An FGF3-BMP Signaling Axis Regulates Caudal Neural Tube Closure, Neural Crest Specification and Anterior-Posterior Axis Extension

During vertebrate axis extension, adjacent tissue layers undergo profound morphological changes: within the neuroepithelium, neural tube closure and neural crest formation are occurring, while within the paraxial mesoderm somites are segmenting from the presomitic mesoderm (PSM). Little is known about the signals between these tissues that regulate their coordinated morphogenesis. Here, we analyze the posterior axis truncation of mouse Fgf3 null homozygotes and demonstrate that the earliest role of PSM-derived FGF3 is to regulate BMP signals in the adjacent neuroepithelium. FGF3 loss causes elevated BMP signals leading to increased neuroepithelium proliferation, delay in neural tube closure and premature neural crest specification. We demonstrate that elevated BMP4 depletes PSM progenitors in vitro, phenocopying the Fgf3 mutant, suggesting that excessive BMP signals cause the Fgf3 axis defect. To test this in vivo we increased BMP signaling in Fgf3 mutants by removing one copy of Noggin, which encodes a BMP antagonist. In such mutants, all parameters of the Fgf3 phenotype were exacerbated: neural tube closure delay, premature neural crest specification, and premature axis termination. Conversely, genetically decreasing BMP signaling in Fgf3 mutants, via loss of BMP receptor activity, alleviates morphological defects. Aberrant apoptosis is observed in the Fgf3 mutant tailbud. However, we demonstrate that cell death does not cause the Fgf3 phenotype: blocking apoptosis via deletion of pro-apoptotic genes surprisingly increases all Fgf3 defects including causing spina bifida. We demonstrate that this counterintuitive consequence of blocking apoptosis is caused by the increased survival of BMP-producing cells in the neuroepithelium. Thus, we show that FGF3 in the caudal vertebrate embryo regulates BMP signaling in the neuroepithelium, which in turn regulates neural tube closure, neural crest specification and axis termination. Uncovering this FGF3-BMP signaling axis is a major advance toward understanding how these tissue layers interact during axis extension with important implications in human disease.


Introduction
Axis extension in vertebrate embryos requires the coordination of numerous genetic and morphological processes within and between the three germ layers: the endoderm, mesoderm and the ectoderm-derived neuroepithelium. This coordination requires numerous signaling factors for regulation of key events that must occur in a proper sequence.
An example of this is the neuroepithelium, which undergoes morphological movements that transform a sheet of cells into the neural tube. Underlying and driving this essential process are a number of signaling pathways, genetic and epigenetic mechanisms and environmental cues [1][2][3]. Considering this complexity, it is understandable that neural tube defects (NTDs) are among the most prevalent birth defects in the human population (1:1000) [4] and quite a large number of mouse NTD models (exceeding 240) have been identified [5][6][7].
Neurulation during vertebrate axis extension is divided into primary and secondary phases. During primary neurulation, the neural tube forms by folding of the neural plate so that the two edges of the plate are spatially near allowing for fusion at the dorsal midline [8]. Closure along the spinal axis initiates at the hindbrain/ spinal cord boundary and proceeds caudally, ending with closure of the posterior neuropore in the trunk region. Failure to close the neuropore results in spina bifida (Latin for "split spine"), which remains the most frequent type of human NTD, despite a decrease in frequency after the mandatory addition of folic acid to enriched grain products in the USA [9]. Spina bifida has a 1-year survival rate of 79 to 96%, depending on race/ethnicity [9]; however, surviving patients often experience lifelong disabilities and mortality increases into young adulthood [10].
Closure of the posterior neuropore marks the end of primary neurulation. Neural tube formation caudal to this region occurs via the process of secondary neurulation, involving a mesenchymal to epithelial transition of cells within the tailbud to generate a tube [11,12]. These cells reside within a region of the tailbud called the "chordoneural hinge" (CNH) that contains not only neural tube progenitors but also cells that generate caudal somites and notochord [13][14][15].
Concomitant with axis extension and neurulation, is the formation of the neural crest: a transient, migratory lineage that delaminates from the dorsal neural tube and forms various cell types that include the melanocytes, craniofacial cartilage and bone, and peripheral and enteric neurons/glia. The timing of trunk neural crest emigration has been shown to be controlled by caudal FGF signals and rostral retinoic acid signals from the adjacent mesoderm [16].
Human neurocristopathies, syndromes caused by dysregulation of the neural crest, occur as often as NTDs [17,18], and there is a limited overlap of the two types of syndromes [17]. Likewise, there is a subset of mouse mutants where defects occur in both neural tube closure and neural crest development [5]. These include mutations in components of the BMP signaling pathway, demonstrating that the proper regulation of this pathway is essential to both events [19][20][21][22]. In the caudal trunk, exogenous BMP will inhibit the necessary morphogenic movements within the neuroepithelium, which are necessary for neuropore closure, while premature closure occurs in Bmp2 null homozygotes [23]. However, after closure, BMP signaling is required for neural crest delamination from this region. Both BMP4 and BMP7 will induce caudal neural crest cell programs [24] and data from manipulation of the chick embryo support the idea that BMP4 in the dorsal neural tube is required for the delamination of caudal neural crest [25][26][27]. Thus, dorsal BMP activity in the caudal neural tube must be countered because it impedes neural tube closure but afterwards is required for neural crest development. One way the vertebrate embryo carefully calibrates BMP activity in this region is by a balance between BMPs and their antagonist, NOGGIN [25,26]. Consistent with this idea, Noggin null homozygotes display both an open neural tube and excessive neural crest development [23,[28][29][30].
Here we demonstrate that FGF3 signals from the presomitic mesoderm (PSM) provide another level of regulation of caudal neural tube BMP signaling. Multiple FGFs are expressed in the presomitic mesoderm and we have recently demonstrated that both Fgf4 and Fgf8 can each encode a classical activity known as the "wavefront" that prevents differentiation of the PSM during somitogenesis [31]. Simultaneous inactivation of both of these Fgf genes, at various embryonic timepoints, causes an AP axis truncation [31,32]; however, inactivation of either of these Fgfs alone causes no apparent extension defect, demonstrating redundancy of this system [31,33]. Furthermore mouse single gene inactivation studies demonstrate that of the five other Fgfs expressed in the PSM (Fgf3, Fgf5, Fgf15, Fgf17, and Fgf18), only Fgf3 has a requirement in axis extension (reviewed in [27]). Our studies reveal that the Fgf3 axis truncation is caused by an upregulation in BMP signals from the caudal neural tube and that these signals also cause an increase in neuroepithelial proliferation, delay in neural tube closure and premature neural crest formation.

Results
Fgf3 expression in the nascent mesoderm is required for normal axis extension Fgf3 Δ/Δ ("Δ" = deletion and null) mice used in this study [34] have a shortened tail (Fig 1A) similar to that in mice homozygous for the first Fgf3 null allele generated by Capecchi and coworkers [35,36]. This defect is 100% penetrant, with variable expressivity in two parameters: the total tail length ( Fig 1A) and the axial position of the anterior-most (and therefore the earliest to develop) malformed vertebra (Fig 1B). To exclude the possibility that genetic background may affect these parameters, in all genetic experiments described below, we always used littermates for control genotypes (Table 1).
To understand the Fgf3 phenotype, we first considered where the gene is expressed. Fgf3 transcripts have been detected in the nascent mesoderm starting in the primitive streak at E7.5 through E9.5 [37], before closure of the posterior neuropore around E10.0 [38]. We confirmed Fgf3 mRNA is detected in the primitive streak (ps, E7.5, E8.5) as well as in the presomitic mesoderm (E8.5-12.5) and tail bud (E9. 5-12.5). Expression is absent at E13.5. An apparent dynamic domain of Fgf3 is observed in the anterior PSM (asterisks) (see S1 Fig). Fgf3 expression is confined to mesoderm and absent in the posterior neural tube (C", with approximate plane of section shown by dotted line in C'. NT, neural tube; psm, presomitic mesoderm, 24 ss); E7.5-E13.5 images taken at the same magnification (scale bar, E13.5); images in C' and C" taken at the same magnification (scale bar, C"). (D-M) WISH assay for Uncx4.1, which stains the somites, and Msgn1, which stains the caudal PSM, in mutants and littermate controls at the stages indicated. Caudal PSM (Msgn1 domain) is depleted at E12.5 in mutants and E13.5 in controls. D-I are dorsal views; J-M are lateral views. Arrowhead in M indicates touching Uncx4.1 domains. Bars in F and G indicate total PSM length, which is quantified in N. Note that a decrease in the amount of mutant PSM begins at 37 ss followed by a total loss of PSM around 54 ss, resulting with fewer somites in mutants. Error bars represent SEM; significance determined by two-tailed t-test.
and extended these analyses and found that Fgf3 expression is maintained in the nascent mesoderm, emerging from the primitive streak or tail bud, throughout development until axis extension ceases at E13.5 ( Fig 1C); Fgf3 expression is restricted to the tail bud mesoderm and not found in the adjacent neuroectoderm ( Fig 1C' and 1C"). We also noted that an Fgf3 expression domain exists in the anterior presomitic mesoderm (PSM) that apparently oscillates because its dimensions varied between embryos of the same embryonic stage (asterisks in Fig 1C and 1C'; S1 Fig).
Given this oscillatory expression domain and the work we and others have published demonstrating that different FGFs play essential roles controlling oscillatory activities during somitogenesis [31,32,39,40], we considered whether the Fgf3 axis defect might be due to dysregulation of normal gene oscillations in the caudal embryonic axis. To address this, we examined Hes7 expression, a key clock component that responds to FGF signals [31,41,42], in mutant and control embryos from E10.5 to E10.75. We found that Hes7 expression oscillates in Fgf3 null homozygotes in an approximately normal pattern although both the caudal somites and the PSM were abnormally shorter in the anterior-posterior axis (S2 Fig).
We examined this size reduction in the mutant PSM and somites in more detail by simultaneously determining the expression of both Uncx4.1, which marks the posterior half of each somite [43] and Msgn1, which marks the posterior PSM [44]. The total PSM is the area between Uncx4.1 expression in the caudal-most somite and the posterior limit of Msgn1 expression (black bars in Fig 1F and 1G). Furthermore, the Msgn1 expression domain is a readout of wavefront activity [45,46] and as such, regresses as the posterior axis extends, eventually dwindling in the growing caudal axis and then disappearing as axis extension ceases [47]. In the developmental time window we examined (E9.5-E13.5) control PSM length is highest at 20 somite stage (ss) (about E9.25), rapidly decreases till~28 ss (about E10.0), and then remains relatively unchanged for the next 2 days (Fig 1D, 1F, 1H and 1N). Then on day E12.5 the amount of PSM declines rapidly so that by E13.5 no Msgn1 expression can be observed because all PSM has been incorporated into somites, forming a total of~64 somites ( Fig 1J, 1L  and 1N). These changes in PSM length resemble those originally described by Tam [48]. Fgf3 mutants have comparable amounts of PSM from 20 ss (~E9.5) (compare Fig 1D and 1E) up to 37 ss (~E10.5) when they possess significantly less PSM (Fig 1N and compare Fig 1F and 1G) and at E12.5 lose all PSM, a full day earlier than controls ( Fig 1K). Because of this aberrantly early PSM loss, in this genetic background, Fgf3 mutants form an average of~10 fewer somites than control littermates ( Fig 1N). The size and shape of the Fgf3 mutant caudal somites are also clearly smaller at E11.5 (compare Fig 1H and 1I). A consistent dorsal curve of the axis causes the dorsal distance between somites to be shorter than the ventral distance (Fig 1K and  1M) and in some cases, the two Uncx4.1 domains are fused (arrowhead in Fig 1M). These changes in mutant somite shape and number sufficiently account for the final vertebral defects observed in adults (Fig 1A and 1B).
We explored whether these changes in somite morphology were preceded by changes in dorsal-ventral (DV) somite patterning by examining expression of Pax1 and Pax3, [49,50]. This analysis revealed that DV patterning was not affected in Fgf3 mutants (S3 Fig). Another consideration is that the simultaneous loss of both Fgf4 and Fgf8 expression in the PSM affects axis extension [31,32]. We found that expression of neither of these Fgf genes was affected at stages preceding PSM loss (S4 Fig), however at stages when the PSM is reduced we did observe a similar loss of Fgf8 expression as a consequence of tissue loss (inset C, D S4 Fig). In addition, expression of Spry2 and Etv4, two "FGF-target" genes responsive to Fgf4 and Fgf8 [31,32], was also not affected (S4 Fig). Thus FGF3 is not necessary for Spry2 and Etv4 expression, an observation that builds on our previous insight that FGF3 was not sufficient, when Fgf4 and Fgf8 were inactivated, for the expression of these FGF target genes [31]. Therefore Fgf3 is uniquely required for caudal axis extension, independent of Fgf4 and Fgf8 function.
Fgf3 mutants display aberrant cell death in the tail bud region Premature loss of Fgf3 mutant PSM could be explained by a decrease in cellular proliferation and/or an increase in cell death. We measured the rate of proliferation in the PSM at 28 ss and 34 ss and found there to be no difference between mutants and controls (S5 Fig). To determine if there were any changes in cell death within the PSM we stained with LysoTracker Red, which marks domains of cell death [51], in 28 ss (E9.5), 36 ss (E10.5), and 44 ss (E11.5) mutant and control embryos. At 28 ss mutants and control littermates had a similar pattern of staining with a scattering of LysoTracker Red-positive cells in the dorsal neural tube and surrounding caudal end of the tail gut (Fig 2A and 2B). However, beginning around the 36 ss and continuing through 44 ss, an aberrant domain of cell death is observed in the dorsal tail bud caudal to the neural tube (Compare Fig 2D with 2C and 2F with 2E).
To more precisely locate this region of aberrant cell death we immunostained sagittally sectioned mutants for cleaved-Caspase3 [52] and found that the aberrant cell death domain occurs in the CNH region, where progenitors reside that give rise to the mesoderm during secondary axis extension [13][14][15]. To examine the state of these progenitors, we took advantage of a unique quality of Fgf8 expression that allows us to specifically label these cells. Although Fgf8 mRNA is found throughout the PSM, transcription only occurs in the tailbud progenitors that generate the PSM (S6 Fig). As new mesoderm is formed from the progenitors, Fgf8 mRNA perdures in the daughter cells, as they are displaced from this region and populate the PSM. Decay of mature Fgf8 transcripts then generates the observed Fgf8 mRNA expression gradient [53]. This observation demonstrates that the cells actively transcribing Fgf8 are the progenitors of the PSM. Hence, whereas an exonic riboprobe marks both the PSM and its progenitors, an intronic probe (Fgf8intron) can be used to label only the progenitors that are located within the CNH (S6 Fig).
At 28 ss, prior to aberrant cell death, no difference was observed in the intronic Fgf8 staining domain between Fgf3 homozygotes and littermate controls (Fig 2J and 2I). However, by 38 ss, the Fgf8intron domain in mutants was reduced in the dorsal posterior tailbud (compare Fig 2L  with 2K) and was absent by 44 ss (compare Fig 2N with 2M), a day before the PSM is prematurely exhausted (Fig 1K and 1N). These data led us to initially hypothesize that Fgf3 mutants were losing PSM progenitors, and therefore PSM, due to aberrant cell death.
It is important to distinguish between the loss of the Fgf8intron signal that is the result of the Fgf3 defect and reflects loss of PSM progenitors and the possible contribution of the loss of FGF8 activity to the Fgf3 axis defect. As mentioned above, analysis of Fgf8 exon staining and downstream FGF8 target genes indicates no loss of Fgf8 expression except that due to loss of the PSM tissue (S4 Fig). This suggests no contribution of Fgf8 loss to the Fgf3 defect. To test this directly, we used TCre [31,33] to inactivate Fgf8 in the gastrula mesoderm in an Fgf3 null background and found that this early inactivation of Fgf8 had no effect on the Fgf3 defect (average number of caudal vertebrae: 20.9 for TCre; Fgf8 flox/wt ; Fgf3 Δ/Δ , n = 7 vs. 22.0 for TCre; Fgf8 flox/Δ ; Fgf3 Δ/Δ , n = 6).

Restoration of cell survival in Fgf3 mutants exacerbates the mutant phenotype and reveals a neural tube defect
To determine if aberrant cell death was responsible for premature progenitor loss and hence Fgf3 caudal defects, we sought to restore cell survival by generating Fgf3 mutants that also lacked Bak and Bax, which encode pro-apoptotic BCL-2 family members [54][55][56]. We scored the two parameters of the Fgf3 defect that we described above in Fig 1A and 1B: 1) the first vertebral body (that is, the most anterior) to be defective and 2) the total number of caudal vertebrae. To exclude genetic background as a factor that might affect variability in these parameters, we intercrossed Fgf3 Δ/Δ ; Bak Δ/wt ; Bax Δ/wt mice to produce progeny that were triple null for all three loci as well as controls that were only Fgf3 null homozygotes (Table 1).
Mice doubly homozygous null for Bak and Bax have no detectable cell death in the tailbud and form normal caudal vertebrae (S7 Fig). Cell death is also absent in all tailbud tissues of E10.5 Fgf3; Bak; Bax triple null homozygotes, including the dorsal neural tube, gut tube and the aberrant PSM domain (compare Fig 3B, with 3A).
Remarkably, removal of Bak and Bax, rather than rescuing the Fgf3 defect, worsened it. The dorsal curve of the posterior axis was more pronounced and occurred more anteriorly in compound mutants at E10.5 (compare Fig 3B with 3A) as did the first vertebral defect at E18.5 ( Fig  3I). The PSM progenitor domain, marked by the Fgf8intron signal, was further reduced in triple Fgf3; Bak; Bax E10.5 mutants (compare Fig 3D with 3C) and, likewise, the total number of caudal vertebrae formed was significantly reduced ( Fig 3I). Lastly, Fgf3; Bak; Bax triple null homozygotes displayed a new defect not immediately apparent in embryos lacking only To generate such mutants at a higher frequency (one in four) we intercrossed Fgf3 Δ/Δ ; Bak Δ/Δ ; Bax Δ/wt mice ( Table 1) and found that of thirty E18. To determine if altered proliferation affected the closure of the neural tube of these mutants, we examined rate of proliferation in E9.5 (26-28 ss) and E10.5 (32-34 ss) triple mutant neural tubes. We noted a trend of greater proliferation at the earlier stages in triple mutants but no difference at the later stages (S5B Fig). Therefore we conclude that an increase in cell number through a lack of cell death, and possibly through increased proliferation, leads to the observed failure of neural tube closure. These defects led us to consider whether there were defects in neural tube closure of simple Fgf3 mutants.
The earliest morphological defect in Fgf3 null homozygotes is a delay in neural tube closure The original characterization of the axis defect in Fgf3 null homozygotes found no phenotype in neural tube closure at E11.5, although it was noted that such mutants had an abnormally large neural tube at this stage [36]. We reexamined the Fgf3 mutant neural tube at earlier developmental stages, motivated by our observations of the severe open NTDs in Fgf3 Δ/Δ ; Bak Δ/Δ ; Bax Δ/Δ mice ( Fig 3F). Prior to 25 ss (E9.5), the width of mutant and littermate-control neural tubes appeared similar ( Fig 4C). However, at 25 ss the Fgf3 homozygous caudal neural tube starts to appear abnormally wide (Fig 4A-4C). We assayed for cell proliferation in Fgf3 mutants and controls at 26 ss and found that, indeed, Fgf3 mutants had an increase in the rate of proliferation (Fig 4D-4F). Similar to Fgf3 Δ/Δ ; Bak Δ/Δ ; Bax Δ/Δ mutants the change in proliferation detected at E9.5 was not observed at E10.5 (32-34 ss, S5 Fig).
To determine exactly when the Fgf3 mutant neural tube closes, we injected India ink into the neural tube at about the position of the third somite rostral to the PSM; in embryos with an open neuropore, the ink is easily viewed escaping the neural tube lumen (Fig 4G). Such experiments revealed that neuropore closure begins around 27 ss in control embryos but is delayed by about 1 somite stage (2 hours) in the Fgf3 mutant ( Fig 4G). Because this defect is the earliest observed and occurs prior to defects in axis extension (Fig 1), it is the primary Fgf3 mutant phenotype. As development proceeds, the mutant neural tube remains abnormally large ( Fig  4C) but the posterior neuropore closes in most mutant embryos; we have observed occasional individuals with an open neuropore at E10.5 (Fig 4H).
NTDs can be affected by diet and environment [1]. In humans, dietary folic acid (FA) intake during pregnancy has been shown to reduce the occurrence of human NTDs [57]. In mice, the effect of FA supplementation has been explored in twenty-three genetic models to date where it reduced NTD occurrence in seven mutants [58][59][60][61][62], had no effect in twelve mutants  [59,61,63-66] and either increased embryonic lethality or induced a more complicated response in four mutants [61,67,68]. To test if the Fgf3 phenotype is folate-sensitive we crossed Fgf3 mutant heterozygotes to homozygotes, and intraperitoneally injected pregnant dames at E7.75 with FA. Such injections had no effect on mutant neural tube closure at E9.5 (Fgf3 heterozygote: average width: 0.882, n = 3; Fgf3 mutant: average width: 1.887, n = 3). However, at E18.5 only 5/21 (24%) mutants were present instead of the expected 50% (which indeed we observed in uninjected control experiments), suggesting that FA injections caused embryonic lethality of homozygotes. Thus Fgf3 apparently falls into the gene subset wherein FA treatment causes embryonic lethality of homozygous null mutants [61,67,68].
BMP signals are increased in the caudal regions of the Fgf3 mutant neural tube BMP activity inhibits morphological movements necessary for neural tube closure. Bmp2, Bmp4 and Bmp7 are all expressed in the dorsal neural tube at the time of posterior neuropore closure with Bmp4 and Bmp7 expression occurring rostral to the posterior neuropore [19,23,69]. Noggin, which encodes a BMP antagonist, is specifically expressed in the caudal regions of the developing neural tube, inhibiting BMP activity and thus allowing closure of the posterior neuropore [1,70].
We found that neural tube-specific Noggin expression was unchanged in Fgf3 mutants starting at 28 ss, when the posterior neuropore is closing, as well as later stages (S9 Fig). However, we found that Bmp4 expression was initiated prematurely in the Fgf3 mutant caudal neural tube at 24 ss (Fig 5A-5B'), and remained aberrantly upregulated and extended caudally from 28 ss through 34 ss (Fig 5C, 5D, 5I, 5J and 5K). At 28 ss expression of Bmp7 was also increased and shifted caudally in mutants (Fig 5E, 5F and 5I). To determine the outcome of BMP activity in mutant neural tubes, we assayed for expression of Msx1, a downstream BMP target gene [71][72][73]. In all Fgf3 mutants, Msx1 expression in the dorsal neural tube was upregulated and shifted caudally (compare Fig 5G, 5H, 5I, 5L and 5M).
Neural Crest is prematurely specified in the caudal regions of the Fgf3 mutant neural tube Neural crest migration within the embryonic trunk is thought to be triggered by BMP4-dependent WNT1 activity [74]. Therefore, based on our observation that BMP activity was precociously upregulated in the absence of Fgf3, we asked whether trunk neural crest specification was affected in these mutants. We found that the caudal limit of Wnt1 expression was extended at 30 ss and remained upregulated through 36 ss (Fig 6A-6D). We also examined Wnt3a expression, which is also a marker of caudal neural crest as well as the tailbud PSM [74]. We found Wnt3a expression, like Wnt1, was notably upregulated in the mutant dorsal neural tube from 28 ss through 34 ss (Fig 6E-6H). Interestingly, we found no change in the PSM domain of Wnt3a in these samples, suggesting that altered WNT signals from this region do not cause the Fgf3 defect (Fig 6E-6H).
To further examine neural crest development in Fgf3 mutants we assayed for expression of Snai1 (neural crest specification), Foxd3 (premigratory and migratory neural crest) and Sox10 (migratory neural crest) [30,75,76,77]. The caudal limit of Snai1 expression was extended in mutants, suggesting premature specification of neural crest in the absence of Fgf3 (Fig 6I and  6J). This idea was confirmed by a caudal extension of both premigratory and migratory neural crest in mutants, as indicated by Foxd3 and Sox10 expression at all stages examined (Fig 6K-6P). We then examined the anatomy of caudal dorsal root ganglia, which are derived from the neural crest, in E11.5 mutants by immunostaining for Neurofilament1. This analysis revealed no changes between mutant and littermate controls (S10 Fig). Therefore we conclude that the increase in BMP activity in the caudal dorsal neuroepithelium of Fgf3 mutants causes a premature specification of neural crest, but without an overall increase in neural crest formation.

Increased BMP activity causes the axis truncation defect in Fgf3 mutants
We noted that Msx1 expression in Fgf3 mutants occurred in the PSM progenitor domain at 34 ss (compare Fig 5M with 5L), suggesting that these cells were receiving BMP signals (a similar Msx1 expression domain occurs in normal controls, but about a day later, as shown in   (Fig 2L), we asked whether increased BMP signals cause this loss. To test this, we grafted BMP4-soaked beads in the wildtype neural tube just rostral to the progenitor domain and determined that we could readily induce expression of the BMP target gene Msx1 (compare Fig 7B with 7A). In all such cases (n = 5), we found a loss of Fgf8intron signals (compare   Fig 7D with 7C), suggesting that increased BMP activity within the progenitors was sufficient for the loss of this population, precipitating shortening of the axis.
To test our hypothesis that an increase in BMP signaling was the cause of these defects in vivo we assessed whether they are worsened by increasing BMP signaling or improved by decreasing BMP signaling. We did this genetically, by generating Fgf3 homozygous embryos and animals that also carried mutant alleles in key BMP signaling genes.
One of these genes is Noggin (Nog), which encodes the BMP antagonist that plays an essential role in the normal inhibition of BMP signals necessary for posterior neuropore closure. To evaluate the consequences of increased BMP signaling, we generated Fgf3 null homozygotes that are also heterozygous for a Nog null allele (Nog lacZ/wt ) [28] (Table 1). It is important to note that such Nog null heterozygotes are phenotypically normal in a wildtype background. Msx1 transcripts were more readily detected within the neural tube of Fgf3 Δ/Δ ; Nog lacZ/wt embryos, when compared to littermates that were simple Fgf3 null homozygotes, demonstrating the expected relative increase in BMP signaling (S12 Fig). At 27 ss, these compound mutants, when compared with control Fgf3 null homozygotes, displayed a neural tube that was on average 35% wider (compare Fig 8B with 8A, quantified in 8C, S12 Fig). This worsening of the neural tube defect is accompanied by an increase in proliferation within the neuroepithelium (Fig 8D). Likewise all other Fgf3 defects are exacerbated in these compound mutants: the PSM progenitor domain, marked by the Fgf8intron signal, is further decreased (compare Fig  8F with 8E) and the caudal limit of neural crest markers is further expanded (compare Fig 8H  with 8G). These aggravations of the Fgf3 null defects in early embryos prefigured changes in later development: in E18.5 Fgf3 Δ/Δ ; Nog lacZ/wt embryos, there was an anterior shift in the position of the first defective vertebra as well as an overall reduction in the total number of caudal vertebrae (compare Fig 8J with 8I, quantified 8K). Therefore, increasing BMP signaling increases the severity of the NTDs, neural crest expansion and PSM loss in Fgf3 mutants.
We then asked the converse question: does reducing BMP signaling improve the Fgf3 defects? To address this, we generated Fgf3 null homozygotes that were also lacking Bmpr1b [78] (Table 1), which is expressed in the developing neural tube [79] and encodes a type-1 BMP-receptor. Mice lacking Bmpr1b have no obvious neural tube or axis extension/vertebral defect [80].
We confirmed that BMP signaling was decreased in Fgf3 Δ/Δ ; Bmpr1b Δ/Δ embryos by comparing the intensity of Msx1 expression in such double mutants to that found in littermates that were null only for Fgf3 (S12 Fig). At 28 ss double mutants displayed a neural tube that was significantly narrower, and therefore more normal, than that found in Fgf3 homozygotes (compare Fig 9B with 9A, quantified 9C, S12 Fig) with a corresponding reduction in cellular proliferation within the neuroepithelium (Fig 9D). We observed a definite restoration of the PSM progenitor domain at 34 ss (compare Fig 9F with 9E). Likewise, the caudal limit of neural crest formation was also restored to a more normal level at 34 ss (compare Fig 9H with 9G). At E18.5 we observed a significant shift toward normality in Fgf3 Δ/Δ ; Bmpr1b Δ/Δ mutants in both the position of the first vertebral defect and the total number of caudal vertebrae (compare Fig  9J with 9I, quantified in 9K). Thus, these data support our hypothesis that BMP signals from the neural tube in Fgf3 null homozygotes cause abnormal axis extension and reducing BMP signaling in these mutants partially rescues all the phenotypic defects.

Discussion
Fgf3 holds a special place in the history of genetics as the first non-selectable locus to be inactivated in the mouse by gene targeting [81]. Mice lacking Fgf3 have defects in ear and tail development; although much has been studied about the former phenotype, heretofore little beyond the initial characterization has been analyzed regarding the tail defect [35,36]. Our findings are summarized in Fig 10. We demonstrate that the tail defect is preceded, at E9.5, by premature initiation of BMP signals (both Bmp4 and Bmp7) from the dorsal neural tube. This has two immediate consequences within the neuroepithelium: a delay in closure of the posterior neuropore and abnormal expansion of neural crest during primary neurulation.
Increased Bmp expression in the dorsal neural tube can be expected to increase Wnt gene expression, and through WNT action, increase both cellular proliferation and neural crest [19,23,26,27,74,82]. Consistent with this, we observe both an increase in Wnt1 and Wnt3a in the caudal neural tube and an increase in proliferation in the Fgf3 neural tube. Excessive proliferation has been associated with neural tube closure defects [83,84] not surprisingly given the complex patterns of proliferation and cell movement that accompany neural tube closure [85].
When we genetically ratchet BMP signaling up or down, which increases or decreases Fgf3 neural tube defects, we observe a concomitant increase or decrease, respectively, in proliferation. Hence increased proliferation, through increased BMP activity, may lie at the heart of the Fgf3 neural tube phenotype.
At E10.5, 24 hours after BMP signals have increased, we observe the first evidence of an axis truncation: loss of PSM progenitor cells (a reduction of the Fgf8intron domain) and a shorter PSM (a reduction in the Msgn1 expression domain). These changes occur during secondary axis formation; thus the Fgf3 phenotypes straddle primary and secondary neurulation and axis formation. We show that exogenous BMP4 is sufficient to reduce the progenitor domain of the PSM. Furthermore, just as our genetic manipulations of BMP signaling in Fgf3 mutants affect neural tube phenotypes, they likewise affect axis extension defects, either increasing them when one Noggin allele is removed or decreasing them when Bmpr1b is inactivated. One way to look at Fgf3 mutants is that they provide a genetic background that is sensitized to differences in BMP signaling.

Does FGF3 signal directly to the neural tube?
Although our data supports a model in which FGF3 regulates BMP signals in the neuroepithelium, this regulation could be achieved through FGF3 signaling directly to the neuroepithelium or indirectly through an intermediate. While the current study does not conclusively prove either of these possibilities, we favor the former idea (direct signaling). This is because we find no evidence FGF3 signals to the PSM, regulating another factor that signals to the neural tube, e.g. there is no reduction in the expression of the known FGF-target genes Spry2 or Etv4 within the PSM of Fgf3 mutants (S4 Fig). We have previously published data demonstrating that these  target genes are downregulated when Fgf4 and Fgf8 are inactivated within the PSM, despite Fgf3 expression, indicating that FGF3 is not sufficient for their expression [31].
Direct FGF3 signals to the neural tube may occur through FGFR1 or FGFR2, as mutations of either result in neural tube malformations. Fgfr1 homozygous null development arrests prior to neural tube closure [86,87], but chimeric embryos with contributions of Fgfr1 null cells in the posterior neural tube displayed spina bifida and kinked tails [88]. Furthermore, embryos that carry FGFR1 mutations that have abrogated signal transduction to the MAPK and PI3K pathways have NTDs [89]. Equally compelling is the case for FGFR2. The IIIb isoform of FGFR2 binds FGF3 and is preferentially expressed in epithelia, such as the neuroepithilium [90]. Embryos homozygous for mutations that delete only this Fgfr2 isoform have tail defects identical to those in Fgf3 mutants (neural development was not examined in these studies) [91,92]. Neural tube-specific deletions of these receptor genes are needed to explore whether FGF3 signals directly to the neuroepithelium.

Fgf3 signaling and genetic redundancy
Although Fgf3 expression is continuous in the emerging mesoderm starting at E7.5 (this study and [37]), the earliest consequence of Fgf3 loss is not observed until 24 ss (about E9.5). Therefore if an earlier FGF3 requirement from this domain exists, it must be redundant with another FGF(s). Genetic studies in the mouse reveal that FGF3 shares function with FGF8 and FGF10 in inner ear development [34,[93][94][95], with FGF10 in cardiovascular development [96] and with FGF8 in pharyngeal morphogenesis [97]. Candidate Fgf genes expressed within or near the PSM that may function along with Fgf3 include Fgf4, Fgf5, Fgf8, Fgf15, Fgf17 and Fgf18 [98], with Fgf4 and Fgf8 arguably being the strongest candidates based on their requirements in axis extension [31,32]. We have recently published that Fgf3 and Fgf4 are functionally redundant for mesoderm maintenance in the tailbud during secondary axis formation; however we do not see additional defects in mesoderm or neural tube patterning in the primary axis [99]. In the present work we demonstrate that no functional redundancy exists between Fgf3 and Fgf8.
We propose that Fgf4 and Fgf8 are required throughout axis extension for PSM gene expression and maintenance whereas FgF3 is required during a window during the transition between primary and secondary axis extension for neuroepithelium gene expression and normal morphogenesis through BMP action. In support of this idea, expression of FGF target genes in the PSM fails when Fgf4 and Fgf8 are inactivated [31,32], even in the presence of persistent Fgf3 expression [31,32]. Although Boulet and Capecchi found neural tube defects when Fgf4 and Fgf8 were inactivated, it is unclear whether these defects were due to a primary or secondary effect of this phenotype. Furthermore, Boulet and Capecchi found that expression of Wnt3a, the only neural crest marker in their study, was not upregulated in the neural tube [31,32], as we observe in Fgf3 mutants (Fig 6).

Cell death in neural tube closure and axis extension termination
We discovered the neuropore closure defect in Fgf3 mutants because it was dramatically exacerbated in embryos lacking the pro-apoptotic factors BAK and BAX in which apoptosis was abolished. In normal development, cells at the neural ridges undergo apoptosis before and after neural tube closure [100][101][102]. Whether this apoptosis is required for normal neural tube closure is somewhat controversial. Inhibiting apoptosis with the pan-caspase inhibitor, zVADfmk, causes a failure of neural tube closure in avians [102] but not in mice [103]. Although mice lacking genes that encode intrinsic apoptotic factors such as Casp3, Casp9 or Apaf1 have neural tube closure defects [104][105][106], Copp and coworkers have found that the forebrain and spinal regions of the neural tube close in Casp3 and Apaf1 mutants [103]. Conflicting data concerning the requirement for these factors may be explained by sensitivities of these defects to genetic background [104]. Also, the requirements for apoptosis may be subtle; Yamaguchi et al demonstrated that when closure occurs in Apaf1 and Casp3 mutants, it does so with delayed kinetics [101], which is what we observe in Fgf3 null homozygotes.
Surprisingly, these defects in Fgf3 mutant neural tube closure were worsened upon abrogating cell death when we additionally deleted BAK/BAX function. In these Fgf3; Bak; Bax triple null mutants, not only was the abnormal PSM cell death prevented, but normal cell death in the dorsal neural tube was also inhibited (Fig 3A, bracket). This may result in retention of neural BMP-expressing cells, which, together with enhanced Bmp expression due to a lack of Fgf3, causes overall high BMP levels causing the observed severe open NTD. Indeed, we find higher Bmp4 expression in the dorsal neural tube of Fgf3; Bak; Bax triple null homozygotes compared to simple Fgf3 mutants (S13 Fig). Furthermore, there are increased migratory and premigratory neural crest cells in the caudal neural tube of these compound mutants (S13 Fig), consistent with the hypothesis of an increase of BMP signals; this interesting observation clearly warrants further study.

One question our data raises is whether the BMP-mediated loss of PSM progenitor cells in
Fgf3 mutants is an acceleration of a process that occurs during normal axis termination. Prior to normal termination of the axis, PSM-specific Fgf3 expression is downregulated (Fig 1C) and BMP signaling increases in the PSM progenitor domain of wildtype embryos (as noted by expression of Msx1 as shown in S11 Fig). This is consistent with the idea that an FGF3-BMP signaling axis may terminate anterior-posterior extension. Genetic proof of this hypothesis would be the production of mice with longer tails due to the inactivation of gene(s) encoding BMP-signaling components. Although we observe no such effect in Bmpr1b null homozygotes (the average number of caudal vertebrae was 30 in both Bmpr1b Δ/Δ , n = 4, and Bmpr1b wt/wt , n = 3), other candidate genes remain to be explored.
In chick and fish embryos, current models of axis termination propose that a downregulation of FGFs and WNTs and rising retinoic acid (RA) levels lead to death of the PSM progenitors [107][108][109][110][111]. However, experiments in the mouse suggest that axis termination occurs in the absence of RA [112]. Furthermore, axis truncations due to PSM-specific inactivation of Fgfr1 or Fgf4/Fgf8 occur without changes in RA levels [32,39]. Whatever the signals involved, our data suggest that, at least in the mouse, cell death is not required for PSM progenitor loss, either normally during axis extension or prematurely in Fgf3 mutants, because we find that if we inhibit apoptosis by Bak/Bax inactivation, axis termination still occurs in both cases (S7 Fig, Fig 3).
Potential FGF3 role in human neural tube closure defects Characterization of mouse tail mutants have led to insights into human syndromes affecting axis extension. For example, mutations in the T locus, which cause a short tail in mice [113], are associated with spina bifida and vertebral malformations in humans [114,115]. Therefore, we speculate that human FGF3 mutations may also contribute to abnormal neural tube closure or vertebral malformation. This would be consistent with Fgf3 targeted mouse alleles, which model defects in human tooth and ear development where FGF3 mutations cause congenital deafness, microtia, and microdontia. (<http://omim.org/entry/164950?search=fgf3&highlight=fgf3>http://omim. org/entry/164950?search=fgf3&highlight=fgf3).
In support of this idea, a genomewide screen for human NTDs found linkage at regions of chromosome 10 (10q25.3) that include FGFR2 [116], which, as discussed above, encodes a likely FGF3 receptor. Interestingly, Hillbertz et al suggested that mutations within the region of chromosome 11 (11q13) that contains the closely linked genes, FGF3, FGF4 and FGF19, may cause human NTDs based on their finding that a duplication of this region causes the hair ridge and predisposition to dermoid sinus in Ridgeback dogs [117]. Dermoid sinus is caused by an incomplete separation of the neuroepithelium and ectoderm during neural tube closure and, in humans, is often associated with spina bifida occulta [118]. We suggest that FGF3 may be an important gene in this region required for normal neural tube closure and caudal development.

Ethics statement
Mice were treated in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals (National Academies Press; 8th edition). The protocol was approved by the Animal Care and Usage Committee of NCI-Frederick (NIH) (Animal Study Proposal: 11-069).
Whole mount in situ hybridization, skeletal preparations and cell death analysis Whole mount in situ hybridization was performed as previously described [31]. In all cases, control and experimental embryos were processed in the same tubes throughout the procedure. The Fgf8intron probe was constructed using the primers: 5'-GGAATTC ACGCAGTCCTT GCCTTTGCCG-3' and 5'-GGAATTCATCTGCATGAACAAGAAGG-3', to amplify añ 700bp product from intron 5 of the Fgf8 gene. For sectioning, embryos were fixed overnight in 4% paraformaldehyde in PBS then dehydrated into 100% methanol. Embryos were then embedded using the JB-4 Embedding kit as per kit protocol (Polysciences, Inc; Catalog #00226). Blocks were sectioned at a thickness of 6μM and counterstained with a neutral red stain. Skeletal preparations were performed as previously described [121]. LysoTracker Red staining was performed as previously described [122].
Phosphohistone H3 positive and negative nuclei were counted from 3 sections from each of 3 embryos for each genotype and averaged. Percentages represent number of positive nuclei divided by total nuclei.

Measurements and statistics
Using the measuring tool within Photoshop, linear measurements were made on images of embryos that were taken at the same magnification; units are in pixels. For all data provided, at least an n = 3 was used unless otherwise noted. Significance was determined using a two-tailed t-test unless otherwise noted in the figure legend.

Neural tube ink injections
For determination of NT closure, India ink was diluted 1:10 in PBS and injected into the NT at the anterior boundary of the hindlimb buds using a pulled glass capillary needle. The posterior neuropore was scored as open if the ink passed through the pore; otherwise it was scored as closed.