Generation of a ciliary margin-like stem cell niche from self-organizing human retinal tissue

In the developing neural retina (NR), multipotent stem cells within the ciliary margin (CM) contribute to de novo retinal tissue growth. We recently reported the ability of human embryonic stem cells (hESCs) to self-organize stratified NR using a three-dimensional culture technique. Here we report the emergence of CM-like stem cell niches within human retinal tissue. First, we developed a culture method for selective NR differentiation by timed BMP4 treatment. We then found that inhibiting GSK3 and FGFR induced the transition from NR tissue to retinal pigment epithelium (RPE), and that removing this inhibition facilitated the reversion of this RPE-like tissue back to the NR fate. This step-wise induction-reversal method generated tissue aggregates with RPE at the margin of central-peripherally polarized NR. We demonstrate that the NR–RPE boundary tissue further self-organizes a niche for CM stem cells that functions to expand the NR peripherally by de novo progenitor generation. The ciliary margin of the eye functions as a source of multipotent progenitor cells in certain organisms but whether it plays this role in humans has not been easy to study. Here the authors culture human embryonic stem cells that self-organize into retinal tissue, and show that ciliary margin-like growth zones emerge from the developing human retinal tissue and contain stem cell niches.

T he retina is the main visual sensory tissue in mammals. Its anlage, the optic cup, is derived from the rostral diencephalon and is composed of the inner and outer walls-the neural retina (NR) and retinal pigment epithelium (RPE), respectively 1,2 . Mammalian NR is known to possess a low regenerative capacity, and spontaneous tissue recovery hardly occurs following substantial damage caused by trauma or degenerative diseases. This stands in sharp contrast to the newt NR, for instance, which can fully regenerate from neighbouring RPE even after total resection (reviewed in refs 3,4).
In vertebrate retinogenesis, the NR has two different groups of progenitors present in distinct locations. During early retinal development, epithelial progenitors are widely present in the NR epithelium and function to generate photoreceptors and other types of retinal neurons 5 . Their differentiation capacity gradually changes as retinogenesis proceeds; they have a tendency to produce ganglion cells and cones at early stages, and then rods and bipolar cells at late stages 6,7 . Like cortical progenitors, these NR progenitors show the radial glia-like morphology with thin apical and basal processes, characteristic of neuroepithelial progenitors, and they undergo interkinetic nuclear migration during mitotic cycles 8,9 . (Hereafter, these progenitors are referred to as 'NR progenitors').
Another kind of progenitor is located at the peripheral margin of the NR, the ciliary margin (CM) 5,10,11 . The teleost (fish) eye contains an active stem cell system in the CM zone (also called circumferential germinal zone) that contributes to NR expansion by de novo generation of the tissue even in adulthood 12 . This type of stem cell robustly self-renews and gives rise to all NR cell types in the extending NR periphery, unlike the nonmarginal-NR progenitor, whose developmental competence gradually becomes confined to the generation of limited cell types during retinogenesis as mentioned above. In the chick embryo, the stem cell system at the CM also plays an active role during eye development, and stem cells isolated from this region can form NR tissue in floating aggregate culture 13 . The mammalian foetal eye also contains a peripheral NR zone resembling the chick embryonic CM in molecular marker expression, and this zone contains a substantial number of stem cells 10,14,15 (Hereafter, these cells are referred to as 'CM stem cells'). The presence of a small number of retinal stem cells at the ciliary region in the adult mammalian eye has also been reported [16][17][18][19][20] , although the extent of their in vivo contribution to the maintenance and regeneration of adult retinal tissue remains elusive.
Recent advancement in stem cell research has enabled in vitro differentiation of retinal progenitors and their derivatives from pluripotent stem cells [21][22][23][24][25][26][27][28][29] . Moreover, mouse and human embryonic stem cells (hESC) aggregate have been shown to possess the potential to self-form optic cups in three-dimensional (3D) culture when retinal differentiation efficiency is sufficiently high [30][31][32] . Notably, the ESC-derived NR self-organizes the formation of multiple retinal layers reminiscent of the postnatal retina. NR progenitors in this culture exhibit radial glia-like epithelial morphology, frequently proliferate with interkinetic nuclear migration and give birth to photoreceptors and other retinal neurons in a stage-dependent manner 30,32 , a process which largely recapitulates the activity of NR progenitors in vivo. Moreover, a recent report using mouse ES cells showed that self-forming NR tissue can give rise to functional photoreceptors capable of light response after grafting in the murine eye 33 .
In the self-organized optic cup, the boundary domain between NR and RPE, called the hinge domain, contains cells with a morphology distinct from that of NR and RPE cells. However, whether CM-type stem cells are generated at this region in longterm culture has remained unexplored. In the present study, we have uncovered the formation of CM-like zones in self-organizing retinal culture. First, we made substantial improvements to the 3D culture method of hESCs for promoting retinal differentiation. As reported previously, the formation of optic cup in hESC culture is less frequent (B10% of aggregates) 32 than in mouse ESC culture. Its efficacy is also influenced by minute changes in culture conditions. In particular, our previous culture included Matrigel, a crude extract of extracellular matrix from mouse EHS sarcoma cells, for promotion of retinal differentiation and epithelialization 32 . The use of Matrigel is problematic because this undefined extract shows substantial batch-to-batch differences in promoting retinal differentiation. Furthermore, the use of animal-derived matrix proteins is also unfavourable from the viewpoint of future clinical applications.
Therefore, we established culture conditions for inducing selective retinal differentiation without the use of Matrigel. Using this new 3D culture technique, we further developed a two-step method that allows a reproducible formation of NR-RPE junctions to generate CM-like zones. Combining these new methods, we demonstrate that CM-like zones self-organize in vitro and function as growth zones in the periphery of hESC-derived NR.

Results
Highly selective self-formation of NR with BMP treatment. We observed in a separate study of cortical self-organization that hESC aggregates selectively differentiated into Foxg1 þ telencephalic neuroepithelium when cultured in growth factor-free chemically defined medium (gfCDM) without Matrigel (Fig. 1a). While retinal differentiation was rarely seen under these conditions, we reasoned that the addition of certain signalling molecules to the medium might cause a shift of positional information from the telencephalic to the retinal direction, given the proximity of their anlages in the embryonic brain.
We therefore examined the effects of several candidate signals that are known to affect retinal differentiation (see Methods for details). Using Venus-knock-in reporter hESCs for Rx (retinal marker) 32,34 expression, we observed that a transient treatment of hESC aggregates with a low concentration of Bone morphogenetic protein-4 (BMP4; starting at 1.5 nM on day 6 and subject to half-dilution every 3rd day) dramatically increased Rx::Venus expression ( Fig. 1b- Supplementary Fig. 1a), reminiscent of the expression profile of NR epithelium. This NR epithelium reproducibly contained phospho-Histone H3 þ mitotic cells on the outer surface, which accumulated the apical marker aPKC (Fig. 1j,k; also showing strong N-cadherin localization there). The basement membrane positive for Laminin and Collagen IV ( Fig. 1l and Supplementary Fig. 1b) spontaneously formed on the luminal side in this culture without the addition of exogenous matrix proteins such as Matrigel. The induction efficacy of Rx::AcGFP (a separately derived reporter line) exceeded 80% at the cellular level ( Fig. 1m-p and Supplementary Fig. 1c,d), which is substantially higher than the efficiency seen in previous Matrigel-based cultures 32  and thick (Fig. 1q), and contained cells expressing NR cell typespecific markers such as Brn3 and Pax6 for ganglion cells (Fig. 1r,s), Crx for photoreceptors (Fig. 1t), Calretinin for ganglion cells and interneurons ( Supplementary Fig. 1e), and Ptf1a for interneurons ( Supplementary Fig. 1f) in addition to Chx10 þ NR progenitors (Fig. 1r). Taken together, these observations show that selective NR differentiation from hESCs without the use of Matrigel is enabled simply by a transient early BMP4 treatment. The strong promotion of NR differentiation by BMP4 treatment was rather unexpected, since the inductive role of BMP signals is known for RPE specification 36 but not for early NR differentiation. The BMP4 treatment was transiently given at the early phase of culture, while the NR-RPE specification occurs later around day 18-20 and optic cup formation around day 24 (ref. 32). Therefore, we speculate that this effect could be relevant to the role of BMP signals in early forebrain patterning (this point is discussed later).
Generation of RPE next to NR by induction-reversal culture. In the embryonic eye, the CM arises from the hinge region joining the NR and RPE. Although the BMP-induced retinal differentiation culture was highly efficient, the retinal epithelium mostly consisted of NR (Fig. 1f), lacking RPE or a NR-RPE junction (Fig. 1i). This is consistent with the idea that NR differentiation is the default direction of retinal differentiation in ESC culture 30 . Our previous hESC study 32 showed that RPE differentiation could be induced from hESC-derived retinal epithelium by treating with canonical Wnt signals from day 18 (RPE-inducing roles of Wnt signals in vivo are also reported and discussed in refs [37][38][39][40]. Conversely, FGF signals have been shown to promote NR differentiation and suppress RPE differentiation 41,42 . Based on these effects (summarized in Supplementary Fig. 2), we next sought to induce RPE differentiation in a limited domain within the hESC-derived retinal epithelium.
When BMP-induced Chx10 þ retinal epithelium was treated with the GSK3 inhibitor CHIR99021 (Wnt agonist) and the FGFR inhibitor SU5402 for 6 days from day 18 (Fig. 2a, condition 2), the epithelium became thin and expressed Mitf but not Chx10 on day 24 (these RPE features were found in 490% of aggregates), whereas most of the nontreated epithelium (Fig. 2a, condition 1) remained thick and Chx10 þ /Mitf À (Fig. 2b). Rx::Venus expression, which was initially strong on day 18, became dim in the epithelium after treatment with GSK3 and FGFR inhibitors (see below; Rx expression is much weaker in RPE than in NR). Thus, Wnt signal augmentation and FGF signal inhibition for a relatively short period was sufficient to induce RPE-type epithelium from Chx10 þ NR progenitors.
Interestingly, the fate of the induced Mitf þ epithelial cells was not irreversibly fixed even after they exhibited these molecular and morphological characteristics of RPE (Fig. 2b). When hESCderived retinal epithelium was treated with CHIR99021 and SU5402 (days [18][19][20][21][22][23][24] and subsequently cultured in NRdifferentiation medium (not containing these inhibitors) from day 24 (Fig. 2a, condition 3), the majority of epithelium (in B70% of aggregates) repressed Mitf expression in their major portions, regained Chx10 expression and increased the epithelial thickness ( Fig. 2b-e). This stands in contrast to the thin Mitf1 þ epithelium cultured continuously in the presence of the CHIR99021 (condition 2; Fig. 2a,b). Live imaging analysis (Supplementary Movies 1 and 2, Fig. 2f-i) showed that the recovery of the Rx::Venus-strong thick epithelium (NR type) in the aggregates did not occur by rapid local regrowth of residual NR-type epithelium but by the transition from Rx::Venus-dim thin epithelium (RPE type) to Rx::Venus-strong thick epithelium. In fact, recovery of epithelial thickness and Rx::Venus levels was already seen within the first few days after removal of GSK3 and FGFR inhibitors (Fig. 2f-i). This suggests that retinal differentiation, even after the appearance of early RPE features, can be reversed and redirected towards the NR fate simply by culturing the tissue in NR-differentiation medium. Hereafter, we refer to this step-wise culture (RPE induction followed by reversal to NR epithelium) as 'induction-reversal' culture (its possible mechanism is interpreted later in Discussion).
On day 35, in contrast to NR-and RPE-inducing conditions 1 and 2 (Fig. 3a,b), B40% of retinal aggregates in the inductionreversal culture (condition 3) exhibited a characteristic twodomain pattern: a large semispherical domain of continuous NR epithelium (with strong Rx::Venus expression) with a small pigmented domain (or a few small domains) of thin and winding RPE in an adjacent location (Fig. 3c). This type of aggregates (two-domain aggregates, hereafter) could be easily selected (even without using microscope) for further culture (Fig. 3d at day 60). Figure 3e shows quantitative comparison of two-domain aggregate generation (shown in orange) between conditions 1, 2 and 3 at day 35.
Immunostaining on day 50 showed that the two-domain aggregate contained continuous NR epithelium positive for Rx::Venus and Chx10 (Fig. 3f). Notably, the NR epithelium exhibited a gradual change in shape along the proximal-distal axis relative to RPE; the closer to RPE, the thinner the NR epithelium. In addition, molecular marker distribution exhibited an intriguing graded pattern. During in vivo development, differentiation of fate-specified retinal cells starts earlier in the central domain and is delayed in the peripheral portion 43,44 . For instance, Pax6 is initially expressed at high levels in all Chx10 þ NR progenitors. However, as neurogenesis begins in the NR, Pax6 expression becomes attenuated in progenitors except in the peripheral domain, and becomes strong in postmitotic ganglion cells (Pax6 þ /Chx10 À ) in the central domain. In hESC-derived retina, strong expression of Pax6 in Chx10 þ progenitors was seen preferentially in the NR margin next to the NR-RPE junction ( Fig. 3f,g; arrow, junction; bracket, NR margin), whereas Pax6 þ /Chx10 À ganglion cells were mostly found in the nonmarginal portion of the NR (Fig. 3g, arrowheads). Consistent with this idea, the nonmarginal NR portion also contained more cells positive for the postmitotic neuronal markers Brn3, TuJ1 and Calretinin ( Fig. 3h-j) than the marginal NR portion. Moreover, differentiated photoreceptors marked by Blimp1, Recoverin or NRL (Fig. 3k-n; Supplementary  Fig. 3a,b, see Methods) were preferentially located in the NR epithelium at a distance from the NR-RPE junction, indicating that differentiation of NR is more advanced in the nonmarginal NR portion (hereafter, the 'marginal'-'nonmarginal' axis is referred to as the peripheral-central axis, in analogy to the in vivo situation).
Collectively, these observations demonstrate that our 3D hESC culture, in which selective retinal differentiation and the induction-reversal process are combined, enables the adjacent formation of NR and RPE tissues. In the two-domain retinal structure generated by this culture, the continuity and integrity of NR epithelium was more efficiently maintained in long-term culture. While the aggregates carrying only NR (those generated typically in condition 1) tended to be discontinuous and break into small rosettes in culture beyond day 60 ( Supplementary  Fig. 3c, left), the NR in the two-domain aggregates (condition 3) more frequently contained continuous epithelial structures ( Supplementary Fig. 3c,d).
The continuous NR exhibited a graded regional difference (or polarity) along the central-peripheral axis. In addition, this NR  Condition (2) Condition (  margin that was adjacent to RPE had a characteristic 'tapering' morphology of epithelium, which is also seen in the periphery of the foetal NR 10,45 . Therefore, we became intrigued by the possibility that this NR-RPE junction may have some similarity to the embryonic CM, and thus examined this region for the expression of CM-characteristic molecular markers. Expression of CM markers in the self-organized retina. In hESC-derived retina (Fig. 4a-h; day 63), the NR-RPE junctional region expressed a set of regional markers characteristic for the embryonic CM in a complex but spatially controlled manner (refs 45-49 and Supplementary Fig. 4a). The thin Mitf þ epithelium (Fig. 4a) near the junction expressed the ciliary epithelium marker Aquaporin1 (Aqp1; Fig. 4b) and the CM marker Otx1 (Fig. 4c). In contrast, the CM markers Zic1, Rdh10 and SSEA1 (Fig. 4d,f,g and Supplementary Fig. 4b-d), as well as the NR marker Sox2 (Fig. 4f) were expressed only in the Chx10 þ domain (Fig. 4e). Interestingly, Aqp1 and Otx1 were expressed across the junction, that is, in both Mitf þ and Chx10 þ epithelia (Fig. 4b,c). The tapered NR epithelium, showing a gradual increase in thickness going distally from the thin Mitf þ epithelium, appeared to have two subdomains. The Chx10 þ NR margin directly next to the thin Mitf þ epithelium (hereafter, subdomain 1 (sd 1)) expressed CM markers such as Otx1, SSEA1, Zic1 and Rdh10, but not fate-specific NR cell markers such as TuJ1 and Crx::Venus (Fig. 4c-h   Sox2 þ progenitors (Fig. 4e,f), but also the fate-specific NR cell markers (Fig. 4g,h), particularly on its distal side. Conversely, no expression of Otx1, SSEA1, Zic1 and Rdh10 was observed in this region. While both sd 1 and sd 2 in the tapering proximal NR contained Ki67 þ mitotic cells (Fig. 4i,j), the two subdomains differed in the expression of CDK inhibitors (p21, p27 and p57), which positively control cell cycle exit 50,51 ; sd 1 was mostly negative for p27 and p57, while sd 2 contained substantial numbers of cells positive for these CDK inhibitors (Fig. 4k,l). Expression of p21 was observed in neither subdomain (Supplementary Fig. 4e-g).
Together, these findings (summarized in Fig. 4m) indicate that the tapered epithelium portion adjacent to the NR-RPE junction, sd 1, has an expression profile resembling that of the in vivo CM zone.  ARTICLE Sphere-forming progenitors from CM-like zones. We next examined whether the hESC-derived CM-like region, like the foetal CM, contains sphere-forming progenitors (Fig. 5a). We manually dissected nonpigmented Rx::Venus-strong tissues in the tapered NR margin (CM-like tissues, hereafter) from hESCderived retina (day 60-85), dissociated them, and then subjected them to retinosphere culture (1,000 cells per well) 17,52,53 . For comparison, we dissected nonmarginal NR (central NR) tissues from the same aggregates. While Rx::Venus þ primary retinospheres formed from both CM-like and central-NR groups (Fig. 5b,c; note that non-CMtype NR progenitors were also present in the central-NR tissues), the efficacy was higher with cells from CM-like tissues (Fig. 5d and Supplementary Fig. 5a). Secondary sphere formation (1,000 cells per well) also showed a clear difference in efficacy ( Fig. 5e; central-NR group oCM-like group). Furthermore, the sizes of both primary and secondary spheres from CM-like tissue were significantly larger than those from central NR (Fig. 5b,f,g and Supplementary Fig. 5b). The majority of cells in the primary and secondary retinospheres from CM-like tissues were Chx10 þ progenitors (Fig. 5c). These findings suggested that the CM-like tissue contains self-renewing progenitors with a robust proliferative capacity.
To confirm that sd 1 cells in the CM-like tissue have the sphereforming ability, we next examined the activity of the SSEA1 þ cells purified from hESC-derived CM-like tissues (Fig. 5h-j and Supplementary Fig. 5c,d). SSEA1 þ cells (purified by fluorescentactivated cell sorting (FACS) or magnetic-activated CS (MACS)) formed retinospheres (Fig. 5h) more frequently than did SSEA1 À cells ( Fig. 5i; MACS tended to have a better cell viability than FACS; Supplementary Fig. 5d). In addition, spheres from SSEA1 þ cells were substantially larger in size than those from SSEA1 À cells (Fig. 5j). Although SSEA1 À cells may still have the potential to form retinospheres, these findings indicate that SSEA1 þ cells play a major role in sphere formation from CM-like tissues.
When cultured under differentiation conditions (adhesion culture) for 2 weeks, retinospheres from CM-like tissues (both primary and secondary) produced cells expressing fate-specific retinal cell markers such as Crx/Blimp1 (photoreceptors), TuJ1/ Pax6 (ganglion cells) and Calretinin (ganglion cells and interneurons; Fig. 5k and Supplementary Fig. 5e). The generation of fate-specific retinal cells from spheres was also observed with SSEA1-sorted cells after 2-week culture ( Supplementary Fig. 5f), while few differentiated cells were seen initially. Taken together, these observations indicate that the NR-RPE junction zone of hESC-derived retina not only has a characteristic CM expression profile, but also contains sphere-forming progenitors resembling CM stem cells.
De novo NR generation at hESC-derived CM-like zones. We next investigated the progenitor dynamics at the CM-like zone of hESC-derived retina, focusing on de novo generation of NR tissue at its margin. To this end, we performed pulse-labelling experiments using nucleotide analogues (BrdU and EdU; Fig. 6). One-day exposure of hESC-derived retina to BrdU heavily labelled most cells in sd 1 (Fig. 6a) and a majority of cells in sd 2. These cells expressed the mitotic marker Ki67 (Fig. 6a, top). When the pulse-labelled tissues were cultured without BrdU for an additional 6 days, the cells in sd 1 rarely retained BrdU labelling (Fig. 6b, sd 1), suggesting that sd 1 mainly contains actively proliferating progenitors in which BrdU was diluted by cell division (the loss of BrdU þ cells was unlikely to be caused by selective cell death, since this subdomain showed no substantial accumulation of cleaved caspase 3, a marker for apoptosis; Supplementary Fig. 6a).
In contrast, BrdU-labelled Ki67-negative cells, representing newly born postmitotic cells, appeared in sd 2 (red cells in Fig. 6b,  sd 2). Next, we cultured the tissue for 13 days after 1-day BrdU labelling and then treated with EdU for 1 day (Fig. 6c). In this case too, EdU uptake was seen in most Ki67 þ cells in both sd 1 and sd 2. In contrast, BrdU label-retaining cells were found in the central side of sd 2, but rarely in sd 1 or the peripheral side of sd 2 (Fig. 6c).
After the tissue was further cultured for an addition 13 days, BrdU-retaining cells were found in the thick portion of the NR but rarely in sd 1 and sd 2, while EdU-retaining cells were seen also in sd 2 (Fig. 6d). These EdU-retaining cells expressed fate-specific markers such as Brn3, Crx and Calretinin ( Supplementary Fig. 6b-d).
Collectively, these findings demonstrate dynamic generation and differentiation of progenitors in distinct subdomains around the CM-like zone of the hESC-derived retina (Fig. 6e). In sd 1, the epithelium consists mostly of actively proliferating Ki67 þ cells and the nucleotide analogue labels in these cells are diluted fast by continuous cell division, suggesting that this is a place for the de novo production of progenitors from stem cells rather than a site for generation of postmitotic cells. In contrast, sd 2 have Ki67 þ cells predominantly on the apical side and gradually accumulates Ki67 À postmitotic cells on the basal side as it goes farther from the NR-RPE junction, suggesting that postmitotic cells are generated in this subdomain, unlike sd 1. This idea is consistent with the observation that the CDK inhibitors p27 and p57 are expressed in sd 2 but rarely expressed in sd 1 (Fig. 4k,l and Supplementary Fig. 4b-d). Within sd 2, label-retaining cells gradually disappear from the peripheral to central direction during a time course of 1-4 weeks, such that cells labelled earlier are found more distally from the NR-RPE junction. Taken together with the observation that sphere-forming SSEA1 þ cells are densely present in sd 1 (Figs 4g and 5i), these findings indicate that the hESC-derived CM-like tissue constitutes a growth zone for de novo generation of retinal progenitors at the NR-RPE boundary in self-organizing retinal culture (Fig. 7).

Discussion
A long-standing question regarding the formation of the human eye is how new cells-both progenitors and postmitotic-are added to the rapidly expanding NR during foetal development. Previous studies in nonhuman model systems, have suggested that the CM of the eye can in certain situations function as a source of multipotent progenitor cells. Determining whether this is the case in human foetal eyes, however, has been difficult due to the limited number of experimental models amenable to investigating human NR growth, expansion and cell contribution.
In this study, to overcome this limitation, we have successfully applied the self-organization culture approach of human ESCs to recapitulate CM development in vitro. Using induction-reversal culture combined with BMP4-triggered highly selective retinal differentiation, we demonstrated two aspects of retinal selforganization in 3D hESC culture (Fig. 7): central-peripheral selfpatterning of NR and self-formation of CM-like growth zones. These two aspects may be closely related: the CM-like growth zone formation occurs in the tapered portion of the NR, which resembles many of the characteristics of the peripheral domain of the embryonic NR.
Thus, the presence of tissue interactions between the NR and RPE appear to foster the self-organization of stem cell nichecontaining CM-like growth zones. Taken together with our previous studies 30,32 , we now demonstrate that self-organizing retinal culture of hESCs is capable of not only generating highly ordered epithelial structures, but also orchestrating bona fide In induction-reversal culture, GSK3 and FGFR inhibitorinduced thin RPE tissue (Rx::Venus-dim) exhibited a transition to Rx::Venus-strong thick NR epithelium within a few days after exposing RPE tissue to NR-differentiation medium ( Supplementary Movies 1 and 2). A similar induction-reversal phenomenon is also seen when only the GSK3 inhibitor was applied. However, the further addition of the FGFR inhibitor (which by itself had only weak effects if any) made RPE  Live imaging indicated that the reappearance of NR-type epithelium in reversal culture is caused by conversion of RPE-type epithelium rather than by rapid expansion of residual NR-type epithelium in the aggregate. In addition, this 'RPE-to-NR' conversion, triggered only by a change in extracellular signals, can occur even after the onset of Mitf expression (bona fide RPE marker) and the repression of Chx10 expression (bona fide NR marker). These observations are consistent with previous studies in vivo showed a profound plasticity between the NR and RPE fates during development 41,42,54 . Interestingly, the temporal window of competence for RPE-NR reversal may be limited in our hESC culture; preliminary experiments have indicated that a longer treatment of GSK3 and FGFR inhibitors (for example, days 18-27 or 18-30, instead of days 18-24) makes RPE less responsive to reversal signals in NR-differentiation medium. This suggests that the length of RPE-specifying culture (which may deepen RPE commitment) affects the reversion efficacy. In addition, an intriguing report has described that a subpopulation of adult human RPE cells, on activation, gain competence to become multipotent stem cells 55 , suggesting that RPE cells, even in adult humans, maintain special state of epigenetic control. In future investigations, our culture method may be useful in elucidating molecular mechanisms for the stage-and species-specific control of retina transdifferentiation competence.
In our hESC culture, some variation was observed in the response to the reversal trigger at the tissue level. In B40% of aggregates, RPE and NR were co-generated (two-domain aggregates) and CM-like zones self-formed between them. In B30% of aggregates, most of epithelia became entirely NR after reversal, whereas 25-30% of aggregates were resistant to the reversal trigger and remained RPE dominant.
The cause for this heterogeneity in responsiveness is an intriguing question. We infer that it may be relevant to a 'bistability' control, which tends to give a 'black-or-white' (not grey) outcome (reviewed in ref. 56). In the signalling network for retinal specification (Supplementary Fig. 2), RPE differentiation is promoted by Wnt signals, and, once the retinal tissue becomes RPE, it enhances local Wnt signals by inducing endogenous expression of Wnts and Mitf 57,58 . Conversely, NR differentiation is suppressed by Wnt signals [37][38][39][40] . Moreover, differentiating NR shuts off local Wnt signals by attenuating Wnt ligand production by the transcription factor Six3 (ref. 40) and secreting Wnt antagonists such as Dkk1 (refs 30,59). Furthermore, specification into NR and RPE is promoted and suppressed by FGF signals, respectively, which are produced by NR 41 . These multiplex regulations of mutual inhibition and autoactivation lead to a 'bistable' system, in which a small difference in endogenous signals could make a pivotal specification of retinal epithelium (or its local domains) into either NR or RPE. In the two-domain aggregate containing both NR and RPE, only a small portion of RPE is resistant to reversal signals promoting NR differentiation. We infer that a bias for the location of the RPE-prone domain  ARTICLE may be related to the co-presence of nonretinal neural tissues (representing B10% of cells; Fig. 1n,o), which was previously shown to promote RPE differentiation within the ESC-derived optic cup 30 . Consistent with this idea, in close visual examinations of two-domain hESC aggregates after inductionreversal culture (Fig. 3c,d), nonpigmented Rx::Venus À tissues (that is, nonretinal neural tissues) were not found within the continuous NR but preferentially located together with pigmented RPE. Another remaining question concerns how the boundary between RPE and NR is determined. In 'reversal' culture method, our preliminary observations suggest that thin RPE tissue (sometimes adjacent to nonretinal tissue) tends to resist the reversal stimuli, whereas the retinal epithelium distally (centrally) located from thin RPE tissue tends to exhibit a more rapid transition to strong Rx::Venus-expressing thick NR epithelium. This raises a possibility that some bias in the differentiation state has already existed in RPE-like tissue before starting the reversal culture method. This hypothesis will be more rigorously tested in future studies using RPE/NR double-labelled cell lines, which might allow a more detailed observation of the dynamics of NR-RPE formation.
The 'induction-reversal culture' method enables the cogeneration of NR and RPE in hESC culture. In this continuous structure, the NR exhibits a central-peripheral polarity. With regard to morphology, a graded change in the epithelial thickness is seen along the central-peripheral (thick-to-thin) axis of hESCderived NR. In the central portion, NR differentiation into fatespecific cells tends to be more advanced than in the peripheral portion, while immature progenitors are more prominent in the NR periphery, particularly in the tapering domain where de novo NR generation occurs.
This self-formation of the central-peripheral polarity appears to correlate with the presence of RPE adjacent to the NR, since such polarity is not obvious within NR epithelia in hESC aggregates without RPE (in the latter case, the entire NR tends towards the central-NR type with regard to differentiation and thickness; Fig. 1q). Moreover, when NR epithelium, by chance, is flanked by more than one RPE domain, a CM-like zone emerges in the NR domain next to each NR-RPE junction. It remains to be clarified whether the effect of RPE is mediated by soluble signals or spreads via planar induction. Another related question for future study is whether the effect of RPE on the centralperipheral polarity in NR is direct on the NR or secondary through the CM-like zone formation between them.
In the human foetal eye, the regional difference along the central-peripheral axis is quite remarkable. For instance, the generation of morphologically recognizable photoreceptors is observed around foetal week 10 at the central-most NR region, while photoreceptor differentiation in the noncentral region occurs very gradually over the following 15 weeks and slowly proceeds in a central-peripheral pattern 44,60 . The central-most NR domain eventually gives rise to the fovea, in which photoreceptors are mostly cones. The generation of the fovea-like structure (a pit structure selectively expressing colour opsins and not Rhodopsin) has not been observed thus far in our culture yet, and is an intriguing and challenging topic for future study. As described in our previous report 32 , both RXR-gamma þ cone precursors and Nrl þ rod precursors are generated in wide domains of hESC-derived NR. Therefore, a first step towards this challenging goal may be region-selective control of cone and rod generation along the central-peripheral axis of the self-organized NR.
In induction-reversal culture, the tapered peripheral end of the hESC-derived NR (sd 1), adjacent to the Mitf þ domain, formed a niche for SSEA1 þ stem-like cells that continuously produce progenitors. Cells in sd 1 were Ki67 þ and were efficiently labelled with nucleotide analogues after 1-day of exposure, while these labels in sd 1 (showing no substantial level of apoptosis) became undetectable within 6 days, indicating that these cells are under active proliferation. sd 1, unlike sd 2, contained few fatecommitted postmitotic cells and did not express CDK inhibitors (promoting cell cycle exit), suggesting that these actively proliferating cells in sd 1 were not directed to produce terminally differentiated postmitotic cells by cell cycle exit.
SSEA1 þ cells are also present in the CM-like zone of the mouse foetal retina. However, the strict identity of SSEA1 þ cells across species could be a matter of argument. For instance, during early mouse retinal development, SSEA1 is widely expressed throughout the optic vesicle. From E13.5 onwards, SSEA1 expression is mainly seen in the peripheral NR and pigmented outer layer of the ciliary epithelium 48 . However, within the mouse peripheral NR, the nonpigmented epithelium at the peripheral end (corresponding to sd 1) is devoid of SSEA1 expression, in contrast to our finding with its expression in the CM-like tissue generated from hESCs. Since early SSEA1 expression in the mouse retina dynamically changes in a stage-specific manner 48 , a careful stage-by-stage investigation would be necessary to examine whether SSEA1 þ cells in the mouse and human foetal retinae represent distinct (albeit related) subpopulations in the growing end of the tissue.
Clarke et al. recently reported the generation of 'retinal stemlike cells' from hESCs, but these stem-like cells may have critical differences from those generated in our study. Namely, the stemlike cells reported by Clarke et al. 53 were derived from a rare population within hESC-derived pigmented cells (stem cells from pigmented ciliary epithelial cells are also described in ref. 61). In contrast, the retinospheres in the present study are derived from retinal tissues from which pigmented tissues were removed. Moreover, the SSEA1 þ cell population did not show obvious pigmentation in the tissue section nor by visual inspection of dissociated immunostained cells (Supplementary Fig. 5g; they did not show strong side scattering in FACS analysis either; not shown). Thus, the stem cell population in the CM-like zone of our hESC culture (within the Chx10 þ peripheral NR) appears to be distinct from the one reported previously, which seems to be more closely related to RPE or pigmented ciliary epithelium. We infer that the 3D formation of retinal tissue with juxtaposed NR and RPE is vital for the generation of CM-type stem cells, which contribute to de novo NR extension in the peripheral direction.
Regarding human retinal stem cells, another set of previous studies have suggested the presence of Müller glia-type stem cells in the marginal zone of human adult retina 19,62 . The similarity and difference between these adult-type cells and CM-type stem cells in our hESC culture (presumably foetal type) is another intriguing topic for future investigation. In the human foetus, the CM appears by foetal week 10 and the ciliary body formation begins at the end of the 3rd foetal month 45 . We have so far failed to observe the morphological appearance of ciliary body-like structures in our hESC culture, whereas Mitf þ /Aqp1 þ epithelia (Fig. 4), reminiscent of ciliary epithelium, was readily generated.
While the present study focused on new aspects of selforganization of hESC-derived retina, the improved culture that is introduced here also serves as an advanced protocol towards future applications of human pluripotent stem cells. First, the improved culture has enabled a more selective retinal differentiation of hESCs (480%) by the use of timed BMP4 treatment. In our preliminary experiments, other BMPs such as BMP2, BMP7 and Gdf7 had similar inductive abilities. Moreover, unlike our and others' previous methods, this BMP-based hESC culture does not require the xenogenic matrix protein mixture Matrigel for the 3D formation of NR. The precise mode of BMP4 action in retinal specification remains an intriguing unanswered question. Previous studies have shown that BMP signals directly suppress the expression of Foxg1, which is essential for cortical development 63,64 . Therefore, one interpretation may be that BMP4 treatment during the early differentiation stage in our culture prevents hESC-derived forebrain neuroepithelium from adopting the telencephalic fate and redirects it into retinal progenitors. In the early embryo, BMP signals are expressed near the anterior neural ridge and in the prechordal mesoderm, and are implicated in the patterning of rostral diencephalon of the early embryo 65 , including hypothalamic (diencephalic) induction 66,67 .
In condition 1 of our new culture method, the continuous NR forms vesicles in apically convex manner, which resemble retinal tissues shown by Meyer et al, 2011 (ref. 27). The vesicles formed in our culture, however, do not undergo further morphogenesis (invagination) to form optic cup structures, since they lack RPE as the outer wall. Regarding the long-term culture, after 60 days, the NR in vesicles gradually looses its integrity and tends to accumulate neural rosettes (differentiation itself proceeds).
In the induction-reversal culture combined with BMP4induced retinal differentiation (condition 3), a portion of RPE tissue forms adjacent to NR (Fig. 3c,d). In this two-domain aggregate, NR is generated as a large continuous epithelium. NR can be also obtained from optic cups cultured in hESC aggregates under our previous conditions 32 . However, our new culture method has clear advantages in the reproducible generation of larger NR epithelia. First, the formation of optic cups in our previous culture was not very frequent (B10%). Second, NR epithelia in these optic cups were typically 500 mm in diameter around day 35. In contrast, NR forms as a fairly large domain (typically 1.3-1.5 mm in diameter at day 35; Fig. 3c) with induction-reversal culture, spanning a majority of the aggregate surface, and continuous NR epithelium grows robustly (for example, typically 2 mm at day 60; Fig. 3d).
Finally, the two-domain structure of NR epithelium has a characteristic thick and transparent appearance, and its boundary to RPE can be easily recognized because of the presence of pigmented tissue adjacent to it. This simple structure may facilitate the preparation of high-purity NR tissues in large quantities (much like removing the stem ends from radishes) and may prove useful in future studies examining optic tissue transplantation, pathogenesis and drug discovery.
Retinal differentiation from hESCs. For serum-free floating culture of embryoid body-like aggregates with quick reaggregation (SFEBq) culture, hESCs were dissociated into single cells in TrypLE Express (Gibco) containing 0.05 mg ml À 1 DNase I (Roche) and 20 mM Y-27632, and quickly re-aggregated using low-celladhesion 96-well plates with V-bottomed conical wells (Sumilon PrimeSurface plate; Sumitomo Bakelite) in 'differentiation medium' (12,000 cells per well, 100 ml) supplemented with 20 mM Y-27632 (ref. 68) under 5% CO 2 at 37°C. The differentiation medium is gfCDM 69 supplemented with 10% KSR, while gfCDM contains 45% Iscove's modified Dulbecco's medium (IMDM, Gibco), 45% Hams F12 (F12, Gibco), Glutamax, 1% chemically defined lipid concentrate (Gibco), monothioglycerol (450 mM, Sigma), 100 U ml À 1 penicillin and 100 mg ml À 1 streptomycin (Gibco). Defining the day on which the SFEBq culture was started as day 0, recombinant human BMP4 (R&D) was added to the culture to the final 1.5 nM (55 ng/ml) on day 6, and its concentration was diluted into half by half medium change every 3rd day. The addition of BMP4, started on day 3 and day 6, induced Rx þ NR epithelium at day 18 with similar efficiency. The number of initial plating of cells (12,000 cells per well) was optimized (in both efficiency and size), but Rx þ NR epithelium was also achieved using 9,000 cells per well.
To establish NR-differentiation culture from hESCs without the use of Matrigel, we examined the effects of the following signal modulators on the induction of Rx::Venus þ cells at various doses and time points: BMP agonists/antagonists (BMP4, dorsomorphin), Wnt agonists/antagonists (Wnt3a, IWR1e), Shh agonists/ antagonists (SAG, cyclopamine) and FGF agonists/antagonists (FGF2, SU5402). Among them, the addition of BMP4 at low concentrations during early phases of SFEBq culture (day 3 or day 6) proved to most effectively induce Rx::Venus þ retinal epithelium on day 24.
Induction-reversal culture. To induce CM-like tissues within NR, NR epithelium generated from hESCs was subjected to two-step 'induction-reversal' culture under the following conditions. For the 'induction' culture, NR-containing aggregates generated from hESCs on day 18 were transferred from a 96-well plate to a 9-cm Petri dish (noncell adhesive), and further cultured in suspension for 4-6 days under 5% CO 2 conditions in DMEM/F12-Glutamax medium containing 1% N2 supplement, 3 mM CHIR99021 (GSK3 inhibitor; Stemgent), 100 U ml À 1 penicillin and 100 mg ml À 1 streptomycin. The addition of 5 mM SU5402 (FGFR inhibitor; Sigma) treatment to this induction medium increased efficiency and reproducibility of RPE induction and formation of CM-like tissues after the reversal culture, while CHIR99021 treatment without SU5402 tended to give higher variations in the level of RPE induction.
For the 'reversal' culture, the floating aggregates with RPE-like thin epithelium (Fig. 2b,c) were cultured in suspension under 40% O 2 /5% CO 2 conditions in NRdifferentiation medium (described above). On day 50, two-domain aggregates with CM-like tissues formed and they were further cultured in NR-differentiation medium. Two-domain aggregates with CM-like tissues stably grew in long-term culture. For the reversal culture, the addition of FBS is essential and its depletion from the reversal culture medium substantially decreased NR differentiation.
CM-like tissues could be also generated by the induction-reversal culture using NR epithelia obtained by our previous method using Matrigel 32 . CM-like tissues generated by using Matrigel and induction-reversal culture were similar to those generated by using induction-reversal culture with BMP4 treatment, as judged by CM marker expression and morphology (data not shown). NR epithelia generated with Matrigel tended to be more sensitive to RPE-induction medium than BMP4induced NR and more reproducibly underwent RPE differentiation in medium containing CHIR99021.
FACS and MACS. For FACS analysis (Fig. 1o,p), cells (from 24 aggregates per each sample, n ¼ 3) were dissociated into single cells using Accumax (Innovative Cell Technologies) and analyzed by using FACSAria with FACSDiva software (Becton Dickinson). For FACS or MACS sorting, dissected tissues were dissociated into single cells by using papain (described above), and then immunostained with an SSEA1-APC antibody (Becton Dickinson; for FACS) or SSEA1-MicroBeads antibody (Miltenyi Biotec; for MACS) at 4°C for 15-30 min. For FACS, cells were sorted into Rx::Venus-positive/SSEA1-APC þ fraction and Rx::Venus-positive/ SSEA1-APC À fractions using FACSAria with the FACSDiva software. For MACS, cells were sorted into SSEA1 þ or SSEA1 À fractions using autoMACS Pro (Miltenyi Biotec) according to manufacturer's instruction.