal endoderm (Fig. 3d) by day 7 of differentiation. Separately, BMP inhibition in conjunction with RA signaling generated PFG (Fig. 3e,f), consistent with how RA regionalizes the PFG in vivo (Stafford and Prince, 2002). AFG and PFG were functionally distinct, as only PFG was competent to subsequently form liver and pancreas (Fig. 3g).

Invoking the above signaling logic, we generated separate AFG, PFG and MHG populations from DE in a mutually-exclusive manner. Global microarray profiling of distinct patterned populations revealed that anteroposterior marker expression was clearly developmentally demarcated (Fig. 3e,f – reproduced in 2 hESC lines). Graded, spatially collinear HOX gene expression (Zorn and Wells, 2009) was observed after in vitro patterning, whereby PFG expressed 3’ anterior HOX genes (e.g., HOXA1) but MHG exclusively expressed 5’ posterior HOX genes and CDX genes (Fig. 3e,f).

TGFβ competes with BMP/MAPK signaling to specify mutually-exclusive bifurcation of pancreatic and hepatic fates

In vivo, liver and pancreas develop from a common PFG precursor (Chung et al., 2008; Deutsch et al., 2001). During PSC differentiation, BMP and FGF are typically used to induce liver, whereas Hedgehog inhibition and FGF are applied to generate pancreas (e.g., Cho et al., 2012; Kroon et al., 2008). We executed a signaling analysis encompassing >500 conditions (Fig. 4a, Fig. S4b) to clarify how pancreas versus liver might be segregated in a mutually exclusive way (Fig. 4b).

We found TGFβ signaling promoted PDX1⁺ pancreas formation whereas BMP and FGF/MAPK signaling specified AFP⁺ liver (Fig. 4a). Importantly, we clarified that each of these signals reciprocally repressed formation of the alternate lineage (Fig. 4a), explaining why the PFG lineage decision is bistable (Chung et al., 2008). Due to such cross-repression, eliminating pro-pancreatic TGFβ reciprocally expanded liver (Fig. 4aiii) whereas inhibition of pro-hepatic FGF/MAPK (Deutsch et al., 2001) diverted differentiation towards pancreas (Fig. 4aiv). Our findings differ from prior work and may explain previous inefficiencies in liver or pancreas induction. Prior use of FGF for pancreatic induction (Cho et al., 2012; Kroon et al., 2008; Nostro et al., 2011) may in fact block pancreas and instead specify liver (Fig. 4aiv), as suggested by embryonic studies (Deutsch et al., 2001). On the other hand, provision of TGFβ for hepatic induction (Cho et al., 2012) may abrogate liver and instead drive pancreas (Fig. 4ai-ii).

In summary, a dichotomy in TGFβ versus BMP in respectively specifying pancreas versus liver (Fig. 4b) has not been previously elucidated and is reminiscent of how these signaling pathways often cross-repress each others’ transduction (Candia et al., 1997). We further identified combinatorial interactions between these morphogens. For example, TGFβ signaling AND MAPK inhibition was essential for pancreas formation, as MAPK inhibition was ineffective if TGFβ was inhibited in parallel (Fig. S4bi). Conversely, hepatic induction cooperatively required TGFβ inhibition AND MAPK signaling (Fig. 4aiv, Fig. S4bi), as TGFβ inhibition failed to efficiently create liver if MAPK was simultaneously inhibited.

hESC-derived hepatic progeny engraft long-term into unconditioned mouse liver

To differentiate DE towards liver while explicitly inhibiting pancreas, we induced DE towards PFG for 1 day (Fig. 4bi, Fig. S4biv) and then employed BMP and other factors together with inhibition of pro-pancreatic TGFβ signaling to direct PFG towards liver over 3 subsequent days with minimal pancreatic contamination (Fig. S4c). We generated 72.3±6.3% AFP⁺ early hepatic progenitors (Fig. 4c) from 4 hESC lines within 7 days of differentiation, which is twice as rapid as prior methods. Moreover liver markers were induced ~60-210 times higher than earlier protocols (Fig. S4d).

To validate the hepatic potential of early AFP⁺ liver progenitors, they were empirically matured in vitro with oncostatin M and dexamethasone (Kamiya et al., 1999) into a mixed albumin (hALB)⁺ hepatoblast population (Fig. S5a), which exhibited some CYP3A4 metabolic activity (Fig. 4di), expressed LDLR and could uptake cholesterol (Fig. 4dii). When transplanted into neonatal mouse livers, early AFP⁺ hepatic progenitors failed to engraft (Fig. S5b), but when their differentiated hALB⁺ progeny were transplanted, human albumin (mean 15.1 ng/mL) was detected in the blood of 47% of recipients 2-3 months post-transplantation, indicating long-term engraftment (Fig. 4e). Indeed, foci of hALB⁺ hESC-derived hepatic cells (marked with constitutively-expressed GFP prior to transplantation) were present in all lobes of the adult liver (Fig. 4e, Fig. S5b). This suggested hALB⁺ hepatic cells had integrated and/or migrated throughout the liver and they were not simply locally persisting at the site of transplantation. Finally, hALB⁺ cells coexpressed human hepatic marker HepPar1 (Fig. S5c) but did not detectably express fetal marker AFP (Fig. S5d), suggesting they had progressed past the fetal stage. To our knowledge this is one of the first demonstrations that hESC-derived hepatic cells could engraft long-term into normal mouse livers that were not compromised by extensive pharmacologic or genetic damage (cf. Yusa et al., 2011).

Comprehensive transcriptional and chromatin state mapping of endoderm induction and anteroposterior patterning

Capitalizing on our ability to obtain rather homogenous populations of hESC-derived endodermal lineages, we captured genome-wide transcriptional and chromatin dynamics during endoderm development by profiling a hierarchy of 6 pure progenitor populations (hESC, APS, DE, AFG, PFG and MHG) using RNA-seq and ChIP-seq for 4 histone H3 modifications (K4me3, K27me3, K27ac and K4me2; Fig. 5-7, Fig. S5-7). This yielded 30 transcriptional and chromatin state maps spanning 4 embryonic stages (epiblast, PS, DE and anteroposterior patterning) totaling >1.3 billion aligned reads (Fig. S5e), thus providing a global view of molecular events driving endoderm development.

Figure 5. Comprehensively mapping transcriptional and epigenetic dynamics during endodermal development.

A) RNA-seq heatmap of stage-specific genes upregulated at indicated lineage transitions (Supplementary Experimental Procedures)

B, D) Compiled ChIP-seq (histone modifications), RNA-seq (gene expression), vertebrate conservation (Phastcons) and coding gene structure at selected genomic loci with cell types and genomic distance indicated. Numbers indicate fold enrichment over input (ChIP-seq) and FPKM values (RNA-seq).

C) Binary heatmap of H3K27ac-marked active enhancers activated at respective differentiation phases (Supplementary Experimental Procedures), each row is an individual enhancer See also Fig. S5-S7

Figure 7. A constellation of diverse “pre-enhancer” states.

A) ChIP-seq signal heatmap of indicated chromatin marks across future DE enhancer regions in hESC, organized by unbiased clustering; each row is a single pre-enhancer

B) Frequency of DE pre-enhancers overlapping with a given chromatin mark in hESC

C) Occupancy of DE enhancers by endoderm TFs in DE cells that were originally either H2AZ-only pre-enhancers (class 1) or latent pre-enhancers (class 5) in hESC

D) Pre-enhancer state summary

See also Fig. S6

Our analyses captured acute developmental transitions. RNA-seq revealed dramatic transcriptional changes within 24 hours during synchronous transit from pluripotency to APS in vitro (Fig. 5a), mirroring how epiblast (~E5.5) and PS (~E6.5) arise within 1 day in the mouse. The BRACHYURY and NODAL promoters were bivalently marked by activation-associated K4me3 and repression-associated K27me3 in hESC. Yet within 24 hours of APS induction they were unilaterally resolved, losing repressive K27me3 and gaining active marks K27ac and K4me3 concomitant with rapid BRACHYURY and NODAL upregulation in APS (Fig. 5b).

Endoderm enhancer activation is associated with EOMES, SMAD2/3/4 and FOXH1 co-occupancy

To map K27ac-marked active enhancers (Rada-Iglesias et al., 2011) throughout all 6 profiled lineages, we employed DFilter (Kumar et al., 2013) to identify distal elements with significant K27ac enrichment. Distinct batteries of active enhancers were invoked during each endodermal lineage transition (Fig. 5c). APS enhancers (e.g.,BRACHYURY and NODAL) were rapidly activated within 24 hours (Fig. 5b). During DE patterning, distinct cohorts of enhancers were commissioned in each anteroposterior domain in AFG (SIX1 and TBX1; Fig. S7a), PFG (HOXA1; Fig. S7b) and MHG (CDX2 and PAX9; Fig. 5d, Fig. S5g).

10,543 enhancers were activated upon DE specification (Table S5), gaining K27ac despite being largely inactive in hESC. Active DE enhancers flanked archetypic DE regulators, e.g. SOX17 (Fig. 6g) and CXCR4 (Fig. S5f). Gene ontology analyses (McLean et al., 2010) associated these enhancers most significantly with endoderm development (P<3.84×10⁻) and gastrulation (P<7.92×10⁻; Fig. 6a), affirming the purity of differentiated DE populations. Genes adjacent to active DE enhancers were upregulated in gastrula-stage endoderm in vivo (P<1.38×10⁻³⁹, Fig. 6a) and upon DE differentiation in vitro (Fig. 6b). Active DE enhancers coincided with euchromatic mark K4me2 (Fig. S6a), were devoid of repression-associated K27me3 (Fig. S6a), were evolutionarily conserved (Fig. 6c) and were broadly inactive in other lineages (Fig. S6b).

Figure 6. TGFβ signaling inaugurates endodermal active enhancers.

A) Top-ranked GO terms associated with DE-specific active enhancers by GREAT (McLean et al., 2010) without pre-selection

B) Boxplot of RNA-seq FPKM expression values of genes adjacent to DE-specific enhancers at indicated in vitro differentiation stages

C) Phastcons score of DE-specific active DE enhancers

D) Top-ranked GO terms associated with DE enhancers identified from a previous dataset (Gifford et al., 2013) using identical analytic methods (Table S6)

E) TF motifs overrepresented in DE-specific active enhancers (Table S5b)

F) Left: ChIP-seq signal heatmap based on all distal EOMES, SMAD2/3/4 and FOXH1 peaks in DE, showing TF overlap with one another and K27ac; each row is a single distal element (6 kB window size). Right: Average H3K27ac tag count at DE distal elements bound by all 1, 2, 3 or 4 DE TFs (EOMES, SMAD2/3, SMAD4 & FOXH1) G) EOMES, FOXH1 and SMAD2/3/4 colocalize at conserved SOX17 enhancer (MTL: multiple TF locus)

See also Fig. S6

DE enhancers previously remained elusive because most prior work only assessed promoter marks (Kim et al., 2011; Xie et al., 2013). However, enhancer profiling of hESC-derived DE was recently reported (Gifford et al., 2013) and therefore we compared our two DE datasets using identical analytic methods (Table S6). Paradoxically, DE enhancers from the former dataset (Gifford et al., 2013) were highly enriched for neural functions (P<3.93×10⁻²⁸; Fig. 6d), as enhancers for neural TFs BRN2 and PAX3 were activated, but SOX17 enhancers were virtually silenced (Fig. S6c). Association of DE enhancers with neural genes led to the prior conclusion that endoderm and ectoderm development are related (Gifford et al., 2013), which contradicts the in vivo order of germ layer segregations (cf. Tzouanacou et al., 2009). By contrast, neural terms were largely absent in SR1-derived DE (Fig. 6a) and ultimately only 4.8% of DE enhancers were shared between our and their datasets. Thus, molecular profiling of mixed DE populations (potentially enriched for ectoderm; Gifford et al., 2013) may have precluded accurate molecular description of endoderm development.

How DE enhancers are inaugurated during differentiation remains obscure. Motifs for multiple TFs, including DE specifiers EOMES and FOXA2 as well as TGFβ signaling effectors SMAD2/3 and FOXH1 (P=10⁻⁵⁹-10⁻¹⁹⁷) were enriched in DE enhancers (Fig. 6e), consistent with how these TFs specify DE in vivo (e.g., Dunn et al., 2004; Teo et al., 2011). Interestingly, we found EOMES, SMAD2/3, SMAD4 and FOXH1 (Kim et al., 2011; Teo et al., 2011) co-occupied an extensive series of DE enhancers (Fig. 6f), including a SOX17 enhancer (Fig. 6g). Although EOMES individually engaged some elements, colocalization of EOMES with TGFβ signaling effectors SMAD2/3/4 and FOXH1 correlated with maximal enhancer acetylation (Fig. 6f, P<10⁻³⁰⁰). Thus, convergence of both lineage-specifying and signaling-effector TFs may propel full-fledged enhancer activation upon differentiation (Calo and Wysocka, 2013).

Endoderm enhancers reside in a diversity of “pre-enhancer” states in uncommitted cells prior to activation

It remains unclear how DE enhancers are swiftly engaged upon hESC differentiation. SMAD2/3/4 and FOXH1 occupy DE enhancers upon differentiation but infrequently do so in the uncommitted state (Fig. S6a). Perhaps these enhancers are instead primed for activation at the level of chromatin. Premarking of developmental enhancers by euchromatic K4me1 in ESC signifies a “window of opportunity” for subsequent enhancer activation (Calo and Wysocka, 2013; Rada-Iglesias et al., 2011). We looked back in developmental progression, assessing occupancy of DE enhancers by >24 histone modifications and chromatin regulators (Ernst et al., 2011) in hESC prior to enhancer activation (Fig. 7a). Unexpectedly, K4me1 labeled less than one third of future DE enhancers in hESC, implying “poising” by K4me1 in hESC is not always essential for immediate enhancer activation (Fig. 7a,b). Thus we sought to systematically discover all possible “pre-enhancer” chromatin states of DE enhancers in hESC.

Unsupervised clustering revealed 25% of DE enhancers existed in a novel pre-enhancer state (cluster 1) in hESC largely defined by histone variant H2AZ and no other known chromatin marks (Fig. 7a, Fig. S6d). Despite virtual absence of K4me1, H2AZ-marked pre-enhancers became rapidly activated within 3 days of DE induction (Fig. 7a). DE enhancers less frequently resided in a repressed state designated by heterochromatic mark K9me3 (cluster 2) (Zhu et al., 2012) or a “latent” pre-enhancer state largely lacking known histone modifications (cluster 5, Fig. 7a) (Ostuni et al., 2013). Only 10% of DE pre-enhancers were marked by K27me3 in hESC (Fig. 7b), suggesting Polycomb (Rada-Iglesias et al., 2011) was not always necessary to repress developmental enhancers in hESC. Instead perhaps absence of K27ac/histone acetyltransferases (HATs) was sufficient to confer inactivity. Only a minority of DE pre-enhancers (10%) were pre-loaded with HAT p300 (Rada-Iglesias et al., 2011) (Fig. 7b), suggesting rapid enhancer acetylation during differentiation may largely involve de novo HAT recruitment.

A pre-enhancer state solely delineated by H2AZ without other detectable distinguishing factors has not been previously described. H2AZ-laden nucleosomes are unstable and are readily displaced by TFs (Jin et al., 2009; Li et al., 2012). This may permit endoderm TFs to rapidly infiltrate DE enhancers upon differentiation, explaining rapid enhancer activation. Indeed, H2AZ-marked DE pre-enhancers in hESC more readily attracted EOMES, SMAD2/3/4 and FOXH1 upon differentiation (Fig. 7c, P=10⁻¹³-10⁻¹⁵) compared to latent pre-enhancers.

In sum, initial K4me1 “poising” is not the only predictor of subsequent enhancer activation. We show there is a diversity of pre-enhancer states characterized by different combinations of chromatin marks (Fig. 7d).

DISCUSSION

PSC differentiation typically yields a range of developmental outcomes that vary between PSC lines. Contaminating lineages may generate undesired tissues upon transplantation and obscure molecular analyses of lineage commitment. To meet this challenge, we delineated the signaling logic for induction and anteroposterior patterning of human endoderm from PSC and for subsequent bifurcation of pancreas versus liver, clarifying separation of alternate lineages at each stage. Such knowledge permitted us to rationally exclude alternate fates at every step following the in vivo hierarchy of germ layer segregations (Tzouanacou et al., 2009). This approach yielded precise induction of a single lineage (endoderm) from diverse hESC and hiPSC lines, without extraneous lineages typically induced by earlier protocols. This level of endodermal purity enabled accurate chromatin analysis of endoderm induction at a resolution previously unattainable due to contaminating lineages. Therefore the highly-homogeneous definitive endoderm populations described here constitute an ideal starting point to efficiently generate downstream endodermal derivatives (McKnight et al., 2010), a notion we validate by producing engraftable liver cells. In summary, this work expounds a coherent view of signaling logic and chromatin dynamics propelling endoderm specification and patterning, thereby availing both developmental biology and hPSC differentiation.

Developmental segregation of mutually exclusive endodermal fates

Throughout 4 successive stages of endoderm development, we accurately defined the signals that instruct or repress a given lineage, thus providing a clearer view of how endodermal lineage bifurcations are driven. In fact, this refined understanding suggested that previous protocols provided incorrect signals that repressed DE formation, thereby leading to inefficient differentiation. For example BMP, FGF, TGFβ and Wnt have been used to elicit both endoderm (Touboul et al., 2010) and mesoderm (Gadue et al., 2006) and therefore the exact lineages induced by these signals has remained ambiguous.

We generated DE in the virtual absence of mesoderm or ectoderm. We found that combined FGF, TGFβ and Wnt together with low BMP signaling (Bernardo et al., 2011; Blauwkamp et al., 2012; Gadue et al., 2006) was necessary to specify APS (>99% MIXL1⁺) and repress ectoderm, abolishing ectoderm competence within 24 hours of APS induction. After ectoderm exclusion, mesoderm was sequentially eliminated by BMP inhibition, which when combined with TGFβ and endogenous FGF signaling (Bernardo et al., 2011; D'Amour et al., 2005) exclusively drove PS towards DE. It was crucial to suppress endogenous mesoderm-inducing BMP and Wnt signaling within PS to achieve pure DE populations. We also clarified nuances in the interpretation of combinations of signals, showing that reception of one signal altered the response to others. For example, while BMP inhibition typically eradicated mesoderm from the PS, if DE-inducing FGF was blocked in parallel, mesoderm formation was re-enabled. Thus, FGF was obligatory to consolidate DE commitment.

Following PFG formation, TGFβ and BMP signaling dueled to specify pancreas versus liver and each bilaterally cross-repressed the alternate fate, reminiscent of in vivo findings (Chung et al., 2008; Deutsch et al., 2001). Therefore, efficient liver induction required TGFβ inhibition to eliminate pancreatic fates in conjunction with BMP and MAPK to positively drive liver and vice versa. In sum, we show that in order to efficiently drive hPSC differentiation down a single developmental route, it is critical not only to provide the relevant positive inductive signals, but it is equally important to inhibit repressive signals that instead drive progression down alternate lineage pathways.

By inhibiting alternate fates at each juncture, we could universally differentiate 9 diverse hESC/hiPSC lines into highly-pure DE populations in defined conditions. This is contrary to the notion that different hPSC lines have distinct differentiation biases and each might require customized signals to drive efficient commitment. Our observations are timely as a prerequisite for cell replacement therapy is the consistent generation of homogeneous lineages from hPSC under defined conditions (Cohen and Melton, 2011; McKnight et al., 2010). Recent strategies to generate “self-renewing” DE (Cheng et al., 2012) or liver buds (Takebe et al., 2013) from hPSC are appealing but require coculture with heterologous feeders and thus suit a different type of application.

Obligatory endodermal signaling inputs are highly temporally dynamic

The precise sequence and kinetics of endoderm signaling transitions remain to be fully elucidated, despite their evident importance in vivo and in vitro (Green et al., 2011; Wandzioch and Zaret, 2009). For example, BMP and Wnt have been associated with mesoderm induction through studies of prolonged treatment over several days (Bernardo et al., 2011; Gadue et al., 2006). However we found that BMP and Wnt initially specified APS, but within 24 hours of differentiation, signaling requirements were reversed such that BMP and Wnt repressed DE from PS and instead induced mesoderm. Prior protocols reduced APS and DE induction into a single lengthy stage and persistently provided BMP for 3-5 days (Nostro et al., 2011; Touboul et al., 2010), likely generating contaminating mesoderm at later stages and inhibiting DE formation. The dynamism with which BMP and Wnt signals are interpreted during hPSC differentiation (within 24 hours) closely tracks how Wnt is initially inactive in E5.5 post-implantation epiblast, transiently elicited in E6.5 PS and then silenced once again in E7.5 DE in vivo (Maretto et al., 2003). Therefore, assigning BMP and Wnt as either proor anti-endoderm is a misnomer because these signals can induce either outcome depending on timing within just 24 hours in vivo and in vitro.

Developmental competence and a diversity of pre-enhancer states

To gain insight into endodermal lineage commitment mechanisms, we globally mapped transcriptional changes and regulatory element redeployment across multiple steps of endoderm induction and patterning. This resource could unveil novel drivers or markers of DE specification by identifying TFs upregulated at distinct stages. Here we exploited the accompanying chromatin data to explore how endoderm competence is preconfigured in pluripotent cells.

Since Waddington's formalism of developmental competence (Waddington, 1940), its molecular basis has remained cryptic. Competence may be foreshadowed by permissive chromatin priming of developmental enhancers in progenitors (Calo and Wysocka, 2013). Various models proposed such enhancers resided in “poised” or “latent” chromatin states prior to activation (Ostuni et al., 2013; Rada-Iglesias et al., 2011). However, the prevalence of “poised” or “latent” pre-enhancer states (and whether they represented all pre-enhancer states) remained uncertain. With a priori knowledge of a catalog of DE enhancers, we systematically determined their antecedent “pre-enhancer” states in hESC. Individual DE enhancers existed in a wide continuum of differentially-marked pre-enhancer states prior to activation, extending beyond “poised” or “latent” states. Only a subset of DE enhancers were pre-marked by K4me1, p300 or other proposed “poising” factors in hESC, showing there is no universal poising signature.

Strikingly, we found many prospective DE enhancers were marked exclusively by H2AZ in the general absence of other chromatin marks. Thus H2AZ is sometimes the earliest recognizable enhancer mark in lieu of K4me1. H2AZ prepositioning at DE enhancers enhanced future infiltration by EOMES, SMAD2/3/4 and FOXH1 upon differentiation and combinatorial occupancy by all these TFs correlated with maximal enhancer activation. Indeed H2AZ is essential for DE induction from mESC and it was shown its presence at promoters increased FOXA2 recruitment (Li et al., 2012). Our related findings with DE enhancers suggest the primordial chromatin state of a DE enhancer in hESC can influence its future engagement upon differentiation. Because some mesoderm enhancers are likewise exclusively marked by H2AZ in hESC (Fig. S6f), H2AZ prepositioning on developmental enhancers may broadly signal future fates available to uncommitted precursors. How H2AZ is deployed to these silent enhancers in ESC remains unclear. It may be targeted by pluripotency TFs (e.g., Oct4), which physically interact with H2AZ depositor p400 (van den Berg et al., 2010) and might guide it to lineage-specification genes in uncommitted ESC to functionally presage future differentiation potential (Loh and Lim, 2011; Teo et al., 2011). Yet half of endoderm enhancers apparently lack H2AZ in hESC: therefore to understand developmental competence we must decipher the whole range of alternative pre-enhancer states.

EXPERIMENTAL PROCEDURES

SR1 definitive endoderm induction and patterning

mTeSR1-grown hPSC (Fig. S2f) were passaged ~1:3 as small clumps using collagenase IV onto fibronectin- or Matrigel-coated plates. 1-2 days later, they were washed and differentiated with Activin A (100 ng/mL), CHIR99021 (2 μM) and PI-103 (50 nM) in serumless CDM2 basal medium for 24 hours to specify APS (day 1), followed by Activin A (100 ng/mL) and DM3189 (250 nM) for 48 subsequent hours to specify DE (day 3). DE was anteroposteriorly patterned into either AFG (A-83-01, 1 μM and DM3189, 250 nM), PFG (RA, 2 μM and DM3189, 250 nM), or MHG (BMP4, 10 ng/mL; CHIR99021, 3 μM; and FGF2, 100 ng/mL) for 4 subsequent days until day 7. For detailed differentiation methods, see Supplemental Experimental Procedures.

Hepatic induction and empirical maturation

Day 3 DE was differentiated for 24 hours into early PFG by DM3189 (250 nM), IWP2 (4 μM), PD0325901 (500 nM) and RA (2 μM) and further differentiated into hepatic progenitors by A-83-01 (1 μM), BMP4 (10 ng/mL), IWP2 and RA for 3 further days until day 7. They were then empirically matured in vitro with BMP4 (2 days) followed by dexamethasone (10 μM) and oncostatin M (10 μg/mL) for 10 days (Kamiya et al., 1999) and then intrahepatically transplanted into newborn NSG mice (Chen et al., 2013).

Chromatin state analysis

For ChIP-seq, H7-derived endoderm lineages were formaldehyde-fixed, lysed to extract nuclei, sonicated and pre-cleared (Supplemental Experimental Procedures). Chromatin was probed overnight using K4me2, K4me3, K27ac and K27me3 antibodies (Table S7) conjugated to Protein G Dynabeads (Invitrogen). Subsequently, chromatin was precipitated, rigorously washed (8 times) and cross-linking undone by overnight 65 °C heating before RNase/Proteinase K treatment and column purification. 10ng chromatin was used to generate libraries (TruSeq Kit, Illumina) for Hi-Seq 2000 sequencing (Illumina, 36bp single-end reads; Fig. S5e). Reads were aligned to hg19 (Bowtie), extended and input-normalized (MACS). DE enhancers (Table S5a) were assigned by DFilter (Kumar et al., 2013) as peaks with ≥4-fold more K27ac tags in DE than hESC and were associated with GO terms via GREAT (McLean et al., 2010). Transcriptional and ChIP-seq data are available under Gene Expression Omnibus accession number GSE52658.