Bioengineered embryoids mimic post-implantation development in vitro

The difficulty of studying post-implantation development in mammals has sparked a flurry of activity to develop in vitro models, termed embryoids, based on self-organizing pluripotent stem cells. Previous approaches to derive embryoids either lack the physiological morphology and signaling interactions, or are unconducive to model post-gastrulation development. Here, we report a bioengineering-inspired approach aimed at addressing this gap. We employ a high-throughput cell aggregation approach to simultaneously coax mouse embryonic stem cells into hundreds of uniform epiblast-like aggregates in a solid matrix-free manner. When co-cultured with mouse trophoblast stem cell aggregates, the resulting hybrid structures initiate gastrulation-like events and undergo axial morphogenesis to yield structures, termed EpiTS embryoids, with a pronounced anterior development, including brain-like regions. We identify the presence of an epithelium in EPI aggregates as the major determinant for the axial morphogenesis and anterior development seen in EpiTS embryoids. Our results demonstrate the potential of EpiTS embryoids to study peri-gastrulation development in vitro.


Introduction
Recapitulating early mammalian development in vitro is an important challenge that could overcome the experimental bottleneck created by intrauterine development and assist in the reduction of the use of animals in research 1 . The control of the differentiation potential of ESCs has been a central feature in these efforts. Constraining the differentiating cells within micropatterned substrates controls the stochastic differentiation of cells in adherent culture and results in patterns that resemble the organization of germ layers in the embryo 2,3 . Three-dimensional embryoid models trigger selforganization programs that mimic much of the early development of the embryo and overcome the heterogeneous and heterochronic patterns characteristic of Embryoid Bodies (EBs) 4 . In particular, gastruloids 5 , aggregates of defined numbers of ESCs, develop derivatives of all germ layers with spatiotemporal patterns characteristic of embryos even though they lack clear brain structures 6 .
Gastruloids have proven useful tools to explore the consequences of gastrulation in the absence of extraembryonic tissues, such as formation of cardiac primordia 7 or somite-like structures 8,9 . Other embryoid models have explored interactions between embryonic and extraembryonic tissues during early development by taking advantage of the existence of trophoblast (TSC) and extraembryonic endoderm (XEN) stem cells to recreate the early conceptus. Co-culture of TSCs and XEN cells with mouse ESCs in Matrigel results in a suite of structures, including blastoids 10 , ETS 11 -and ETX 12embryos, that recapitulate events and interactions of the pre-gastrulation embryo. A challenge with ETSand ETX-embryos has been the reliance on the self-organizing activity of the cellular compartments that results in their stochastic occurrence; only ~ 20% of the starting aggregates will form patterned structures 11 , which limits their broader applicability. Furthermore, their developmental potential is not yet clear, as there are no reports of their development beyond early gastrulation stages 11 . Altogether, these pioneering in vitro models of early embryo development highlight the remarkable capacity of embryonic and extraembryonic cells to organize themselves but thus far have not yet been ideal to explore the fate and derivatives of cells in the emerging structures.
Here, we used a bioengineering approach combining features of ETS embryos and gastruloids to build an embryoid model that could more realistically capture the events leading to and resulting from gastrulation. We hypothesized that to achieve this, it would be essential to optimize the starting culture conditions in multicellular aggregates to assemble, in the same embryoid, an anterior neuroepithelium with a posterior T/Bra expressing cell ensemble. We used arrays of cell-repellent hydrogel microwells to separately generate epithelialized or non-epithelialized epiblast-like (EPI) aggregates, as well as TSC aggregates mimicking extraembryonic ectoderm, in a scalable manner. When assembled together in low-attachment wells in serum-free medium, EPI and TSC aggregates rapidly merged and underwent symmetry breaking similar to gastrulating embryos, as demonstrated by polarized and restricted Brachyury (T/Bra) expression. Importantly, in these structures, T/Bra expression dynamics was strictly dependent on the epithelial architecture of the EPI aggregates. Subsequently, hybrid EPI/TSC structures, termed EpiTS embryoids, underwent axial morphogenesis to display patterning along anterior-posterior, dorsal-ventral and medio-lateral axes. Strikingly, we observed that embryoids formed from nonepithelialized EPI aggregates predominantly generated mesendodermal tissue, whereas epithelialized ones formed cell types that are present in the developing midbrain/hindbrain. Our approach enables the generation of various tissues in a stereotyped and scalable manner, with independent modulation of physical (e.g., size, epithelial architecture) and biological (e.g., provision of signaling molecules) properties of EPI and TSC aggregates, such as to systematically parse out the role of these parameters in promoting key steps in embryogenesis.

Scalable formation of homogeneous EPI aggregates
We aggregated ESCs in round-bottom microwell arrays composed of non-adhesive PEG hydrogels 13 in epiblast induction medium comprising Activin-A, bFGF and KSR, supplemented with low percentage (3%) Matrigel to induce epithelialization 14 (Fig. 1a). The microwells allowed us to titrate the average number of cells seeded in each well, resulting in aggregates of defined size (Fig. 1b, Supplementary   Fig. 1a, upper panel). The starting number of cells determined the size of the aggregate at 72 h of culture: an average of ~25 ESCs per microwell yielded aggregates with ~180 µm diameter, whereas seeding ~100 cells reached a diameter of ~230 µm (Fig. 1c). These aggregates featured a single lumen surrounded by an E-cadherin-positive polarized epithelium, displaying apical Pdx and Par6 expression ( Fig. 1d-f, lower panel; Supplementary Fig. 1c). Smaller aggregates exhibited a discontinuous apical expression of aPKC surrounded by a multi-layered E-cadherin-positive epithelium and multiple F-actin labeled cavities, suggesting poor epithelialization and incomplete lumenization ( Fig. 1d-f, upper panel).
These results demonstrate that a critical aggregate size needs to be reached to form an apico-basally polarized epithelium with a central lumen, a phenomenon reminiscent of epiblast maturation 15 . When cells were aggregated in the absence of Matrigel, we detected increased cell shedding ( Supplementary   Fig. 1a, lower panel, white arrows) but no significant difference in aggregate diameter ( Supplementary   Fig. 1b). However, these aggregates composed of E-cadherin-positive cells demonstrated inverted Factin polarity and featured multiple Pdx-positive and Par6-positive foci with no clear epithelium or lumen (Fig. 1g, Supplementary Fig. 1d). These results show that polarized epithelial aggregates of defined size can be readily generated from ESCs using hydrogel microwell arrays, and confirmed that the provision of Matrigel is critical for their lumenization and epithelialization 16 .
Of note, affecting cytoskeletal activity by inhibiting stress fiber formation (Y-27632) and disrupting the actin network (CK-666) resulted in loss of a continuous apicobasally polarized epithelium. Interestingly, inducing stress fiber formation with LPA did not have any effect on epithelialization (Supplementary Fig. 2a,b). Moreover, blocking actin-myosin interaction (Blebbistatin), myosin activity (ML-7 and Calcyculin-A) or actin polarization dynamics (Cytochalasin-D/Latrunculin-A and Jasplakinolide) prevented formation of compact aggregates and, in some cases (ML7 and Cyto-D), epithelialization. Interestingly, treatment with the PTEN inhibitor BpV showed a similar effect, in line with previous reports highlighting the role of PTEN in epiblast polarization 17 .
These results suggested that actin polymerization and actin-myosin interaction play crucial roles in the polarization and epithelialization of EPI aggregates.
To test the identity of the aggregates after 72 h, we performed immunostaining for pluripotency and epiblast markers. In the absence of Matrigel, the aggregates showed a uniform expression of Otx2 and E-cadherin (Fig. 1h). Interestingly, Sox2 expression was detected in a polarized fashion in cells that showed high WNT and low TGF-β activity (Fig. 1h, Supplementary Fig. 3a), suggestive of an axis establishment similar to 48-72 h gastruloids 18 . Epithelialized aggregates were comprised of E-cadherin+ cells organized in a columnar epithelium and were uniformly positive for Oct4 and Otx2.
Expression of Sox2 and Nanog were detected at lower levels, often in a salt-and-pepper fashion. In general, Sox2+ cells in epithelialized aggregates demonstrated high WNT and low TGF-β activity (Fig.   1i, Supplementary Fig. 3b). Bulk RNA sequencing revealed low expression levels for naïve pluripotency factors Klf2/4, Dppa3, Tbx3 and higher levels of epiblast-specific genes Fgf5, Otx2, Utf1 in EPI aggregates. At this stage, transcripts that mark further differentiated states such as T, Sox1 or Sox17 were not upregulated (Fig. 1i). Altogether, these observations suggested that ESCs could be coaxed into EPI aggregates that morphologically and transcriptionally resemble post-implantation epiblast.

Scalable formation of homogeneous TSC aggregates
In the early embryo, the epiblast develops in conjunction with extraembryonic tissues, and signaling from the extraembryonic ectoderm is involved in specifying the onset of gastrulation 19 . To mimic these interactions, we used the same hydrogel microwell array technology to generate TSC aggregates (Fig.   1k). On average, TSC aggregates were slightly smaller than EPI aggregates but their size at 72 h was also found to be dependent on the initial cell seeding concentration: aggregates composed of 25 cells or 100 cells reached a diameter of ~140µm or ~200µm, respectively (Fig. 1l). We detected expression of Cdx2, Elf5 and Eomes transcripts and high levels of Tfap2c (Fig. 1m), suggesting that aggregated TSCs maintain stem-cell identity and get primed for differentiation 20 . At this timepoint, genes expressed in more differentiated cell types such as spongiotrophoblast or giant cells were detected at relatively low levels (Fig 1m) Unlike EPI aggregates, TSC aggregates grown in the presence of low percentage Matrigel did not lumenize (Fig. 1n). Immunostaining showed that TSC aggregates could interact with the surrounding laminin-based extracellular matrix (ECM) and deposited fibronectin at the aggregate core ( Fig. 1o). Expression of the stem cell markers Cdx2, Tfap2c and Eomes were primarily maintained on the periphery, suggesting initiation of differentiation from the core of the TSC aggregates 24 (Fig. 1o,p).

Formation of EpiTS embryoids
Next, we assembled the EPI and TSC aggregates into structures that we termed EpiTS embryoids.
Individual EPI and TSC aggregates were transferred at 72 -75 h to U-bottom low-attachment wells of a 96-well plate, where they fused together within a few hours (Fig. 2a, top panel). Time-lapse imaging revealed that by 120 h, T/Bra expression appeared at the aggregate interface. The size of EPI and TSC aggregates strongly influenced the timing of T/Bra expression (Fig. 2b). Embryoids composed of smaller EPI aggregates initiated T/Bra expression generally before 110 h; by 120 h, all embryoids had T/Bra-expressing cells regardless of the size of TSC aggregates they were fused to (Fig. 2a, bottom panel). On the other hand, EpiTS embryoids formed from larger EPI aggregates showed a delayed onset of T/Bra expression (Fig. 2b), with 50-70% being T/Bra-positive at 120 h (Fig. 2a, bottom panel). In addition to differences in timing, we observed a strong dependence of the initial EPI aggregate size on the T/Bra expression domain (Fig. 2c). At 120 h, embryoids formed from smaller EPI aggregates acquired a dispersed expression of T/Bra, covering almost the entire EPI compartment, whereas EpiTS embryoids formed from bigger EPI aggregates featured a restricted T/Bra expression (Supplementary

Movies 1-4).
To test whether the appearance of T/Bra expression was dependent on the epithelial architecture of the aggregates, or the presence of an extraembryonic compartment, we generated epithelialized (with Matrigel) and non-epithelialized (w/o Matrigel) EPI aggregates, and cultured them in the presence or absence of TSC aggregates. Almost all non-epithelialized EPI aggregates displayed T/Bra expression by 120 h, independent of the extraembryonic compartment. In contrast, epithelialized EPI aggregates alone did not initiate T/Bra expression and required co-culture with TSC aggregates (Fig. 2d,e), similar to embryo explants cultured without extraembryonic ectoderm 25 . Moreover, the T/Bra expression domain was found to be more restricted in epithelialized embryoids (Epith.+TS) compared to nonepithelialized (Non-epith.+TS) ones (Fig. 2f).
To test whether T/Bra induction is specific to TSC aggregates, we co-cultured epithelialized EPI aggregates with aggregates composed of mouse embryonic fibroblasts (MEFs). MEF aggregates could occasionally induce T/Bra expression, albeit at much lower levels compared TSC aggregates ( Supplementary Fig. 4a,b). This suggests that cellular interaction alone is not sufficient to induce T/Bra expression and that TSC aggregate-specific signaling may be necessary.
In vivo, signaling from the extraembryonic ectoderm is crucial for the initiation of gastrulation 26 .
To better understand the signaling influence of the TSC compartment in our embryoids, we replaced the cellular aggregates with cell-adhesive hydrogel microbeads (composed of denatured collagen crosslinked to dextran) coated with various (diffusible) morphogens involved in early mouse development.
When coupled to EPI aggregates at 72 h, plain beads or beads coated with Fgf2, Activin-A or Wnt3a could induce T/Bra expression, although at lower levels than TSC aggregates. Strikingly, Bmp4-coated beads reached the highest T/Bra expression level, outperforming the TSC aggregates ( Supplementary   Fig. 4a,b). Similarly, Bmp4 delivery increased WNT activity, even to higher levels compared to Wnt3a ( Supplementary Fig. 4a,c). Notably, we did not detect any significant change in TGF-β activity for the proteins tested ( Supplementary Fig. 4a,d). Transwell co-culture of epithelialized EPI aggregates with varying numbers of TSC aggregates did not induce T/Bra expression (Supplementary Fig. 4c,d), showing that physical interactions between the two cell compartment is crucial for symmetry breaking.
Altogether, these results highlight a role of the epithelial architecture of EPI aggregates for T/Bra expression dynamics. EPI aggregates that are small (i.e., poorly epithelialized), or formed in the absence of Matrigel (i.e., not epithelialized), showed an extended T/Bra expression domain when coupled to TSC aggregates, suggesting uniform mesodermal differentiation, while in EPI aggregates the epithelium limits T/Bra expression to one side of the embryoids. Furthermore, TSC aggregate-specific signaling, likely via Bmp4, and physical EPI-TSC interaction are both important for the reproducible symmetry breaking and induction of T/Bra expression in epithelialized EpiTS embryoids.

Cell fate assignments in EpiTS embryoids
Next, we performed immunostaining and confocal microscopy to characterize the tissues present in  (Fig. 2g). We believe that the Oct4+ T-Bra+ Nanog+ domain in these early EpiTS embryoids resembles the posterior end of the early to mid-streak embryo 27 . Conversely, the Oct4+ Otx2+ Sox2+ domain marked the opposite end of the embryoids, suggesting maintenance of the anterior pluripotent epiblast. We did not detect Sox17-positive cells in embryoids that initiated T/Bra expression, but we could detect some Dppa3-positive cells (Supplementary Fig. 5b), likely marking primordial germ cell (PGC) progenitors 28 .
Embryoids formed from non-epithelialized EPI aggregates showed dispersed T/Bra expression and did not demonstrate E-cadherin+ rosettes. Interestingly, Pdx expression was downregulated in the embryonic compartment. Similarly, laminin was completely absent, and the majority of cells demonstrated Snai1 expression, suggesting an expanded primitive streak-like domain ( Supplementary   Fig. 5c). T/Bra expression mostly co-localized with Oct4/Nanog, however, Otx2/Sox2 expression spanned a smaller domain compared to epithelialized embryoids, suggesting the loss of anterior neural progenitors at the expense of expanded posterior tissue (Fig. 2h). Moreover, we detected a higher number of Sox17-positive endoderm-like cells and Dppa3-positive PGC progenitors ( Supplementary   Fig. 5d), indicating an increased differentiation of posterior cell types.
Collectively, these results show that the epithelial architecture of the EPI aggregates directly influences the type of tissue that is created in embryoids. The presence of an epithelium favors limited T/Bra induction and a larger domain of anterior epiblasts that resemble E6.5 embryos. Conversely, in the absence of an epithelium in the initial EPI aggregate, the embryoids consist largely of posterior tissues and show premature differentiation towards the endoderm lineage.

Gastrulation-like events are orchestrated by key early developmental signaling pathways
The onset of gastrulation is the output of asymmetric signaling across the post-implantation epiblast and is tightly regulated by WNT, TGF-β and BMP signaling pathways 29 . Embryos that lack functional βcatenin 30 or cleavable Nodal 31 fail to upregulate T/Bra and are unable to undergo gastrulation. Time-lapse analysis showed that in non-epithelized embryoids, T/Bra expression begins as early as 96 h, in contrast to epithelized ones, which show a delayed initiation at 120 h ( Fig. 3a-b, upper panels;

Fig. 3c).
In epithelialized embryoids, we could detect an increase of WNT activity from 72 to 96 h, followed by a reduction and restriction to the embryonic-extraembryonic interface by 120 h. In contrast, non-epithelial embryoids demonstrated a progressive increase in WNT activity that was detected throughout the EPI domain ( Fig. 3a-b, middle panels; Fig. 3d). On the other hand, TGF-β signaling was continuously active in both conditions, but its expression domain became restricted over time. Of note, compared to epithelialized embryoids, non-epithelialized ones showed significantly higher TGF-β pathway activity from 72 to 120 h, (Fig. 3a-b, bottom panels; Fig. 3e). These results showed that in EpiTS embryoids, both WNT and TGF-β pathways are active and their dynamics depend on the epithelial architecture of the EPI aggregate. In epithelialized embryoids, WNT and TGF-β signaling preceded T/Bra induction and are co-localized by 120 h, resembling the initial primitive streak-like domain. The signaling polarization rendered the anterior of the embryoids low in WNT and TGF-β activity ( Fig. 3a-b, white arrows), potentially preserving the potential for anterior neural progenitors 35,36 .
To test whether gastrulation-like processes observed in epithelialized EpiTS embryoids involve mechanisms similar to those observed in the embryo, we inhibited the above pathways by adding Dkk1 (i.e., a WNT inhibitor), Lefty (TGF-β inhibitor) and Noggin (BMP inhibitor) between 72-96 h or 96-120 h, respectively. Inhibition of WNT and TGF-β pathways in both time windows greatly reduced the number of embryoids that upregulate T/Bra. However, BMP inhibition by Noggin between 72-96 h significantly lowered the percentage of T/Bra+ embryoids by 120 h (Fig 3f,g). Dkk1-treated EpiTS embryoids exhibited a delayed T/Bra expression compared to non-treated or Noggin-treated ones (median ~140 h versus ~125 h) (Fig 3h). Furthermore, we observed more restricted T/Bra expression in the presence of the inhibitors, which was more pronounced in the cases of Dkk1 and Lefty (Fig. 3i).
Interestingly, early treatment with these inhibitors significantly reduced the total embryoid size, an effect that was more prominent when BMP signaling was inhibited (Fig 3j). Treatment with the small molecule inhibitors XAV939 (i.e., a WNT inhibitor), SB431542 (TGF-β inhibitor) and LDN193189 (BMP inhibitor) resulted in a more pronounced reduction of T/Bra expression; SB431542 and XAV939 treatment led to a complete loss of T/Bra expression, while a small percentage of LDN193189-treated EpiTS embryoids still expressed T/Bra, yet with a domain that was spatially restricted (Fig. 3k,l).
Altogether, these results demonstrate a crucial role of WNT and TGF-β signaling in the induction and level of T/Bra expression in EpiTS embryoids between 72 and 120 h. In contrast, the BMP pathway was shown to have an effect on the overall growth of embryoids, with a smaller effect on T/Bra expression between 96 to 120 h.

Spontaneous axial morphogenesis of EpiTS embryoids
Previous embryoid models involving ESCs and TSCs in ETS-embryos 11 did not recapitulate the axial patterning of the early embryo. In our hands, ETS-embryos could be generated at frequency (corresponding to multicellular aggregates comprising an embryonic and extraembryonic compartment) of ~6% (n=34/560) (Supplementary Fig. 6a). The ETS-embryos expressed epiblast markers Oct4 and Otx2 at 120 h (Supplementary Fig. 6b). To test whether post-gastrulation developmental stages could be achieved in ETS-embryos, we adopted our EpiTS embryoid culture approach by placing individual ETS-embryos in U-bottom 96-wells. T/Bra expression was found to be highly variable, with some tissues showing a dispersed and others a more localized expression by 120 h (Supplementary Fig. 6c).
When cultured until 240 h, ETS-embryos with small embryonic-to-extraembryonic ratio were overgrown by TSC aggregates, and ETS-embryos that were small in size (< 200µm) did not grow further. Interestingly, we observed that ETS-embryos that upregulated T/Bra and had a larger embryonic compartment occasionally underwent axial elongation, suggesting that when grown in ECM-and serumfree culture conditions, these embryoids might reach post-gastrulation stages (Supplementary Fig. 6d).
When 120 hours old EpiTS embryoids were cultured in N2B27 medium with no additional growth factors, the embryonic compartment began to elongate (Fig. 4a,b). While epithelialized EPI aggregates alone remained spherical, non-epithelialized ones elongated efficiently, similar to classical gastruloids 39 (Supplementary Fig. 7a). The EPI compartment in EpiTS embryoids grew significantly over time, while the size of the TSC aggregates did not change markedly (Fig. 4c), suggesting that growth of extraembryonic compartment is not required for the elongation of the EPI domain. To quantify the dynamics of elongation, we performed image analysis to determine the axial length and elongation index of the embryoids (Supplementary Fig. 7b). This analysis revealed that in both epithelialized and non-epithelialized embryoids, the axial length and the elongation index significantly increased from 120 to 168 h (Fig. 4d, Supplementary Fig. 7c). EpiTS embryoids formed from non-epithelialized (or smaller) EPI aggregates were significantly more elongated at 144 h (Fig. 4d, Supplementary Fig. 7d,e), suggesting a role for the size and epithelial architecture of the starting EPI aggregate on the elongation dynamics (Supplementary Movies 5,6).
In embryoids generated from epithelialized tissues, T/Bra expression peaked around 144 h and gradually declined until 168 h, when it became restricted to the elongating tip and then trailed behind in a short stripe of cells. In contrast, non-epithelialized embryoids showed a constant decrease in the intensity and size of the T/Bra domain with increasing elongation (Fig. 4e, Supplementary Fig. 7f).
WNT signaling was active in almost all embryoids and marked the T/Bra-expressing tip with no significant change in activity levels until 168 h (Supplementary Fig. 7g,h). In epithelialized embryoids, WNT activity reappeared at the anterior domain (Fig. 4a,b, middle panels, white arrows), resembling the TCF/LEF reporter activity in the tail and midbrain/hindbrain of E8.5 embryos 33,40 . Overall, TGF-β pathway activity was found to be lower and in close proximity to the extraembryonic compartment, excluded from the elongating tip (Fig. 4a,b, bottom panels; Supplementary Fig. 7i,j).

Signaling and axial patterning in EpiTS embryoids
To better understand the underlying mechanism of axial morphogenesis in our embryoids, we investigated the roles of the WNT, TGF-β and BMP pathways by inhibiting their activity between 96 and 120 h. In general, treatment with inhibitors resulted in significant size reduction by 168 h (Supplementary Fig. 8a). Dkk1 and Lefty led to a reduction of T/Bra-positive embryoids, compared to Noggin. Embryoids treated with Lefty exhibited small clumps of T/Bra-positive cells, while Dkk-treated ones showed a more scattered expression throughout the body of the structure (Supplementary Fig.   8b). In Noggin-treated embryoids, T/Bra expression was retained within a larger domain compared to untreated or Dkk1/Lefty-treated ones (Supplementary Fig. 8c). Notably, the elongation frequency was found to be significantly higher in the presence of Noggin (Supplementary Fig. 8d, Supplementary   Movies 7-10). It is conceivable that excessive BMP signaling from the extraembryonic ectoderm could impair axial elongation, and treatment with Noggin could mitigate this effect, consistent with the antagonistic role of BMP4 on axial elongation of gastruloids 18 . In these embryoids, T/Bra-positive cells were often organized into long stripes that co-expressed FoxA2, suggestive of a notochord or an anterior mesendoderm (AME)-like structure (Supplementary Fig. 8e). In support of this, we detected higher expression levels of T/Bra, Shh, FoxA2 and Gsc with Noggin-treated compared to untreated embryoids ( Supplementary Fig. 8f), in line with previous reports defining a role for Noggin in proper notochord formation 41 .
At 168 h, epithelialized embryoids acquired an expanded neural identity, as evidenced by Sox1 expression spanning from the elongated tip towards the opposite end. Interestingly, we could detect polarized expression of Pax6 and T-Bra/Foxa2, all partially overlapping with Sox1 expression, suggesting a dorsal-ventral organization similar to the developing neural tube 42 . Furthermore, the neural tissue was flanked by Six1+ Eya1+ paraxial mesoderm derivatives (Fig. 4f), suggesting an organization across a medio-lateral axis. We detected numerous Tuj1-positive neurons within Sox1-positive rosettes located on the opposite end of the elongated tip. The rosettes and neurons showed high WNT activity co-localized with Otx2 expression, indicating establishment of an anterior-posterior axis to include a 'brain-like' tissue. Moreover, only few Sox17-positive cells were found scattered in the vicinity of the extraembryonic domain, suggesting poor endoderm differentiation (Fig. 4g). In contrast, nonepithelialized embryoids showed a smaller domain of Sox1 expression comprising comparable numbers of Tuj1-positive neurons with relatively short axonal projections. Compared to epithelialized embryoids, the Otx2 expression domain was more restricted and showed low levels of WNT activity. Tuj1-positive neurons were also located away from the WNT-positive elongated tip. Interestingly, the expression domain of Sox17 in these embryoids was much larger and located at the anterior end of the embryoids co-localizing with Foxa2 (Fig. 4h, Supplementary Fig. 8g), similar to endoderm tissue seen in gastruloids 6 . Collectively, these data demonstrate the capacity of EpiTS embryoids to spontaneously undergo axial morphogenesis to establish multi-axial patterning. Importantly, the presence or absence of an epithelium in the starting EPI aggregate directly influences the elongation dynamics as well as the proportions of the tissue types that are generated.

Single cell RNA-sequencing reveals distinct differentiation trajectories in EpiTS embryoids
To shed light on the cell type composition and diversity of EpiTS embryoids, we performed single cell RNA sequencing (scRNA-seq) analysis between 120 and 192 h (Fig. 5). We could identify that from 120 h onwards, cells mapping to an epiblast-state gradually differentiated into ectoderm or mesendoderm lineages (Fig. 5a, germ layers). Primordial germ cells and extraembryonic tissue had a distinct transcriptional identity that segregated away from the germ layers observed in our embryoids.
RNA velocity on extraembryonic cells showed no clear temporal differentiation trajectory; however, we identified distinct cell populations that suggested a compartmentalization within the extraembryonic tissue (Fig. 5a, extraembryonic). At 120 h, the majority of the extraembryonic tissue had chorion identity (Cdx2, Elf5, Eomes, Esrrb) and expressed Bmp4, supporting involvement of BMP signaling for the initiation of gastrulation-like events in our embryoids. Interestingly, we detected Furin and Pcsk6 expression that could suggest the processing of embryonic Nodal by the extraembryonic tissue 43 . We further identified clusters of cells that map to ectoplacental cone (Fgfbp1, Ascl2) or labyrinth progenitors (Gjb2, Dlx3, Ovol2) that demonstrate multipotentiality of chorionic ectoderm. At later stages of culture, we detected trophoblast giant cells (Hand1, Ctsq, Prl2c2) and a few spongiotrophoblast cells (Tpbpa, Flt1) (Fig. 5b, extraembryonic markers, Supplementary Fig. 9a-d), in line with previous in vitro TSC differentiation data 44 .
We then used the scRNA-seq dataset to systematically assess the effect of the initial epithelial architecture of embryoids on the resulting cell type diversity. A comparative analysis of epithelialized and non-epithelialized embryoids showed that several tissue types, including cardiac mesoderm, definitive endoderm and pharyngeal mesoderm derivatives, were almost exclusively found in nonepithelialized embryoids (Fig. 5e, Supplementary Fig. 13a,b). Moreover, the neural tissues in nonepithelialized embryoids were shown to be composed of spinal cord tissue. Conversely, midbrain/hindbrain tissue, brain-specific neural stem cells and their respective neurons were only detected in epithelialized embryoids (Supplementary Fig. 12e). These results demonstrate that the epithelial architecture of initial EPI aggregates has a direct impact on embryonic developmental trajectories and promotes neural differentiation in embryoids.

Discussion
We have developed a novel bioengineered embryoid culture system for the study of early mouse postimplantation development that combines aggregates of ESCs and TSCs. Unlike previous studies 11,12,47 , this culture method allows simple independent modulation of the embryonic and extraembryonic compartments and makes our system both experimentally tractable and efficient for quantitative and mechanistic studies. A significant finding of our study is the effect that an epithelium has on EpiTS embryoid development. Under the same culture conditions, T/Bra expression is much more restricted in epithelialized than in non-epithelialized EPI aggregates. This suggest that the formation of an epithelium raises the threshold for WNT and TGF-β signal response such that, after the onset of T/Bra expression at one end, the probability that a second event happens in the aggregate end is low and this contributes to the restricted patterning of the EPI. The culture conditions we use here sensitize the EPI compartment to T/Bra expression which, in non-epithelialized aggregates, initially spreads throughout the aggregate.

The controlled interaction between EPI and TSC aggregates leads to
However, in the original gastruloid protocol, ESC aggregates undergo spontaneous symmetry breaking in the absence of any extraembryonic signals, localized to one end of the aggregate. Altogether, these observations suggest that one role of the extraembryonic ectoderm is to bias the spontaneous symmetry breaking event and that the epithelial organization of the EPI contributes to this by raising the threshold for signal response. The mechanism for this is unclear at the moment but the junctional organization of signaling receptors in the epiblast 48,49 is likely to play a role. The situation we observe in EpiTS embryoid phenocopies aspects of the activity of the AVE where BMP, TGF-β and WNT signaling inhibitors secreted from the Anterior Visceral Endoderm (AVE) inhibit the spread of T/Bra expression, contributing to the anterior-posterior patterning of the embryo, suggesting that the epithelial organization of the epiblast contributes to this effect.
Long-term culture of EpiTS embryoids in growth factor-free medium resulted in axial elongation (originating from the T/Bra expressing pole) and subsequent multi-axial patterning. The extent of this process was dependent on the proportions of cell types generated as a function of the initial tissue architecture of the EPI aggregates. For example, we detected cardiac mesoderm and definitive endoderm derivatives almost exclusively in non-epithelialized embryoids, resembling post-occipital patterning in gastruloids 6 . In both cases the EPI compartment displayed expanded WNT and TGF-β signaling at early stages of culture, and this is likely to cause the observed loss of anterior neural progenitors. In contrast, epithelialized embryoids, although also exhibiting T/Bra expression initially, later are depleted for mesodermal derivatives and display cell fates associated with midbrain/hindbrain and, in some instances, spinal cord levels, supporting the importance of an epithelium for the formation of brain tissues 50,51 . The organization of epithelialized embryoids can be construed as an 'early embryo' in which the primitive streak is initiated but immediately truncated, as they lacked most of the mesendodermal derivatives. This could suggest that neural induction has taken place in these embryoids.
In support of this, late EpiTS embryoids had T/Bra+Foxa2+ tissue extending along the embryoids suggesting the existence of a notochord/anterior mesendoderm (AME)-like structure which in the embryo is responsible for the induction of the midbrain/hindbrain 52,53 .
A surprising feature of the emergence of brain structures in the EpiTS embryoids is that it occurs in the absence of an AVE which, in the embryo, is thought to be needed to preserve neural potential in the epiblast 23 . An earlier study has shown that gastruloids derived from a combination of XEN cells and ESCs trigger an epithelium 12 and develop midbrain/hindbrain tissue 52 . In the embryoids that lack extraembryonic endoderm cells, we do not observe any forebrain tissue, supporting this possibility. This could potentially be overcome by WNT inhibition at later stages of the culture, since Dkk1 secreted from the prechordal plate is known to be instructive for forebrain formation 53 . Together with our results, these observations suggest that the AVE might be involved in maintaining an intact epithelium in the anterior region of the epiblast 54 , and that this might be sufficient to preserve the neural potential which is later developed by the AME during neural induction.
Our experiments show how bioengineering-inspired approaches can be exploited to standardize complex embryoid models and allow for the modeling of embryonic-extraembryonic cell/tissue interactions involved in the generation of precise patterns during development. EpiTS embryoids should serve as a canvas to generate developmentally relevant organ-specific structures of either neural or mesendoderm origin, through an in vivo-like gastrulation process. Traditional approaches to identify tissue origins in the embryo have relied largely on genetic manipulations which are labor intensive.
EpiTS embryoids thus provide an alternative approach to decipher the roles of mechanics and gene expression, which is difficult to perform in vivo. Ultimately, the adoption of our approach to generate human embryoids holds significant potential to shed light on early human development as experimenting with early human embryos is challenging.  -test (c,l) or multiple t-tests followed by Holm-Sidak multiple comparison test (j,m) were performed. Following P-value style was used: P****<0.0001, P***<0.0002, P**<0.0021, P*<0.0332.    conditions in (a-e), number of mCherry-positive embryoids over total number of embryoids analyzed are indicated at bottom right of (a-b). Large symbols indicate mean values of each replicate. Black lines indicate median and quartiles. For all statistical analysis, one-way ANOVA followed by Tukey multiple comparison test was performed. Following P-value style was used: P****<0.0001, P***<0.0002, P**<0.0021, P*<0.0332. Scale bars: 200µm. was prepared. 35µl of the suspension was added dropwise on microwell arrays to have 100-150 cells/well. Seeding was done at 37 0 C for 15 minutes. Growth factor reduced Matrigel (#356231) was diluted in cold EPIdiff or TSdiff to 3% or 2% (v/v), respectively. The medium was vortexed and 1ml was slowly added from the side of the well, avoiding direct addition from the top of the microwell arrays.

Main Figures
Plates were kept at 37 0 C in 5% CO2 for at least 72 hours before further processing.
For transwell experiments, EPI aggregates were placed in 190µl to low adherent U-bottom 96 well plates and 1, 3 or 5 TSC aggregates were placed in transwells on top (#CLS3374). 40µl medium was added to cover TSC aggregates. At 96h, 180µl of medium was replaced with fresh N2B27.

Immunostaining and confocal microscopy
EpiTS embryoids were washed with PBS and fixed with 4% PFA for 2 hours at 4 0 C. PFA was removed by three serial washes of 20 minutes at room temperature. Blocking was performed in blocking solution (PBS+10%FBS+0.3% Triton-X) for 1 hour at room temperature. Primary antibodies (see supplementary  table 1) were incubated for at least 24 hours at 4 0 C in blocking solution. Next day, primary antibodies were removed by three serial washes of 20 minutes at room temperature. Secondary antibodies were incubated for 24 hours and next day embryoids were washed and mounted on glass slides in mounting medium. Confocal images were taken using an LSM700 inverted (Zeiss) with EC Plan-Neofluar 10x/0.30 or Plan-Apochromat 20x/0.80 air objectives.

Image analysis
All images were processed using algorithms developed in Image J (version 2.0.0-rc-69/1.52n).
Brightfield, GFP (for TS cells), and mCherry (for T/Bra) channels were used as input. Thresholding and segmentation was performed sequentially for each channel. Area of EPI domain was calculated by subtraction of the area of the object identified in GFP channel from the area of the object identified in brightfield channel. For the calculation of the anterior EPI domain, area of the object identified in mCherry channel was subtracted from the EPI area. T/Bra coverage index was calculated by dividing area of the object identified in mCherry channel to the EPI area. For morphology measurements, brightfield images were thresholded and segmented. Maximum inscribed circle function was used to fit circles in the identified object. Axial length was determined by connecting centers of the fit circles. Elongation index was calculated by dividing axial length to the diameter of the maximum inscribed circle. All image analysis codes are available upon request.

RNA isolation and qPCR
RNA was extracted with the RNeasy Micro kit (QIAGEN), according to manufacturer's instructions and quantified with a spectrophotometer. 1 μg of RNA was reverse transcribed with the iScript cDNA Supermix (#1708890). cDNA was diluted 1:5 and amplified by using Power SYBR Green PCR Master Mix. qPCR was run with a 7900HT Fast PCR machine (#4329001), using Power SYBR Green PCR Master Mix (Applied Biosystems), with an annealing temperature of 60°C. Gene expression was normalized on Gapdh expression. Relative fold expression was calculated with the 2−ΔΔCT method. Table 2.

Bulk RNA-seq
EpiTS embryoids were lysed with 200 µl trizol, followed by addition of 70 µl of chloroform to trigger phase separation and then the aqueous phase was collected. The extraction process was repeated a second time, and an equal volume of isopropanol was added to precipitate the RNA, which was collected by centrifugation at 20000 g for 30 min. The pellet was washed with 15 mM sodium acetate in aqueous 70% ethanol, followed by salt-free 70% ethanol, before picking up in RNAse free water. RNA quantity and quality were assessed on nanodrop, qubit, and Agilent TapeStation 4200 profiling, and showed absorbance ratios 260/280 of 1.85 ± 0.12 and RNA integrity numbers (RIN) of 9.9 ± 0.2 (average ± SD), supporting good purity and absence of degradation. TruSeq stranded mRNA LT libraries were prepared according to Illumina protocol 15031047 Rev. E, starting from 300 ng of RNA, quantified by qubit DNA HS, profiled on TapeStation 4200, and sequenced on an Illumina HiSeq 4000 at a targeted depth of 36 Mreads/sample and paired-end read length of 81,8i,8i,81. The reads were trimmed for adapters with bcl2fastq v2.20.0, aligned to the mouse genome mm10 with STAR 2.7.0e, and a count matrix was assembled using the cellranger v4.0 curation of ENSEMBL annotations. In the manuscript, "Gene expression" refers to natural logarithm of counts per million for bulk RNA-seq data. The data was collected from 4 independent experiments.

Single-cell RNA-seq
Single cell RNA-seq of EpiTS embryoids was performed with 10x single cell 3' gene expression reagent kit chemistry v3.1, according to the protocol CG000204 Rev D. Library QC was performed on an Agilent TapeStation 4200, and sequencing was performed on Illumina HiSeq 4000, with a paired-end read length of 28, 8i, 110, targeting approximately 3000 cells and 300 Mreads per library (which included two replicates per condition, 4 timepoints, epithelialized and non-epithelialized embryoids, for a total of 16 libraries). Recovered cells were 3300 ± 900, and reads per cell after UMI collapse of 29 000 ± 4500 (average ± SD). The reads were trimmed for adapters and polyA with cutadapt 57 v2.1, and aligned to mm10 with cellranger 3.1. The count matrices were imported in R v3.6 with Seurat 58 v3.1.
Live cells were selected based on detection of more than 2000 genes, and mitochondrial gene content of 1.5 to 15%, and counts were normalized as natural logarithm of counts normalized to 10k per cell (referred to as "Gene expression" in the manuscript). A mini-analysis including dimensionality reduction by PCA based on centered and scaled most variable genes and cell type annotation transfer from an in vivo atlas 59 with scmap 60 was performed on individual datasets for the needs of doublet removal with DoubletFinder v2.0.3 61 . After data filtering, all datasets were merged, variable genes, scaling and PCA repeated, and batch correction was then done with Harmony v1 62 . A first Louvain clustering yielded clusters that were assigned to germ layers. UMAPs 63 were computed with uwot v0.1.5, first with 100 cells per cluster and spectral initialization, and then with all cells, initialized on the previous cluster centroids plus noise. The data was then split by germ layers, and a refined Louvain clustering and UMAP view was computed for each subset. Louvain clusters were finally assigned to cell types based on canonical markers. Custom code was used for visualizing cell type proportions vs day and/or epithelialization, normalizing for heterogeneous dataset sizes to avoid bias. A further closeup was similarly performed on the neurons of the epiblast+ectoderm subset. Splicing data was estimated with velocyto 64 v0.17, and RNA-velocity was computed with scVelo v0.2.1 65 in deterministic mode, using python v3.7.2 and scanpy v1.5 66 .