We previously observed that pharmacological inhibition of MEK (the kinase upstream of ERK) or FGF receptors leads to a loss of differentiation to primitive streak-like and primordial germ cell-like cells, suggesting a crucial role for FGF/ERK signaling³⁴. Moreover, a ring of active ERK (pERK) has been observed in 2D gastruloids but its connection to FGF signaling or differentiation remains unclear³³. Therefore, we wanted to investigate the role of FGF/ERK signaling further.

To determine the dynamics of pERK and relate it to PS-like differentiation, we first performed time series immunofluorescence co-staining for pERK and the PS marker TBXT after BMP4 treatment in 700um diameter colonies. This revealed the pERK ring first emerges around 24h, around the same time as TBXT expression, and gradually gets brighter and wider (Fig. 1a). By 42h the outer edge of the ERK ring coincided with TBXT expression, while the inner edge was less sharp and extended further towards the center than TBXT, reminiscent of the domains of WNT and NODAL signaling that extend further in than the PS-like region^3,6 (Fig. 1a). Cell fate patterning in 2D gastruloids has a fixed length scale from the edge, with the central part absent in smaller colonies due to the fixed length scale of the BMP signaling gradient from the edge. Consistently, we found that ERK signaling also has a fixed pattern from the edge regardless of colony size, so that in smaller colonies the ERK ring becomes a domain of high ERK signaling extending through the colony center (Supp. Fig. 1a). MEK or FGFR inhibition for 30 minutes after 41.5h eliminated ERK activity, suggesting ERK activity depends entirely on FGF and that continuous FGF signaling is necessary to maintain the ERK signaling pattern (Fig. 1bc). Inhibition for the full duration of differentiation eliminated TBXT expression as well as other PS markers (Fig. 1bc, Supp. Fig. 1b). However, we were able to rescue TBXT expression by directly and uniformly activating ERK (Fig. 1d-f, Supp. Fig. 1cd) in cells stably expressing the doxycycline-inducible membrane-targeted catalytic domain of SOS (dox-SOS^cat)³⁷. Despite uniform induced ERK activity, TBXT expression remained excluded from the center, showing that ERK activation alone is not sufficient for TBXT induction. However, TBXT levels were much higher in the rescue (Supp. Fig. 1cd), possibly reflecting the much higher level of ERK activity induced by dox-SOS^cat. MEK and FGFR inhibition also led to a reduction in cell number which staining for BrdU and cleaved Cas3 suggested was primarily due to reduced proliferation (Supp. Fig. 1ef). In contrast to TBXT expression, reduced cell number caused by FGFR inhibition was only partially rescued by dox-SOS^cat, indicating additional downstream pathways regulate cell number (Fig. 1e). Altogether, these data show the formation of a PS-like ring is dependent on FGF signaling through ERK.

This raised several questions. First, is FGF/ERK signaling required directly for PS-like differentiation, or does it act indirectly? These possibilities are not mutually exclusive. For example, FGF/ERK could be required directly but also act indirectly by controlling Wnt or Nodal expression, as both Wnt and Nodal signaling are required for PS-like differentiation. Second, which are the FGFs activating the ERK pathway? The pathway could be activated by exogenous FGF2 present in the pluripotency medium mTeSR1 or by endogenously expressed ligands. Third, how is ERK activity restricted to a ring?

To determine if there is a direct requirement of FGF/ERK signaling and whether this depends on exogenous FGF, we performed differentiation to PS-like cells in standard culture by exogenous activation of Wnt and Nodal signaling, bypassing the need to form endogenous Wnt and Nodal gradient as in 2D gastruloids³⁸. We used a minimal base medium (E6) without exogenous FGF or other growth factors. To rule out any additional requirement for endogenous Nodal and Wnt which may be FGF/ERK-dependent, we used Nodal knockout cells³ treated with the Wnt secretion inhibitor IWP2, thereby eliminating any endogenous activity of these pathways. We also inhibited BMP receptors using LDN193189 to rule out any indirect effect through endogenous BMP. Under these conditions, with exogenous Wnt and Nodal activation for 24h in the absence of exogenous FGF, cells efficiently differentiated to TBXT+ PS-like cells (Fig. 1g, Supp. Fig. 1g). These PS-like cells had high active ERK levels just as in the 2D gastruloids. Inhibition of either FGFR or MEK abrogated ERK activity and TBXT expression while FGFR inhibition also severely reduced proliferation. In both cases significant expression of the pluripotency and ectoderm marker SOX2 was maintained (Supp. Fig. 1g). Given that FGF/ERK is known to be involved in pluripotency maintenance, this suggests inhibition may not be complete and that much lower ERK activity is required for pluripotency maintenance than PS-like differentiation, consistent with lower ERK activity in the pluripotent center of gastruloids.

MEK and FGFR inhibition reduce proliferation and thereby cell density, raising the possibility that their lack of differentiation is an indirect effect that depends on cell density: a critical density may be required for differentiation due to a community effect. To rule this out we performed the experiment across a range of densities, which confirmed that PS differentiation in the control was efficient even at the lowest densities while FGFR and MEK inhibition effectively blocked differentiation when compared to controls with similar final cell densities (Supp. Fig. 1h). We also tested whether proliferation and differentiation require similar levels of ERK signaling by determining the MEKi dose response of gastruloids. This revealed that lower doses of MEKi effectively block differentiation without strongly affecting cell number, further supporting that the effects of ERK signaling on proliferation and differentiation are independent (Supp. Fig. 1i). Altogether, our results show that endogenous FGF expression is both necessary and sufficient for PS-like differentiation by exogenous Wnt and Nodal stimulation and demonstrate a direct requirement for ERK signaling in PS-like differentiation.

The above results suggest endogenous FGF expression may contribute to elevated ERK signaling in the PS-like ring. However, FGF/ERK signaling may be modulated at many levels (Fig. 2a). A spatial pattern of ERK activation could reflect a concentration gradient of FGF. Alternatively, differential expression of receptors and HSPG-modifying enzymes could lead to spatially patterned receptor activation and downstream ERK activity even in the absence of an FGF gradient. Even if receptor activation is uniform, downstream ERK activity could be spatially patterned by downstream negative regulators of ERK activity such as Sprouty-family proteins (SPRY) and dual-specificity phosphatases (DUSPs)⁴⁰.

To understand which genes modulating FGF/ERK signaling could be responsible for the elevated ERK signaling in TBXT positive cells and required for TBXT expression (Fig.1, Fig. 2b), we used single cell RNA-sequencing data from one previously published sample³⁴ and one newly collected replicate (Supp. Fig. 2a) to determine expression of FGF pathway components in human 2D gastruloids, extending earlier work describing FGF expression^25,35,36,41. TBXT-positive PS-like cells give rise to nascent mesoderm expressing TBX6 as well as definitive endoderm expressing SOX17 and FOXA2, both of which may still express TBXT at 42h⁴². To understand how TBXT expression is initiated in an FGF-dependent manner, we focused on differential expression in the PS-LCs relative to pluripotent cells, reasoning that FGF modulation specific to nascent mesoderm or endoderm is a consequence rather than a cause of PS-LC differentiation. We identified different cell types based on marker genes with the PS-like cells defined as TBXT+MIXL1+POU5F1+TBX6-SOX17-FOXA2-(Fig. 2cd, Supp. Table 4). We then analyzed expression of different categories of FGF/ERK regulators.

We found four FGF ligands with significant expression: FGF2, 4, 8, and 17 (Fig. 2e, Supp. Fig. 2b). We then compared the expression of these ligands in PS-like cells to reference data of CS7 human, CS8 monkey, and E6.5-E7.5 mouse gastrulation (Fig. 2f, Supp. Fig. 2cd). We also included FGF3 and FGF5, which are expressed during mouse gastrulation^32,43,44. High expression of FGF4 and FGF17 and low expression of FGF8 was consistent between human, monkey, and 2D gastruloid. FGF2 expression was also high in the PS-like populations in 2D gastruloids and the CS8 monkey embryo, but much lower in the human embryo data. Furthermore, the monkey and human embryos expressed FGF3, which was minimally expressed in gastruloids. A similar comparison in nascent mesoderm showed much lower FGF2/4 and much higher FGF17 compared to PS across all primate samples (Supp. Fig. 2e). Expression in 2D gastruloids was consistent with published RNA-seq data for hPSC-derived PS-like cells, which expressed high FGF2 and FGF4, some FGF8 and FGF17, and no FGF3 (Supp. Fig. 2f). Comparison with mouse confirmed significant differences from primates, with high expression of FGF5, 8, much lower expression of FGF4,17, and no expression of FGF2 in the mouse^21,43,44.

Of the FGF receptors, only FGFR1 and FGFR4 were strongly expressed, with FGFR1 expression about 5-fold higher (Fig. 2g, Supp. Fig. 2g). Both receptors showed graded expression that decreased from pluripotent to PS-like cells and much lower expression in amnion-like cells (Fig. 2g). Therefore, elevated ERK activity in PS-like cells cannot be accounted for by overall receptor expression. Differential expression of genes regulating heparin-sulfate proteoglycans which modulate FGF receptor binding was mild except for glypicans 3 and 4 which were downregulated in PS-like cells relative to pluripotent cells (Fig. 2h, Supp. Fig. 2g). In contrast, several intracellular negative regulators of ERK signaling were strongly differentially expressed between cell clusters, including SPRY1 and PEBP1 which were significantly higher in pluripotent cells (Fig. 2i, Supp. Fig. 2g). Surprisingly, however, expression of SPRY4 was not elevated in PS-like cells where ERK signaling is high, suggesting that transcription of SPRY4 may not be a good readout of ERK activity in human hPSCs like it is in the mouse⁴⁵. On the other hand, expression of the ERK negative feedback regulator IL17RD (SEF) does seem to approximately reflect expected ERK activity in each population. Components of the core ERK pathway including ERK itself were highly expressed across all clusters (Supp. Fig. 2g).

Thus, while scRNA-seq identified specific genes involved, it is consistent with several non-exclusive mechanisms for generating an ERK pattern. For example, uniform FGF2 could form an ERK ring by equally activating receptors across the colony but with ERK response reduced by elevated expression of negative regulators in the pluripotent center. Alternatively, based on gene expression we cannot rule out HSPGs increasing receptor binding specifically in the ring. Finally, localized expression of FGF4 or FGF17 in the PS-like ring could lead to a concentration gradient responsible for increased receptor activation.

To distinguish between the different possibilities, we asked whether the spatial pattern of pERK signaling is controlled primarily below, at, or above the level of the receptors. To address this, we first visualized phosphorylated FGFR1 (pFGFR1) and found its spatial pattern strongly resembles that of pERK (Fig. 3a, Supp. Fig. 3a). As expected, pFGFR1 was lost upon brief FGFR inhibition (Fig. 3a, Supp. Fig. 3a). In contrast, after MEK inhibition, high pFGFR1 remained while pERK was greatly reduced (Supp. Fig. 3bc). This suggests the spatial pattern of ERK reflects a spatial pattern of receptor activity.

The spatial pattern in receptor activity could either be due to differential receptor expression or differential receptor activation. There was no significant upregulation of FGF receptors in PS-like cells according to our single cell RNA-sequencing data (Fig 2e). In fact, the most highly expressed receptor, FGFR1, was elevated in pluripotent cells. This suggests the ERK activity pattern is not due to a receptor expression pattern with two caveats. First, mRNA expression may not reflect protein expression. Second, multiple isoforms exist for FGF receptors 1-3 (but not 4) that have different affinity for FGF ligands, leaving open the possibility that PS-like cells could express a different FGFR1 isoform than pluripotent cells, despite overall FGFR1 expression being approximately uniform.

To determine if the spatial pattern in pFGFR1 could be explained by a spatial pattern of FGFR1 protein, we performed immunofluorescence staining of FGFR1. Consistent with the scRNA-seq data, and in sharp contrast to the pFGFR1 stain, FGFR1 protein expression was graded from center to edge with its highest levels in the pluripotent center and therefore could not explain the signaling pattern (Fig. 3b, Supp. Fig. 3d). To determine if different isoforms of FGFR1 are differentially expressed in PS-like vs. pluripotent cells, we designed qPCR primers to detect the IIIb and IIIc isoforms of FGFR1 (Supp. Fig. 3e-g). These are the most common isoforms, also known as the epithelial and mesenchymal isoforms, and therefore appeared the mostly likely candidates for differential expression. We then analyzed expression of these isoforms in 42h micropatterns versus pluripotent cells and PS- like cells differentiated directly as in Fig. 1d. However, the change in expression between samples was similar for each isoform, with all isoforms decreasing in 2D gastruloids and directly differentiated PS-like cells relative to pluripotent cells (Fig. 3c). To corroborate this and also determine absolute FGFR isoform expression in pluripotent and PS-like cells, we analyzed our single cell RNA-sequencing data and previously published bulk RNA-sequencing data for isoform expression in these two cell types. This revealed that for both pluripotent and PS-like cells FGFR IIIc is the dominant isoform with on average almost 900-fold higher expression than IIIb, while IIIa was not expressed at all (Fig. 3d). Altogether, these data suggest that spatial patterning in ERK is due to differential activation of FGFR receptors, either due to FGF concentration gradients or possibly spatially patterned HSPGs. Immunostaining for HS showed no clear correlation with TBXT or the expected ERK signaling pattern which is similar but extends further inward (Supp. Fig. 4). This does not rule out a role for HSPGs in spatial modulation, since FGF signaling depends on specific sulfation patterns rather than overall HS levels^46,47. Nevertheless, it led us to focus on FGF gradients as the most likely cause for the ERK signaling pattern.

To maintain an endogenous FGF concentration gradient on the apical side of the colony which faces the medium, would require ligands to remain tethered to the surface. However, FGF has been reported to diffuse freely in the intercellular space⁴⁸. Moreover, there is FGF2 present in the medium that would be expected to uniformly activate FGF receptors. The situation is similar for BMP4, which is present uniformly in the medium for 2D gastruloids but does not cause uniform signaling activity⁴. In the case of BMP, that is because receptors are localized basolaterally⁴ and tight junctions seal the basolateral side from the medium. Consequently, response to exogenous ligands in the medium is restricted to the colony edge, while endogenous ligands are confined to the basal side and prevented from diffusing away into the medium everywhere but the edge. This was found to be critical for patterning in both 2D gastruloids and the mouse embryo^4,5. To explain the pERK pattern, we therefore hypothesized FGF receptors were similarly restricted to the basolateral side.

To test our hypothesis, we first starved pluripotent colonies of FGF by maintenance in FGF-free E6 medium for 12h and then treated with FGF2. Consistent with our hypothesis, phospho-FGFR and phosphor-ERK were both restricted to the colony edge (Fig. 4a). To determine if this could be explained by diffusion of exogenous FGF from the medium, we repeated the experiment with micropattern pluripotent cells and added fluorescent dextran with a similar molecular weight as FGF. We observed fluorescent dextran diffusing into the intercellular space from the edge inward over the same distance that ERK signaling was observed (Fig. 4b-d). We then acquired high resolution z-stacks of FGFR1 co-stained with the tight junction protein TJP1 (ZO-1) which revealed the majority of FGFR1 was below the tight junctions (Fig. 4e). To directly test that cells are more responsive to basal FGF than apical FGF, we then performed a transwell assay, where pluripotent cells were grown on filters to enable separate apical and basal stimulation. We found that only basal stimulation with FGF led to a significant increase in pERK (Fig. 4f). Finally, to determine if the center of the micropatterned colony is similarly able to respond to basal FGF stimulation and pERK levels there can be attributed to reduced FGF ligand concentration, we performed a scratch assay. Strong ERK response to FGF2 in the media was seen in the colony center around a scratch, but this response was blocked by FGFR inhibitor (Fig. 4gh). If HSPGs rather than FGF ligand were limiting ERK response in the center, we would not have expected strong response around a scratch. The initial ERK levels around the scratch greatly exceeded those anywhere in the control. However, we found this was due to adaptive ERK response to FGF. After six hours ERK levels around the scratch were similar to those in the ERK ring around the PS-like cells (Fig. 4ij). We therefore concluded that the pERK pattern is most likely due to a basolateral FGF ligand gradient which has its maximum in the PS-like ring.

To identify which FGF ligands are required for PS-like differentiation, we first determined the requirement for exogenous FGF2 in the media. Differentiation of 2D gastruloids in FGF-free medium was severely reduced but restored by addition of FGF2 (Fig. 5a, Supp. Fig. 5a). However, if FGF2 was removed 24h after BMP treatment, differentiation was normal. The fact that removal 2h after BMP treatment significantly reduced differentiation suggests this was not a failure to wash out the FGF2. Furthermore, MEK inhibition at 24h severely reduces TBXT expression³⁴ while FGF2 is known to have a half-life on the order of hours⁴⁹ and no media changes are performed during standard differentiation. Altogether this suggests that exogenous FGF2 is required at the time of BMP treatment but that endogenous FGFs are sufficient for the ERK activity ring and PS-like differentiation that start a day later.

We reasoned that FGFs whose expression correlates with the presence of PS-like cells are more likely to be responsible for the observed ERK ring and PS-like differentiation. Our analysis in figure 2 had implicated FGF4 and FGF17 as most strongly expressed in the PS-like cells and nascent mesoderm, respectively (Fig. 2c). However, it is unclear whether expression of these ligands is a cause or consequence of PS-like differentiation, or both, in the case of a feedback loop. To get more information about the relationship between FGF ligands and PS versus mesoderm differentiation, we performed pseudotime analysis of our scRNA-seq data (Fig. 5b, Supp. Fig 5b). Along a pseudotime trajectory from pluripotent cells to nascent mesoderm, FGF2 expression was nearly constant. FGF8 was upregulated at lower levels prior to TBXT and declined as TBXT was upregulated. FGF4 upregulation approximately coincided with TBXT, but FGF4 began decreasing as TBX6 expression increased. In contrast, FGF17 expression started after FGF4 but continued to increase after FGF4 started decreasing. Direct comparison of FGF and TBXT expression also showed strong positive correlation between TBXT and both FGF4 and FGF17 but opposing relations with TBX6 (Supp. Fig. 5c). Together this suggested to us initial TBXT expression may depend more on FGF8 or FGF4, while FGF17 could play a later role in maintaining or expanding the PS-like population or facilitating subsequent mesoderm differentiation.

Pseudotime may not accurately predict true dynamics. Therefore, we verified expression at two different times using RNA FISH (fluorescent in situ hybridization). We found that FGF4 expression was similar at 32h and 40h (Fig. 5cf, Supp. Fig. 5d). This is consistent with its decrease in pseudotime as cell differentiate to nascent mesoderm, since in real time differentiation is asynchronous and new PS-like cells form while older ones differentiate to nascent mesoderm, which would yield a flatter temporal profile. In contrast, peak FGF8 levels decreased while FGF17 strongly increased over time, both matching their pseudotime trends (Fig. 5d-f, Supp. Fig. 5ef). As expected, the spatial profiles of both FGF4 and FGF17 resembled TBXT, although the peaks were displaced slightly inward, while FGF8 expression was highest in the inner part of the ERK ring where TBXT expression was low (Fig. 5c-f). Co-staining showed FGF4 and FGF8 expressed in concentric rings with little overlap at 42h (Supp. Fig. 5g). Scatterplots further confirmed that high expression of both FGF4 and FGF17 but not FGF8 was restricted to TBXT positive cells (Supp. Fig. 5h).

To test the functions of these FGFs we decreased their expression. Knockdown (KD) of FGF4 with small interfering RNAs (siRNA) led to a strong decrease in TBXT and pERK (Fig. 5g, Supp. Fig. 6a). We repeated this in a different cell line and also created stable cell lines expressing short hairpin RNA (shRNA) to knock down FGF4 in both cell lines with similar effect (Supp. Fig. 6b-d). Strikingly, FGF8 KD had the opposite effect and led to an increase in TBXT expression without a strong effect on pERK visible at 42h (Fig. 5h, Supp. Fig. 6e). Finally, FGF17 KD caused only a small decrease in TBXT or pERK, again consistent between different cell lines (Fig. 5i, Supp. Fig. 6f-i). We further interrogated the endoderm and PGC marker SOX17 since both cell types require ERK signaling to differentiate and transiently express TBXT, as well as the mesoderm marker TBX6. TBX6 expression at 42h was near completely lost upon either FGF4 KD or FGF17 KD (Fig. 5j, Supp. Fig. 6j). This suggests that while FGF17 is not required for PS induction (Fig. 5i), it is needed for subsequent mesoderm differentiation. In contrast, there was only a small reduction of SOX17 (Fig. 5j). Since we previously found with our protocol most SOX17 positive cells at 42h represent PGC-LCs, while most endoderm differentiates later³⁴, this suggests FGF4/17 are not required for PGC-LC induction but does not rule out a requirement for these FGFs in endoderm differentiation. A double KD of FGF4 and FGF17 had a similar phenotype as FGF4 KD alone, suggesting FGF17 may be downstream of FGF4 (Supp. Fig. 6k). Combined, these data support distinct functions for FGF4, FGF8, and FGF17 in our human gastrulation model, where FGF4 and FGF8 may play opposing roles in PS induction while FGF17 is important during continued differentiation to mesoderm.

The cell lines used were the embryonic stem cell line ESI017 (XX), and the induced pluripotent stem cell line PGP1 (XY). The pluripotency of these cells was confirmed by immunostaining of pluripotency markers OCT3/4, SOX2, NANOG. All cells were routinely tested for mycoplasma contamination, and negative results were recorded. ESI017 cells stably expressing doxycycline inducible SOS^cat were made using a published Piggybac plasmid gifted by Jared Toettcher³⁷.

Human pluripotent stem cells were cultured in the commercially defined pluripotency maintaining media mTeSR1 (STEMCELL Technologies #85850) on Cultrex (R&D Systems)-coated tissue culture plates. For FGF2 starvation we used Essential 6 Medium (Thermofisher, A1516401), or mTeSR1 without bFGF, TGF-beta, LiCl, GABA and pipecolic acid (STEMCELL Technologies #05896). For routine cell maintenance, cell passaging was performed every 3 days. For whole-colony passaging, L7⁶² was used, while single-cell passaging was performed with Accutase or TrypLE (Gibco). Single-cell suspensions were used to seed for all the differentiation experiments to control initial cell number.

To directly differentiate to primitive streak-like cells, cells were resuspended in a single-cell suspension and seeded in cultrex coated wells in mTeSR with ROCK inhibitor (RI) Y-27632 (MeChemExpress, cat# HY-10583). Cells were then maintained in mTeSR with RI for 24h before adding CHIR (3-5uM) and Activin (100ng/ml) treatment for 24-30h. Experiments were performed in 18-well Ibidi slides (cat# 81818).

To differentiate cells for 2D gastruloids by micropatterning, we followed the protocol in⁴². Briefly, cells were resuspended in a single-cell suspension and seeded in laminin-coated micropatterned wells in mTeSR with ROCK inhibitor (RI) (MeChemExpress, cat# HY-10583). Micropatterned wells were washed with PBS-/- 30 minutes after the initial seeding to wash off the non-attached cells. Cells were then maintained for 24 hours in mTeSR before adding the BMP4 treatment by a full medium change. For time series micropatterned differentiation, full media changes with designated treatment(s) were performed every 24 hours. Experiments were performed in micropatterned 18-well Ibidi slides (cat# 81818) prepared as previously described⁶³. Cell signaling reagents and doses used are listed in Supplementary Table 1.

Antibodies can be found in Supplementary Tables 2 and 3. For stains including pERK, we used methanol fixation, for all other stains we used PFA fixation. PFA fixation: Samples from the 18-well Ibidi slides were rinsed with PBS, fixed for 20 min in 4% paraformaldehyde, rinsed twice with PBS, and blocked for 30 min at room temperature with 3% donkey serum and 0.1% Triton X-100 in 1× PBS. After blocking, cells were incubated with primary antibodies at 4°C overnight, followed by three washes in PBST (PBS with 0.1% Tween 20). They were then incubated with secondary antibodies and DAPI for 30 min at room temperature and washed twice in PBST at room temperature. Methanol Fixation: Samples from the 18-well Ibidi slides were rinsed with PBS, fixed for 20 mins in cold methanol, then rinsed with PBS for 5 min two times, after which 10% phosphate buffered formalin was added for 20 min at room temperature. Cells were then rinsed with TBS twice followed by 5 mins of incubation in cold methanol and rehydrate in TBS for 20 mins. Blocking was done for 30 mins at room temperature with 5% donkey serum, 1% BSA, 0.2% Triton-X-100 in 1x TBS after which staining proceeded the same way as with PFA fixation.

Fixed sample imaging was performed with an Andor Dragonfly/Leica DMI8 spinning disk confocal microscope with a 40x, 1.1NA water and 20x 0.8NA air objectives using Andor Fusion software version 2.3.0.31 and a Nikon/Yokogawa spinning dish confocal microscope with a 40x silicon oil objective using NIS Elements AR software version 5.41.02. Live-cell imaging was performed with an Andor Dragonfly/Leica DMI8 spinning dish confocal microscope under the controlled temperature (37°C), CO₂ concentration (5%), and humidity (>60%).

We segmented nuclei in individual z-slices based on nuclear fluorescence stained with DAPI in fixed cells using a pipeline we previously described, which integrates two machine learning approaches: Ilastik pixel classification⁶⁴ and Cellpose⁶⁵ (v1). Fluorescence intensities were then calculated per nucleus as mean intensities in the 3D nuclear again with our established custom image-processing pipeline. To calculate the radial fluorescence distributions in micropatterned colonies, colonies were subdivided into radial bins with equal numbers of cells, mean or median and variance were then calculated within these bins. To calculate positive fractions, markers were first thresholded based on their intensity distribution, after which the fraction of cells positive for each marker was calculated within different bins. Standard deviations were then calculated over multiple replicate colonies. For visualization purposes only, background subtraction using grayscale opening on a 100 micron scale and 5 pixel wide median filter was applied slice by slice to z-stacks of micropatterned colonies before making maximal intensity projections.

Cells were collected using accutase and resuspended in ice-cold PBS. Single-cell RNA-sequencing was performed by the University of Michigan Advanced Genomics Core. Cells were barcoded using the 10X Genomics Chromium system (part numbers 1000268, 1000120, 1000215). For quality control, cDNA was quantified by Qubit High Sensitivity DNA assay and Agilent TapeStation. Sequencing was performed on the Illumina NovaSeq 6000 with NovaSeq S4 flowcell and Control Software version 1.7.0. Reads were aligned using cellranger-4.0.0 with the GRCh38 reference. Further processing was mostly done in Python, primarily using the Scanpy⁶⁶ and SCVI⁶⁷ packages. Analysis was performed on log transformed data. Library size normalization was performed on the intersection of the genes present in our data and the reference CS7 human embryo data to facilitate comparison of expression levels. We integrated the two scRNA-seq datasets for 42h 2D human gastruloids using SCVI based on the top 2000 highly variable genes, which were selected using the default function in Scanpy using the seurat_v3 flavor. Cell cycle genes were regressed out during integration. Dimensional reduction for visualization was performed using PHATE³⁹ on the SCVI latent space. Gene expression was smoothened with MAGIC⁶⁸ for visualization on PHATE plots in Fig. 2e,g-i. To annotate different cell types, we used MAGIC to impute that data and get smooth distributions of gene expression. We then thresholded marker genes to assign cell fate (Table 4), which produced better results than the more common unsupervised clustering using Leiden, and in particular more consistent results between datasets from different species. For interspecies comparison of FGF ligands we normalized the to the maximal FGF expression within each species, where for the mouse data, the three different time points were first integrated using SCVI. Thus we compare between species the relative expression of FGF ligands within each species, e.g., FGF8 expression is much higher than FGF4 at any time in the mouse, but much lower than FGF4 in human.

For pseudotime analysis (Fig. 5), we applied the trajectory inference tool Slingshot⁶⁹ in R to compute the pseudotime trajectory in the SCVI latent representation. Slingshot determines the membership of a cell to a lineage at each branching point by its projection distance to the fitted principal curve of that lineage. We projected the mesoderm developmental trajectory onto the PHATE map for visualization. We further used TradeSeq⁷⁰, which fits a generalized additive model (GAM), to infer the pseudotime gene expression trends along the mesoderm lineage trajectory from Slingshot. For visualizing the correlation between the expression of FGF ligands and marker genes (Supp. Fig. 5), we used our SCVI model to denoise the expression and mutual nearest neighbor to further harmonize the difference between samples.

To quantify isoforms from RNA sequencing salmon v1.10.0 was used⁷¹. First, gencode.v45.transcripts.fa was downloaded from gencodegenes.org. The reference was customized by removing lines for FGFR1 and replacing them with the transcript sequence for FGFRIIIa, FGFRIIIb, FGFRIIIc (sequences included as supplementary file). Three samples from GEO dataset GSM2257301 were downloaded through SRA (SRX1725562 (Pluripotent), SRX1725577 (APS), SRX1725643 (MPS))³⁸. Adapters and the first 10 bases were trimmed from the FASTQ files using trim_galore. Finally, the Salmon “quant” command with the --validateMappings option was used to quantify transcripts from trimmed FASTQ files. For single cell datasets, https://timoast.github.io/sinto/ was used to subset the 10x aligned file by cluster. Reads for clusters of interest were converted back to raw FASTQ files and run through the same alignment-based mode of salmon but as single end reads.

RNA were extracted using RNAqueous™-Micro Total RNA Isolation Kit (CAT# AM1931, Thermo Fisher Scientific) and then cDNA synthesis was prepared from it using SuperScript™ VILO™ cDNA Synthesis Kit (CAT# 11754250, Thermo Fisher Scientific) following the manufacturer‘s protocol with adjustment for optimization. Measurements were performed with SYBR green and the primers in the Table 5. GAPDH was used for normalization in all experiments.

Fluorescent dextran essays were performed by adding 10 ug/ml fluorescein isothiocyanate-dextran (10KD, Sigma Aldrich) to the media. To calculate fluorescence intensity as a function of distance from the colony edge, Otsu thresholding followed by several morphological operations was used to create a mask for the colony in each z-slice, thereby excluding bright dextran fluorescence outside the colony. Background subtraction on each z-slice of the image was performed to further remove background signal from dextran outside the colony while maintaining signal in the intercellular space. The radial intensity profile was then calculated from the maximal intensity projection of the masked, background subtracted image.

Human embryonic or induced pluripotent stem cells were differentiated on micropattern wells for 41h, then each micropattern colony was manually scratched with a serological needle. After 60mins, they were fixed with PFA or methanol followed by immunostaining with appropriate antibodies. To calculate the intensity profile as a function of distance from a scratch, we created a mask for the scratch based on the difference between an overall mask for the colony and the convex hull, calculated the distance transform of the scratch mask, then binned pixels by distance from the scratch and calculated their mean ERK intensity.

Human embryonic or induced pluripotent stem cells were seeded into Corning 6.5 mm transwell with 0.4 µm pore polyester membrane insert (cat# 3470) with E6 supplemented with TGF1b, FGF2, and Y-27632 for 16h, then media was changed to E6 with TGF1b. After 24h, the media was changed to E6 supplemented with FGF2 from the top or the bottom side of the transwell. 30 mins later, cells were fixed followed by immunofluorescence staining. The transwell cartoon in Fig 4 was generated with Biorender.

The fluorescence in situ hybridization protocol (FISH) was performed according to the manufacturer’s instructions (ACDbio, RNAscope multiplex fluorescent manual). A list of probes and reagents can be found in Supplementary Table 6.

We used the constructs shRNA F4: pPB[shRNA]-EGFP-U6>hFGF4[shRNA#1] and shRNA F17: pPB[3shRNA]-U6>hFGF17[shRNA#1]-U6>hFGF17[shRNA#2]-U6>{hFGF17_shRNA3}-hPGK>Bsd both purchased from VectorBuilder in a PiggyBac Transposon Vectors with puromycin resistance which was stably integrated into ESI17 hESCs and PGP1 hiPSCs. Transfections were performed using Lipofectamine™ Stem Transfection Reagent (CAT# STEM00001, Thermo Fisher Scientific), following the manufacturer‘s protocol.

We purchased predesigned siRNA from Millipore Sigma; NM_002007 (FGF4), NM_006119 (FGF8), NM_008004 (FGF17). The siRNA transfections were performed using Lipofectamine™ RNAiMAX Transfection Reagent (CAT# 13778075, Thermo Fisher Scientific), following the manufacturer‘s protocol with adjustment for optimization.

All experiments were performed twice or more. All the attempts in replicating experiments yielded consistent results. Unless specifically noted, all the quantifications were performed with N=4 colonies within the same experimental condition.

We comply with all the ethical regulations in our research. We were granted approval by the Human Pluripotent Stem Cell Research Oversight (HPSCRO) Committee at the University of Michigan to work with human embryonic stem cells and human induced pluripotent stem cells.