To classify a particular single cell as a developing RGC based upon its gene expression profile, genes strongly associated with the RGC marker neurofilament light (NF68) were determined. The single cell expression profiles used to generate this list of genes included the potential RPCs profiled for this study, developing RGCs and ACs [36], single mature bipolar cells (Kim et al., in press), single mature amacrine cells (Cherry et al., in preparation) and single mature Müller glia (Roesch et al., in press). Using a Fisher's exact test (see Materials and Methods and [36]), the probability that there was a correlation between the distribution of any given gene and NF68 was calculated. Setting a p-value cutoff of 10⁻³ yielded 81 genes highly associated with NF68 (Table S4). Many genes previously shown to have significant RGC expression, such as GAP43 [36], were included in this list. At least one gene known to be expressed in RGCs, Brn3b, was not found to be strongly associated with NF68. The reason for the absence of an association between these two genes most likely is that temporally Brn3b is turned on in newborn RGCs prior to NF68. In fact, two of the RPCs (E14 cell B1 and E16 cell F1) identified as transitional cells in the process of deciding upon a final fate (see below) possessed significant levels of Brn3b, but were devoid of NF68. Since all of the cells profiled in this study were isolated during retinal development, genes with such temporally distinct windows of expression would not necessarily be expected to associate together. However, NF68 and Brn3b were strongly associated by hierarchical clustering [36]. This lack of correlation again demonstrates the increased utility of the Fisher's exact test over other clustering methods for analyzing gene expression data from single cells.

The relative expression levels for each of the 81 genes associated strongly with NF68 (as well as NF68 itself) were calculated by dividing the signal from each single cell by the maximum signal for that gene across the entire data set of single cells. The scaled values for each of the 82 genes within each cell were summed. All sums were then scaled, so that the maximum score was 10, to generate an RGC score for each single cell (Figure 1). This operation was carried out for markers of all cell types on all cells, as described further below. Scaling of the summed scores was required due to the fact that, for each cell type (below), the number of genes defining that cell type differed. An examination of the scores for the 13 previously characterized developing RGCs revealed that, for 12 of these cells the RGC score was considerably higher than the RPC, AC or PR score (Figure 1). For one cell, E14 cell E1, the RPC score and the RGC score were almost the same. Since this cell was a developing RGC and not isolated from adult tissue, the most likely explanation for this result is that this is a cell that was transitioning from an RPC to an RGC. This cell added to the six RPCs designated as transitioning means a total of 7 cells were identified as transitional cells, those having characteristic gene expression of multiple cell types, and these will be discussed in more detail below.

In a similar manner to that used for RGCs, classification scores were generated for ACs and PRs using gene clusters built around the transcription factors TCFAP-2β and Nrl respectively (see Table S5 and S6 for full clusters). Many of the genes identified as strongly associated with either TCFAP-2β or Nrl were predicted based upon previous work that characterized them as having either AC expression (Lrrn3, TCFAP-2α and Bruno-like 4 for example [36]) or rod photoreceptor cell expression (Crx, IRBP, Pde6a, Rom1, and Tulp1 for example [53]). In the TCFAP-2β associated genes, at least one previously known AC gene, glycine transporter 1, was not identified (Table S5) because the AC cells isolated in this study were all GABAergic ACs (as assessed by Gad1 expression [36]). Using these sets of associated genes to generate classification scores revealed that 4 out of 4 rod photoreceptor cells (2 adult and 2 P0) had significantly higher PR scores than the RGCs and ACs (Figure 1). However, the TCFAP-2β associated genes only yielded considerably higher AC scores for 3 out of the 6 ACs (Figure 1). This result demonstrates the sensitive nature of this classification scheme since it had been previously noted that these single ACs appeared to fall into 2 distinct classes based upon analysis of their gene expression using other methods [36]. Additionally, one of these groups of 3 ACs scored approximately the same for ACs as they did for RGCs (Figure 1). Again, this points to the robust nature of this classification scheme as these cells were also previously observed to have many similarities in gene expression to developing RGCs [36].

Given the success of this classification scheme in sorting out the different types of retinal neurons, it was used to distinguish the profiles of cycling RPCs from those of the developing, but more committed, retinal cell types. Cyclin D1 has been characterized as a gene expressed broadly in cycling RPCs [27], [44] and, therefore, this gene was chosen to generate a list of associated genes for classifying profiled single cells as RPCs. The distribution of cyclin D1 expression was compared pairwise to the signal levels for every other gene on the array across 128 single cell profiles in exactly the same manner as for the RGC, AC and PR markers. This yielded 94 associated genes whose expression was significantly similar in distribution to cyclin D1 (Table S7). Included in this list were several ribosomal protein genes and other known RPC expressed genes such as Fgf15 [27]. The relative expression levels for each of these genes and scaled scores were calculated (data not shown). Upon inspection, however, the distribution of these scores was observed to be very narrow owing to the high levels of expression for many of these genes in the profiled single cells and the persistence of many of these transcripts in newborn neurons (data not shown). Therefore, to improve the classification of cycling RPCs, additional gene clusters were added to generate a composite RPC score.

To generate a composite RPC classification score, three additional genes (Fgf15, Sfrp2 and μ-crystallin) were chosen to generate gene clusters. These genes have been observed previously in the outer neuroblastic layer (ONBL) of the retina, where the RPCs reside [27]. These 3 genes were also chosen as they together accommodate some of the temporal heterogeneity of the RPCs, as described below. Using the Fisher's exact test and a cutoff p-value of 10⁻³ as before, associated genes were identified for each of these three genes (Table S7). The relative expression levels were calculated and scaled RPC scores generated. As shown in Figure 1, 42 cells displayed a significant RPC score. For 36 of these cells, this score was considerably higher than that for RGC, AC, or PR, establishing these single cell profiles as coming from cycling RPCs (Figure 1). For the additional 6 cells (denoted with an * in Figure 1 and E14 cell E1, see above), while the RPC was the highest, at least one other classification score was significant as well. These cells are most likely transitional cells, RPCs that are in the process of generating a postmitotic daughter and a full analysis of their gene expression will be presented elsewhere (Trimarchi and Cepko, in preparation). Since transcripts expressed in RPCs would not be expected to disappear immediately, it was predicted that some cells would possess profiles containing genes expressed in one or more neuronal cell types, together with RPC genes that are in the process of being downregulated. Such transitional cells are of interest as they provide a window into cells that might still be in the process of deciding upon a final fate [54]. If this state was plastic, it might be revealed through the expression of markers of multiple neuronal cell types.

In order to develop assays to more fully explore the potential heterogeneity of gene expression in RPCs, two ISH methods were used to examine expression, with an initial test using a probe for Cyclin D1, the gene expressed in the highest percentage of RPCs profiled. Confirmation of expression in RPCs was possible on section ISH since RPCs reside in a distinctive layer, the outer neuroblastic layer (ONBL), though they are not the only cell type in the ONBL as migrating neurons also are present in this layer. In addition, expression of a gene in an RPC is detectable in cells acutely labeled with [³H]-thymidine, using the quantitative method of dissociated ISH (DISH). ISH using a cyclin D1 riboprobe on retinal cryosections from E12.5, E16.5 and P0 revealed strong staining in the ONBL, and diminished staining in the inner neuroblastic layer (INBL), where most postmitotic neurons are located (Figure 2B–D). DISH was performed on dissociated retinas that had been incubated with [³H]-thymidine for 1 hour before dissociation. This method primarily labels cells in S-phase and to a minor degree, cells in the early portion of G2. DISH performed at E16.5 or P0 with a cyclin D1 riboprobe showed that greater than 90% of [³H]-thymidine⁺ cells were also cyclin D1⁺ (Figure 3A,C), indicating that most S-phase progenitor cells express cyclin D1. As a control, GAP43, a gene expressed in developing RGCs [36], was never observed in [³H]-thymidine⁺ cells (data not shown). Additionally, only ~1/3 of cyclin D1⁺ cells were [³H]-thymidine⁺ (Figure 3A,C), consistent with the microarray data that suggested that cyclin D1 was present in RPCs in other cell cycle phases, as well as S phase. DISH for other genes broadly expressed in RPCs showed that 93% of [³H]-thymidine⁺ cells at E16.5 expressed Pax6 and 93% of [³H]-thymidine⁺ cells expressed Chx10. It should be noted that the microarray data reflects expression patterns of RPCs in all cell cycle phases, not just S phase, so an exact match of the percentage of RPCs positive for a given gene in the microarray analysis would not necessarily be expected in the [³H]-thymidine/DISH experiments. Nonetheless, these data demonstrate that a very high percentage of at least S phase RPCs express these genes. At the same time, they also show that a clear minority of [³H]-thymidine⁺ cells was negative for these RPC markers. This finding corroborates the microarray results, where not all RPCs were positive for these RPC marker genes (see Figure 2A). Although the possibility exists that the [³H]-thymidine⁺ population does not appear to be 100% for these genes for technical reasons, DISH performed for Ubiquitin B at E16.5 did show that 100% of the [³H]-thymidine⁺ cells expressed Ubiquitin B, demonstrating that at least for one probe this degree of co-staining was achievable. Taken together, these data indicate that most, but not all, RPCs do indeed express classical markers of RPCs at all times. The mild degree of heterogeneity observed likely reflects some dynamic processes within these cells. A similar suggestion was made based upon observations of the expression of a transgene encoded by a Chx10 BAC, in which the onset of expression was not uniform throughout the retina [58].

In addition to examining the RPC profiles for the expression of classical RPC marker genes, we wished to identify new genes that were consistently expressed among most or all RPCs, and thus could serve as new markers of RPCs and perhaps reveal new findings regarding RPC biology. To identify such genes, three different approaches were used. First, the RPC genes identified using the Fisher's exact test that were used for classifying the single cells were examined. Second, hierarchical clustering was performed on the set of single cells including the 42 RPCs and the 21 developing neurons previously characterized [36] using Gene Cluster software [37]. Finally, genes with potentially interesting RPC expression patterns were identified by visual inspection of the microarray data in Microsoft Excel (for details see Materials and Methods and [36]). Many different types of genes were found to be broadly expressed in the single RPCs and some representative examples are shown in the heatmap generated by Treeview in Figure 2A. The types of genes identified ranged from transcription factors (Plagl1 and Lhx2) to secreted molecules (Fgf15 and Dkk3). The presence of these transcripts in many of the RPCs continued to demonstrate the robustness of the single cell profiling method since these genes have been shown to be expressed in the retina by other means [27], [46], [59], [60]. As observed earlier for genes such as Sox2 and Chx10, these additional RPC genes were broadly expressed throughout individual RPCs (i.e. found in >50% of RPCs and in many cases >75% of RPCs), but most of these genes also were not observed in all RPCs, displaying some heterogeneity of expression.

To investigate the heterogeneous expression of the newly identified broadly expressed RPC markers, section ISH and DISH were performed for one such gene, Fgf15. According to the microarray data, Fgf15 was expressed in 9 out of 11 cyclin D1⁺ E12.5 RPCs, 9 out of 11 E16.5 cells and 10 out of 13 P0 cells (Figure 2A). Section ISH was performed at E12.5, E16.5 and P0. Fgf15 was expressed in the ONBL at E12.5 (Figure 2E) and E16.5 (Figure 2F), but was already less intense in the center of the retina at E16.5 (Figure 2F), despite the continued presence of RPCs at this stage. By P0, strong staining for Fgf15 was only observed in the most peripheral portion of the ONBL (Figure 2G). At all of the timepoints observed, it appeared that Fgf15 expression was much patchier than that of cyclin D1 (compare Figure 2B to Figure 2E for instance), confirming the heterogeneity in Fgf15 expression. DISH also was performed on E12.5, E16.5 or P0 retinas that had first been labeled for 1 hr with [³H]-thymidine. When compared with the results for cyclin D1, the percentage of Fgf15⁺ cells was consistently lower, whether among the entire population of retinal cells or among the [³H]-thymidine⁺ population (compare Figure 3A and 3B). In fact, the expression of Fgf15 was heterogeneously expressed in the S phase population of RPCs at all the timepoints examined (Figure 3D). Two color DISH on P0 retinal cells revealed that while >90% of Fgf15⁺ cells also expressed cyclin D1, only 50% of cyclin D1⁺ cells expressed Fgf15. The section ISH and DISH provide a confirmation that the heterogeneity of Fgf15 expression shown in the microarray analysis is unlikely to be a consequence of the single cell method, but reveals bona fide heterogeneity in expression among RPCs, even from a given age and in the same portion of the cell cycle, and even for a gene that is generally broadly expressed in RPCs.

In contrast to small number of genes expressed only at early timepoints in the developing retina, a tightly regulated cluster of late expressed genes emerged, centered around μ-crystallin. This cluster was apparent in both hierarchical clustering methods and a Fisher's exact test (Table S7) and a representation of the μ-crystallin cluster is shown in the bottom portion of the heatmap in Figure 4A. To examine the expression of these genes more thoroughly, section ISH was performed at E12.5, E16.5 and P0. Consistent with the microarray results, many of these genes were not detected in E12.5 retinal cryosections (Figure 5). At least one, retinaldehyde binding protein 1 (Rlbp1), was detected strongly in the pigment epithelium layer (RPE) at E12.5 (Figure 5M), but not in the retina itself. At E16.5, the kinetics of expression of these genes split into two groups. The first set was turned on by E16.5 and this group was represented by Crym (Figure 5B), Carbonic anhydrase 2 (Car2) (Figure 5E) and Patched 1 (Ptch1) (Figure 5H). Interestingly, Car2 was clearly detected in the ONBL in these ISH experiments as well as in the microarrays, whereas in previous studies Car2 expression in RPCs was contentious [27], [62]. The second group of late expressed RPC genes were not detected by ISH until P0 (Figure 5J–L and M–O).

Neurogenic basic helix-loop-helix (bHLH) transcription factors have been shown to play crucial roles in the generation of many postmitotic retinal cell types [75], [76], [77]. Recently the loss of one bHLH, Math5, was shown to lead to deficiencies in cell cycle progression in RPCs, revealing a possible additional coordinating role for this class of TF in RPCs [78]. Understanding the mechanism of action of these bHLH factors requires a detailed knowledge of their expression patterns. In the 42 single RPC profiles, the neurogenic bHLH genes were found in subsets of cells (Figure 7). To verify that these bHLH factors were expressed in RPCs, retinas were pulse labeled with [³H]-thymidine and DISH was performed for Math5 or NeuroD1. At E16.5, 18% of [³H]-thymidine⁺ cells were Math5⁺ while at P0 6% of [³H]-thymidine⁺ cells were NeuroD1⁺ and 25% were Ngn2⁺ (Figure 8A). These results indicate that while the bulk of cycling RPCs are not expressing these bHLHs, these genes most likely begin their expression either in late S phase or early G2. Interestingly, when one bHLH transcript was observed in an RPC, other bHLH transcripts were present as well, and some RPCs expressed as many as 4 different bHLH genes (Figure 7). For example, at least three cells at three different timepoints expressed significant levels of Math5, Ngn2 and NeuroD1 (see E13 cell B3, E16 cell D6 and P0 cell C1). This result was surprising given previous reports that both Ngn2 and NeuroD1 were suppressed in Math5 expressing cells [78]. However, these previous conclusions were based upon upregulation of these bHLHs in the absence of Math5 [78] and not a direct observation of their co-expression. Interestingly, single cell RT-PCR in the chick retina revealed that a few cells could co-express certain bHLHs [79]. Additionally, 7 RPC single cell profiles showed co-expression of Ngn2 and Mash1, including cells isolated from 3 different timepoints (Figure 7). Again this result is in contrast to previous observations that Ngn2 and Mash1 were never expressed in overlapping cells [55]. However, as before, this prior result was not based upon direct detection of Ngn2 and Mash1 transcripts, but instead relied upon GFP-based reporters for both genes [55]. It is possible that these reporters did not fully recapitulate the entire spectrum of expression for these genes, perhaps due to differences in the regulation of transcription or translation, as has been shown for certain homeobox TFs in Xenopus [80]. These single cell profiles demonstrate the expression of multiple neurogenic bHLHs in single RPCs and suggest that the interplay among these TFs is perhaps not as simple as previously postulated. These data provide a potential explanation for the observed redundancy of these bHLH factors in retinal development [81]. Furthermore, Xenopus NeuroD1 has been shown to be regulated by phosphorylation and if similar regulatory mechanisms exist in the mouse [82], this could provide a method for independently controlling bHLHs that are co-expressed.

Hairy and enhancer of split 6 (Hes6) is an additional bHLH gene whose protein product antagonizes the activity of other Hes family members and thereby facilitates the action of neurogenic bHLHs [83]. The single cell expression profiles showed Hes6 expression in RPCs across developmental time and revealed a significant overlap in expression with Mash1 at P0 (Figure 7). In situ hybridizations with a Hes6 riboprobe revealed a subset of cells in the ONBL stained at E12.5 and E16.5 (Figure 9A,B), with broader ONBL staining at P0 (Figure 9C). DISH performed on [³H]-thymidine pulse labeled retinas showed that 25% of [³H]-thymidine⁺ cells also stained for Hes6, confirming the presence of this transcript in RPCs (Figure 8A). As this is a time when many RPCs are producing postmitotic neurons, it is consistent with previous work showing that Hes6 has a positive role in this process [83].

Many homeodomain (HD) containing transcription factors have been found to play crucial roles in retinal development [32], [56], [84], [85]. In the analysis of TF expression among single RPCs, many HD containing TFs were observed. Some of these TFs showed a broad expression pattern in the single RPCs (for example Pax6, Chx10, Sox, Lhx2, Six3) (See Figure 2, ​,7,7, 9D–F and Figure S4), while others showed more heterogeneity of expression (Otx2, Rax, Six5). Otx2 expression was confirmed to begin in RPCs by DISH on [³H]-thymidine pulse labeled retinas. In these experiments, 15% of [³H]-thymidine⁺ cells were also Otx2⁺ (Figure 8A), indicating that the expression of Otx2 mRNA most likely begins in the late S to early G2 phase of the cell cycle. Pulse-chase experiments at E18.5 further validated this idea by showing that the number of [³H]-thymidine⁺ cells that were also Otx2⁺ increased from 9% 4 hours after labeling to 28% by 24 hours after labeling. Interestingly, Crx, another HD containing TF that is involved in photoreceptor development and maintenance [86], [87], showed similar, but slightly delayed kinetics of onset to that of Otx2. Crx was not present in [³H]-thymidine⁺ cells after a 1 hour pulse (Figure 8A) or a 4 hour chase, but was only first observed in 15% of [³H]-thymidine⁺ cells 24 hours after labeling. This observation fits with the predicted regulation of Crx by Otx2 [88].

Two additional TFs, Gli3 and Etv5, exhibited a pattern in the single RPC profiles that was somewhat similar to Mash1 and Lhx2 in that there appeared to be more RPCs at P0 expressing these genes that at the earlier timepoints (Figure 7). However, this correlation was not exact and it was difficult to ascertain the precise number of cells expressing these genes at different timepoints on section in situ hybridizations (see Figure 9G–L). In DISH experiments, only the riboprobe for Etv5 yielded significant signal and showed that Etv5 was in ~2/3 of the [³H]-thymidine⁺ population of RPCs at P0 (Figure 8A). Additionally, these TFs are part of much larger families of factors and many other family members were found in subsets of RPCs as well (Figure 4A and Figure S4). Therefore, at present any possible overlapping roles for these factors in currently unknown. However, recently, Gli3 was shown to have a genetic interaction with Pax6 in the developing retina [89] demonstrating that the interplay among all these TFs is probably critically important for retinal development.

The forkhead/winged helix family of transcription factors is considerable in number and these genes have been shown to play important roles in development through the control of cellular proliferation, apoptosis and metabolism [90], [91], [92]. There were numerous forkhead transcription factors expressed in the single cell profiles of RPCs and many of these TFs showed profound heterogeneity in their expression (Figure 4A, ​,77 and Figure S4). These forkhead factors ranged from those for which there is little known about their functional roles in general and nothing understood of their part in retinal development (Foxj3 and Foxk2) to Foxn4, which has been shown to play a crucial role in the ability of RPCs to generate amacrine and horizontal cells [93]. Foxm1 was also found in a subset of RPCs and ISH revealed expression of this in the ONBL (Figure 9M–O). Intriguingly, at P0, Foxm1 was confined to the vitreal side of the ONBL where S phase RPCs are located (Figure 9O). In serum starved cells induced to re-enter the cell cycle, Foxm1 mRNA has been shown to initiate at the onset of S phase and remain on from that point [94]. It is interesting that Foxm1 may be more tightly regulated in the developing retina, but its expression is consistent with a role in cell cycle control. Foxo3a was another family member expressed in subsets of RPCs (Figure 7). Although the Foxo3a transcript was detected in cells from early and later developmental timepoints, there were more single RPCs expressing this gene at P0 (Figure 7). Section ISH confirmed this later expression, as signal was observed in more cells in the ONBL at P0 (Figure 9R) than at E12.5 (Figure 9P). The significance of this Foxo3a expression in the retina is unclear. In other organisms and contexts, Foxo3a has been shown to induce either cell cycle arrest or apoptosis [95], [96]. While it is conceivable that Foxo3a is sensitizing the later retina to apoptotic signals, Foxo transcription factors in general have been linked to the cyclin kinase inhibitor p27Kip1 [92]. Therefore, given that more RPCs at P0 are generating postmitotic daughter cells and p27Kip1 is involved in cell cycle exit in the retina [31], perhaps Foxo3a plays an upstream role in the decision to generate a postmitotic daughter cell.

Many additional TF families are present in the single cell profiles, with different family members represented in distinct subsets of RPCs. For instance, there were at least 5 different Krüppel-like factors (KLFs) detected in subsets of RPCs (Figure 7 and Figure S4). KLF15 has previously been shown to be capable of inhibiting the rhodopsin promoter [97], [98], but otherwise, no function has been ascribed to these factors during murine retinal development. The fact that these KLFs are expressed in RPCs in the mouse is interesting given the important role played by Krüppel in the temporal progression of neuroblasts in Drosophila [24]. It has also been shown that Castor, the final gene in the Drosophila cascade, is expressed in photoreceptors in mouse retina [27]. Further examination of the single cell gene expression profiles revealed the presence of both an ikaros-like zinc finger (Ikzf5) in subsets of RPCs and several POU domain containing genes (Figure S4). The presence of such homologs for other genes in the Drosophila neuroblast cascade in subsets of RPCs during mouse retinal development suggests that a cascade similar to that in Drosophila might be playing a role in the progression of competence states in the murine retina as well.

One final TF family expressed in subsets of RPC profiles is the nuclear hormone receptor family of gene regulators. Many members of this large family were observed in distinct subsets of RPCs (Figure S4), including both orphan nuclear receptors such as the CoupTFs (Nr2f1 and Nr2f2) and those for which the hormone is characterized such as RXRα (Figure 7, 10A–C). This latter gene is interesting in light of the critical role that retinoic acid signaling plays in dorsoventral patterning of the chick and zebrafish retinas [99], [100], as well as the expression of rod genes, including rhodopsin [101], [102], [103]. Another set of TFs expressed in RPC subsets is the Tcfs that are downstream of the Wnt signaling pathway (Figure 7, 10D–I). Tcf3 was present in more RPC profiles from early timepoints, while Tcf7l2 was detected more in P0 RPCs (Figure 7). The possible significance of this early/late expression difference is unclear. Also, the function of these TFs in retinal development remains elusive since the Wnt factors themselves were not strongly detected in the retina (data not shown and [27]) and Wnt inhibitors such as Sfrp2 were found, at least in the early retina (Figure 4). Finally, some uncharacterized TFs were also identified as expressed in RPCs. One of these factors, Tcf19 was found in a small subset of RPCs at E12.5 (Figure 7, 10J) and then an increasing subset at subsequent stages (Figure 7, 10K–L). At P0, Tcf19 transcripts were localized near the vitreal edge of the ONBL (Figure 10J), suggesting this gene may be cell cycle regulated (see below). In total, many different classes of TFs were found expressed in subsets of RPCs throughout retinal development. The challenge in the future will be to use these single cell data in concert with functional studies to understand the precise combinations of TFs that RPCs use to generate all the distinct subtypes of retinal neurons.

Through the use of the different clustering methods employed in this study, some genes were uncovered as potentially cell cycle regulated that were not previously well characterized with respect to the cell cycle. In addition, their expression patterns in the developing retina were completely unknown. One such gene was Karyopherin alpha 2 (Kpna2). Kpna2 was observed in a subset of RPCs (Figure 11A). ISH on retinal cryosections revealed an expression pattern in a scattered subset of ONBL cells at E12.5 (Figure 11H) that progressed to a staining pattern that was concentrated on the vitreal side of the ONBL where the S phase cells reside at P0 (Figure 11J). At E16.5, the expression of Kpna2 was in a repeated pattern of radial stripes across the retina (Figure 11I). The striped expression pattern is not a predicted pattern, but it may indicate a dynamic regulation of Kpna2 that is coordinated with neighboring cells. Neighboring cells might share a history of recent cell divisions and thus might be somewhat synchronized with respect to the cell cycle. DISH for Kpna2 revealed that this gene is almost exclusively expressed in S phase, as 95% of all Kpna2⁺ cells at P0 were also [³H]-thymidine⁺ after a 1 hour pulse (Figure 8B,C). The protein product of this gene has been linked to the nuclear import of a protein important for DNA repair and checkpoint function [104], [114]. It is possible that this novel finding of restriction of Kpna2 expression to cells almost exclusively in S phase points to an important role in regulation of the normal cell cycle in a developing tissue. In addition to this insight into cell cycle function, this observation means that Kpna2 can be used as a marker for RPCs in S phase in future studies.

One cell cycle gene that was found in the fewest RPCs and might, therefore, be tightly regulated in its cell cycle expression was Cyclin E1 (Figure 11A). Section ISH for Cyclin E1 demonstrated that while it showed weak signal in the retina, it was expressed in a small subset of cells near the vitreal surface of the ONBL (data not shown). Hierarchical clustering experiments revealed that the gene Rassf1 was significantly correlated with Cyclin E1 (data not shown). ISH on retinal cryosections revealed a similar expression pattern to Cyclin E1, both in the intensity of signal and in location (Figure 11K–M and data not shown). While Rassf1 has been shown to be a tumor suppressor and to affect the ras pathway, its exact function, especially in development, remains poorly understood [115]. Additionally, in tissue culture experiments, Rassf1 has been shown to be capable of impinging on the regulation of the cell cycle at multiple points [115], but an interaction with Cyclin E1 is yet to be explored.

The analysis described above for G1/S phase genes was also done for genes that had previously been shown to have their expression concentrated in the G2 phase of the cell cycle [116], [117], [118], [119], [120]. The heatmap in Figure 12A shows some genes that are representative of the G2/M class. These genes were present in a smaller number of RPCs than those associated with the G1/S phases. As reported by Young [121], RPCs spend less time in G2/M than they do in G1/S, and thus fewer cells will show the G2/M pattern of gene expression than the G1/S pattern. Many RPCs showed robust expression of markers for both G1/S and G2/M (note E12 cell A4, E13 cell B3 and P0 cell F3 as examples in Figure 11A and Figure 12A). These results may point to a tighter transcriptional regulation of genes that play a role in G2/M than those genes involved in the G1 and S phases. Section ISH for some of these G2/M genes revealed some variation in the expression patterns (Figure 12B–M). Cdc20 was found in subsets of cells in the ONBL at E12.5 (Figure 12B), a radial like pattern at E16.5 (Figure 12C), and throughout the ONBL at P0 (Figure 12D). At all timepoints, though, the expression did appear to be enriched toward the scleral side of the ONBL, where mitosis occurs, but it was never absent from the vitreal side, indicating the possibility of broader expression. DISH performed at both E16.5 and P0 showed significant overlap between [³H]-thymidine and Cdc20 after a 1 hour pulse (Figure 8B), showing that the expression of this gene most likely begins while cells are still in S phase. However, the overall percentage of cells expressing Cdc20 was always much lower than those expressing a G1/S marker such as Rrm2 (10% for Cdc20 versus 30% for Rrm2 at P0, based upon DISH). Cyclin B1 and Cyclin B2 showed a similar expression pattern to that of Cdc20, with the exception of being more biased toward the scleral edge of the retina, especially at earlier stages such as E12.5 and E16.5 (Figure 12E–G and data not shown). Ubiquitination and subsequent degradation of many substrates is important for driving cells through anaphase and allowing the completion of mitosis [122]. One enzyme that participates in the ubiquitination is Ube2c and this gene also was found strongly associated with other characterized G2/M markers (Table S9). Section ISH revealed that Ube2c was in a subset of cells in the ONBL at all three stages examined (Figure 12H–J) and also displayed a pattern of radial stripes reminiscent to Cdc20 at E16.5 (Compare Figure 12C with 12I). Surprisingly, Spbc25, a component of the kinetochore, used to set up and maintain proper chromosome alignment during mitosis, displayed more intense staining toward the vitreal (S phase) edge of the ONBL than the scleral (M phase) edge (Figure 12K–M). This may indicate that even some of the G2/M class of genes is less tightly regulated in the developing retina than previously appreciated, and/or that there is differential regulation of the RNA and the protein. Finally, the cyclin dependent kinase inhibitor P27Kip1 was also found strongly associated with G2/M marker genes by hierarchical clustering (data not shown). This suggests that P27Kip1 is involved in the normal transition from G2 through M in RPCs and this is most likely a strong factor as to why in P27^−/− mice there is a cell cycle defect, but no corresponding perturbation of the distribution of retinal cell fates [31].

Exit from the cell cycle is intimately coordinated with the cell fate decision making process in RPCs [2]. Since the G2/M phases of the cell cycle would most likely be when this link would occur, it was of interest to identify those RPCs with the most characteristic gene expression of G2/M. The classification scheme described above for the identification of cell types was adopted to generate G2/M scores for the RPCs, transitional cells and postmitotic neurons. These scores were a composite based upon the clusters of genes generated around three G2/M markers, Cdc20, Aurora kinase B (AurkB), and Ube2c (Table S9). Fourteen of the 42 RPCs displayed high scores for G2/M while none of the postmitotic neurons scored highly (Figure 13). Only 2/6 of the RPCs denoted as transitional cells scored highly for G2/M markers (Figure 13), suggesting that these cells are most likely in different windows of the transition from RPC to postmitotic neuron. Examination of the RPCs with high G2/M gene expression was performed with the hopes of identifying novel genes involved in the process of exiting from the cell cycle. Both hierarchical and Fisher's exact test based clustering methods yielded many significantly associated genes, but given the nature of these genes (many kinesins and microtubule associated proteins), the majority of them probably play generic roles in cytokinesis and other mitotic processes (Table S9 and data not shown). Visual inspection of the single cell profiles in Microsoft Excel, with an extra focus on the cells receiving the highest G2/M scores, revealed several interesting candidate genes that may play a role in the ability of an RPC to exit from the cell cycle (Figure 14A). The notch ligand, delta-like 1(Dll1), whose expression had previously been shown to be well correlated with the timing of retinal neurogenesis [123], was found to be highly expressed in several G2/M cells (see P0 cell A8, P0 cell C8 and P0 cell H3 for examples, Figure 14A). Section ISH showed Dll1 staining in subsets of cells in the ONBL at all three stages (Figure 14B–D) and DISH performed with P0 retinas that were labeled with [³H]-thymidine for 1 hour revealed that 1/3 of Dll1⁺ cells were also [³H]-thymidine⁺. These data are consistent with expression of Dll1 at the right place and the right time to be correlated with the production of a postmitotic cell(s) by an RPC. Three additional genes (B-cell translocation gene 2 [Btg2], Rhomboid veinlet-like 3 [Rhbdl3], and Sprouty protein with EVH-1 domain 1 [Spred1]) were discovered in subsets of G2/M RPCs (Figure 14A). Btg2 was found in subsets of ONBL cells at E12.5 (Figure 14E) and E16.5 (Figure 14F), but its expression appeared to become more broadly expressed at P0 by section ISH (Figure 12G). This result is consistent with Btg2 playing a role in cell cycle exit since many more retinal neurons are being generated at P0 than at E12.5. Btg2 has been shown to enhance neural differentiation upon overexpression in PC12 cells [124] and its expression in the neural tube correlates with those cells that will generate a postmitotic neuron [125], further suggesting a role for this gene in the control of cell cycle exit in the retina. Both Rhbdl3 and Spred1 were found in subsets of RPCs and both were in more RPC profiles at P0 than at the earlier stages (Figure 14A). Section ISH for both genes confirmed that their expression increased as the number of retinal neurons generated increased (Figure 14H–M). Neither gene has been extensively characterized in general, nor does any information exist as to the possible functions of these genes during retinal development. Drosophila homologues of Rhbdl3 have been shown to modulate both the EGF pathway and the notch pathway and, in that manner, play specific roles in the cell fate specification of neuroblasts [126], [127]. Alternatively, Spred1 has been shown to be a negative regulator of the ras pathway and perturbation of Spred1 function interfered with neural differentiation in tissue culture [128]. Given that both of these genes have been shown to be possible regulators of important pathways that are also at work in the developing retina, they make excellent candidates for regulators of the cell cycle exit and cell fate decision-making processes of RPCs. The fact that these different candidate genes are expressed in distinct RPCs suggests that the machinery utilized by RPCs to exit the cell cycle may vary among exiting cells.

We would like to thank Trent Rector and Sarah Annis at the Harvard Medical School Biopolymers Facility for assistance with the Affymetrix arrays and the members of the Cepko, Tabin and Dymecki labs for valuable comments. CLC is an investigator of the Howard Hughes Medical Institute.