Signaling mechanisms and dynamics governing the myocardial-epicardial fate switch during human cardiogenesis

Abstract

The signaling mechanisms and developmental dynamics that govern the divergence of myocardial and epicardial lineages during human heart development remain poorly understood. Here, we developed a human pluripotent stem cell-based cardiac development model and employed time-course single-cell RNA sequencing to delineate cardiac lineage specification trajectories. We identified retinoic acid (RA) as a critical fate switch at the cardiac mesoderm stage. RA instructs epicardial lineage commitment of cardiac mesoderm through a primed-epicardium to proepicardium-like population and finally to epicardium, a process requiring precise BMP modulation. Conversely, RA absence directs cardiac mesoderm along a default myocardial pathway, yielding developing and mature cardiomyocytes. Both trajectories are governed by the hierarchical activation of key transcription factors. Our study integrates signaling and dynamics to elucidate the temporal regulatory network of the RA-BMP axis in human cardiac fate determination. These findings provide fundamental insights into human cardiogenesis and a crucial roadmap for modeling heart disease and advancing regenerative strategies.

Uncover RA-BMP axis as a switch controlling cardiac mesoderm toward epicardial or mycardial fate using hPSC model and scRNA-seq.

INTRODUCTION

The heart is the first functional organ formed during human embryogenesis, underscoring the fundamental importance of understanding its early development. However, direct study of early human cardiogenesis in vivo remains challenging due to limited tissue accessibility. Human pluripotent stem cell (hPSC) models offer a powerful alternative (1–3), having already elucidated key mechanisms such as WNT-BMP-HAND1 axis-mediated cardiac cavity formation (4) and NRP2-driven epicardial epithelial-mesenchymal transition (EMT) (5). Despite these progresses, a fundamental limitation persists: these models fail to reconstruct the initial segregation of myocardial and epicardial lineages, a pivotal early event in heart development.

Cardiac development is orchestrated by multiple spatiotemporally distinct progenitor populations that originate from MESP1-positive mesoderm during gastrulation (6, 7). Within this complex process, the emergence of cardiomyocytes (CM) and epicardial cells (Epi)—two functionally indispensable lineages—has been a particular focus of research. CMs, responsible for cardiac contraction, are known to derive primarily from the first and second heart fields (FHF, SHF) (8–12), as well as the juxta-cardiac field (JCF) (13, 14). In parallel, the epicardium contributes essential roles in myocardial growth, gives rise to non-myocardial lineages including cardiac fibroblasts, and mediates injury responses (15, 16). While JCF (13, 14) and Tbx18⁺ progenitors (17–19) have been implicated in the Epi lineage, a complete understanding of its developmental origins has remained less clear (20–25).

A growing body of evidence suggests that the segregation of myocardial and epicardial progenitors occurs remarkably early. For instance, JCF progenitors are detectable as early as the head-fold stage, coincident with the initial specification of cardiomyocytes (13). More definitively, clonal analysis has demonstrated that unipotent Mesp1⁺ epicardial progenitors diverge from the myocardial lineage during the earliest phases of cardiac mesoderm formation (26). These observations collectively point to an initial, fate-determining branch point at the early stage of heart development. Despite these advances, the regulatory logic underlying this early lineage bifurcation has remained obscure. Thus, a central and unresolved question persists: what are the precise signaling factors and transcriptional mechanisms that govern the initial separation of myocardial and epicardial progenitor lineages? Resolving this early fate decision is critical to a mechanistic understanding of heart development and the cell lineage relationships that underpin its form and function.

Here, to investigate the early mechanisms of human heart development, we analyzed the published single-cell sequencing dataset of a gastrulating human embryo—a stage corresponding to the initiation of cardiogenesis (27). This analysis revealed differential expression and reception of retinoic acid (RA) signaling across distinct germ layers. Notably, high RA responsiveness was observed in both nascent mesoderm and advanced mesoderm populations, which are known to give rise to human cardiogenic mesoderm (27, 28). Additionally, RA has been applied as an inductive cue in certain in vitro studies aimed at generating epicardial cells from PSCs (5, 29, 30). To functionally examine the role of RA in cardiac lineage specification, we established a human embryonic stem cells (hESC)-based model that recapitulates the developmental progression from pluripotency to key functional lineages: cardiomyocytes and epicardial cells. Using time-course single-cell genomics analysis integrated with in vivo reference data, we reconstructed the developmental trajectories of myocardial and epicardial cells, and identified the earliest transcription factors associated with their fate divergence. Our results establish RA as a critical mediator directing cardiac mesoderm cells away from a myocardial fate and toward an epicardial lineage. Furthermore, mechanistic studies revealed that RA exerts this fate-switching function primarily through the precise modulation of BMP signaling.

RESULTS

Retinoic acid is a critical fate switch of epicardial and myocardial cells

To better simulate in vivo developmental process and signaling dynamics, we produced embryoid bodies (EBs)in ultra-low attachment 96-well round-bottom plates using human embryonic stem cell line H9 (Fig. 1A). These spheroids were subsequently exposed to differentiation cues through combined stimulation with the WNT pathway agonist CHIR-99021 and BI-1347 [a small molecule inhibitor of Cyclin-dependent kinase 8, reported to favor cardiac mesoderm specification (31)] for 48 hours, sequentially inducing mesoderm and cardiac mesoderm formation. Following the 48-hour CHIR-99021/BI-1347 treatment, spheroids were cultured in basal medium without additional inductive factors, designated as the spontaneous differentiation (SD) group (Fig. 1A). By day 8, spontaneous localized beating was observed in some spheroids of SD (Fig. 1B and movie S1). Immunofluorescence analysis at day 8 revealed the presence of both TNNT2⁺ cardiomyocytes and WT1⁺ epicardial cells within the spheroids (Fig. 1C). By day 10, all spheroids exhibited localized beating under a microscope (Fig. 1B). Quantitative analysis revealed that while all day 8 spheroids contained both cardiomyocytes and epicardial cells, the cell number of these lineages varied randomly among individual spheroids (Fig. 1D).

Fig. 1. Retinoic acid is a critical fate switch of epicardial and myocardial cells.

(A) Schematic of hESCs towards cardiac differentiation under the SD, CM or EPI protocols. (B) Kinetics of beating spheroids across differentiation conditions at the indicated days. Data are shown as mean ± SD. n = 24 spheroids for each condition, 3 independent differentiations. (C) Representative bright-field images and immunostaining for the myocardial marker TNNT2 and epicardial marker WT1 in spheroids of SD, CM and EPI group at day 8. Scale bars, 100 μm. (D) Quantification of TNNT2⁺ cells and WT1⁺ cells from (C). Data are shown as mean ± SD, analyzed by one-way ANOVA with a Tukey’s multiple comparisons test. Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 6 spheroids for each condition, 6 independent differentiations. ∼1,300 cells per confocal field. (E) Immunostaining for TNNT2 and WT1 in day 8 spheroids treated with indicated RA concentrations. Scale bars, 100 μm. (F) Quantification of TNNT2⁺ cells and WT1⁺ cells from (E). Data are shown as mean ± SD, analyzed by one-way ANOVA with a Dunnett’s multiple comparisons test, using 0 μM RA as control group. Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 6 spheroids for each condition, 3 independent differentiations. (G) Expression of representative pluripotency, mesoderm, cardiac mesoderm, myocardial and epicardial marker genes during CM and EPI differentiation (days 0–10). The heatmap displays row-scaled z-scores of log₂(FPKM +1) values. n = 3 independent samples per time point. E8, TeSR-E8 medium; Advanced 1640 + B27, advanced RPMI 1640 supplemented with B27; EB, embryoid body; SD, spontaneous differentiation; CM, cardiomyocyte differentiation; EPI, epicardial differentiation.

Next, we attempt to control the myocardial-epicardial lineage allocation in the spheroids with exogenous signals. Established knowledge has shown that inhibition of WNT signaling following mesoderm formation is critical for cardiomyocyte specification. From days 2 to 4, we administered the WNT inhibitor IWP2 to enhance myocardial differentiation (Fig. 1A). By day 5, some spheroids exhibit global beating, and by day 6, global beating can be observed in all spheroids (Fig. 1B and movie S2). We named this group the cardiomyocyte (CM) group (Fig. 1A). Immunofluorescence staining at day 8 revealed over 80% TNNT2⁺ cardiomyocytes with very few or absent WT1⁺ epicardial cells in each spheroid of the CM group (Fig. 1, C and D).

In research on the differentiation of PSCs into epicardial cells, multiple inductive factors have been mentioned, including BMP4, FGF2, CHIR-99021, and retinoic acid. To select suitable exogenous signals for promoting epicardial cell development, we analyzed a published human gastrula-stage single-cell sequencing dataset (27). During the analysis, retinoic acid signaling attracted our attentions. Consistent with findings from in vivo animal studies, we observed that human embryonic mesoderm cells begin to express RA synthetase endogenously during gastrulation (32–34) (fig. S1A). Benefiting from single-cell resolution, we precisely localized the expression of embryonic development-related RA synthetase gene ALDH1A2 (35) to a portion of nascent mesoderm, while the cell differentiation-associated RA receptor RARB (36, 37) was highly expressed in some advanced mesoderm cells (fig. S1. A and B). Both nascent and advanced mesoderm populations contribute to cardiogenic mesoderm (27, 28). We observed declining CYP26 expression from emergent/nascent to advanced mesoderm transition (fig. S1A), underlying a window of altered RA sensitivity that likely guides lineage progression. The RA nuclear transporter CRABP was abundant in advanced mesoderm (fig. S1, A and B). Collectively, these observations lead us to hypothesize that RA signaling plays a vital regulatory role during early heart development. This hypothesis is further supported by the application of RA in studies on the in vitro differentiation of PSCs into epicardial cells, which also suggests its potential regulatory effects (5, 29, 30).

When a certain concentration (0.5 μM) of RA was added into CM condition (IWP2 presence) during days 2 to 4, we were surprised to observe that no spontaneous beating occurred even at day 10 (Fig. 1, A and B). Immunofluorescence analysis at day 8 revealed approximately 60% WT1⁺ epicardial cells within the spheroids, with no or very few TNNT2⁺ cardiomyocytes detected (Fig. 1, C and D). We named this group the epicardial group (EPI) (Fig. 1A). Brightfield and staining images showed clear morphological differences between EPI and CM group, and the cells in the EPI group are more tightly aggregated (Fig. 1C). Temporal sampling of both groups to observe the expression of cell-specific genes revealed distinct gene expression patterns. In the CM group, a small subset of cells began expressing TNNT2 at day 4, with predominant TNNT2 expression observed by days 6and 8 (fig. S2C), and sarcomere-like structures detected by day 10 (fig. S2B), indicating cardiomyocyte maturation. In contrast, the EPI group showed rare WT1⁺ cells at day 6, with robust WT1 expression detected in the majority of cells by day 8 (fig. S2C).

To further elucidate the role of RA in myocardial and epicardial development, we conducted RA concentration gradient experiments and RA inhibition assays. During days 2 to 4 of differentiation, varying concentrations of RA were added into the CM medium with IWP2, and analyses performed at day 8. The results revealed a dose-dependent reduction in cardiomyocyte number within the spheroids as RA concentration increased (Fig. 1E). For epicardial cells, cell number exhibited a dose-dependent increase before RA concentration up to 0.5 μM, followed by a decrease with further increase of RA doses (Fig. 1, E and F). For RA inhibition assays, we employed BMS493, which is a pan-retinoic acid receptor inverse agonist commonly used to block RA activity. During days 2 to 4 of differentiation, treated the cultures with both IWP2 and 0.5 μM RA, without or with 1 μM or 2 μM BMS493. Compared to the RA-treated group without the inhibitor (the EPI group), supplementation with 1 μM BMS493 significantly reduced epicardial cells while increasing cardiomyocytes and localized beating by day 8 (fig. S2, D and E, and movie S3). In 2 μM BMS493 group, epicardial cells were nearly undetectable, and spheroids exhibited global beating (fig. S2, D and E, and movie S4), with TNNT2⁺ cardiomyocyte number approaching levels observed in the CM group (fig. S2E).

We performed bulk RNA sequencing (RNA-seq) on samples collected from CM and EPI groups at days 0, 2, 3, 4, 5, 6, 8, and 10 of differentiation to monitor developmental dynamics across time points. Prior to day 3, both groups shared identical induction protocol and samples. Analysis revealed progressive downregulation of pluripotent-related genes such as NANOG and OCT4as differentiation progressed, and concurrent upregulation of mesoderm markers TBXT and MIXL1 was observed at day 2. By day 3, both groups exhibited elevated expression of cardiac mesoderm-associated genes including GATA6, PDGFRA, and HAND1. At day 4, cardiac lineage transcription factors GATA4 became prominently expressed in both groups (Fig. 1G). Immunofluorescence staining of stage-specific markers corroborated transcriptomic findings (fig. S2F). By day 5, CM group showed robust expression of cardiomyocyte-related genes, with significant enrichment of the early myocardial key transcription factor NKX2–5. In contrast, EPI group exhibited very few NKX2–5⁺ cells at this stage (Fig. 1G and fig. S2F). At day 6, CM group demonstrated high expression of contractile genes TNNT2 and TTN, which continuously expressed or upregulated through later stages (Fig. 1G). ACTN2 and MYL3 were highly expressed in CM group at days 8 and 10, reflecting early ventricular differentiation, and the ventricular cardiomyocyte specific gene IRX4 was highly expressed at day 10 (Fig. 1G). In EPI group, early epicardial development-related gene NR2F2 (COUP-TFII) (38) began to be detected on day 4, followed by progressive upregulation of representative epicardial genes including SEMA3D, TBX18, TCF21, WT1, and BNC1 as development proceeded (Fig. 1G).

We also verified our differentiation protocols with embryonic stem cell lines h1 and h2 that isolated by our laboratory (39). The generation of cardiomyocytes and epicardial cells were highly reproducible across these two hESC lines (fig. S2G), with h1 exhibited slightly lower but no significant differentiation efficiency compared to H9 and h2 (fig. S2, G and H).

Single-cell transcriptomics reveals the repertoire of distinct cardiomyocyte and epicardial subtypes

To assess our system’s capacity to reproduce the spectrum of in vivo cardiomyocyte and epicardial subtypes, we performed single-cell RNA sequencing (scRNA-seq) of day-10 CM and EPI cultures. Subclustering analysis resolved distinct myocardial and epicardial subtypes within each population. CM-d10 sample segregated into six transcriptionally distinct subclusters (Fig. 2A). The vast majority of cells (86.2%) fell into subclusters 1–3, which uniformly exhibited robust expression of the core cardiomyocyte marker MYH6 (Fig. 2, A to C, and fig. S3). Subcluster 4 (∼10% of cells) was characterized by low expression of MYH6 and TTN but high levels of VIM, COL3A1, and COL1A1, suggesting that they are cardiac fibroblasts—a critical myocardial support population in vivo (40). Minor populations of endoderm-derived cells (subcluster 5, expressing TTR and FOXA2) and endothelial cells (subcluster 6, expressing KDR, CDH5, and CD34) were also identified, likely representing residual lineages (Fig. 2, B and C, and fig. S3). Further dissection of the cardiomyocyte pool (subclusters 1–3) revealed subtype diversification. Subcluster 1 highly expressed ventricular cardiomyocyte markers HEY2 and IRX4, indicative of ventricular identity. Subcluster 2 was enriched for the atrial natriuretic peptide precursor gene NPPA and atrial-specific markers MYL7 and NR2F2, indicating an atrial cardiomyocyte lineage. In contrast, subcluster 3 exhibited elevated expression of BMP2, TBX2, and RSPO3—genes associated with the atrioventricular canal cardiomyocytes (41, 42), suggesting a distinct developmental origin. In the CM condition, ventricular cardiomyocytes account for the highest proportion (Fig. 2, A and C).

Fig. 2. Single-cell transcriptomics reveals the repertoire of distinct cardiomyocyte and epicardial subtypes.

(A) UMAP projection of cells under the CM condition at day 10, colored by six annotated cell clusters. (B) Dot plot displaying the expression and prevalence of cell-specific marker genes used to define each cell cluster in (A). (C) Bar plot showing the proportions of the indicated cell types identified in the CM-d10 dataset. (D) UMAP projection of cells under the EPI condition at day 10, colored by seven annotated cell clusters. (E) Dot plot of cell-specific markers defining each cell cluster shown in (D). (F) Bar plot showing the proportions of the indicated cell types in the EPI-d10 dataset. (G) Integration of CM/EPI-d10 cells with corresponding populations from a human fetal heart reference, visualized on a UMAP. The inset indicates the dataset origin of each cell. (H) Pearson correlation heatmap assessing transcriptional similarity between the in vitro-derived clusters (CM/EPI-d10) and their in vivo fetal heart counterparts. To minimize bias from imbalanced cell type abundances, cells were downsampled to 500 per cell type prior to correlation analysis. vCM, ventricular cardiomyocytes; aCM, atrial cardiomyocytes; avc-CM, atrioventricular canal cardiomyocytes; Fibro, fibroblasts; Epi, epicardial cells; EPDCs, epicardium-derived progenitor cells; Epi-Fibro, epicardium-derived fibroblast; ncCM, non-chambered cardiomyocyte; AVC, atrioventricular canal; IFT, inflow tract; aCM-LA, atrial cardiomyocytes-left atria; aCM-LA, atrial cardiomyocytes- right atria; vFibro, ventricular fibroblast; aFibro, atrial fibroblast.

Analysis of the EPI-d10 sample identified seven subclusters (Fig. 2D). The majority of cells (88.8%, subclusters 1–4) expressed canonical epicardial markers WT1, TCF21, and TBX18, confirming their epicardium identity (Fig. 2, D to F, and fig. S4). The remaining minor subclusters were annotated as fibroblasts, endoderm-derived cells, and endothelial cells respectively based on established marker genes (Fig. 2, E and F, and fig. S4). Subclusters 1 and 2 exhibited similar epicardial marker profiles (WT1, TCF21, and TBX18); however, subcluster 1 was uniquely marked by high expression of proliferation-related genes (CDK1, MKI67, and TOP2A), leading us to designate them as proliferating Epi (subcluster 1) and Epi (subcluster 2), respectively. Subcluster 3 showed elevated expression of FRZB and MOXD1, reminiscent of epicardium-derived progenitor cells (EPDCs). Subcluster 4 was characterized by upregulated EMT-related genes TWIST1 and SNAI2, cardiac fibroblast markers POSTN and ALDH1A2, and epicardial genes NR2F2 and TCF21, pointing toward an epicardium-derived fibroblast (Epi-Fibro) fate. To rigorously benchmark the identity of these populations, we integrated the CM, Epi, and cardiac fibroblasts from our in vitro model with corresponding cell types from a human fetal heart reference (42) (Fig. 2G). Transcriptomic correlation analysis demonstrated that the in vitro-derived cardiomyocytes and epicardial cells closely aligned with their native in vivo counterparts, validating the physiological relevance of our differentiation model (Fig. 2H). Together, the in vitro human cardiomyocyte and epicardial development model recapitulates in vivo cardiogenesis.

Bifurcating cardiac lineage trajectories revealed by scRNA-seq

To further dissect the developmental processes of the myocardium and epicardium, we performed whole-transcriptome analysis by time-course scRNA-seq at the following stages: days 0 and 2 (pluripotency and mesoderm progenitors), day 3 (the progenitors from SD, CM, EPI conditions), and days 4, 6, and 10 (the committed CM and EPI lineages) (Fig. 3A). Following quality control, a total of 193,637 high-quality single-cell transcriptomes were retained for subsequent analysis. The dataset comprised cells from baseline timepoints (day 0: 17,149; day 2: 14,555) and from specified populations across later stages: day 3 (SD: 17,679; CM: 19,612; EPI: 13,298), day 4 (CM: 20,288; EPI: 17,988), day 6 (CM: 20,430; EPI: 19,463), and day 10 (CM: 15,671; EPI: 17,504) (Fig. 3B). Following unsupervised clustering analysis, all cells were annotated into 11 major cell types based on established lineage-specific markers (Fig. 3, C and D, fig. S5, and table S1). Cells from day 0 (d0) were ESCs, characterized by high expression of pluripotency markers such as NANOG and OCT4 (Fig. 3, B and D, and fig. S5). By day 2 (d2), the majority had differentiated into mesodermal cells, marked by the expression of representative mesodermal genes like TBXT, MSGN1 and MIXL1, with a minor fraction of undifferentiated cells remaining. At day 3 (d3), cells from all three groups (SD, CM, EPI) robustly expressed early cardiac mesoderm markers including GATA6, MEIS1 and PDGFRA, and were thus uniformly annotated as cardiac mesoderm. By day 4 (d4), the CM and EPI groups began to diverge. While both still expressed general cardiac mesoderm markers (GATA6, PDGFRA), they exhibited distinct transcriptional profiles. Cells in CM-d4 group specifically upregulated myocardial development genes such as MEF2C and PLN, with a subset beginning to express the early key myocardial transcription factor (TF) NKX2–5. As cardiomyocyte markers (MYH6 and MYL7) were not yet broadly expressed, this population was defined as cells primed for cardiomyocytes (CPCM) (Fig. 3, B to D). In contrast, EPI-d4 cells did not express myocardial genes but initiated expression of proepicardial related genes NR2F2 and CDH18 (43). Given the absence of the canonical epicardial marker WT1, this population was designated as cells primed for epicardium (CPEpi) (Fig. 3, B to D, and fig. S5). From day 6 (d6) onward, lineage commitment became firmly established. Cells in CM-d6 abundantly expressed representative myocardial genes (TNNT2, NKX2–5, MYH6, MYL3), a finding consistent across immunofluorescence, bulk RNA-seq, and scRNA-seq data (fig. S2A; Figs. 1G and 3D). Co-expression of ACTA1 and NPPA indicated an immature phenotype, leading to their classification as developing cardiomyocytes (DCM). By CM-d10, upregulation of subtype-specific genes (e.g., IRX4) suggested further maturation, and these cells were conclusively annotated as cardiomyocytes. The majority of EPI-d6 cells showed upregulation of proepicardium gene NR2F2, with a subset beginning to express TBX18, and were therefore defined as early proepicardium. By EPI-d10, cells demonstrated co-expression or individual expression of canonical epicardial markers (TBX18, TCF21, SEMA3D, WT1 and BNC1), confirming their identity as epicardial cells. Throughout the developmental process, a small number of endoderm-derived cells were intermittently detected from d3 onwards, and a minor population of endothelial cells was detected in d6 and d10 samples (Fig. 3, B to D, and fig. S5).

Fig. 3. Single-cell transcriptomics reconstructs the developmental trajectories of epicardial and myocardial lineages.

(A) Experimental timeline and sampling strategy for scRNA-seq across differentiation conditions and time points. (B) Integrated UMAP visualization of all single cells, colored by their collection time point. (C) Integrated UMAP visualization of the same dataset as in (B), now annotated and colored by the eleven identified cell types. (D) Dot plot illustrating the expression levels and prevalence of canonical markers used to define the cell types in (C). (E) Pseudotime trajectory (DDRTree) showing cell states (State 1–3) and cluster distribution. (F) Expression patterns of key lineage markers projected onto the trajectory, illustrating the transition from pluripotency to specified cardiac lineages. (G) Heatmap of the top 500 genes dynamically regulated along pseudotime, grouped into five co-expression clusters (C1–C5). The most significantly enriched KEGG pathways for each gene cluster are indicated. ESCs, embryonic stem cells; CPCM, cells primed for cardiomyocytes; DCM, developing cardiomyocytes; CM, cardiomyocytes; CPEpi, cells primed for epicardium; Epi, epicardial cells.

We next employed pseudo-temporal ordering to reconstruct the developmental trajectories of cardiac lineages. Monocle 2 algorithm revealed a bifurcating path originating from a common progenitor state (State 1), which subsequently diverged into two distinct branches corresponding to the cardiomyocyte (CM, State 2) and epicardial (Epi, State 3) lineages (Fig. 3E). Expression of key marker genes along this trajectory confirmed its biological relevance (Fig. 3F). The pluripotency marker OCT4 and mesodermal markers TBXT and MIXL1 were highly expressed at the beginning of state 1 (Fig. 3F). The cardiac mesoderm marker PDGFRA was expressed at the bifurcation point and persisted along both the CM and Epi branches. Further along the branches, lineage-specific markers MYH6 (in State 2) and TBX18 (in State 3) confirmed the successful specification of CM and Epi, respectively (Fig. 3F). Analysis of transcription factor dynamics provided further resolution of the lineage divergence (fig. S6). Mesodermal factors (TBXT and MIXL1) and cardiac mesoderm markers (GATA6 and MEIS1) were sequentially activated in State 1. The early cardiac factor GATA4 was upregulated at the onset of both branches, while the myocardial regulator MEF2C was specific to early State 2. Myocardial-related TFs NKX2–5 and TBX5 were highly enriched in the terminal phase of State 2, whereas the epicardial markers WT1 and TCF21 were predominant in the late stage of State 3 (fig. S6). To dissect the molecular programs underlying lineage specification, we identified five distinct clusters of differentially expressed genes (DEGs) (Fig. 3G and table S2). Genes associated with calcium signal, myocardial transcription factors, cardiac contraction, and energy metabolism transformation (C1) were progressively upregulated along the CM branch, illustrating the path from specification to functional maturation. Concurrently, a concerted upregulation of glycolysis, pyruvate metabolism and oxidative phosphorylation was observed (C2), accompanied by enhanced TGF-β signaling and matrix remodeling (C3), reflecting changes of metabolism and ECM in cardiac mesoderm progressing from a proliferative to a differentiated state. As development proceeded along the Epi branch (State 3), we observed a sequential enhancement of gene programs: First, those governing cytoskeletal organization and adherens junction (C4), followed by pathways mediating ECM-receptor interaction, cell adhesion, and chemokine signaling (C5) (Fig. 3G). This stepwise activation demonstrates the structural and functional maturation of the epicardial lineage. In summary, our pseudo-time analysis delineates the developmental trajectory from pluripotent stem cells to two key cardiac functional lineages—cardiomyocytes and epicardial cells—and reveals the distinct transcriptional and metabolic programs that guide their specification and maturation.

Retinoic acid specifies epicardial fate within a defined temporal window of cardiac mesoderm development

To determine the precise developmental window during which RA governs lineage specification, we first conducted a preliminary RA time-shift assay. In the control group, cells were differentiated toward CM without RA supplementation. Based on the CM differentiation protocol (CHIR-99021 and BI-1347 treatment from days 0–2, followed by IWP2 from days 2–4), RA was administered for a 48-hourinduction across sequential 24-hour intervals: days 2–4 (Group 1, equivalent to the EPI group), days 3–5 (Group 2), days 4–6 (Group 3), days 5–7 (Group 4), and days 6–8 (Group 5) (Fig. 4A). Immunofluorescence analysis revealed that only groups 1 and 2—RA was supplied between days 2–4 or 3–5—efficiently generated significant WT1⁺ epicardial populations (∼60%), while TNNT2⁺ cardiomyocytes were nearly absent (Fig. 4, B and C). In contrast, RA supplementation from day 4 onward (Groups 3–5) failed to induce epicardial specification and instead yielded predominantly cardiomyocytes, similar to the CM control (Fig. 4C). We also evaluated the temporal requirement for WNT inhibition by IWP2 in myocardial commitment. Following 48-hour induction with CHIR99021 and BI-1347, IWP2 was applied during days 2–4, 3–5, and 4–6, respectively (fig. S7A). Earlier IWP2 treatment (days 2–4 or days 3–5) consistently yielded high proportion of TNNT2⁺ cardiomyocytes, whereas delayed IWP2 inhibition(days 4–6) led to variable cardiomyocyte yields and sporadic epicardial cell formation—phenocopying the SD group(fig. S7, B and C). Principal component analysis (PCA) of bulk RNA-seq data from CM and EPI time-course samples indicated that transcriptional trajectories diverge markedly by day 4 of differentiation (Fig. 4D). Together, these results pinpoint day 3 as the critical period during which RA directs cardiac mesoderm toward an epicardial fate, thereby defining the developmental window for myocardial–epicardial lineage commitment.

Fig. 4. Retinoic acid specifies epicardial fate within a defined temporal window of cardiac mesoderm development.

(A) Schematic of the RA treatment timeline applied to cardiac differentiation cultures. (B) Representative immunostaining images for TNNT2 and WT1 in day 8 spheroids from the treatment groups outlined in (A). Scale bars, 100 μm. (C) Quantification of TNNT2⁺ and WT1⁺ cells from (B). Data are mean ± SD; analyzed by one-way ANOVA with a Dunnett’s multiple comparisons test, using the CM condition as control. Significance levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 6 spheroids for each condition, 3 independent differentiations. (D) Principal component analysis (PCA) of bulk RNA-seq data from CM and EPI time-course samples. (E and F) Single-cell clustering (E) and marker gene expression (F) for samples from days 2–4. (G and H) Developmental trajectory analysis via diffusion map, showing cell fates (G) and key marker expression (H). (I) Volcano plot of differentially expressed transcription factors in CPEpi versus myocardial progenitors (FHF-, aSHF-, and pSHF-like cells) (|log₂FC| > 0.25, adjusted P < 0.05).

Immunofluorescence and transcriptomic data confirmed that cardiac mesoderm constitutes the dominant population on day 3 (Fig. 1G and fig. S2F). To gain deeper insight into how RA influences early lineage segregation, we reanalyzed the scRNA-seq data from days 2 to 4. Unsupervised clustering identified 18 distinct subpopulations (clusters 0–17) (Fig. 4E). Based on established cell-specific markers, we annotated nine major cell types (Fig. 4, E and F, and fig. S7D). Clusters 5, 9, and 17—originating from day 2—expressed high levels of mesodermal markers such as TBXT and MSGN1. Cluster 9 displayed a marked down-regulation of MESP1 and an up-regulation of HOXA1 relative to clusters 5 and 17, supporting its annotation as advanced mesoderm, whereas clusters 5 and 17 were designated as nascent mesoderm (Fig. 4, E and F). Cells from day 3 (clusters 0, 1, 2, 6, and 11) across all three groups showed high transcriptional homogeneity and uniformly expressed cardiac mesoderm markers (GATA6, MEIS1, and PDGFRA) (fig. S7E), while lacking expression of early myocardial (NKX2–5) or proepicardial (NR2F2) transcription factors (Fig. 4, E and F, and fig. S7E). In addition, we found these cells exhibit relatively specific expression of HAPLN1 and COL1A2 (Fig. 4F). These results confirm that by day 3, myocardial and epicardial lineages have not yet diverged. All three groups generated cardiac mesoderm in similar proportions and with comparable transcriptomic profiles on day 3 (fig. S7, D and E), indicating that both IWP2 and RA exert their fate-specifying functions primarily at the cardiac mesoderm stage. By day 4, cells from the CM group (clusters 3, 7, and 10) strongly upregulated myocardial-associated genes (e.g., NKX2–5, MYH6, MEF2C, and MYL7), whereas those from the EPI group (clusters 4, 8, and 12) did not (Fig. 4E and fig. S7F), confirming that RA suppresses myocardial gene activation. Cluster 10 expressed TBX5 and BMP2 together with high levels of NKX2–5 and MYH6, suggesting a FHF-like identity (Fig. 4, E and F, and fig. S7F). Cluster 7 showed high ISL、FGF10 and HAND2, but lower NKX2–5 and MYH6, resembling aSHF progenitors. Cluster 3 exhibited further reduced myocardial marker expression but elevated levels of pSHF-related genes NR2F1 and NR2F2 (44), consistent with a pSHF-like identity (Fig. 4, E and F). Thus, we annotated these clusters as FHF-like, aSHF-like, and pSHF-like cells, respectively (Fig. 4E). KEGG enrichment analysis of DEGs further revealed the distinctions among these three cell populations. FHF-like cells exhibited significant upregulation of pathways associated with myocardial development, including cardiac muscle contraction, calcium signaling, and cGMP-PKG signaling. aSHF cells mainly showed upregulation of the Apelin, Hippo, and cell adhesion-related pathways. In contrast, pSHF-like cells were significantly enriched in pathways including ribosome, spliceosome, glycolysis, cell cycle, and RNA degradation, reflecting relatively high proliferative and metabolic activity (fig. S7G and table S3). Due to the lack of established reference datasets defining epicardial progenitors at this developmental stage, clusters 4, 8, and 12 from the EPI-d4 sample were conservatively retained as CPEpi (cells primed for epicardium). This cell population specifically highly expresses genes including DACH2, NR2F2, NR2F1, and HOXA3 (Fig. 4, E and F). Meanwhile, we investigated whether juxta-cardiac field (JCF) progenitor cells—a cell type identified in mice that can give rise to both myocardial and epicardial cells during subsequent development (14)—were generated at this developmental stage. By examining the expression of MAB21L2, a representative marker gene for JCF (13, 14), and BNC2, which labels JCF derived from PSCs in vitro (5, 44), we found that MAB21L2-positive cells were extremely rare and scattered. Accordingly, a typical JCF population was not produced at this developmental stage in our model (fig. S7H).

Diffusion map and diffusion pseudo-time (DPT) analyses revealed a continuous developmental continuum from mesoderm toward two distal endpoints corresponding to FHF-like cells and CPEpi (Fig. 4G). In vivo, FHF progenitor cells are the first to be induced and exit the primitive streak to form the heart tubes, followed by aSHF and then pSHF (45). Consistent with the temporal progression of heart field patterning in vivo, pSHF-like cells, aSHF-like cells, and FHF-like cells were sequentially distributed along the same differentiation branch in pseudotime (Fig. 4G). Expression levels of key lineage markers across major cell type are visualized in Fig. 4H.

We next sought to determine the in vivo correlates of the key developmental stages and major cell types in our model. We hypothesized that the d2–4 stage in our model likely corresponds to the human gastrulation stage, when cardiac development initiates. To test this, we integrated our single-cell dataset (d2–4) with published human gastrula single-cell datasets from PCW2–3 (27) and PCW3 (28), focusing on cardiac lineage-related populations (fig. S8A). Our analysis revealed alignments among the datasets. Nascent mesoderm from PCW2–3 largely overlapped with our annotated nascent mesoderm. Some lateral plate mesoderm 1 (LPM1) cells from PCW3 mapped to both our nascent mesoderm and advanced mesoderm clusters; a subset of LPM1 cells and myocyte progenitor from PCW3, together with a portion of advanced mesoderm and nearly all emergent mesoderm cells from PCW2–3, coincided with our cardiac mesoderm population. Most cells overlapping with our CPEpi originated from lateral plate mesoderm 2 (LPM2) at PCW3. All cardiac myocytes defined in PCW3 mapped to our FHF-like and aSHF-like cells, and a subset of advanced mesoderm cells from PCW2–3 also clustered within our FHF-like and aSHF-like populations. Only a very small number of in vivo-derived cells overlapped with our pSHF-like cells (fig. S8A). Pseudotime analysis further revealed that our FHF-like and aSHF-like cells aligned developmentally with cardiac myocytes from PCW3 and a subset of advanced mesoderm cells from PCW2–3. LPM2 from PCW3 shared a developmental period with our CPEpi. We also observed that the PCW2–3 advanced mesoderm population contained cells at distinct stages of progression (fig. S8B). We next compared gene expression between our in vitro cell types and their in vivo counterparts. The cardiac mesoderm marker PDGFRA and our newly identified marker HAPLN1 were highly expressed in both our cardiac mesoderm and its in vivo counterparts. FHF-associated genes, including SORBS2 and TBX5, were specifically expressed in our FHF-like cells and their matching cells in PCW2–3 and PCW3. In contrast, genes marking more mature cardiomyocytes were predominantly expressed in our FHF-like, aSHF-like cells and their corresponding cells in PCW3 counterparts, but not in PCW2–3, supporting an earlier developmental stage of the latter. ISL1 and FGF10 were expressed in our aSHF-like cells and their in vivo counterparts. DACH2, NR2F2, and NR2F1 showed relatively specific expression in our CPEpi and their matching PCW3 cells (fig. S8C).

We next identified TFs differentially expressed between CPEpi and myocardial progenitors (FHF-, aSHF-, and pSHF-like cells). A volcano plot highlights the top 15 upregulated TFs in each population (Fig. 4I). The top 15 up-regulated TFs in myocardial progenitor cells included the well-known myocardial development-related TFs NKX2–5, MEF2C and TBX20 (Fig. 4I). In contrast, the top 15 TFs upregulated by CPEpi are associated with early proepicardial development TF NR2F2, RA receptor RARB, and RA target genes such as HOXA1 and HOXA3 (Fig. 4I). This population expressed TFs not previously linked to early epicardial specification—such as SOX11 and DACH1/2. These differential expression patterns were recapitulated in the corresponding in vivo cell populations: the in vivo counterparts of CPEpi showed upregulated expression of SOX11, HOXA1, MYCN, and DACH1, while the in vivo populations corresponding to FHF- and aSHF-like cells exhibited elevated expression of DPF3, ESRRG, SOX6, and MEF2C (fig. S9A). To further identify potential regulators underlying the specification of these two lineages, we performed SCENIC analysis. This revealed distinct gene regulatory networks: myocardial progenitors were characterized by networks centered on the key myocardial developmental regulators NKX2–5 and MEF2C, whereas CPEpi were dominated by networks involving BCLAF1 and PBX1 (fig. S9, B to D). Taken together, our findings establish that a brief temporal window at the cardiac mesoderm stage (day 3) gates epicardial lineage commitment. This fate decision is orchestrated by RA, which initiates a hierarchical regulatory network involving the stepwise activation of key transcription factors.

Retinoic acid attenuates BMP signaling during early cardiac lineage specification

To elucidate the mechanisms by which RA directs cardiac mesoderm toward epicardial rather than myocardial fate, we focused on days 3 and 4 of differentiation—the critical window when RA exerts its fate-determining influence and transcriptional trajectories of the CM and EPI groups begin to diverge. The DEGs between EPI and CM groups of day 3 and day 4 were analyzed via Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis (Fig. 5A and table S4). KEGG pathway analysis of DEGs between EPI and CM groups revealed that by day 4, the EPI group exhibited significant downregulation of pathways related to myocardial development, including calcium signaling, cardiac muscle contraction, and adrenergic signaling in cardiomyocytes (Fig. 5A). Gene Ontology (GO) enrichment analysis corroborated these findings, showing suppression of myocardial developmental processes in EPI. Notably, multiple BMP-related GO terms—such as BMP signaling pathway, cellular response to BMP stimulus, BMP receptor binding—were also significantly downregulated in EPI at day 4 (Fig. 5A). A parallel analysis of DEGs at day 3 showed similar, though less pronounced, enrichment of these pathways, suggesting early pathway modulation (fig. S10A and table S4). Given the established importance of BMP signaling in cardiogenesis (46), we examined key BMP pathway components. Bulk RNA-seq data showed that major BMP ligands—BMP2, BMP4, and BMP7—were significantly downregulated in EPI compared to CM at day 4, with BMP2 and BMP7 already reduced in EPI at day 3 (Fig. 5B). qPCR validated these findings, confirming the suppression of BMP ligand expression during lineage specification in EPI (fig. S10B). scRNA-seq analysis further revealed that compared to cardiomyocyte progenitors (FHF, aSHF, and pSHF-like cells) derived from CM-d4, CPEpi derived from EPI-d4 significantly downregulated the expression of BMP2 and BMP7, whereas BMP4 downregulation was less pronounced. Notably, BMP4 remained relatively high expressed in the mesoderm and cardiac mesoderm populations (Fig. 4E and fig. S10C).

Fig. 5. Retinoic acid attenuates BMP signaling during early cardiac lineage specification.

(A) Gene Ontology (GO) and KEGG pathway enrichment analysis of genes downregulated in EPI-d4 versus CM-d4 cells (bulk RNA-seq; P < 0.05). (B) Expression levels of key BMP ligands (BMP2, BMP4, BMP7) in CM and EPI differentiation time courses (bulk RNA-seq). Data are shown as mean ± SD; unpaired two-tailed t-test. Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. (C) Expression of BMP2, BMP4 and BMP7 in day 4 spheroids treated with the indicated concentrations of RA and/or the pan-RA antagonist BMS493. Data are mean ± SD; one-way ANOVA with Dunnett’s test vs. CM control. Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 12 spheroids per sample, 3 technical replicates. (D) Representative immunostaining images of phosphorylated Smad1/5/9 (pSmad1/5/9) in day 4 spheroids of the treatment groups in (C). Scale bars, 100 μm. (E) Quantification of pSmad1/5/9⁺ cells from (D). Data are shown as mean ± SD; one-way ANOVA with Dunnett’s test vs. CM control. Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 6 spheroids for each condition, 3 independent differentiations. ∼1000 cells per confocal field.

To investigate whether RA directly modulates BMP signaling, we performed a dose-response experiment. Following 48-hour induction with CHIR-99021 and BI-1347 combined treatment, spheroids were treated from days 2–4 with IWP2 combined with varying RA concentrations, with or without the RA inhibitor BMS493. qPCR analysis of samples at day 3 and day 4 demonstrated that RA dose-dependently reduced transcript levels of BMP2, BMP4, and BMP7 (Fig. 5C and fig. S10D). Inhibition by 0.5 μM RA was partially reversed by 1 μM BMS493 and fully abolished by 2 μM BMS493 (Fig. 5C). Consistent trends were observed at both day 3 and day 4 (Fig. 5C and fig. S10D). We next assessed BMP signaling activity at the protein level by immunostaining for phosphorylated Smad1/5/9 (pSmad1/5/9), key downstream effectors of BMP signaling. A marked reduction in pSmad1/5/9 was detected in EPI versus CM at day 4, but not at day 3 (fig. S10E). In RA dose–response samples collected at day 4, we observed a progressive decrease in pSmad1/5/9-positive nuclei with increasing RA concentrations, an effect that was reversed by co-treatment with BMS493 (Fig. 5, D and E). Together, these results demonstrate that RA attenuates BMP signaling during cardiac mesoderm patterning, revealing a mechanism by which RA promotes epicardial over myocardial fate.

The epicardial effect of RA depends on finely tuned BMP activity

We observed that exogenous RA signaling during the cardiac mesoderm stage downregulates BMP activity, with moderate RA concentration (0.5 μM)—rather than high levels (2 or 5 μM)—optimally supporting specification. We therefore hypothesized that during cardiac mesoderm development, WNT inhibition alone sustains BMP signaling at levels permissive for cardiomyocyte formation, whereas additional RA exposure further suppresses BMP signaling, with moderately reduced BMP activity favoring epicardial commitment. To test this, we used Noggin—a specific BMP antagonist that binds BMP2, BMP4, and BMP7—to inhibit BMP signaling. Following 48-hour induction with CHIR-99021 and BI-1347, spheroids were treated from days 2–4 with IWP2 plus varying concentrations of Noggin, and cardiomyocyte/epicardial output was assessed at day 8 (Fig. 6A). Quantitative analysis showed that TNNT2⁺ cardiomyocytes decreased dose-dependently with increasing Noggin, becoming nearly undetectable at 100 ng/ml. In contrast, WT1⁺ epicardial populations exhibited a biphasic response: their numbers increased with Noggin titration, peaking at approximately 40% of total cells under 50 ng/ml Noggin, then declining at higher concentrations (Fig. 6, A and B). This dose-response profile closely mirrors the epicardial induction pattern observed with RA treatment. These results showed that direct titration of BMP signaling is sufficient to phenocopy the pro-epicardial effect of RA, supporting a model in which moderate BMP inhibition biases cardiac mesoderm toward epicardial commitment (Fig. 6C).

Fig. 6. Fine-tuned BMP activity is required for RA-mediated epicardial specification.

(A) Representative immunofluorescence images of TNNT2 and WT1 in day 8 spheroids treated with increasing concentrations of the BMP antagonist Noggin. Scale bars, 100 μm. (B) Quantification of TNNT2⁺ cells and WT1⁺ cells from (A). Data are mean ± SD; analyzed by one-way ANOVA with a Dunnett’s multiple comparisons test versus0 ng/ml Noggin control (CM group). Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 6 spheroids for each condition, 3 independent differentiations. (C) The functions of RA and BMP signaling for epicardial fate decision. (D) Representative immunofluorescence images of TNNT2 and WT1 in day 8 spheroids treated with increasing BMP4 concentrations. Scale bars, 100 μm. (E) Quantification of TNNT2⁺ and WT1⁺ cells from (D). Data are mean ± SD; analyzed by one-way ANOVA with Dunnett’s test versus 0 ng/ml BMP4 control (EPI condition). Statistical levels: ns, not significant; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. n = 6 spheroids for each condition, 3 independent differentiations. (F) The effects of RA, WNT and BMP signaling on the myocardial-epicardial fate decision. (G) Heatmap of selected differentially expressed genes between CPEpi and myocardial progenitors (FHF-, aSHF-, and pSHF-like cells), with representative marker genes and enriched KEGG pathways indicated.

We next asked whether restoring BMP signaling under EPI conditions could redirect epicardial progenitors back toward a cardiomyocyte fate. To reconstitute BMP activity suppressed by RA, we supplemented EPI condition with varying concentrations of recombinant BMP4 protein from days 2 to 4. Adding recombinant BMP4 protein to the EPI protocol from days 2–4 partially restored cardiomyocyte generation, with TNNT2⁺ cells increasing initially but plateauing at ∼20% of total cells despite further BMP4 elevation (Fig. 6, D and E). This incomplete rescue suggests that RA suppresses myocardial development through mechanisms beyond BMP inhibition. Transcriptomic analysis supported this view: the EPI group at day 4 significantly downregulated cardiomyocyte proliferation and development related pathways, including PI3K-Akt signaling (47, 48) (Fig. 5A). Meanwhile, rising BMP4 concentrations progressively reduced epicardial cell numbers (Fig. 6, D and E); reinforcing that excessively high BMP activity during the cardiac mesoderm stage is incompatible with epicardial specification. Taken together, our data define a precise BMP signaling level, calibrated by WNT and RA that promotes cardiomyocyte formation while higher, non-physiological BMP activity disrupts epicardial development (Fig. 6F). DEGs analysis further showed that genes involved in cardiac contraction, myocardial development, and signaling pathways such as Hippo, oxytocin, and cGMP–PKG were significantly upregulated in myocardial progenitors (FHF-, aSHF-, and pSHF-like cells) at day 4 (Fig. 6G and table S5), supporting the notion that RA influences cardiac lineage bifurcation through multiple parallel mechanisms.

To further investigate BMP signaling requirements, we used a BMP4/7 double-knockout (BMP4/7 KO) H9 cell line generated in our laboratory (49). Pluripotency markers NANOG and OCT4 were normally expressed in BMP4/7 KO cells (fig. S11A). When subjected to SD, CM, or EPI differentiation protocols, BMP4/7 KO spheroids showed markedly impaired cardiogenesis: beating rates were reduced in the SD group and absent in the CM group at day 10 (fig. S11B). Immunofluorescence analysis at day 8 revealed sparse cardiomyocytes and epicardial cells in some SD spheroids, near-absent TNNT2⁺ cells in the CM group, and limited WT1⁺ cells in the EPI group (fig. S11, C and D). Analysis of day 3 samples indicated that HAND1⁺ cardiac mesoderm formation was significantly compromised in BMP4/7 KO lines (fig. S11, E and F), consistent with reported roles of BMP signaling in mesoderm induction and cardiac specification (50, 51). These results confirm that complete BMP loss-of-function disrupts early cardiac lineage development.

Considering that BMP signaling is essential for early mesoderm development, we used BMP4/7 knockout (KO) cell lines to determine whether the inhibition of mycardial and epicardial differentiation caused by BMP4/7 KO is mediated through impaired mesoderm development. Specifically, we added exogenous BMP4 from day 0 to day 2 and withheld supplementation from day 3 to day 4, thereby effectively mimicking the inhibition of BMP signaling during days 3–4. Subsequently, we conducted differentiation experiments for CM and EPI condition based on this scheme (fig. S11G). First, we conducted a BMP4 concentration gradient experiment during its exogenous supplementation from day 0 to day 2. We observed that, at this stage, the number of HAND1-positive cells generated on day 3 in BMP4/7 KO was positively correlated with the dose of BMP4 added. At 2 ng/ml BMP4, the number of HAND1-positive cells most closely matched that of the wild-type group. Higher BMP4 concentrations resulted in more HAND1-positive cells than the wild-type control (fig. S11, H and I). We selected the 2 ng/ml BMP4 group for subsequent cardiomyocyte and epicardial cell differentiation experiments. We observed that in the CM group, all spheroids simultaneously generated both cardiomyocytes and epicardial cells (fig. S11J), a phenotype similar to that induced by low-concentration RA (0.2 μM) (Fig. 1, E and F) or moderate-concentration Noggin (20 ng/ml) (Fig. 6, A and B) treatment. In contrast, in the EPI group, approximately 35% of WT1-positive epicardial cells were detected, while no cardiomyocytes were observed (fig. S11, J and K). This phenotype reminded us of the effects induced by excessive RA or Noggin. These results further underscore the crucial role of precise regulation of BMP signaling in the development of cardiomyocytes and epicardial cells.

DISCUSSION

Although single-cell transcriptomic analyses of human embryonic heart tissues have provided valuable snapshots of early cardiac development, isolated temporal samples are insufficient for reconstructing dynamic developmental processes. To fill the gap, we established a human pluripotent stem cell-based model that faithfully models the stepwise progression from nascent and advanced to cardiac mesoderm, and subsequently captures the initiation of myocardial and epicardial lineage separation starting from the cardiac mesoderm stage. Through time-course single-cell sequencing of our model, we not only delineated the developmental trajectories of both lineages but also uncovered the gene regulatory programs governing their fate determination. Importantly, our findings show that RA promotes epicardial commitment via BMP-dependent priming and stepwise maturation, whereas in the absence of RA, cardiac mesoderm undergoes default myocardial differentiation. Both fate trajectories involve the hierarchical activation of lineage-specific transcription factors. These results deepen our understanding of human heart development, thereby enabling more accurate disease modeling and paving therapeutic avenues for cardiac regeneration.

The epicardium, first identified in chick embryos, originates from a conserved vertebrate structure—the proepicardial organ (PEO) which located on the dorsal heart tube (22–25, 52, 53). Proepicardial cells migrate to cover the heart tubes, forming the epicardium, with both stages expressing markers like TBX18 and WT1 (54). Furthermore, JCF progenitors emerge as early as the headfold stage in mice (13), and lineage tracing confirms that unipotent Mesp1⁺ epicardial progenitors segregate during early cardiac mesoderm specification (26). Collectively, these findings support a model in which the epicardium arises from multiple spatiotemporally distinct progenitor sources, with myocardial and epicardial lineages separating early in development. Our model recapitulates key in vivo cardiac cell types and developmental dynamics, thus serving as a discovery system to unravel the mechanism of the myocardial-epicardial fate switch. Using this system, we demonstrated that myocardial and epicardial lineages begin to diverge during the early cardiac mesoderm stage under specific signal guidance. Interestingly, we identified a CPEpi population that diverges early from the myocardial lineage, preceding known in vivo progenitors. This population likely represents a distinct “pre-epicardial” progenitor stage, revealing an earlier developmental origin for the epicardial lineage. Furthermore, their subsequent developmental trajectory confirms epicardial lineage commitment, making CPEpi valuable resource for identifying transcription factors involved in early epicardial determination.

The complexity of signal networks in regulating early embryonic development has obscured how individual signals govern specific processes. To address this, we developed an early cardiac mesoderm spheroid model employing a reductionist approach that eliminated confounding signals and complex components like Matrigel, while enabling precise single-factor manipulation of human pluripotent stem cells. Concentration gradient and time-course experiments revealed that RA signal alone during the cardiac mesoderm stage determines lineage fate. Although RA is known to influence FHF/SHF progenitor cell fate and cardiogenic mesoderm boundaries in vivo and in vitro (55–58), its specific role and action temporal window in controlling myocardial versus epicardial fate remained undefined. Surprisingly, our findings demonstrated that RA dosage during early cardiac mesoderm development directly controls lineage patterning: low levels favor cardiomyogenesis, while moderate levels drive epicardial development. These findings align with zebrafish studies showing that RA deficiency causes cardiomyocyte overproduction and cardiomegaly through expansion of myocardial progenitor pool (59), and corroborate reports that RA promotes epicardial differentiation in stem cell models (5, 30). As a critical developmental regulator, RA signaling coordinates somitogenesis and the development of diverse tissues and organs derived from all germ layers, including the hindbrain, eyes, heart, and pancreas (60–62),with its specification role of myocardial versus epicardial lineage fate further underscoring its pleiotropic roles.

The isolation of early-diverged myocardial and epicardial lineages permitted identification of lineage-specific transcription factors, paving the way for the mechanistic investigation of RA-mediated fate patterning. While previous reports demonstrated cross-talk among BMP, FGF, RA, and WNT signaling during cardiac lineage development (19, 55, 63, 64), our single-factor experiments establish that RA signaling directly modulates BMP signaling levels at the cardiac mesoderm stage. Similarly, Sheng et al. (65) found RA also exerts an inhibitory effect on the BMP signaling during neural tube patterning. In our model, under conditions of WNT inhibition, increasing RA suppresses BMP signaling and reduces cardiomyocyte output, consistent with BMP’s recognized role in initiating FHF lineage (66). Although coexisting RA and BMP signaling was previously considered beneficial for epicardial formation (38, 67), we demonstrate that RA regulates BMP signaling to direct fate choices, and that independent manipulation of either pathway can drive specific differentiation—resolving previous ambiguities arising from signal interference. Beyond BMP signals, RA also affects other myocardial development pathways, illustrating the efficient multiplexing of developmental signals during embryogenesis (68).

This study has several limitations. While we provided functional and mechanistic evidence supporting the critical role and regulatory mechanism of RA via temporal chemical perturbation, lineage marker analysis, and scRNA-seq analysis, more direct functional validations—such as gain- and loss-of-function experiments using gene editing—are still lacking. Furthermore, after confirming the essential function of RA in fate determination of myocardial and epicardial lineages, we did not perform in-depth functional exploration, including precisely tuning the timing of RA treatment based on the key fate-decision window and optimal dosage of RA identified in this study, combined with previously reported signals governing cardiac morphogenesis (1), to further optimize the structure and function of in vitro cardiac organoids and establish more application valuable in vitro models. These aspects represent important directions for our future investigations.

In summary, the myocardial-epicardial developmental model presented here provides a valuable framework for decoding early lineage separation and offers in-depth insights into human heart development. Understanding how RA specifically regulates myocardial-epicardial fate decisions will inform strategies for directing cardiovascular progenitors toward specific cardiac cells, with significant implications for developing in vitro models and producing therapeutic cells for treating heart diseases.

MATERIALS AND METHODS

Culture of hESCs

The human embryonic stem cell (hESC) lines used in this study included H9 (39), h1, h2, and BMP4/7 double knockout (KO) H9. h1 and h2 were previously isolated by our research team from human blastocysts (39), and the BMP4/BMP7-KO H9 was generated in a prior study (49). For routine maintenance, hESCs were seeded on Vitronectin (Gibco, A14700) coated plates and cultured in TeSR-E8 medium (Stemcell Technologies, 05990). The culture medium was changed daily and cells were passaged every 3 to 4 days. For sub-culturing, cells were dissociated into aggregates using ReLeSR (Stemcell Technologies, 100–0483) according to the manufacturer’s protocol, and seeded at a ratio of 1:15 to 1:20. 10 μM Y27632 (Sellcek, S1049) supplementation was required only on the day of passaging. For cardiac differentiation models, hESCs were treated with 33% TrypLE (Gibco, 12605028) at 37°C for 3–4 minutes to generate single-cell suspensions prior to subsequent embryoid body formation and differentiation experiments.

Induction of three-dimensional cardiac lineage spheroids

Embryoid body (EB) formation

On day 4 of routine culture (or when 80% confluency was reached), hESCs were dissociated into single cells using 33% TrypLE. The cells were resuspended in Ca²⁺/Mg²⁺-free Dulbecco’s phosphate buffered saline (DPBS) (Sigma, D5652), counted, and resuspended in appropriate volumes of TeSR-E8 medium supplemented with 10 μM Y27632. The cell suspension was seeded into ultra-low attachment 96-well round-bottom plates (Corning, 7007) at a density of 4000 cells per 100 μl per well. The plates were centrifuged at 300g for 3 minutes to facilitate cell aggregation, then incubated in a standard CO₂ incubator under normoxic conditions. This timepoint was recorded as day −1 (d-1). EBs usually formed after 24 hours of culture, namely day 0 (d0), and were ready for subsequent differentiation experiments.

The first stage of differentiation

At day 0, after hESCs formed EBs, the TeSR-E8 medium was replaced with basal medium composed of advanced RPMI 1640 (Gibco, 12633012) supplemented with 1× B27 (Gibco, 17504044). This basal medium was used for all subsequent differentiation and maintenance cultures. For the first stage of differentiation, 5 μM CHIR-99021 (Selleck, S2924) and 0.5 μM BI-1347 (Selleck, S1058) were added to the basal medium, and maintained for 48 hours.

The second stage of differentiation

Spontaneous Differentiation (SD) Group. At day 2 (end of the first stage), the medium was replaced with basal medium without any factors. Cultures were maintained with medium changed every 2–3 days.

Cardiomyocyte (CM) Group. At day 2, the medium was replaced with basal medium supplemented with 5 μM IWP2 (Selleck, S7085). After 48 hours of incubation, the medium was replaced with basal medium without any factors, with medium refreshed every 2–3 days.

Epicardial (EPI) Group. At day 2, the medium was replaced with basal medium containing 5 μM IWP2 and 0.5 μM retinoic acid (Sigma, R2625). After 48 hours of incubation, the medium was replaced with basal medium without any factors, with medium refreshed every 2–3 days.

Cell culture treatments

For RA concentration gradient assay, after the first stage of differentiation, replace the medium with basal medium containing 5 μM IWP2 and varying concentrations (0, 0.02, 0.05, 0.1, 0.2, 0.5, 1, 2, or 5 μM) of RA. After 48 hours of culture, the medium was replaced with basal medium without any factors for continued maintenance.

For RA inhibition experiment, after the first stage of differentiation, replace the medium with basal medium containing 5 μM IWP2, 0.5 μM RA, and 1 μM or 2 μM BMS493 (Selleck, E1627). After 48 hours of culture, the medium was replaced with basal medium without any factors for continued maintenance.

For RA time-shift assay, after the first stage of differentiation, six treatment groups were designed as follows:

Control group (the CM group): 5 μM IWP2 for 48 hours, followed by basal medium.

Group 1: 5 μM IWP2 and 0.5 μM RA for 48 hours, followed by basal medium.

Group 2: 5 μM IWP2 for 24 hours, 5 μM IWP2 and 0.5 μM RA for 24 hours, 0.5 μM RA for 24 hours, followed by basal medium.

Group 3: 5 μM IWP2 for 48 hours, 0.5 μM RA for 48 hours, followed by basal medium.

Group 4: 5 μM IWP2 for 48 hours, basal medium for 24 hours, 0.5 μM RA for 48 hours, followed by basal medium.

Group 5: 5 μM IWP2 for 48 hours, basal medium for 48 hours, 0.5 μM RA for 48 hours, followed by basal medium.

For IWP2 time-shift assay, after the first stage of differentiation, three treatment groups were designed as follows:

Group 1: 5 μM IWP2 for 48 hours, followed by basal medium.

Group 2: Basal medium for 24 hours, 5 μM IWP2 for 48 hours, followed by basal medium.

Group 3: Basal medium for 48 hours, 5 μM IWP2 for 48 hours, followed by basal medium.

For BMP inhibition experiment, after the first stage of differentiation, replace the medium with basal medium containing 5 μM IWP2 and varying concentrations (0, 5, 10, 20, 50, or 100 ng/ml) of Noggin (Novoprotein, CB89). After 48 hours of culture, the medium was replaced with basal medium without any factors for continued maintenance.

For BMP4 concentration gradient assay, after the first stage of differentiation, replace the medium with basal medium containing 5 μM IWP2, 0.5 μM RA and varying concentrations (0, 0.5, 1, 2, 5, or 10 ng/ml) of BMP4 (PeproTech, 120–05). After 48 hours of culture, the medium was replaced with basal medium without any factors for continued maintenance.

Cryosections preparation

The spheroids at designated time points were transfer from ultra low attachment 96-well round-bottom plates into 200 μl microcentrifuge tubes. Aspirated residual medium, and washed samples once with DPBS. Then fixed the spheroids with 4% paraformaldehyde (PFA) (Biosharp, BL539A) at 4°C for 4–5 hours. After fixation, the PFA was removed, and samples were dehydrated using 30% sucrose in PBS until spheroids sank to the bottom of the tube. Spheroids were subsequently transferred into embedding cassettes, embedded in O.C.T. (Sakura Finetek, 4583), and frozen. The frozen samples were stored at −80°C until sectioning on a Leica cryostat. Sections (8–10 μm thickness) were collected onto adhesive glass slides and stored at −20°C prior to immunofluorescence staining.

Immunofluorescence staining

Cryosections were washed three times with phosphate-buffered saline (PBS) to remove O.C.T.. Permeabilization and blocking were performed overnight at 4°C with PBS containing 0.2% Triton X-100 (Sigma, 9036-19-5) and 3% bovine serum albumin (BSA) (Solarbio, A8020). Samples were then incubated overnight at 4°C with primary antibodies diluted in PBS containing 1% BSA. After washing three times for 5 min with PBS containing 0.05% Tween-20 (Sangon, A443678) (PBST), samples were incubated for 2 hours at room temperature with secondary antibodies and DAPI (Sigma, 32670) diluted in PBS containing 1% BSA, protected from light. Following three additional washes with PBST, samples were mounted and stored at 4°C protected from light. Images were acquired using a Nikon AX confocal laser scanning microscope. For quantitative analysis, a minimum of three confocal planes were captured per sample, and were consistently replicated in ≥3 technical repeats. Fiji software was used to count marker-positive cells. Antibodies used are listed in table S6.

Quantitative real-time PCR(qPCR)

Total RNA was isolated from cell spheroids using the Flash Cell/Tissue Total RNA Kit (Yeasen, 19221ES50) following the manufacturer’s instructions. RNA concentration was quantified using NanoDrop (Thermo Fisher Scientific). Complementary DNA (cDNA) was prepared using the cDNA Synthesis Kit (Yeasen, 11141ES60) according to the manufacturer’s instructions. qPCR was performed using the Universal Blue qPCR SYBR Green Master Mix (Yeasen, 11184ES08) on a Real-Time PCR instrument (BIO-RAD). Relative RNA expression levels of target genes were normalized to GAPDH. Primer sequences are listed in table S7.

Cell preparation for bulk RNA sequencing

All the samples used for bulk RNA sequencing and single-cell sequencing were derived from H9. The following samples were collected for Bulk RNA sequencing: Day 0 (d0) spheroids; d2 spheroids after the first stage of differentiation; d3, d4, d5, d6, d8, and d10 spheroids from the CM group; d3, d4, d5, d6, d8, and d10 spheroids from the EPI group. Three replicates were performed for each sample.

Spheroids from suspension cultures at designated time point were collected, washed twice with DPBS, and lysed with TRIzol Reagent (Thermo Fisher Scientific, 15596018) to obtain total RNA. Library preparation and sequencing were performed by Annoroad Gene Technology (http://www.annoroad.com/). Sequencing was conducted on the Illumina NovaSeq 6000 platform.

Bulk RNA-seq data analysis

Following quality control, sequencing reads were aligned to the GRCh38 human reference genome (Ensembl Release 108) using HISAT2 (version 2.2.1) with default parameters. Alignment files were processed with SAMtools (version 1.19.2) for sorting and index generation. Transcript assembly and quantification were carried out using StringTie (version 2.2.1), and gene expression levels were calculated as fragments per kilobase per million mapped reads (FPKM). Differentially expressed genes (DEGs) were identified using DESeq2 (version 1.40.2), with the absolute value of log₂-transformed fold change ≥1, adjusted P value < 0.05, and FPKMs ≥ 1 in at least one sample. Functional enrichment of DEGs was performed using the ClusterProfiler package (version 4.8.3) in R (version 4.3.1). Gene Ontology (GO) terms and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways were assessed to identify overrepresented biological processes and signaling pathways. Enrichment significance was defined as P < 0.05, and only terms meeting this threshold were reported.

Cell preparation for single-cell sequencing

Spheroids were collected at the following time points for single-cell sequencing: d0; d2 after the first stage of differentiation; d3 of the SD group; d3, d4, d6 and d10 of the CM group; d3, d4, d6 and d10 of the EPI group.

For single-cell dissociation, spheroids were treated with 100% TrypLE at 37°C for 6–15 minutes (digestion time appropriately increased with spheroid cultivation duration). After digestion, cells were resuspended in DPBS, followed by gentle pipetting to form single-cell suspension. Cellular aggregation was monitored under a microscope, and cell suspensions were filtered through 40 μm cell strainers (Falcon, 352340) if necessary.

scRNA-seq

Library preparation and sequencing were performed by Annoroad Gene Technology (http://www.annoroad.com/). Briefly, single-cell suspensions were counted with trypan blue, and samples with cell viability exceeding 85% were proceeded for subsequent library construction and sequencing. Subsequently, single-cell suspensions of each sample were individually loaded into channels to generate gel beads-in-emulsion on the MobiNova-100 (MobiDrop, China) platform. The 3′ RNA-seq libraries were prepared using the MobiCube Single-Cell 3′ RNA-seq Kits (MobiDrop, China) in accordance with the manufacturer’s instructions. The libraries were then sequenced on the MGI DNBSEQ-T7 platform using 150-base pair (bp) paired-end sequencing, with a median depth of 30,000 raw reads per cell.

Single-cell RNA-seq data analysis

Processing and annotation of single-cell transcriptomes

The raw single-cell transcriptomic data were processed using MobiVision (version 1.1) provided by MobiDrop. Raw FASTQ files were demultiplexed by cell barcode, error-corrected, aligned to the GRCh38 human reference genome (GRCh38–2020-A), and quantified using UMI-based correction to generate a digital gene expression matrix. The resulting expression data were further processed using CellRanger (version 7.0.1), with the default target of 10,000 cells per sample. The Seurat package (version 4.4.0) in R was applied for downstream analyses, incorporating essential quality control steps. Low-quality cells were excluded based on the following criteria: gene counts outside the range of 2500–8000 or mitochondrial RNA content exceeding 15%. After quality control, all samples were merged into a single Seurat object. Highly variable genes were identified using the “vst” method, and the data were subsequently scaled and subjected to principal component analysis (PCA), retaining 30 principal components based on the elbow plot. These principal components were then used for graph-based clustering using the Louvain algorithm. Dimensionality reduction for visualization was achieved using Uniform Manifold Approximation and Projection (UMAP) with the parameters: dims = 10, seed.use = 70, and min.dist = 0.4. Cell types were annotated based on canonical marker genes.

Identification of differentially expressed genes

Differentially expressed genes (DEGs) between cell types were identified using the FindMarkers function in the Seurat package, with parameters set to min.pct = 0.25 and logfc.threshold = 0.25. To ensure statistical significance, genes with an adjusted P value (Padj) < 0.05 were considered as differentially expressed. KEGG pathway enrichment was assessed via the KOBAS web server (http://bioinfo.org/kobas). Pathways achieving a Benjamini–Hochberg adjusted p-value of less than 0.05 were regarded as statistically significant.

Integration with published single-cell data

To compare cardiac lineage similarities across datasets, myocardial and epicardial cells derived from our in vitro differentiation model were integrated with publicly available human heart single-cell data from Farah et al. (Accession number phs002031) (42). Data integration and dimensionality reduction were performed using canonical correlation analysis (CCA), and the top 3000 highly variable genes were selected with the “vst” method. PCA was then conducted using the top 25 principal components, followed by UMAP visualization with parameters n.neighbors = 30, min.dist = 0.35, and seed.use = 80 to ensure reproducibility. To minimize the potential bias introduced by imbalanced cell type abundances, we downsampled each cell type to 500 cells prior to calculating Pearson correlation coefficients. Besides, we integrated specific cell populations from our single-cell dataset (days 2–4) and those of Tyser et al. (27) and Zeng et al. (28). This integration assessed the transcriptional fidelity of in vitro-derived cardiac lineage-related populations against their in vivo counterparts. Integration was performed using Seurat (version 4.4.0) with CCA. The top 2000 highly variable features were identified per dataset. For UMAP visualization, the first 15 principal components were used with parameters set as follows: n.neighbors = 30, min.dist = 0.4.

Single-cell trajectory inference

Developmental trajectories were reconstructed using Monocle2 (version 2.28.0), Monocle3 (version 1.3.7), and destiny (version 3.14.0, for diffusion pseudotime analysis). Gene expression matrices were first normalized and transformed into a CellDataSet object. Highly variable genes were selected based on a mean expression of 0.15–0.6 and an empirical dispersion ≥4.5 × dispersion fit. Dimensionality reduction was performed using the DDRTree method, and cells were then ordered along the trajectory according to the minimum spanning tree (MST) constructed in the reduced space. The root state was defined according to the expression pattern of known lineage markers. To identify branch-dependent dynamic genes, the Branch Expression Analysis Modeling (BEAM) method was applied, and the top 500 genes ranked by BEAM q-values were selected, and visualization was achieved through the plot_genes_branched_heatmapfunctions. To investigate the developmental temporal order of cell types within each cell population in the integrated dataset (Tyser, Zeng, and our dataset), we performed trajectory analysis using Monocle 3. Given that Monocle3 employs a semi-supervised pseudotime algorithm, the Nascent Mesoderm was defined as the trajectory root based on prior biological knowledge. Cell trajectories were inferred using the orderCells function and visualized with the plot_cells function. For Destiny-based diffusion pseudotime analysis, log-transformed normalized expression dataNormalized expression data were used to compute the diffusion map via the DiffusionMap() function with n_eigs = 25. Diffusion pseudotime (DPT) values were computed using the DPT() function with the undifferentiated cluster as the root, to reflect cell progression along the inferred trajectory.

Gene regulatory network analysis

To reconstruct gene regulatory networks and evaluate regulon activities, we performed pySCENIC (version 0.12.1) analysis on the single-cell expression matrix. We applied GRNBoost2, implemented in pySCENIC, to infer the gene regulatory network. The cisTarget step was performed to perform transcription factor (TF) motif enrichment, prune the co-expression modules, and identify direct transcriptional targets. A motif rankings database based on the human genome (hg38, 500 bp centered at TSS) and the corresponding motif-annotation file were employed for this step. Subsequently, AUCell was applied to calculate regulon activity scores for each cell, generating a regulon activity matrix (cells × regulons). The resulting activity profiles served as distinguishing features to facilitate the identification of typical regulatory architectures of different cell types.

Statistics

Statistical analysis was performed using GraphPad Prism Software (version 9.0). Data was presented as means ± SD. Comparisons among multiple groups were performed with a one-way analysis of variance (ANOVA) and comparisons between two groups were evaluated by a two-tailed unpaired Student’s t test. P < 0.05 were considered statistically significant.

Supplementary Materials

The PDF file includes:

Figs. S1 to S11

Legends for tables S1 to S7

Legends for movies S1 to S4

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S7

Movies S1 to S4