Retinal progenitor cells (RPCs) express basic helix-loop-helix (bHLH)factors in a strikingly mosaic spatiotemporal pattern, which is thought to contribute to the establishment of individual retinal cell identity. Here, we ask whether this tightly regulated pattern is essential for the orderly differentiation of the early retinal cell types and whether different bHLH genes have distinct functions that are adapted for each RPC. To address these issues, we replaced one bHLH gene with another. Math5 is a bHLH gene that is essential for establishing retinal ganglion cell (RGC) fate. We analyzed the retinas of mice in which Math5 was replaced with Neurod1 or Math3, bHLH genes that are expressed in another RPC and are required to establish amacrine cell fate. In the absence of Math5, Math5Neurod1-KI was able to specify RGCs, activate RGC genes and restore the optic nerve, although not as effectively as Math5. By contrast, Math5Math3-KI was much less effective than Math5Neurod1-KI in replacing Math5. In addition, expression of Neurod1 and Math3 from the Math5Neurod1-KI/Math3-KIallele did not result in enhanced amacrine cell production. These results were unexpected because they indicated that bHLH genes, which are currently thought to have evolved highly specialized functions, are nonetheless able to adjust their functions by interpreting the local positional information that is programmed into the RPC lineages. We conclude that, although Neurod1 and Math3 have evolved specialized functions for establishing amacrine cell fate, they are nevertheless capable of alternative functions when expressed in foreign environments.
In vertebrates, postmitotic retinal neurons are generated from a pool of multipotent retinal progenitor cells (RPCs) in a temporal order: retinal ganglion cells (RGCs) are produced first, followed immediately by amacrine cells, horizontal cells and cone photoreceptor cells in highly overlapping waves of cell differentiation. Subsequent to the formation of these early neuronal cell types, the later cell types form: bipolar cells, rod photoreceptor cells and, lastly, Müller glial cells. Distinct RPCs integrate their intrinsic programs with local environmental signals to define their competence states and to commit to individual retinal cell fates(Livesey and Cepko, 2001). Many transcription factors have been implicated in setting up the competence states of RPCs (Hatakeyama and Kageyama,2004; Ohsawa and Kageyama,2008). In the mouse, Math5 (Atoh7 - Mouse Genome Informatics), an ortholog of the Drosophila bHLH proneural gene Atonal, is expressed in a subpopulation of RPCs and is essential for establishing RGC competence (Brown et al.,1998; Brown et al.,2001; Wang et al.,2001; Yang et al.,2003; Mu et al.,2005).
A remarkable feature of retinal development is that RPCs are capable of simultaneously producing multiple cell types, suggesting the presence of subpopulations RPCs with each possessing a distinct genetic makeup. Unfortunately, these genetically distinct RPC subpopulations have not been clearly defined by conventional cell lineage tracing experiments (reviewed by Mu and Klein, 2004; Mu and Klein, 2008). The bHLH genes Math5, Mash1 (Ascl1 - Mouse Genome Informatics), Math3 (Neurod4 - Mouse Genome Informatics) and Neurod1 are expressed in the developing retina at overlapping times but in largely distinct, interspersed RPC subpopulations(Vetter and Brown, 2001; Akagi et al., 2004; Hatakeyama and Kageyama, 2004; Le et al., 2006; Ohsawa and Kageyama, 2008; Trimarchi et al., 2008). The bHLH factors collaborate with homeobox factors to specify particular retinal cell fates at the expense of others (Wang and Harris, 2005; Cayouette et al., 2006; Ohsawa and Kageyama, 2008). However, in early developing retina, many key homeobox genes, such as Pax6 and Six3, are expressed in virtually all RPCs (Lagutin et al.,2001; Bäumer et al.,2003) and have multiple functions in specifying distinct cell fates (Ohsawa and Kageyama,2008). It is therefore likely that the mosaic expression pattern of bHLH genes more accurately mirrors the state of competency within each individual RPC for the early retinal cell types(Cayouette et al., 2006). This concept implies that a unique bHLH gene expression pattern regulates the competence state of each RPC fate for the early differentiating cell types,with more widely expressed homeobox factors acting in conjunction with the bHLH factors. Thus, replacing one bHLH gene with another might be expected to redirect the RPC to assume the competence state defined by the replacing bHLH gene. However, it is also possible that this type of replacement would not be tolerated because the replacing bHLH gene, which might have evolved specialized functions in the retina, would be incapable of integrating into the intrinsic program of a foreign RPC. A final possibility is that replacing one bHLH gene with another would restore the original RPC lineage. If this were the case, it would suggest that retinal bHLH genes might not be highly specialized and therefore are susceptible to the intracellular environment of the foreign RPC.
To determine which of these possibilities actually occurs, we replaced Math5 with either Neurod1 or Math3, which are required together for establishing amacrine cell fate, and are capable of producing excess numbers of amacrine cells when each of them is ectopically co-expressed with Pax6 or Six3 (Inoue et al., 2002). Here, we demonstrate that Neurod1 can partially rescue the functions of Math5 in RGC production. By contrast, Math3 could only modestly rescue Math5 mutant defects by activating some RGC genes. In addition, Neurod1 and Math3 co-expression at the Math5 locus does not lead to the overproduction of amacrine cells. Our results demonstrate that although Neurod1 and Math3 have evolved specialized functions, they are nevertheless capable of alternative functions when expressed in a foreign environment. These results suggest that RPC heterogeneity is largely programmed by intrinsic mechanisms that are not solely dependent on a specific bHLH gene.
MATERIALS AND METHODS
Gene targeting and animal breeding
To construct the targeting vector, genomic DNA from G4 ES cells(George et al., 2007) was used for the polymerase chain reaction (PCR) amplification of 5 kb fragments representing the left and right homologous replacement arms from the Math5 locus (Fig. 1B). Complete coding regions for Neurod1 and Math3 are located in single exons, and their sequences were amplified using G4 ES cell genomic DNA. These fragments were sequentially subcloned into a targeting vector(Mao et al., 2008). The resulting constructs were linearized and electroporated into G4 ES cells,after which G418-resistent ES cells were selected to identify homologous recombination events. A 5′ probe from outside the homologous recombination region was used to detect 20.5 kb wild-type and 15.6 kb knock-in fragments produced by BamHI digestion of ES cell DNA(Fig. 1B,C). Two targeted ES cell lines for each construct were expanded and injected into B6(GC)-Tyrc-2J/J blastocysts, and the injected blastocysts were transferred into the uteri of pseudopregnant C57/BL/6J female mice. Chimeric males resulting from the injected blastocysts were bred to B6(GC)-Tyrc-2J/J females (Jackson Laboratory) to generate Math5Neurod1-KI/+ and Math5Math3-KI/+heterozygotes. The two lines were subsequently bred to Math5lacZ-KI/+ mice(Wang et al., 2001) to generate Math5Neurod1-KI/lacZ-KI and Math5Math3-KI/lacZ-KI lines. Math5Neurod1-KI, Math5Math3-KI,Math5lacZ-KI and Math5 wild-type alleles were distinguished using a Southern blot genome hybridization probe and EcoRV-digested genomic DNA (see Fig. 1D). Embryos were designated as embryonic day 0.5 (E0.5) at noon on the day in which vaginal plugs were observed.
All animal procedures in this study followed the US Public Health Service Policy on Humane Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee at The University of Texas M. D. Anderson Cancer Center.
Histology, in situ hybridization and immunohistochemical analysis
Embryos and eyes dissected from embryos or adults were fixed,paraffin-embedded and sectioned into 7 μm or 12 μm slices for immunohistochemical analysis or in situ hybridization, respectively, as described by Mao et al. (Mao et al.,2008). After de-waxing and rehydration, the sections were stained with Hematoxylin and Eosin before further analysis. In situ hybridization was performed as described by Mu et al. (Mu et al., 2004).
For immunohistochemical analysis, sections were placed in a microwave oven at 600 W in 10 mM sodium citrate for 15 minutes to expose the antigen epitopes. Microwave-treated sections were then incubated with the primary antibodies listed below. For indirect immunofluorescence, a tyromide signal amplification kit (TSA biotin system, PerkinElmer) was used for Neurod1 and Eomes to optimize signal intensity. For double immunofluorescence,Alexa-conjugated secondary antibodies (Invitrogen) were used. The primary antibodies were anti-Brn3b/Pou4f2 (Santa Cruz Biotech, 1:200 dilution),anti-Chx10 (Exalpha, 1:1000 dilution), anti-GSK3β (Cell Signaling, 1:400 dilution), anti-NFL (Invitrogen, 1:250 dilution), anti-NF160 (DSHB, 1:1000 dilution), anti-Isl1 (DSHB, 1:250 dilution), anti-melanopsin (provided by Satchidananda Panda, Salk Institute), anti-SMI32 (Covance, 1:1000 dilution),anti-Tbr2/Eomes (Chemicon, 1:1000 dilution), anti-Sox9 (Chemicon, 1:200 dilution), anti-TUJ1 (Covance, 1:500 dilution), anti-ChAT (Chemicon, 1:100 dilution), anti-opsin(R/G) (Chemicon, 1/200 dilution), anti-calrectinin(Chemicon, 1:2500 dilution), anti-calbindin (Swant, 1:5000 dilution),anti-Pax6 (DSHB, 1:200 dilution) and anti-p57 (Santa Cruz, 1:40 dilution). Horseradish peroxidase-conjugated secondary antibody for tyromide signal amplification was obtained from Jackson ImmunoResearch. To detect RGC axons,anti-NFL antibody was used to stain flat-mounted adult retinas(Mao et al., 2008). The number of axonal bundles and the axonal density within each bundle were analyzed with the same settings using SimplePCI software (Compix, Sewickley, PA) to automatically select regions of interest from the peripheral retinal flat-mount images. Flat-mount immunostaining was also used to monitor the distribution of melanopsin and SMI-32 positive RGCs. For quantifying RGC specification during embryonic stages, Pou4f2-positive cells were used to estimate RGC number. Three retinal sections (three sections apart) collected from littermates of different genotypes were stained with anti-Brn3b/Pou4f2 antibody, and the number of Pou4f2-positive cells on each section was counted on an Olympus Fluoview 1000 confocal microscope.
Quantitative reverse transcriptase-PCR analysis
Total RNA were collected from two E13.5 retinas using TRIZOL reagent(Invitrogen, CA). RNAs were reversed transcribed using Superscript First-Strand Synthesis System for reverse transcriptase (RT)-PCR (Invitrogen)following the manufacturer's instruction. One twentieth of the total cDNAs was amplified for quantitative (q)PCR using SYBR green PCR master mix (Applied Biosystems, CA). The relative expression levels were normalized to that of GAPDH and calculated using the comparative Ct method (7500 Fast Real-time PCR systems SDS software, Applied Biosystems). DNA sequences of PCR primers indicated were: Math5 5′UTR forward,5′-TCCGTCTGTGTCTTATTCACTC-3′; Math5 reverse,5′-TTTGCAGGCCGACTTCATCCTC-3′; Math3 reverse,5′-ATATACATTTTTGCCATGGCCGC-3′; GAPDH forward,5′-AGGTCGGTGTGAACGGATTTG-3′; GAPDH reverse,5′-TGTAGACCATGTAGTTGAGGTCA-3′.
Expression of Neurod1 and Math3 in Math5Neurod1-KI and Math5Math3-KIembryonic retinas
The bHLH sequences from Math5, Neurod1 and Math3 are similar but not identical, clustering together with the other mouse bHLH genes(Fig. 1A)(Ledent et al., 2002). Outside of the bHLH domain, significant differences in Math5, Neurod1 and Math3 sequences indicate that these bHLH genes have undergone extensive divergence since their emergence from a common ancestral gene. A sequence tree shows that Math5 is rooted in a clade distinct from that of Neurod1 and Math3, and that Neurod1 and Math3 are closely related to each other in their bHLH domains(Fig. 1A). To determine the effects of replacing Math5 with Neurod1 or Math3,we used targeting constructs to target embryonic stem (ES) cells in which we removed the entire Math5 sequence and replaced it with either a Neurod1 or a Math3 sequence(Fig. 1B). Germline mice containing Math5Neurod1-KI or Math5Math3-KI alleles were bred to Math5lacZ-KI/lacZ-KI mice (Math5-null mice) and to each other to generate mice with the following genotypes: Math5Neurod1-KI/+ (Fig. 1C), Math5Neurod1-KI/lacZ-KI,Math5Neurod1-KI/Neurod1-KI(Fig. 1D), Math5Math3-KI/+, Math5Math3-KI/lacZ-KI,Math5Math3-KI/Math3-KI and Math5Neurod1-KI/Math3-KI.
We first determined whether Math5Neurod1-KI and Math5Math3-KI alleles were expressed in a pattern mimicking that of Math5. At E12.5, we detected low expression levels of Neurod1 protein in the retinas of wild-type mice(Fig. 1E). By contrast, high levels of Neurod1 expression were observed in Math5Neurod1-KI/Math3-KI retinas at E12.5(Fig. 1F). Similar to Neurod1 protein expression, Math3 transcript expression,although weak, was readily detectable near the ventricular region in E12.5 wild-type retinas, and in a pattern similar to that of Neurod1 expression in Math5Neurod1-KI/Math3-KI retinas at the same developmental time (Fig. 1G,H). Between E12.5 and E15.5, the expression of Neurod1 and Math3 from the Math5Neurod1-KI and Math5Math3-KIalleles closely resembled endogenous Math5 expression, indicating that ectopic Neurod1 and Math3 expression was under the control of the Math5 regulatory region. Furthermore, Neurod1expression did not differ between retinas with Math5Neurod1-KI/+ or Math5Neurod1-KI/Neurod1-KI genotypes, indicating the accurate replacement of Math5 by Neurod1 (data not shown). To determine whether the knock-in Math3 allele expressed transcripts at the same level as the wild-type Math5 allele, we compared the expression levels of Math5 and Math5Math3-KIalleles at E13.5. Fig. 1I shows that Math5 and Math5Math3-KI alleles were expressed at similar levels in Math5Math3-KI/+ retinas,and that Math5Math3-KI in Math5Math3-KI/Math3-KI retinas was expressed at levels corresponding to those of Math5 in wild-type retinas.
Partial restoration of RGC axons, optic nerves, and RGC subtypes with Math5Neurod1-KI and Math5Math3-KIalleles
Math5-null mice have severe optic nerve hypoplasia and in extreme cases lack optic nerves entirely (Fig. 2A) (Brown et al.,2001; Wang et al.,2001). Eyes from adult Math5Neurod1-KI/lacZ-KIand Math5Neurod1-KI/Math3-KI mice 4-5 weeks old (P30) were often found attached to well-developed optic nerves with diameters that were 50-60% those of wild-type mice (Fig. 2A,B). Histological sections from Math5Neurod1-KI/lacZ-KI and Math5Neurod1-KI/Math3-KI mice also revealed substantial restoration of optic nerves (Fig. 2A1,A2,B). Immunostaining retinas with a Pou4f2/Brn3b antibody,which detects RGCs, showed that 30-40% of the cells in the ganglion cell layer(GCL) of Math5Neurod1-KI/Math3-KI retinas expressed Pou4f2 compared with wild-type retinas (Fig. 2B1,B2). By contrast, we never observed optic nerves in the eyes of Math5Math3-KI/lacZ-KI mice, despite the close relationship between the Neurod1 and Math3 genes. These results indicated that Neurod1, but not Math3, was capable of partially restoring optic nerves in Math5-deficient mice.
The presence of optic nerves in Math5Neurod1-KI/Math3-KI and Math5Neurod1-KI/lacZ-KI mice suggested that RGC axons were forming and extending into the optic disk. Immunostaining of flat-mount retinas from P30-P40 mice with an anti-neurofilament light chain (NFL)antibody revealed that one-third (4/12) of Math5Neurod1-KI/Math3-KI and two-thirds (4/6) of Math5Neurod1-KI/lacZ-KI retinas had significantly greater numbers and higher density of RGC axon bundles compared with Math5-null retinas (Fig. 3A-D). Although some axons appeared misoriented, the majority of axons were well bundled and oriented towards the optic disk(Fig. 3C,D).
The lack of a complete rescue of the optic nerve in the knock-in retinas might be explained by enhanced apoptosis; a significant increase in cell death was detected in Math5Neurod1-KI/lacZ-KI and Math5Neurod1-KI/Math3-KI retinas from E11.5 to E16.5 compared with wild-type controls (data not shown). No differences in RGC axon number or density were detected when Math5Neurod1-KI/+ and Math5Math3-KI/+ retinas were compared with wild-type controls (data not shown). The results demonstrated that Neurod1 could partially replace Math5 in restoring RGC axons and that Math3 alone could not replace Math5.
In the mouse retina, RGCs have been categorized into ∼12-14 subtypes by morphological criteria (Sun et al.,2002; Coombs et al.,2006). Across the GCL, RGC subtypes display regular spacing between cell bodies, and arborize within precise strata of the inner plexiform layer (IPL). To determine whether RGC subtypes formed in Math5Neurod1-KI/Math3-KI retinas, we selected two previously characterized RGC subtype markers, melanopsin and SMI-32(Lin et al., 2004), and performed immunostaining with flat-mount retinas. Melanopsin-expressing RGCs have small somata (∼20 μm) and their dendrites run through the IPL,terminating at layer 1 immediately underneath the inner nuclear layer (INL). SMI-32 is expressed in RGCs with large somata (∼30 μm). Melanopsin- and SMI-32-expressing RGCs both displayed a nonrandom mosaic pattern across wild-type and Math5Neurod1-KI/Math3-KI retinas(Fig. 3E-H). Additionally, the dendrites of melanopsin-expressing RGCs always arborize to layer 1 of the IPL(insets in Fig. 3E,F; n>40). These data suggest that Neurod1 can replace Math5 to specify different RGC subtypes evenly across retina, and these RGC subtypes differentiate normally with proper dendritic arborization.
Expression of RGC genes in embryonic retinas of Math5Neurod1-KI and Math5Math3-KImice
The earliest sign of RGC differentiation begins at E12.5, when the expression of Pou4f2 and Isl1 is first apparent(Gan et al., 1999; Elshatory et al., 2007; Mu et al., 2008). These genes encode POU domain and LIM domain transcription factors, respectively, and both are required for RGC differentiation (Gan et al., 1999; Mu et al.,2008; Pan et al.,2008). We used anti-Pou4f2/Brn3b and anti-Isl1 antibodies to determine the expression pattern of these proteins in E13.5 retinas. In Math5lacZ-KI/+ (wild-type) retinas, Pou4f2 and Isl1 were co-expressed in differentiating RGCs, as we and others have shown previously(Fig. 4A,A1,A2)(Rachel et al., 2002; Mu et al., 2008). Expression was largely absent in Math5lacZ-KI/lacZ-KI retinas(Fig. 4B,B1,B2). We found that retinas expressing Math5Neurod1-KI allele in the absence of Math5 significantly restored the expression of Pou4f2 and Isl1(Fig. 4C-C2,D-D2), whereas in Math5Math3-KI/lacZ-KI retinas, the expression of these early RGC markers, although detectable, was appreciably lower(Table 1). We had shown previously that expression of the neurofilament protein NF160 is strongly dependent on the presence of Pou4f2 (Mu et al., 2004). Retinas expressing the Math5Neurod1-KI allele in the absence of Math5had significantly higher expression of NF160(Fig. 4A3-D3). The expression of Pou4f2, Isl1 and NF160 is indicative of RGC differentiation, and,therefore, of the number of RGCs present in the retinas of the Math5Neurod1-KI and Math5Math3-KImice. According to this criterion, we estimated that the numbers of RGCs present in the Math5Neurod1-KI/LacZ-KI and Math5Math3-KI/LacZ-KI retinas were ∼40% (173/434) and 10% (40/413), respectively, the number of RGCs in wild-type retinas(Table 1). Furthermore, the expression of Pou4f2 in Math5Neurod1-KI/lacZ-KI and Math5Math3-KI/lacZ-KI retinas was significantly lower than that in wild-type controls at E12.5, but recovered to higher levels at E14.5(Table 1), suggesting a delayed RGC differentiation in the Math5Neurod1-KI/LacZ-KI and Math5Math3-KI/LacZ-KI retinas.
|Math5Neurod1-KI/+ × Math5Math3-KI/+||E12.5||E13.5||E14.5|
|Math5Neurod1-KI/LacZ-KI × Math5LacZ-KI/+||E12.5||E13.5||E14.5|
|Math5Math3-KI/LacZ-KI × Math5LacZ-KI/+||E12.5||E13.5||E14.5|
|Math5Neurod1-KI/+ × Math5Math3-KI/+||E12.5||E13.5||E14.5|
|Math5Neurod1-KI/LacZ-KI × Math5LacZ-KI/+||E12.5||E13.5||E14.5|
|Math5Math3-KI/LacZ-KI × Math5LacZ-KI/+||E12.5||E13.5||E14.5|
Embryos from the same litter were used for comparison.
NA, not available.
We recently identified the T-box-containing transcription factor,eomesodermin (Eomes), as an essential factor for RGC differentiation and optic nerve formation (Mao et al.,2008). Eomes is a direct target gene of Pou4f2, and its expression would therefore be indicative of a functional Pou4f2 protein. At E14.5, Eomes protein is co-expressed with Pou4f2 in the innermost RGCs in wild-type retinas (Fig. 5A1-A3)(Mao et al., 2008). At this same time, Eomes is also weakly expressed in non-RGCs in the RPC proliferation layer; expression of Eomes in the innermost layer is absent in Math5-and Pou4f2-null retinas (Mao et al., 2008). Substantially more Eomes-expressing cells were observed in Math5Neurod1-KI/Math3-KI E14.5 retinas than in Math5-null retinas (Fig. 5B1). Similarly, more Eomes-expressing cells could be found in Math5Neurod1-KI/lacZ-KI retinas than in Math5-null retinas (Fig. 5C1). However, we did not detect Eomes-expressing cells in Math5Math3-KI/lacZ-KI retina(Fig. 5D1). These results suggest that Neurod1 alone can rescue Eomes expression in RGCs, but Math3 lacks this activity. By contrast, both Neurod1 and Math3 contribute to the restoration of Pou4f2 expression observed in the absence of Math5, and that Neurod1 has stronger restoration capabilities than Math3(Fig. 5A2,B2,C2,D2; Table 1).
At E13.5 and E14.5, the numbers of Pou4f2-positive cells in the Math5Neurod1-KI/lacZ-KI (n>20) or the Math5Math3-KI/lacZ-KI (n>20) retinas were always much greater than in the Math5lacZ-KI/lacZ-KIretinas. However, we noticed that many Pou4f2-expressing cells in these knock-in retinas were abnormally positioned when compared with wild-type controls, residing in the upper-most region of the RPC layer (compare Fig. 5A2 with Fig. 5C2,D2). This suggested that RGCs expressing Neurod1 or Math3 in the absence of Math5 were unable to properly migrate to the GCL. However, in Math5Neurod1-KI/Math3-KI retinas, this abnormality was partially corrected (Fig. 5B2). This result suggested that, together, Neurod1 and Math3 were slightly more effective in replacing the functions of Math5 than was Neurod1 alone.
The RGC gene regulatory network is restored in Math5Neurod1-KI/Math3-KI retinas
The partial restoration of RGC axons and optic nerves in adult retinas and the expression of the early RGC markers Pou4f2, Isl1, NF160 and Eomes in Math5Neurod1-KI/Math3-KI retinas indicated that, together,Neurod1 and Math3 could replace Math5 to activate the entire RGC gene regulatory network (Mu et al.,2004; Mu et al.,2005; Mu et al.,2008; Mu and Klein,2008). We therefore determined the expression levels of a number of RGC-expressed genes in wild-type and Math5Neurod1-KI/Math3-KI E14.5 retinas that had various roles in RGC integrity and physiology, transcriptional regulation and extracellular signal transduction. The selected genes included those whose expression was dependent on the presence of Math5 and Pou4f2 (Persyn,Gap43 and Shh) (Mu et al.,2004), those whose expression was dependent on Math5 but not Pou4f2 (Myt1, Stmn2 and TuJ1)(Brown et al., 2001; Mu et al., 2005), and two genes whose dependence on Math5 and Pou4f2 has not been determined[GDF11 (Kim et al.,2005), Gsk3β(Tokuoka et al., 2002) (see Table S1 in the supplementary material for details)].
Fig. 6 shows that all of the genes were expressed in RGCs of Math5Neurod1-KI/Math3-KIretinas in a similar pattern but at lower levels than in wild-type controls(compare A1-H1 with A2-H2). These results strongly suggest that the RGC gene regulatory network was activated in its entirety in Math5Neurod1-KI/Math3-KI retinas.
Because Neurod1 and Math3 are required together for amacrine cell development (Inoue et al.,2002), we determined whether increased numbers of amacrine cells were present in Math5Neurod1-KI/Math3-KI retinas. Staining with antibodies against markers for amacrine cells, including ChAT, p57Kip2,calrectinin and calbindin, revealed no differences in the numbers of cells between wild-type and Math5Neurod1-KI/Math3-KI retinas(see Fig. S1 in the supplementary material). However, we detected 20% more opsin-positive photoreceptors in Math5Neurod1-KI/Math3-KIretinas than in wild-type retinas (see Fig. S1 in the supplementary material). The significance of this modest increase was unclear, but it might reflect a re-direction of Math5Neurod1-KI/Math3-KI-expressing RPCs to a photoreceptor cell fate. This might occur as a result of restored Ngn2 expression in Math5Neurod1-KI/Math3-KI RPCs,as happens in Math5-null RPCs(Brown et al., 2001; Le et al., 2006). Math3,together with Mash1, is essential for bipolar cell formation(Tomita et al., 2000). We therefore determined whether increased numbers of bipolar cells could be detected in Math5Neurod1-KI/Math3-KI retinas. Fig. S1 shows that there is no difference in Chx10-positive cells in wild-type and Math5Neurod1-KI/Math3-KI retinas. The number of Sox9-positive Müller glial cells also did not change. Math5Math3-KI/LacZ-KI retina displayed histological phenotypes reminiscent of Math5-null retina, in which the numbers of all cell types normally found in the INL are reduced owing to the reduced thickness of the INL, but the ratio of each cell type was not significantly different from wild-type controls (Moshiri et al.,2008). We found that the proportion of amacrine, bipolar,horizontal and Müller cell types in the INL of Math5Math3-KI/LacZ-KI retina was similar to that of Math5LacZ-KI/+ retinas (see Table S2 in the supplementary material). These data suggest that Math3 alone does not significantly influence cell fate determination when expressed at the Math5locus.
Our results demonstrate that Neurod1, and to a lesser extent Math3, can replace Math5 to direct RPCs towards an RGC fate. Expressing Neurod1at the Math5 locus in the absence of Math5 resulted in a substantial restoration of all aspects of RGC differentiation, including axonogenesis and formation of the optic nerve. Neurod1 was better at replacing Math5 than was Math3, but the presence of both factors appeared to have an optimal effect. One possible explanation for this is that Neurod1 and Math3 may have somewhat different DNA-binding propensities and thus activate distinct sets of Math5 target genes. In the retinas of Neurod1 knock-in embryos, fewer cells expressed genes that mark the onset of RGC differentiation; only 30-40% of the number of wild-type RGCs was optimally observed in the knock-in retinas. Although we did not perform a quantitative analysis of RGC gene expression,both in situ hybridization and immunostaining analyses indicated that on a per cell basis, gene expression levels between knock-in and wild-type embryos were comparable. This suggests that a threshold level of Neurod1 protein is required to replace Math5 and that only a minority of RPCs expressing the knock-in alleles achieve this threshold. Accordingly, it might be expected that Math5Neurod1-KI/Neurod1-KI would have a stronger restorative effect than Math5Neurod1-KI/lacZ-KI. However,we did not observe significant differences between these two genotypes (data not shown). In fact, Math5Neurod1-KI/Math3-KI resulted in slightly better effects, suggesting that one copy of Neurod1 was sufficient to produce optimal restoration. These data imply that the threshold level of Neurod1 is modulated by other activity-limiting intrinsic factors,and one copy of Neurod1 exhausts such factors.
Our study suggests that Neurod1 can be inserted into the RPC program to re-establish RGC competence and assume the roles that normally require Math5. However, the lack of over-production of amacrine cells in Math5Neurod1-KI/Math3-KI retinas suggests that, in addition to the ubiquitously expressed homeobox genes Pax6 and Six3, other factors are required to act with Neurod1 and Math3 to reprogram an RPC towards amacrine cell production. Precocious formation of amacrine cells may have occurred in Math5Neurod1-KI/Math3-KI retinas but may not have been tolerated in this slightly earlier environment. In fact, we have detected significant cell death within the central region of Math5Neurod1-KI/Math3-KI retinas as early as E11.5,whereas no increase in cell death was seen in E11.5 Math5-null retina(Le et al., 2006), suggesting improper and precocious cell differentiation may have taken place. The modest increase in the number of cone cells in Math5Neurod1-KI/Math3-KI retina indicates that a few Math5Neurod1-KI/Math3-KI-expressing cells assume a cone cell fate - the next cell type to appear within the Math5-expressing cell lineage (Cayoutte et al., 2006). The results presented here support the view that an intrinsic program within the RPC dictates the functions of Neurod1 and Math3, when they are expressed at the Math5 locus. Therefore, the specialized functions that have presumably evolved for these bHLH factors may not be as crucial in determining early retinal cell fates as is currently thought.
The intrinsic program necessary for a naïve RPC to advance to a specific competence state is thought to arise from a dynamic local external environment. This environment changes through time and continually provides instructions to RPCs to assume successive competence states(Cayouette et al., 2006; Wallace, 2008). The discovery of numerous transcriptional regulators essential for retinal cell fate specification and differentiation has lead to the elucidation of detailed genetic regulatory pathways that define the intrinsic programs of most RPCs(reviewed by Ohsawa and Kageyama,2008; Mu and Klein,2008). However, the mechanisms connecting the dynamic local environment within the developing retina to the intrinsic genetic programs that operate in distinct RPCs remain elusive.
Although we have emphasized the evidence that Neurod1 is more capable of adapting to a foreign environment than Math3, our results also show that neither Neurod1 nor Math3 can fully restore RGCs in the absence of Math5. Neurod1 and Math3 together were slightly more effective that Neurod1 alone, which in turn was more effective than Math3 alone. Recently, it has been shown that a Math5Mash1-KI allele has only modest restorative ability (Nadean Brown, personal communication). The obvious explanation for these differences is that all of these bHLH factors have amino acid sequence differences within and outside of their bHLH domains that are likely to reflect differences in protein-protein interactions,post-translational modifications, and promoter-enhancer preferences. For example, three amino acids within the basic domain of chicken Ath5 are reported to be crucial for protein-protein interactions that confer DNA binding specificity to Ath5(Skowronska-Krawczyk et al.,2005). These residues are found in Math5 but not in Neurod1, Math3 or Mash1 (see Table S3 in the supplementary material). The helix 1 and helix 2 domains within Mash1 and Math1 are crucial in determining neuronal differentiation (Nakada et al.,2004). Thus, the sequence differences in helix 1 and helix 2 domains among Math5, Neurod1, Math3 or Mash1 might account for their differential function in the same environment (see Table S3 in the supplementary material).
In their normal cellular context, Math5, Neurod1 and Math3 are likely to regulate the expression of distinct sets of target genes during retinal development. Recent reports have identified downstream target genes for Math5,but there is little information on the overlap of these target genes with the genes regulated by Neurod1 and Math3 (Mu et al., 2005; Del Bene et al.,2007). Moreover, differences in cellular context could result in post-translational modifications that affect DNA-binding site choice. In Xenopus, a post-translational mechanism mediated by GSK-3βphosphorylation negatively regulates the ability of NeuroD to promote RGC differentiation (Moore et al.,2002). Although mouse Neurod1 lacks the GSK-3βphosphorylation site, other kinases such as ERK may modulate the function of Neurod1 when it is expressed at the Math5 locus in Math5-expressing RPCs (Dufton et al., 2005).
Several studies have reported on the effects of replacing one related transcription factor with another in a developmental context. In many cases,gene swapping demonstrates a large degree of functional redundancy and indicates that the timing of expression is perhaps more crucial than specialized functions that might have evolved. For example, in the retina, the closely related POU domain factors Pou4f1 and Pou4f2 appear to be interchangeable in their ability to function as regulators of RGC differentiation if they are expressed at the Pou4f2 locus(Pan et al., 2005). In mid-hindbrain development, the lethal En1 mutant phenotype can be rescued by replacing En1 with closely related En2(Hanks et al., 1995). However,two related bHLH genes, Mash1 and Ngn2, have been shown to maintain their divergent functions in the specification of neuronal subtype identity in the dorsal telencephalon and ventral spinal cord(Parras et al., 2002). In skeletal muscle, the myogenic bHLH regulatory factor myogenin can substitute for the closely related Myf5 factor in promoting myogenesis, although less efficiently (Wang and Jaenisch,1997). The same swap leads to complete rescue of the lethal Myf5 mutant rib phenotype (Wang et al.,1996). Similarly, in the sensory nervous system of Drosophila, the proneural bHLH factor Amos, a bHLH factor closely related to Atonal, can substitute for Atonal in specifying R8 photoreceptor fate (Maung and Jarman, 2007),whereas another bHLH factor, Sc, cannot(Sun et al., 2000). By contrast, Amos cannot rescue the chordotonal phenotype seen in Atonalmutants (Maung and Jarman,2007), suggesting that the developmental context is critical for distinct bHLH factors to exert their specific activity. Our current study suggests that developmental time and the intrinsic properties within distinct RPCs largely dictate the roles of bHLH factors in specifying early retinal cell fate.
In the developing retina, bHLH factors and other transcriptional regulators produce a highly complex combinatorial state that defines each RPC subpopulation (Ohsawa and Kageyama,2008; Mu and Klein,2008; Mu et al.,2008). Thus, each subpopulation is under the control of a specific gene regulatory network composed of hierarchical tiers of transcription factors connected to their cis regulatory sites on target regulatory genes (Ben-Tabou de-Leon and Davidson,2007). Although each network is distinct, its underlying framework is likely to be similar to that of other RPC subpopulation. The crucial nodes in the different RPC networks are likely to be represented by transcription factors of the same class. There is little question that bHLH factors have evolved specialized features for each lineage, but the fact that they are sometimes interchangeable reflects the flexibility of RPC gene regulatory networks. A given network can tolerate the replacement of one bHLH factor with another, provided the other factor has the capability of fitting into the network at the correct hierarchical level to receive the inputs and transmit the outputs that are required for successful network operation.
We thank Jan Parker-Thornburg, Sandra Rivera and the Genetically Engineered Mouse Facility at The University of Texas M. D. Anderson Cancer Center for generating the knock-in mouse lines. We also acknowledge the assistant of Jodie Polan for confocal microscopy and imaging, and Eric Meadows for qRT-PCR analysis. We thank Satchidananda Panda for providing the anti-melanopsin antibody and acknowledge the M. D. Anderson Cancer Center DNA Analysis Facility for DNA sequencing. We are grateful for insightful comments on the manuscript provided by Xiuqian Mu. The Genetically Engineered Mouse Facility and DNA Analysis Facility are supported in part by a National Cancer Institute Cancer Center Core Grant (CA016672). The work was supported by grants to W.H.K. from the National Eye Institute (EY011930 and EY010608-139005) and from the Robert A. Welch Foundation (G-0010), and by a grant from the E. Matilda Ziegler Foundation for the Blind to S.W.W.