Mapping the pairwise choices leading from pluripotency to human bone, heart and other mesoderm cell-types

SUMMARY

Stem-cell differentiation to desired lineages requires navigating alternating developmental paths often leading to unwanted cell-types. Hence comprehensive developmental roadmaps are crucial to channel stem-cell differentiation towards desired fates. To this end, here we map bifurcating lineage choices leading from pluripotency to twelve human mesodermal lineages, including bone, muscle and heart. We defined the extrinsic signals controlling each binary lineage decision, enabling us to logically block differentiation towards unwanted fates and rapidly steer pluripotent stem cells towards 80–99% pure human mesodermal lineages at most branchpoints. This strategy enabled the generation of human bone and heart progenitors that could engraft in respective in vivo models. Mapping stepwise chromatin and single-cell gene expression changes in mesoderm development uncovered somite segmentation, a previously-unobservable human embryonic event transiently marked by HOPX expression. Collectively this roadmap enables navigation of mesodermal development to produce transplantable human tissue progenitors and uncover developmental processes.

INTRODUCTION

Waddington’s developmental landscape drawings (Waddington, 1940) depicted how differentiating stem cells negotiate a cascade of branching lineage choices, avoiding alternate fates at each juncture to decisively commit to a single lineage (Graf and Enver, 2009). To navigate this brachiating landscape and efficiently differentiate stem cells into desired cell-types for regenerative medicine, one must (i) catalog transitional lineage intermediates, (ii) map the sequence of pairwise lineage choices through which such intermediates are formed and (iii) discover the positive and negative signals that specify or repress cell fate at each lineage branchpoint. Despite successes in charting lineage intermediates in mammalian tissues, key lineage branchpoints remain controversial and it has been impossible to systematically identify the extracellular signals that control cell fate at each exact branchpoint.

With the three above goals in mind, here we map the landscape of human mesoderm development in order to coherently guide stem-cell differentiation (Fig. 1a). Mesoderm development begins with differentiation of pluripotent epiblast cells into the primitive streak, which then segregates into paraxial and lateral mesoderm, amongst other lineages (Lawson et al., 1991; Rosenquist, 1970; Tam and Beddington, 1987). Paraxial mesoderm segments into somites, which are fundamental building blocks of trunk tissue (Pourquié, 2011) (Fig. 1a, purple shading). Somites are then patterned along the dorsal-ventral axis; the ventral somite (sclerotome) generates the bone and cartilage of the spine and ribs, whereas the dorsal somite (dermomyotome) yields brown fat, skeletal muscle and dorsal dermis (Christ and Scaal, 2008). Separately, lateral mesoderm (Fig. 1a, red shading) gives rise to limb bud mesoderm (Tanaka, 2013) and cardiac mesoderm (Später et al., 2014), the latter of which subsequently generates cardiomyocytes and other heart constituents. Various transcription factors (TFs) and signaling molecules regulating mesoderm development in model vertebrates have been identified, broadly outlining the developmental landscape (Kimelman, 2006; Schier and Talbot, 2005; Tam and Loebel, 2007).

Figure 1. Formation of human primitive streak and its bifurcation into paraxial and lateral mesoderm.

A. Each lineage step labeled with a circled number, corresponding to respective sections in the main text and Fig. 7a

B. FACS of MIXL1-GFP hESC (Davis et al., 2008) after 24 hrs in anterior or mid PS induction (left); all cells coexpress BRACHYURY and MIXL1 by scRNA-seq; each dot depicts a single cell (right)

C. BMP induces, whereas WNT inhibits, lateral mesoderm from the PS on D1-2. (i) qPCR of D1 PS treated with BMP4 or a BMP inhibitor (DM3189) for 24 hours (in the context of A8301 + FGF2); (ii) qPCR of D1 PS treated with WNT agonists (CHIR99021 or WNT3A) or WNT inhibitors (300 ng/mL Dkk1 or 1 μM C59) for 24 hrs (in the context of A8301 + BMP4 [AB]); error bars = S.E.M. for this and all other qPCR experiments

D. BMP inhibits, whereas WNT induces, paraxial mesoderm from the PS on D1-2. (i) qPCR of D1 PS treated with BMP4 or a BMP inhibitor (DM3189) for 24 hrs (in the context of A8301 + FGF2 [AF]); (ii) qPCR of D1 PS treated with WNT agonist (3 μM CHIR99021) or WNT inhibitors (2 μM IWR1, 1 μM XAV939 or C59 or 300 ng/mL Dkk1) for 24 hrs (in the context of A8301 + DM3189 + FGF2 [ADF])

E. Lateral versus paraxial mesoderm bifurcation

F. CDX2 and HAND1 staining of day 2 H7-derived paraxial or lateral mesoderm populations or undifferentiated hESCs (scale bar = 100 μm), with Hoechst nuclear staining

G. scRNA-seq of day 2 lateral mesoderm or DLL1⁺ sorted paraxial mesoderm; each dot depicts a single cell; % of marker-positive cells above the dotted TPM (transcripts per million) threshold

See also Fig. S1, Fig. S2

Yet gaps in our understanding have been revealed by efforts to differentiate human pluripotent stem cells (hPSCs) to various mesoderm cell-types in a dish. Human mesoderm has remained obscure because it first forms in gestational weeks 2–4 (O’Rahilly and Müller, 1987), when it is impermissible to access human embryos. There has been some success in generating human mesoderm derivatives from PSCs, including paraxial (Cheung et al., 2012; Mendjan et al., 2014; Umeda et al., 2012) and heart (Ardehali et al., 2013; Burridge et al., 2014; Chong et al., 2014; Lian et al., 2012; Mendjan et al., 2014) cell-types. However because the sequence of lineage branchpoints and the identity of inductive or repressive signals at every developmental step remain incompletely understood, some mesodermal differentiation protocols take weeks to months and generate heterogeneous mixtures of cell-types comprising a subset of the desired lineage along with other contaminating lineages. Prior studies indicated ACTIVIN/NODAL/TGFβ (henceforth referred to as TGFβ), BMP, FGF and WNT broadly induce mesoderm from PSCs (Cheung et al., 2012; Gertow et al., 2013), the importance of dynamic WNT signaling during cardiac induction (Burridge et al., 2014; Lian et al., 2012; Ueno et al., 2007) and that BMP inhibits paraxial mesoderm formation (Cheung et al., 2012; Umeda et al., 2012). Nonetheless, the challenges faced by current differentiation strategies provide an impetus to better understand the complex process of mesoderm development.

Here we delineate a roadmap for human mesoderm development and define the sequential steps through which pluripotent cells elaborate a diversity of mesodermal progeny. At many developmental steps, we discovered the minimal combinations of signals sufficient to efficiently induce each human mesodermal fate and showed that it was key to define both inductive and repressive cues at each step (Table S1). It was critical to define how “unwanted” cell fates were specified in order to logically block their formation and steer stem-cell differentiation down a singular developmental path.

Collectively this knowledge guided the efficient differentiation of PSCs into a variety of human mesoderm fates within several days, without recourse to gene modification or serum treatment. The authenticity of the induced cells was confirmed by their ability to engraft in vivo and by single-cell RNA-seq to test for lineage identity and homogeneity. Global RNA-seq and ATAC-seq analyses also revealed stepwise changes in gene expression and sequential opening and closing of chromatin elements at each developmental transition. Collectively, we chart the developmental landscape of human mesoderm formation and uncover the sequential signaling, transcriptional and chromatin changes at each lineage step. We directly demonstrate the utility of this reference map in guiding stem-cell differentiation, producing transplantable cells for eventual use in regenerative medicine, improving our understanding human development, and uncovering the putative origins of certain human congenital malformations.

RESULTS

Induction of anterior and mid primitive streak

Primitive streak (PS) formation from pluripotent cells is the first step in mesoderm development (Fig. 1a, step ①). We generated a >98% pure MIXL1-GFP⁺ human PS population within 24 hours of PSC differentiation (Fig. 1bi, Fig. S1a,b) by activating TGFβ, WNT and FGF and inhibiting PI3K signaling, in the presence or absence of exogenous BMP (Fig. S1c–e) (Gertow et al., 2013; Loh et al., 2014; Schier and Talbot, 2005; Tam and Loebel, 2007). Attesting to the uniformity of differentiation, single-cell RNA-seq (scRNA-seq) of bulk PS populations revealed that 100% of analyzed cells coexpressed PS TFs BRACHYURY and MIXL1 (Fig. 1bii).

In the vertebrate embryo, different anterior-posterior regions of the PS produce distinct mesoderm derivatives (Lawson et al., 1991; Rosenquist, 1970; Tam and Beddington, 1987). Likewise hPSC-derived anterior PS induced in the presence of an anteriorizing TGFβ signal was competent to form paraxial mesoderm (Fig. S1f). By contrast, mid PS induced in the presence of both anteriorizing TGFβ and posteriorizing BMP harbored maximal potential to form lateral mesoderm/cardiac progenitors (Fig. S1f–i). Thus as in model organisms, human PS is not a singular lineage, but comprises several subtypes each already partially committed to one type of downstream mesoderm (Fig. S1j–l). Altogether, these >98% pure human PS populations provided a starting point to understand the subsequent divergence of distinct mesoderm subtypes.

Bifurcation of paraxial versus lateral mesoderm from primitive streak by competing WNT and BMP signals

In vivo, the PS forms definitive endoderm, paraxial mesoderm and lateral mesoderm, but how these lineages are segregated is not well understood (Fig. 1a, step ②). After PS induction (day 0–1), TGFβ specified endoderm (Loh et al., 2014) (Fig. S2a,b) while TGFβ inhibition blocked endoderm formation and instead induced mesoderm (Fig. S2b–e). Since TGFβ inhibition and FGF/ERK activation (Fig. S2b,f) for 24 hours (day 1–2) created a permissive context for both paraxial and lateral mesoderm formation, we sought how these mutually-exclusive mesoderm subtypes become distinguished.

Countervailing BMP and WNT signals respectively induced human lateral versus paraxial mesoderm and each repressed the formation of the mutually-exclusive lineage on day 1–2 of PSC differentiation, driving the bifurcation of these two mesoderm subtypes from the PS (summarized in Fig. 1e). Exogenous BMP induced lateral mesoderm and repressed paraxial mesoderm (Fig. 1ci, Fig. 1di, Fig. S2g,h). By contrast, blocking BMP signaling abrogated lateral mesoderm and instead expanded paraxial mesoderm (Fig. 1ci, Fig. 1di, Fig. S2h). This reveals a key function of BMP in patterning human mesoderm akin to its activity in chick (Tonegawa et al., 1997).

Conversely, WNT promoted human paraxial mesoderm and repressed lateral mesoderm. WNT activation (by GSK3 inhibition) induced paraxial markers while suppressing lateral/cardiac markers (Fig. 1cii, Fig. 1dii). By contrast, WNT inhibition elicited lateral mesoderm while blocking the paraxial fate (Fig. 1cii, Fig. 1dii). Therefore WNT controls the allocation of human paraxial vs. lateral mesoderm, logically linking the requirement for WNT in mouse paraxial mesoderm formation (Aulehla et al., 2008) to the ability of WNT to repress early cardiac mesoderm in Xenopus (Schneider and Mercola, 2001).

In summary, on day 1–2 of hPSC differentiation, BMP inhibition and WNT activation induced paraxial mesoderm whereas conversely, BMP activation and WNT inhibition specified lateral mesoderm from the PS within the permissive context of TGFβ inhibition/FGF activation (Fig. 1e). These two mutually-exclusive signaling conditions produced either CDX2⁺HAND1⁻ paraxial mesoderm or CDX2^lo/−HAND1⁺ lateral mesoderm by day 2 of differentiation (Fig. 1f, Fig. S2i). Tracking the bifurcation of paraxial versus lateral mesoderm fates by scRNA-seq confirmed the mutually-exclusive marker expression in the two diverging populations at the level of single cells (Fig. 1g). scRNA-seq showed that both human paraxial and lateral mesoderm populations were essentially uniform: 98.5% of paraxial mesoderm cells expressed DLL3 and MSGN1 whereas 98.1–100% of lateral mesoderm cells expressed HAND1 and FOXF1 (Fig. 1g).

Maturation of paraxial mesoderm into early somites by combined BMP, ERK, TGFβ and WNT inhibition

Having generated 91.2±0.1% pure TBX6⁺CDX2⁺ human paraxial mesoderm by day 2 of PSC differentiation (Fig. S3a), we next sought to drive these cells into early somite progenitors (Fig. 1a, step ③; Fig. 2a). During embryogenesis, the U-shaped sheet of paraxial (presomitic) mesoderm is progressively segmented at its anterior edge to generate spherical early somites (Fig. 2a), due to lower anterior levels of FGF and WNT at the “wavefront” (Aulehla et al., 2008; Dubrulle et al., 2001).

Figure 2. Human paraxial mesoderm differentiation into early somites passes through an ephemeral somitomere-like state.

A. Paraxial mesoderm segmentation into somites in vivo

B. To reveal how WNT and FGF/ERK control paraxial mesoderm progression to early somites, day 2 H7-derived paraxial mesoderm was treated with RA (2 μM) for 24 hrs, in combination with a WNT agonist (CHIR, 3 μM), a WNT inhibitor (C59, 1 μM), FGF2 (20 ng/mL), an ERK inhibitor (PD0325901, 500 nM), or combined WNT/ERK inhibition (CPR: C59+PD0325901+RA) and qPCR was conducted (*p<0.05, **p<0.01), showing WNT/ERK blockade enhances early somite induction (it was later found that exogenous RA was dispensable for early somite formation; Fig. S3e)

C. CDX2 and FOXC2 staining of BJC1-derived paraxial mesoderm (day 2) and early somite (day 3) populations (left) and quantification (right)

D. FGF and WNT activation, followed by inhibition, induces human paraxial mesoderm and then early somites

E. Differentially expressed genes in day 2 paraxial mesoderm vs. day 3 early somites (bulk-population RNA-seq)

F. qPCR timecourse comparison of H7 hESCs differentiated into somites using previous protocols (Cheung et al., 2012; Mendjan et al., 2014) or the current method

G. PCA of human somitogenesis scRNA-seq; colors designate cell populations harvested at different timepoints; each dots is a single cell

H. scRNAseq of day 2, day 2.25 and day 3 hESC-derived populations; dots depict single cells; line indicates mean gene expression in all cells at each timepoint

I. Timecourse qPCR of H7-derived cells

See also Fig. S3

Whereas paraxial mesoderm was specified on day 1–2 by FGF and WNT, subsequently the inhibition of FGF/ERK and WNT signaling on day 2–3 strongly downregulated paraxial mesoderm genes (e.g., TBX6, MSGN1) and upregulated early somite markers (e.g., FOXC2, PARAXIS, MEOX1; Fig. 2b, Fig. S3b). Early somite markers were further upregulated when TGFβ and BMP were inhibited (Fig. S3c,d); we therefore employed quadruple inhibition of these 4 pathways to drive near-complete conversion of day 2 CDX2⁺ paraxial mesoderm into 96.8±5.7% pure FOXC2⁺ early somite precursors by day 3 (Fig. 2c–e). Taken together, this identified the minimal signaling cues sufficient to efficiently generate human PS, paraxial mesoderm and subsequently early somite progenitors from PSCs within 72 hours of differentiation (Fig. 2d), more robustly and rapidly than was previously possible (Cheung et al., 2012; Mendjan et al., 2014) (Fig. 2f).

scRNA-seq reaffirmed the homogeneity of these various in vitro-derived populations at different steps of human somitogenesis. 98.5–100% of human paraxial mesoderm cells (sorted for DLL1 expression; Supplemental Procedures) expressed paraxial markers MSGN1 and DUSP6 at day 2, yet these markers were sharply suppressed within 6 hours of FGF/WNT inhibition (by day 2.25) during differentiation towards somites (Fig. 2hi). Conversely, somite TFs MEOX1 and FOXC2 became expressed in the majority (83.3–88.9%) of human early somite cells by day 3 (Fig. 2hiii). Hence, this indicated that human somitogenesis was efficiently reconstituted in culture. We exploited this system to uncover developmental features of this process.

Single-cell RNA-seq identifies a conserved, segmentation-like process in human somitogenesis and infers transcriptional cofactor HOPX as a marker

In vertebrate model organisms, somitogenesis entails transient expression of segmentation genes (for 30–120 minutes in fish and mice, respectively) at the wavefront, inducing paraxial mesoderm to segment into somitomeres (prospective somites) and then into somites (Pourquié, 2011). Human somitogenesis begins at week 3 of gestation (O’Rahilly and Müller, 1987), but whether it also entails a transient segmentation molecular program has remained unclear because it is ethically impermissible to retrieve and analyze week 3–5 human embryos.

Indeed paraxial mesoderm cells in our reconstituted human somitogenesis system passed through an intermediate somitomere stage before differentiating into somites. scRNA-seq of hESCs (day 0) differentiating into anterior PS (day 1), paraxial mesoderm (day 2) through an intermediate step (day 2.25) into early somites (day 3) showed that single cells uniformly moved through a single inferred differentiation trajectory without overtly diverging branch points (Fig. 2g), implying the homogeneity of cells at each lineage step. Principal component analysis (PCA) positioned the D2.25 population between D2 paraxial mesoderm and D3 early somites (Fig. 2g), suggesting that it is a true intermediate between presomitic/paraxial and somitic states.

Indeed a subset of D2.25 cells transiently expressed MESP2, RIPPLY2 and HEYL (Fig. 2hii, Fig. S4f), whose homologs are somitomere segmentation markers in model organisms (reviewed by Pourquié, 2011). qPCR confirmed a brief pulse of MESP2, RIPPLY2 and HEYL expression for several hours (~D2.25-D2.5) in the interval between paraxial and somitic marker expression (Fig. 2i). Hence, there exists a transient somitomere-like transition point in human somitogenesis, arguing that human development entails an evolutionarily-conserved segmentation process.

To discover additional somite segmentation markers, single-cell transcriptomes were aligned in “pseudotime” to infer the temporal order with which they arose along an inferred developmental path (Supplemental Procedures, Fig. 3a). This led to 44 candidate genes with transient somitomere-specific expression (Fig. 3b), including HOPX, a homeodomain-containing transcriptional cofactor. By scRNA-seq, HOPX was specifically expressed in a subset of D2.25 somitomere cells but neither D2 paraxial mesoderm nor D3 early somites (Fig. 3ci). qPCR (Fig. 3cii) and immunostaining (Fig. 3ciii) corroborated the transient expression of HOPX mRNA and protein during in vitro somitogenesis, paralleling the transitory expression of known somitomere markers (Fig. 2i). Additionally, open chromatin analysis by ATAC-seq (Buenrostro et al., 2013) revealed that the HOPX locus was accessible in D2.25 somitomere populations (Fig. 3d). Thus HOPX is briefly expressed in somitomeres and marks human somite segmentation.

Figure 3. Single-cell analysis captures a transient HOPX⁺ human somitomere progenitor state.

A. Heatmap of normalized scRNA-seq gene expression across the inferred trajectory of human somitogenesis. Each column reflects a single cell, with scRNA-seq paraxial mesoderm, somitomere and early somite transcriptomes (colored blocks) ordered in pseudotime along the y-axis (Supplemental Procedures). Genes were clustered into 10 clusters (rows) by virtue of their expression kinetics across this pseudotime timecourse; line indicates smoothed mean expression of all genes in the cluster across pseudotime

B. Mean expression (bold line) of all genes in each temporal cluster across pseudotime (with contours representing density of individual gene expression), with representative genes in each cluster noted

C. Transient HOPX expression during H7 hESC differentiation towards somites, shown by scRNA-seq (i), qPCR (ii) and immunostaining (iii); scale bar = 50 μm

D. ATAC-seq shows the HOPX locus is accessible in D2.25 hESC-derived somitomeres (signal track: −log₁₀ P values)

E. Fate mapping progeny of Hopx⁺ cells in E14.5 Hopx-IRES-Cre;Ai9 embryos reveals contribution to the spine and ribs (labeled by type II collagen); scale bar = 50 μm

Consistent with the notion that Hopx also marks somitomeres in mouse embryos, genetic lineage tracing showed that Hopx⁺ cells contributed to the ribcage, cervical and lumbar vertebrae and tail (Fig. 3e). Sparse labeling reflects the noted inefficiency of the Hopx-IRES-Cre driver allele (Jain et al., 2015). Collectively, this uncovers a transient molecular signature of somite segmentation conserved between human and other vertebrates and suggests HOPX as a possible component of this signature.

Bifurcation of human early somite progenitors into sclerotome and dermomyotome by HH and WNT signals

After human somites are formed, how do they diverge into distinct derivatives (Fig. 1a, step ④)? In vivo, early somites are patterned along their dorsal-ventral axis to generate sclerotome (ventral somite; precursor to bone, cartilage and smooth muscle) and dermomyotome (dorsal somite; precursor to skeletal muscle, brown fat and dorsal dermis; Fig. 4a).

Figure 4. Dorsal-ventral patterning of somite precursors into sclerotome and dermomyotome and downstream progeny.

A. Somite patterning in vivo

B. qPCR heatmap of hESCs (D0), early somite progenitors (D3) or those differentiated into sclerotome (D4, D5 or D6, using 21K+C59) or dermomyotome (D4 or D5, using BMP4+CHIR+Vismodegib)

C. SOX9 and TWIST1 staining of day 6 H7-derived sclerotome; scale bar = 100 μm

D. PCA of scRNA-seq from indicated populations; each dot depicts a single cell

E. EF1A-BCL2-2A-GFP expressing H9-derived sclerotome was subcutaneously injected into NSG mice; 2 months later, ectopic GFP⁺ human bones formed (left); bioluminescent imaging of mice 1 month post-transplantation by UBC-Luciferase-2A-tdTomato H9-derived sclerotome

F. Russell-Movat’s Pentachrome staining of 2-month-old sclerotome grafts revealed zones of chondrogenesis and ossification, with cartilage stained blue; black line denotes the edge of the graft; white line denotes boundary of the ossifying region; scale bar = 1 mm (bottom)

G. COL2A1 (top left) and Safranin-O staining (top right) of D6+2 or D6+6 hESC-derived cartilage, respectively; scale bars = 0.1 mm (left) and 1 mm (right); SMAα intracellular FACS of hESCs or D8 fibroblast-like cells (bottom)

H. Somite patterning into dermomyotome or sclerotome and downstream differentiation

See also Fig. S4, Fig. S5

Starting from day 3 human early somite progenitors, HEDGEHOG (HH) induced sclerotome while WNT specified dermomyotome, and each cross-antagonized the effect of the other on days 3–5 of differentiation (Fig. 4b, Fig. S4a; summarized in Fig. 4h). By activating one signal while inhibiting the other it was possible to specify one somitic derivative while blocking formation of the alternate fate.

HH activation together with WNT inhibition blocked dermomyotome formation, inducing a fairly uniform SOX9⁺TWIST1⁺ sclerotome population that expressed multiple sclerotome markers (PAX1, PAX9, NKX3.2/BAPX1, FOXC2, SOX9 and TWIST1) (Fig. 4b,c, Fig. S4a,b). Conversely, WNT activation together with HH blockade blocked sclerotome formation and instead exclusively specified dermomyotome (Fig. 3b, Fig. S4a). If HH and WNT were simultaneously activated, neither sclerotome or dermomyotome was elicited (Fig. S4a), indicating the importance of their mutually-exclusive activation. Collectively this shows that HH and WNT pattern human somites, analogous to their role in mice (Fan et al., 1997; Fan et al., 1995; Fan and Tessier-Lavigne, 1994), and demonstrates how their manipulation can separately generate human sclerotome or dermomyotome. Indeed scRNA-seq showed that hPSC-derived sclerotome and dermomyotome populations formed in these two mutually-exclusive signaling conditions constituted largely distinct clusters (Fig. 4d) that diverged downstream of a common early somite progenitor state.

Human dermomyotome induction from early somites relied on WNT activation and initial HH inhibition (to block sclerotome induction) in addition to BMP (to induce PAX7; Fig. S4c) and later-stage HH activation (to induce EN1; Fig. S4d–f). In vivo, dermomyotome gives rise to skeletal muscle and accordingly, human dermomyotome was also capable of generating MYH3⁺ skeletal muscle cells in vitro (Fig. S4g).

Human sclerotome forms an ectopic bone in vivo

Strikingly, human sclerotome (ventral somite) progenitors could generate an ectopic human bone in vivo (Fig. 1a, step ⑤), reflecting the skeleton-forming fate of embryonic sclerotome (Christ and Scaal, 2008). Upon subcutaneous injection into immunodeficient mice, hESC-derived sclerotome formed ectopic bone-like structures containing both bone and cartilage (Fig. 4ei). Other tissues (i.e., epithelia) were not present (Fig. S5a), indicating the developmental lineage restriction of sclerotome. Labeling sclerotome with constitutive GFP (Fig. S5b), BCL2-2A-GFP (Fig. S5c) or Luciferase (Fig. S5d) reporters prior to transplantation confirmed that the subcutaneous GFP⁺ or luciferase⁺ bones (Fig. 4e) were not derived from resident mouse cells.

Sclerotome-derived ectopic human bones self-organized themselves even though they were implanted without a patterned matrix. Specifically, they harbored proliferative chondrocytes that progressed to hypertrophic chondrocytes and finally, underwent ossification in a spatially-choreographed fashion within column-shaped streams of cells (Fig. 4f); they also recruited host blood vessels (Fig. S5e), recapitulating a developmental endochondral ossification program. In summary, hESC-derived sclerotome harbored bone/cartilage progenitor activity in vivo and formed bones in a process mimicking natural bone development.

Human sclerotome could also generate cartilage and fibroblasts in vitro (Fig. 1a, step ⑥). Exposure to BMP (Murtaugh et al., 1999) upregulated cartilage structural genes, yielding a fairly-uniform COL2A1⁺ population that later secreted glycosaminoglycans (Fig. 4gi, Fig. S5f). By contrast, PDGF and TGFβ treatment (Cheung et al., 2012) for 3 days yielded a >90% pure SMAα^hi fibroblast-like population (Fig. 4gii, Fig. S5g). Having defined the signaling logic for PS differentiation into human paraxial mesoderm and somitic derivatives, next we focused on the parallel lineage branch: PS differentiation into lateral mesoderm and cardiac fates (Fig. 1a, step ⑦).

Lateral mesoderm patterning into heart vs. limbs by competing FGF and WNT signals

We sought to define how human lateral mesoderm is diversified into multiple derivatives, including NKX2.5⁺ heart-forming anterior lateral mesoderm and limb-forming PRRX1⁺ posterior lateral mesoderm (Tanaka, 2013) (Fig. 5a; step ⑦ in Fig. 1a).

Figure 5. Lateral mesoderm patterning into cardiac vs. limb mesoderm fates.

A. Cardiac vs. forelimb bifurcation

B. To assess the role of WNT in lateral mesoderm patterning, D1 PS was differentiated to lateral mesoderm (30 ng/mL BMP4 + 1 μM C59 + 2 μM SB505124) for varying lengths of time (until D2, D2.5 or D3) and for the last 12 hrs was treated with C59 or 3 μM CHIR (in addition to BS) and qPCR was conducted

C. To assess the role of FGF in lateral mesoderm patterning, day 2 NKX2.5-GFP lateral mesoderm was treated with BMP4 + C59 + SB505124 with or without FGF2 (20 ng/mL) or FGFR inhibitor PD173074 (100 nM) for 24 hrs and FACS was conducted on day 3

D. Timepoint FACS of NKX2.5-GFP hESC (Elliott et al., 2011) differentiation using cardiac mesoderm protocol

E. Comparison of NKX2.5-GFP⁺ cell percentages (determined by FACS) on days of differentiation, using the current protocol or a previous method (Burridge et al., 2014)

F. Intracellular TNNT2 FACS of H7-derived cardiomyocytes (bottom)

G. Electrocardiogram of human fetal heart implanted in the mouse ear, >1 month post-implantation

H. 2.5 months post-transplant of EF1A-BCL2-2A-GFP;UBC-tdTomato-Luciferase H9 hESC-derived cardiac lineages into human fetal heart grafts, luciferase⁺ donor cells were detected (i); engrafted hESC-derived cardiomyocytes were TROPONIN⁺ and CONNEXIN 43⁺, scale bar = 40 μm (ii)

See also Fig. S6

Bifurcation of day 2 lateral mesoderm into day 3 anterior (cardiac) vs. posterior (limb bud) fates was respectively induced by FGF and WNT signals (summarized in Fig. 5a). WNT posteriorized lateral mesoderm, inducing limb markers PRRX1 and HOXB5 while suppressing heart field markers NKX2.5 and TBX20 on day 2–3 of hESC differentiation (Fig. 5b, Fig. S6a). Reciprocally, WNT inhibition suppressed posterior lateral mesoderm and instead induced cardiac mesoderm (Fig. 5b, Fig. S6a). Thus, our findings corroborate the clear requirement for WNT blockade for cardiac specification in vivo (Schneider and Mercola, 2001) and in vitro (Burridge et al., 2014; Lian et al., 2012; Mendjan et al., 2014), and further explain that WNT inhibitors induce heart precursors by restraining limb formation (Fig. S6d).

Conversely FGF anteriorized human lateral mesoderm. Exogenous FGF enhanced NKX2.5-GFP⁺ cardiac mesoderm induction on day 2–3 of hESC differentiation, whereas FGF inhibition abolished cardiac induction (Fig. 5c). This demonstrates a conserved role for FGF in cardiac mesoderm specification from humans to zebrafish (Reifers et al., 2000).

Hence, activating pro-cardiac FGF signaling and inhibiting pro-limb WNT signaling (in the permissive context of BMP activation and TGFβ inhibition) efficiently directed day 2 lateral mesoderm towards cardiac mesoderm, as tracked using an NKX2.5-GFP knock-in hESC line (Elliott et al., 2011). We respectively obtained >80% and >90% NKX2.5-GFP⁺ cardiac mesoderm by days 3 and 4 of hPSC differentiation (Fig. 5d), which was more rapid and robust than a WNT modulator-only cardiac induction protocol (Burridge et al., 2014) (Fig. 5e).

Directing cardiac mesoderm into human cardiomyocytes can that engraft human fetal heart in vivo

Having rapidly generated a >90% pure NKX2.5⁺ cardiac mesoderm population by day 4 of hESC differentiation (Fig. 5d,e), we next sought to drive these progenitors towards cardiomyocytes (Fig. 1a, step ⑧). Single cardiac mesoderm progenitors can form both myocardium (cardiomyocytes) and endocardium in mice (Devine et al., 2014).

BMP activation together with low FGF levels preferentially induced cardiomyocytes from cardiac mesoderm (Fig. S6e,f) at the expense of (pro)epicardium or endocardium (Fig. S6g). WNT activation seemed to sustain undifferentiated ISL1⁺ cardiac progenitors and inhibited maturation into cardiomyocytes (Fig. S6g). Therefore, WNT blockade in conjunction with BMP activation enhanced cardiomyocyte differentiation (Fig. S6h), which was enhanced by vitamin C (Burridge et al., 2014) (Fig. S6i). Treating day 4 cardiac mesoderm with these factors yielded a 72.2±5.6% and a 77.8±1.6% pure TROPONIN⁺ cardiomyocyte population by days 6 and 8 of hPSC differentiation, respectively (Fig. 5f), which spontaneously contracted and expressed cardiomyocyte structural genes (Fig. S6j).

The authenticity of hESC-derived cardiac lineages was confirmed by their ability to engraft human fetal heart tissue. hESC-derived cardiomyocytes can engraft model organisms (e.g., guinea pigs and non-human primates (Chong et al., 2014)) but therapies will require evidence that such cells can engraft human heart tissue. To this end, we employed an experimental system whereby ventricular fragments from week 15–17 human fetal heart (Ardehali et al., 2013) were subcutaneously implanted into the mouse ear (Fig. 5g). These human heart fragments were revascularized and continued beating for months in vivo, as shown by QRS electrocardiogram signals (Fig. 5g). Upon transplantation of luciferase⁺/GFP⁺ hESC-derived heart populations, both day 3 cardiac mesoderm and day 8 cardiomyocytes engrafted human ventricular fragments for at least 10 weeks, as indicated by bioluminescence imaging (N = 10 successfully engrafted human heart fragments obtained from 2 fetal donors; Fig. 5hi). Within human fetal heart tissue, GFP⁺ hESC-derived cardiomyocytes expressed cytoskeletal protein TROPONIN/TNNT2 and membranous gap junction protein CONNEXIN 43 (Fig. 5hii). Altogether, these data show that deciphering the signaling logic for lateral and cardiac mesoderm induction can lead to the rapid generation of a ~75% pure cardiomyocyte population capable of engrafting human fetal heart tissue.

Identifying cell-surface markers to allow purification of diverging human mesoderm subtypes

After generating this hierarchy of mesoderm lineages, we also defined lineage-specific cell-surface markers to track different mesoderm lineages and enable purification of each mesoderm lineage for assessment of biological function and fate, and for potential therapeutic purposes in the future. Screening 332 cell-surface markers across hESCs and 7 mesodermal lineages (Fig. 6a, Fig. S7a, Table S2) revealed that certain previously-described mesoderm markers were broadly expressed in both paraxial and cardiac mesoderm (Fig. S7b,c). Therefore we sought lineage-specific markers.

Figure 6. High-throughput screen for lineage-specific mesoderm surface markers.

A. Clustered heatmap of surface marker expression in hESCs and 6 mesoderm derivatives. Each row represents an individual surface marker and color denotes the percentage of cells positive for a given marker. For PS and cardiac mesoderm, marker expression was analyzed after pre-gating on MIXL1-GFP⁺ and NKX2.5-GFP⁺ fractions, respectively.

B. GARP and DLL1 FACS in hESCs, Day 2 paraxial mesoderm cultures or Day 3 NKX2.5-GFP⁺ pre-gated cardiac mesoderm

C. lrrc32 in situ hybridization in 24 hours post fertilization zebrafish embryo (arrows denote heart)

D. DLL1 FACS of day 2 paraxial mesoderm culture (left); qPCR of sorted populations (right)

E. scRNA-seq of sorted DLL1⁺ human paraxial mesoderm; each dot is a single cell

F. PDGFRα FACS of day 5 sclerotome population (left); qPCR of sorted PDGFRα⁺ and PDGFRα⁻ populations (center); in situ hybridization for pdgfra expression (right) in 22 hpf zebrafish embryo (arrowheads denote ventral staining in sclerotome)

G. scRNA-seq of sorted PDGFRα⁺ human sclerotome; each dot is a single cell

See also Fig. S7, Table S2

Surface markers DLL1 and GARP respectively marked paraxial mesoderm and cardiac mesoderm in a mutually-exclusive fashion. NOTCH ligand DLL1 was specifically expressed in day 2 paraxial mesoderm, whereas conversely GARP marked day 3 NKX2.5-GFP⁺ cardiac mesoderm; neither marker was expressed by undifferentiated hESCs (Fig. 6b, Fig. S7d), thus tracking a clear bifurcation of PS into paraxial or cardiac mesoderm fates.

DLL1 and GARP were mesoderm markers conserved from human to zebrafish. GARP/LRRC32, a transmembrane protein that tethers TGFβ ligands to the cell surface (Tran et al., 2009), was likewise expressed in the heart tube of zebrafish embryos (Fig. 6c). Conversely, deltaC (a homolog to human DLL1) was likewise expressed in zebrafish paraxial mesoderm (Fig. S7e), as reported previously (Smithers et al., 2000).

These surface markers enabled the purification of desired mesoderm lineages from admixed cultures, providing tools to refine stem-cell differentiation. DLL1 was expressed by 91.6±5.4% of cells after 2 days of paraxial mesoderm induction. Paraxial mesoderm-specific TFs (TBX6, MSGN1) were exclusively expressed by the DLL1⁺ fraction (Fig. 6d). Sorted DLL1⁺ human paraxial mesoderm cells were essentially uniform as shown by scRNA-seq: 97.0% of cells coexpressed archetypic paraxial markers MSGN1 and DLL3 (Fig. 6e). TBX6 mRNA expression in all but a few cells (Fig. 6e) may reflect technical dropout in scRNA-seq (Marinov et al., 2014). Hence, sorting for DLL1⁺GARP⁻ cells purifies human paraxial mesoderm attained from either differentiating ESCs or iPSCs (Fig. S7f), providing a method to isolate pure human paraxial mesoderm and interrogate its characteristics.

Downstream of paraxial mesoderm during the bifurcation of sclerotome vs. dermomyotome fates, surface marker PDGFRα enabled the purification of sclerotome. PDGFRα was expressed by 85.2±8.4% of cells in day 5–6 sclerotome cultures, and only the PDGFRα⁺ fraction expressed sclerotome markers (FOXC2, PAX1 and PAX9), consistent with pdgfrα expression in the sclerotome but not dermomyotome of zebrafish embryos (Fig. 6f) (Liu et al., 2002). In vitro, PDGFRα was indeed expressed at higher levels in hPSC-derived sclerotome relative to dermomyotome (Fig. S7g), thereby helping to distinguish ventral from dorsal somite fates.

scRNA-seq of PDGFRα⁺ human sclerotome revealed 86.2% of cells coexpressed chondroprogenitor markers SOX9 and COL2A1 (Fig. 6g), reaffirming the skeletal stem cell-like nature of these cells and consistent with their phenotypic ability to form ectopic bone grafts (see above). However PAX9 and FOXC2 expression was more heterogeneous (Fig. 6g), which may reflect distinct anterior-posterior and medial-lateral sclerotome subdomains (Christ and Scaal, 2008).

In summary, these cell-surface markers define a roadmap for mesoderm development (Fig. 7a) by identifying mutually-exclusive types of mesoderm progenitors, thus enabling one to track the products of key developmental branchpoints. These markers are likewise expressed by the same cell types in zebrafish embryos (Fig. 6, Fig. S7). These markers allowed us to confirm that hPSC differentiation to various mesoderm lineages was efficient (Fig. 6) and enabled further purification of desired lineages for transcriptional and chromatin analysis.

Figure 7. The landscape of mesoderm development.

A. Lineage steps with circled numbers correspond to respective sections in the main text and Fig. 1a

B. RNA-seq expression of human congenital scoliosis genes

C. RNA-seq profiling; color intensity depicts gene expression (log₂ TPM) normalized to the expression of that gene in all populations profiled, with the highest-expressing lineage assigned the most intense color value

D. ATAC-seq heatmap; each horizontal line depicts a single chromatin element (left, non-binarized in Fig. S8b), with motifs representative of the top 4 lineage-enriched motifs shown (right)

E. Inferred trans-regulatory lineages programs (left); heatmap of the 4 FOX TFs most highly expressed in hESC-derived somites (RNA-seq; right)

F. ATAC-seq of the MEOX1 locus, with FOX motifs centered in two somitic enhancer elements shown

See also Fig. S8

Global transcriptional profiling of the mesoderm lineage hierarchy confirms lineage demarcations and suggests a cell-of-origin for congenital malformations

To chart a molecular roadmap for mesoderm development, using bulk-population RNA-seq we captured global transcriptional dynamics during the commitment of ESCs to nine diverging human mesodermal fates spanning multiple developmental stages (Table S3). This mesoderm gene expression atlas provided insight into the potential cell-of-origin of various human congenital malformations.

Congenital scoliosis is a genetically-heterogeneous disease mapping to diverse genes in human patients, where it causes malformations in the spine, scapulae and/or ribs (reviewed by Giampietro, 2012; Pourquié, 2011). Due to the inaccessibility of early human embryos, it has been difficult to assess when and where scoliosis susceptibility genes are expressed during development to uncover the origins of this disease.

RNA-seq analysis inferred at least two independent cells-of-origin for different subtypes of congenital scoliosis. For spondylocostal dysostosis (mapped to DLL3, HES7, TBX6 and LFNG) and sacral agenesis with vertebral anomalies (BRACHYURY), their causative genes were largely expressed in PSC-derived human paraxial mesoderm but not other mesodermal cell-types (Fig. 7b, Fig. 7ci). This implies a paraxial mesoderm cell-of-origin for these two types of scoliosis. By contrast, for six other types of congenital scoliosis, their causative genes were strongly expressed in human sclerotome, but not paraxial mesoderm (Fig. 7b). Hence congenital scoliosis may have at least two independent cells-of-origin (paraxial mesoderm or sclerotome) depending on the specific genetic lesion.

RNA-seq analyses of cells diverging across multiple lineage branchpoints also provided a clear view of how fates segregate across consecutive developmental steps. After PS formation, there was a clear partitioning of gene expression patterns in paraxial mesoderm (TBX6) and lateral/cardiac mesoderm (GARP/LRRC32) (Fig. 7cii). Downstream of paraxial mesoderm, PAX1 and PAX3 (Fan and Tessier-Lavigne, 1994) were respectively restricted to either sclerotome or dermomyotome (Fig. 7ciii). These lineage-specific expression patterns demarcate differences in developmental fate and lend confidence to our transcriptional dataset.

The human mesoderm gene expression atlas also uncovered lineage-specific long noncoding RNAs (lncRNAs), nominating them for further study of lncRNA function. By way of example, Fendrr is a lncRNA critical for mouse heart field development (Grote et al., 2013) and human FENDRR was likewise expressed in hESC-derived cardiac mesoderm (Fig. 7biv).

Mesodermal distal regulatory elements reflect the impact of dynamic signaling and trans-regulatory influences on chromatin

To track how chromatin is dynamically remodeled during development of hESCs into 9 distinct types of mesodermal progeny, we charted open chromatin using ATAC-seq to identify putative regulatory elements. We also inferred active TFs for each mesodermal lineage from TF motifs enriched in stage-specific open chromatin regions and TFs that were expressed according to RNA-seq (Supplemental Procedures).

Though accessible chromatin in pluripotent cells was enriched for OCT/SOX-binding motifs, upon 24 hours of differentiation, the PS chromatin landscape became dominated by motifs of TCF/LEF TFs (the effectors of WNT signaling; Fig. 7d), reflecting how WNT drives PS induction (Fig. 1b). As differentiation progressed, the trans-regulatory landscape of day 2 paraxial mesoderm seemed to be built on that of day 1 PS, consistent with how both lineages experience WNT activation (Fig. 2): TCF/LEF continued to engage paraxial mesoderm chromatin, but was apparently joined by paraxial mesoderm-specific TFs MSGN1 (P<10⁻³⁸²) and TBX6 (P<10⁻³⁵) (Fig. 7d, Fig. S8c). However, within 24 hours of WNT inhibition and the segmentation of paraxial mesoderm into somites, the landscape transitioned from a TCF/LEF-dominated state to one significantly enriched for FOX motifs (P<10⁻⁶¹; Fig. 7d). This was evidenced by predicted FOX motifs in two upstream MEOX1 enhancers that were accessible in somites (Fig. 7f). By virtue of RNA-seq expression patterns (Fig. 7e), multiple FOX TFs could account for the FOX-driven somite regulatory state, including FOXC2, which was indeed expressed at the protein level in somites (Fig. 2).

Along the alternate lineage pathway, the 24-hour progression from PS to lateral mesoderm involved an transition from a TCF-driven chromatin landscape to a GATA-dominated one (Fig. 7d). This reflects the importance of WNT repression in lateral mesoderm specification (Fig. 1e) and expression of multiple GATA TFs in lateral mesoderm (Fig. S8c). Upon 24 hours of further differentiation into GARP⁺ cardiac mesoderm, GATA motifs became accompanied by HAND1 (P<10⁻¹⁴⁹) and NKX2.5 (P<10⁻³⁹) (Fig. 7d). These findings provide insight into control of mesoderm development: there is no monolithic “pan-mesodermal” program but instead chromatin is substantially remodeled every 24 hours even as closely-related mesodermal lineages segue into one another (Fig. 7d). Furthermore the distal element landscape reflects how changes in signaling influences and trans-acting regulatory states become physically imprinted on chromatin. Altogether this sketches a model for how regulatory states change during mesoderm subtype diversification (Fig. 7e).

DISCUSSION

A roadmap for human mesoderm development

Waddington’s developmental landscape drawings (Waddington, 1940) illustrated how, during development, stem cells navigate brachiating lineage choices, entering a single lineage path while avoiding alternate paths leading to illegitimate fates. To navigate this complex landscape and coherently guide stem-cell differentiation, we must map (1) the identity of transitional intermediate states, (2) the sequence of branchpoints through which these intermediates are produced, and (3) the extracellular cues that specify cell fate at each exact branchpoint. However, due to incomplete knowledge of the underlying developmental landscape, in vitro stem-cell differentiation often yields admixed lineages, possibly due to incomplete suppression of alternate fates or passage through incorrect lineage intermediates. If we wish to systematically block production of unwanted lineages in preference to desired fates during stem-cell differentiation, we must precisely map the underlying landscape.

To meet this challenge with respect to the three above goals, here we chart a roadmap for human mesoderm development and describe how twelve different human cell-types including bone, muscle and heart emerge from pluripotent cells (Fig. 7a). Through a reductionist approach, we used single-cell RNA-seq to systematically catalog the diversity of intermediate cell states formed during differentiation and we tested the minimal combinations of positive and negative signals that were sufficient for differentiation between each of these intermediate states. Though vertebrate mesoderm development was broadly outlined by pioneering analyses in model organisms that identified certain key genes and signaling pathways (reviewed by Kimelman, 2006; Schier and Talbot, 2005; Tam and Loebel, 2007), it has been difficult to precisely map mesoderm formation due to the large number of mesodermal subtypes and the finely-graded, temporally-dynamic transitions between them. Throughout consecutive pairwise lineage branches in human mesoderm development, we clearly defined (i) the diverging cell states through single-cell RNA-seq, (ii) positive and negative signals inducing each of the mutually-exclusive lineages, (ii) specific cell-surface markers that identified key mesoderm intermediates and (iv) the chromatin landscapes of the diverging fates (Fig. 7a). Besides providing a broad reference map for developmental biology and regenerative medicine, we directly demonstrate the applications of this roadmap to produce engraftable human tissue progenitors and provide insight into developmental signaling dynamics, chromatin remodeling and congenital disease.

Extrinsic signals: logically blocking alternative lineage formation to guide stem-cell differentiation

Vertebrate embryology has identified certain signals required for mesoderm formation in model organisms (reviewed by Kimelman, 2006; Schier and Talbot, 2005; Tam and Loebel, 2007) and here we have tested whether we truly understand mesoderm development at the level of causation by reconstituting aspects of this process from cultured stem cells. At each lineage transition from human pluripotency to terminally-differentiated mesoderm fates, we could rigorously identify and test the minimal signaling conditions needed to induce each lineage. The resultant understanding of the underlying signaling logic guided the rapid differentiation of hPSCs into desired mesoderm intermediates (>98% pure MIXL1⁺ primitive streak; >90% pure NKX2.5⁺ cardiac mesoderm; >90% pure DLL1⁺ paraxial mesoderm; >95% pure FOXC2⁺ early somite progenitors) within several days of differentiation in serum-free, monolayer conditions (without recourse to gene modification). Such efficient, rapid induction relied on the following two principles.

Firstly, the principal finding of this work is that at each lineage bifurcation, stem cells could be exclusively differentiated down a single lineage path by simultaneously providing the positive signal(s) to induce a given fate while repressing inhibitory signal(s) that instead induced the alternate fate. Blocking the formation of undesired fates was imperative for efficient differentiation. By way of example, efficient differentiation of day 1 PS into day 2 paraxial mesoderm required WNT activation (to specify paraxial mesoderm) together with the simultaneous inhibition of BMP and TGFβ pathways (to block lateral mesoderm and endoderm formation, respectively) in order to block differentiation towards unwanted fates and to consolidate differentiation down only a single path. Therefore hPSC differentiation to a desired lineage cannot solely rely on knowledge of the requisite inductive signal(s), but also an understanding of signals that induce mutually-exclusive fates at each step of the way. This underscores the need for systematic developmental roadmaps.

Secondly, another highlight was the rapidity with which developmental signals were re-interpreted during hPSC differentiation and the consequent importance of controlling temporal signaling dynamics. In the gastrulating mouse embryo, lineage transitions occur every 12–24 hours, for example: E5.5 epiblast → E6.5 primitive streak → E7-7.5 paraxial mesoderm → E8 early somites → E8.5 sclerotome. In vitro, we found that WNT activation on d0-1 drove hPSC towards PS; WNT activation on d1-2 then specified paraxial mesoderm; WNT inhibition on d2-3 differentiated paraxial mesoderm into early somites; and finally, WNT activation on d3-4 specified dermomyotome. Thus, over the course of 4 days in vitro, WNT was interpreted 4 different ways as lineages segued into one another every 24 hours. By contrast, some differentiation methods continuously provide the same signal for days or weeks, potentially explaining why a mélange of lineages is produced. Our system therefore constitutes a venue to mechanistically understand how extrinsic signals are dynamically interpreted in the context of changing windows of developmental competence.

Single-cell RNA-seq: cataloging mesodermal lineages and transition states in between them

Mapping a developmental hierarchy hinges on cataloging its constituent progenitor states, which we have done here for the human mesodermal lineage hierarchy. We proposed that single-cell RNA-seq sampling (Table S4) would be a complete method to test the lineage and homogeneity of cells at each developmental step. Indeed early human primitive streak, lateral mesoderm and paraxial mesoderm lineages were highly uniform. Starting from human paraxial mesoderm, cells initiated somitogenesis along a single continuous trajectory (Fig. 2g) and snapshots of this process uncovered the formation of a transient HOPX⁺ human somitomere intermediate that arose for several hours during differentiation. This argues that human somite development entails passage through an ephemeral segmentation-like state (as described for other vertebrate model organisms; Pourquié, 2011), which has hitherto been impossible to assess in vivo due to the unavailability of early human embryos and its transient nature.

Navigating mesoderm development

Identifying lineage-specific cell-surface markers for major mesoderm subtypes (e.g., DLL1 for paraxial mesoderm and GARP for cardiac mesoderm) enables the purification of desired mesoderm subtypes to investigate the biological characteristics of these cells (as embodied by our RNA-seq and ATAC-seq analyses) or potentially for therapeutic transplantation in the future. Moreover the major surface markers defined here for human mesoderm progenitors were correspondingly expressed in zebrafish embryos indicating that they are conserved developmental markers.

Collectively we delineate a clear lineage hierarchy for mesoderm development with prospectively-isolatable lineage intermediates at each step which should be key for understanding human mesoderm development as well as the clinical purification of hPSC-derived tissue stem and progenitor cells for regenerative medicine in the future (Fig. 7a). The ability to produce highly-homogeneous populations of human mesodermal progenitors now opens the gateway to the rapid generation and purification of a wealth of different mesodermal cell-types from hPSCs—including the engraftable sclerotome and cardiomyocyte populations described here—providing a potential future foundation for regenerative medicine. Artificially reconstituting aspects of mesoderm development from hPSCs should provide a facile system to study basic developmental processes in vitro, including how developmental signals are temporally re-interpreted and combinatorially integrated and how chromatin dynamics are linked to changing windows of developmental competence.

Yet the roadmap remains incomplete. It does not include lineage paths to human axial mesoderm, intermediate mesoderm or mediolateral derivatives of sclerotome and dermomyotome (Christ and Scaal, 2008). Finally, though we have identified extracellular signals that specify human mesoderm cell-fate in vitro, to accompany the jump of complexity from 2D culture dish to 3D embryo we must in turn map the niche cells that produce these signals during embryogenesis and where they are located, thus unraveling differentiation in 3D space.

EXPERIMENTAL PROCEDURES

Mesoderm differentiation

Monolayer, feeder-free differentiation (detailed in Supplemental Procedures) was conducted in serum-free CDM2 basal medium. hPSCs (mainly H7) were passaged ~1:12–1:20 as fine clumps (using Accutase) onto Geltrex-coated wells and cultured overnight in mTeSR1 + 1 μM thiazovivin. The next morning, hPSCs were differentiated towards anterior PS (30 ng/mL Activin + 4 μM CHIR + 20 ng/mL FGF2 + 100 nM PIK90; for downstream paraxial differentiation) or mid PS (30 ng/mL Activin + 40 ng/mL BMP4 + 6 μM CHIR + 20 ng/mL FGF2 + 100 nM; for downsteam lateral differentiation) for 24 hrs. D1 anterior PS → D2 paraxial mesoderm (1 μM A8301 + 3 μM CHIR + 250 nM LDN193189 [DM3189] + 20 ng/mL FGF2); 24 hrs. D1 mid PS → D2 lateral mesoderm (1 μM A8301 + 30 ng/mL BMP4 + 1 μM C59); 24 hrs.

D2 paraxial mesoderm → D3 early somites (1 μM A8301 + 1 μM C59 [or alternately, 1 μM XAV939] + 250 nM LDN193189 + 500 nM PD0325901; 24 hrs). D3 early somites → either D5-6 sclerotome (5 nM 21K + 1 μM C59) or D5 dermomyotome (3 μM CHIR + 150 nM Vismodegib, sometimes with 50 ng/mL BMP4); 48–72 hrs. D5 sclerotome → D8 fibroblast-like cells (10 ng/mL TGFβ1 + 2 ng/mL PDGF-BB (Cheung et al., 2012)); 72 hrs. D6 sclerotome → D9-D12 cartilage (20 ng/mL BMP4); 3–6 days.

Ectopic human bone formation

~1.5×10⁷ D6 sclerotome cells were subcutaneously transplanted in 1:1 CDM2/Matrigel mixture into NOD-SCID Il2rg^−/− mice, which were sacrificed in ~2–3 months.

Human fetal heart graft construction

Week 15–17 human fetal hearts were dissected into 2–7mm strips and subcutaneously implanted into the ear of NOD-SCID Il2rg^−/− mice using a trocar. 1 month later, 1.5–2×10⁶ D3 cardiac mesoderm or D8 cardiomyocytes were directly injected into the fetal heart tissue (in 1:1 CDM2/Matrigel mixture). Experiments were terminated ~2.5 months later.

RNA-seq

RNA was purified from H7-derived mesoderm lineages, either from bulk populations or from 651 single cells spanning 10 lineages (Fluidigm C1 system). RNA-seq libraries were prepared (bulk; Ovation RNA-seq System V2 and NEBNext Ultra DNA Library Prep Kit and single-cell; SMARTer Ultra Low RNA Kit) and sequenced (Next-Seq 500) to obtain 150bp paired-end reads, which were processed using the ENCODE long RNA analysis pipeline (Supplemental Procedures). Collated data viewable at http://cs.stanford.edu/~zhenghao/mesoderm_gene_atlas.