Bimodal control of Hoxd gene transcription in the spinal cord defines two regulatory subclusters

Correct innervation of peripheral muscles by spinal cord motor neurons is a pre-requisite to initiate and coordinate body movements in all vertebrate species (Landmesser, 1978). In higher vertebrates, motor neurons are organized in different columns, such as the median motor column (MMC), the hypaxial motor column (HMC), the pre-ganglionic column (PGC) and the lateral motor column (LMC) (see Fig. 1A). It is believed that, in an ancestral system, an MMC-like column may have controlled an active swimming behaviour, as seen nowadays in jawless vertebrates, that essentially relies upon the undulating activation of the axial musculature all along the anterior to posterior (AP) axis (Fetcho, 1992). The appearance of a specialized column of motor neurons at the hindbrain-spinal cord boundary of ray-finned fishes may have set the ancestral stage for appendage innervation, the pectoral fin in this particular case (Ma et al., 2010). Subsequently, these columns and associated neurons functionally diversified and shifted to more caudal positions to meet the increased neuronal complexity required for controlling the musculature of tetrapod limbs (Landmesser, 2001; Ma et al., 2010). Neurons connecting to this complex muscular system are found in two major columns, the LMCs, which form two plexii at the brachial and lumbar levels. Within these columns, a stereotypic organization into discrete pools of motor neurons reflects their axonal trajectories, as well as their patterns of target innervation (Hollyday, 1980; Gutman et al., 1993).

Transcription factors of the Hox gene family play essential roles in assigning motor neurons both their columnar fates and pools affiliations, along the AP axis (see Fig. 1A) (Dasen and Jessell, 2009). In mammals, 39 Hox genes are distributed into four distinct genomic clusters (HoxA to HoxD) and members of paralogy groups Hox6 and Hox10 have been functionally associated with the brachial and lumbar LMCs, respectively (Wahba et al., 2001; Dasen et al., 2003). By contrast, Hox9 genes seem to determine PGC fates at thoracic levels (Jung et al., 2010). Within these columns, discrete patterns of different Hox proteins are thought to further define the subdivisions into the various motor neuron pools (Dasen et al., 2005).

The mechanisms leading to the required differential expression patterns of these genes along the rostral to caudal axis of the central nervous system (CNS) are less clear. Some of the upstream activating factors have been identified (Liu et al., 2001; Bel-Vialar et al., 2002; Nordstrom et al., 2006), yet whichever signal controls these genes, it will have to be interpreted in the context of their genomic clustering and translated into a collinear response along the AP axis, as an integral part of their transcriptional regulation (Kmita and Duboule, 2003). The mechanisms underlying this spatial collinearity in the CNS have long been studied in a transgenic context (e.g. Sharpe et al., 1998), yet they only recently started to be addressed using experimental approaches whereby a gene cluster was modified at the endogenous locus (Tarchini et al., 2005; Tschopp et al., 2009).

We set out to test and define which part of the global transcriptional outcome of the HoxD gene cluster in the developing spinal cord was due either to gene-specific features, or to regulatory modalities shared between various genes and thus linked to the genomic clustering of this gene family in vertebrates. To this end, we used tiling array-based transcriptome analyses to systematically assess the transcription of the HoxD gene cluster on a series of deletions stocks as well as in mice carrying contiguous Hoxd transgenes combined with the genomic deletions of the exact same loci. We report the presence of two major anterior boundaries, which define independent blocks of transcriptional regulation, associated with the future plexii. Within each block, a delicate balance of promoter competition and sharing of ‘enhancer mini-hubs’ fine-tunes the correct establishment of spatial collinear Hox domains in the developing spinal cord.

Deletion alleles were produced in vivo using targeted meiotic recombination (TAMERE) (Herault et al., 1998b) and have been described elsewhere: Del(4-9) (Tschopp et al., 2009); Del(4-11) (Zacchetti et al., 2007); Del(9), Del(10), Del(i-9), Del(8-13)XII, Del(10) (Tarchini et al., 2005); Del(i), Del(8) (Tarchini and Duboule, 2006); Del(11) (Kmita et al., 2002). The novel Del(1-i) and Del(1-4) alleles (generously provided by B. Mascrez, University of Geneva) were generated using TAMERE with parental alleles Del(1-13)d11lac and L7 or L8, respectively, the latter strains carrying a single loxP site next to Hoxd8 or Hoxd4 (Tarchini et al., 2005). A novel loxP site 6kb telomeric to Hoxd1 (tel1; L. Thévenet, F. Gonzalez and D.D., unpublished) was used to generate Del(tel1-13)d11lac. Timed matings were carried out using animals heterozygous for the respective alleles. Noon on the day of the vaginal plug was considered as day 0.5 of gestation (E0.5).

Custom-made oligonucleotide tiling arrays spanning a 2 Mb large region centred on the murine HoxD cluster were designed and manufactured by Affymetrix (see Soshnikova and Duboule, 2009). Embryos were dissected at E12.5 and stored in RNAlater (Qiagen) at –20°C until genotyping procedures were finished. Subsequently, spinal cords of positive embryos were dissected out and subdivided into three anatomically defined pieces along the AP axis: +2/–2 dorsal root ganglia (DRG) of the forelimb (brachial), +3/–2 DRG of the hind limb (lumbosacral) and the intervening fragment (thoracic). For each RNA extraction, spinal cords of two embryos were pooled, and total RNA was extracted using RNeasy Micro kit (Qiagen). Two micrograms of total RNA was used as starting material for rRNA depletion using the RiboMinus Human/Mouse Transcriptome Isolation kit (Invitrogen). RNA was subsequently reverse transcribed using random hexamers with a T7-promoter for one round of in vitro amplification of cDNA following the GeneChip Whole Transcript Amplified Double-Stranded Target Assay kit (Affymetrix). The resulting cDNA was fragmented and labelled using the GeneChip WT Double-Stranded DNA Terminal Labeling kit (Affymetrix) and hybridized onto our custom-made oligonucleotide tiling arrays. Genomic DNA used for normalization was fragmented with DNase I digestion. Arrays were hybridized, processed and scanned according to manufacturer’s instructions (Affymetrix). Array data were normalized within cDNA/genomic DNA replicates using TAS software and scaled to medial feature intensity of 10 (Affymetrix, www.affymetrix.com). A sliding window of 80 bp was used to generate a data set consisting of all (PM-MM) pairs mapping within it. The averaged ratios were plotted along the genomic DNA sequence using Integrated Genome Browser (IGB) software (Affymetrix, www.affymetrix.com).

Whole-mount in situ hybridization (WISH) and in situ hybridization (ISH) on slides were performed according to standard protocols. DIG-labelled probes for in situ hybridizations were produced by in vitro transcription (Promega) and detection was carried out using an alkaline phosphatase-conjugated anti-digoxigenin antibody (Roche). Probe templates were either as described elsewhere, or newly PCR-cloned from embryonic cDNA pools (the first and last ten nucleotides are given for orientation): Hoxd3 (Condie and Capecchi, 1993), Hoxd4 (Featherstone et al., 1988), Hoxd8 (Izpisua-Belmonte et al., 1990), Hoxd9 (Zappavigna et al., 1991), Hoxd10 and Hoxd11 (Gerard et al., 1996), Hoxd12 (Izpisua-Belmonte et al., 1991), HoxD-ig1 (CCAACCAGCT–ATGTCTGGAT), HoxD-ig2 (GACCTATTTG–CTAGGCTCAA), HoxD-ig3 (GGCATCTCTG–ATGTCTGGAT) HoxD-ig4 (CCTAGGCATG–CTTGGTTCCT), HoxD-ig5 (ATCTCTGCTG–ATGATGAGGG), HoxD-ig6 (CAAGGAATGA–TGGTGATGGG). For WISH, embryos were dissected in PBS and fixed from 4 hours to overnight in 4% paraformaldehyde (PFA), washed in PBS, dehydrated and stored in 100% methanol at –20°C. Both mutant and control embryos were processed in the same well to maintain identical conditions throughout the WISH procedure and spinal cords were dissected out and flat-mounted. Images were globally processed for colour balance and brightness using Adobe Photoshop. For ISH, embryos were fixed for 90 to 120 minutes in 4% PFA in 0.1 M PBS at 4°C, washed extensively in PBS and cryoprotected in 30% sucrose, before embedding in OCT TissueTek compound. Cryoblocks containing mutant and representative wild-type control animals were fixed side by side and sectioned together, to guarantee identical hybridization conditions on the slide.

For the generation of the lacZ transgene series spanning the HoxD clusters, genomic fragments corresponding to each loxP-loxP interval inside the complex were isolated and a lacZ cDNA was inserted in frame with the respective coding sequence of each Hoxd gene (see Fig. 4A and supplementary material Fig. S3). Fragment isolation was done with long-range PCRs (Expand Long Template PCR System, Roche) on purified genomic or bacterial artificial chromosome (BAC) DNA, using two pairs of oligos to generate a 5′ and 3′ fragment. In between, a unique restriction site was introduced to enter the lacZ cDNA with conventional cloning techniques. Moreover, a single loxP site was added to the distal primer of either 5′ or 3′ fragment, to singularize potential tandem copy arrays in vivo using Cre-deleter strains. Alternatively, lacZ cDNA was introduced by ET recombination into the mouse BAC ‘RPCI23-400H17’, leaving behind a single flippase recombination target (FRT) site after the removal of the selection cassette to be used for transgene singularization using an Flp-deleter strain (Yu et al., 2000). The resulting lacZ-tagged Hoxd genes were retrieved from the BAC by gap repair, using modified pBSK or pBR322 vectors containing homology arms derived from PCR products. For TgNHoxd9lac and TgNHoxd11lac transgenes previously generated in the lab were trimmed to match the cluster-internal loxP-loxP intervals and were modified with loxP or FRT sites to meet the requirements for singularization outlined above (Renucci et al., 1992; Gerard et al., 1993). Schematic overviews of the resulting plasmids are depicted in supplementary material Fig. S3. Final constructs were released from their respective backbone plasmids by restriction digest, gel-separated and dialysis-purified before pro-nuclear injection. At least two stable lines were established per transgenic construct and were crossed to either Cre- or Flp-deleter strains. Embryos were assessed for the regulatory potential of the respective constructs using whole mount detection of β-gal reporter activity (Zakany et al., 1988). For some constructs showing unexpected expression patterns, additional F0 embryos were stained for confirmation of the observed results if only two stable lines had been established.

For immunohistochemistry (IHC), embryos were fixed for 90 minutes at 4°C in 4% PFA in PB and cutting was performed as for ISH (see above). IHC was performed following standard protocols. Primary antibodies were used at the following dilutions: goat anti-Hoxd10 (1:500, Santa Cruz); rabbit anti-FoxP1 (1:64,000, kind gift of Dr Jeremy Dasen, NYU School of Medicine). Secondary antibodies were donkey anti-rabbit Alexa Fluor 488 (1:1000, Invitrogen), biotin-conjugated donkey anti-goat followed by peroxidase-conjugated streptavidin (1:300, all Jackson ImmunoResearch) using for detection the Perkin Elmer TSA-Cy3 Kit. FoxP1 quantification was done using biological triplicates. Six individual sections per replicate spanning the brachial plexus were stained for FoxP1 and positive nuclei were counted. Excluding the highest and lowest values of these series, the values of the remaining four sections were averaged and counted as one replicate.

We assessed the transcriptional status of the HoxD complex in the developing spinal cord by using custom-made tiling arrays covering a two megabase large DNA interval centred on the gene cluster. We dissected out spinal cords at E12.5 and subdivided the cord into three anatomically defined fragments (see Materials and methods): firstly, an anterior piece containing the brachial plexus, which innervates the forelimbs; secondly, a posterior piece including the lumbosacral plexus that innervates the hind limbs; and, finally, the piece linking the two former fragments, referred to as the ‘thoracic region’ (Fig. 1A,B). RNA was extracted and processed for the three pieces at the same time and in the same conditions (Fig. 1B).

At brachial levels, only Hoxd3 to Hoxd8 were transcribed at detectable levels, with Hoxd1 being silent at this stage (Fig. 1C). In addition to coding sequences, large stretches of intra- and intergenic DNA around Hoxd3 and Hoxd4 were also transcribed, starting from an alternative promoter of Hoxd3 located in the Hoxd4/Hoxd8 intergenic region (subsequently termed region ‘i’). However, the intronic regions separating the coding exons of Hoxd3 and Hoxd4 were largely devoid of any transcripts (Fig. 1C). In the thoracic sample, no major changes were observed in the anterior transcription profile, whereas exon 2 of Hoxd9 was now transcribed at a clearly detectable level. At lumbosacral levels, all genes from Hoxd3 to Hoxd11 were expressed at high levels, with some trace amounts of transcription detectable in more posterior parts of the cluster (Fig. 1E and supplementary material Fig. S1A-C). Here again, high amounts of intergenic transcription were scored, now also for the posterior part of the cluster (Fig. 1E), whereas throughout the complex we detected only minor amounts of transcripts originating from the ‘non-Hox’ DNA strand (supplementary material Fig. S1E,F).

The transcript distribution at lumbosacral levels seemed to define two separated blocks of transcription. An anterior block covered approximately 40 kb, spanning from region ‘i’ to the 3′UTR of Hoxd3, whereas a ∼25 kb large region, extending from Hoxd12 to Hoxd9, also showed widespread transcription in the posterior half of the complex (Fig. 1E, shaded pale blue). Between these two blocks, Hoxd8 was transcribed at all three anatomical levels. Strikingly, however, Hoxd8 transcription could not be associated to either the anterior or the posterior block. Instead, this transcription unit showed a clear separation from both blocks, by 5 kb large DNA stretches on either side, devoid of any detectable transcription coming from the Hox-coding strand (Fig. 1E and supplementary material Fig. S1E).

We investigated the expression patterns associated with these two transcription blocks by WISH, in embryos at E12.5 with their spinal cord opened and flat-mounted after staining. We used a battery of probes complementary either to known Hoxd transcription units or to unannotated transcripts, as defined by our tiling array analyses. Again, a clear dichotomy appeared in the expression territories of RNAs derived from the two blocks. Transcripts of the anterior block all had a rostral limit of expression either within, or immediately below the hindbrain (Fig. 1F, right). Hoxd4 and Hoxd3 were expressed strongly up into the hindbrain, with the ig6 fragment located in between also being expressed there, though much weaker. Its major transcription boundary, as for all other fragments isolated from this block (ig2 to ig5; Fig. 1F), nicely aligned with the position of the brachial LMCs.

By marked contrast, RNAs derived from the posterior block displayed anterior boundaries within or immediately anterior to the lumbosacral plexus (Fig. 1F, left). Hoxd8, however, showed a columnar restricted expression, with signals extending up to mid-brain regions, setting the gene apart from the rest of HoxD cluster, in agreement with its isolated transcription profile as seen on tiling arrays. Therefore, rather than displaying a progressively staggered distribution of transcripts along the AP axis, as e.g. in mesoderm derivatives, the collinear boundaries in Hoxd genes transcription are built upon two major territories. They coincide with the functional subdivision of the spinal cord; an anterior block of transcription, labelling the future brachial plexus up into the hindbrain, and a posterior block observed at the level of the lumbosacral plexus.

To gain insight into the regulatory logic underlying these two transcription blocks, we used a series of deletion alleles, produced via the TAMERE approach (Herault et al., 1998b), which remove various parts of the HoxD gene cluster. We first challenged the regulatory interdependency of the two apparent blocks of transcriptional activities, by using two deletions removing entirely either one of them; in the Del(1-i) mice, the anterior part of the gene cluster was deleted from Hoxd1 included to downstream Hoxd8, whereas Del(8-13) mice carried a deletion of the posterior half, i.e. from Hoxd8 included to Hoxd13 (Fig. 2A-F, shaded grey). E12.5 homozygous mutant spinal cords were isolated, their RNAs extracted and their HoxD transcription profiles analyzed.

In the Del(1-i) embryonic spinal cord, weak ectopic signals appeared around the breakpoint, anterior to Hoxd8 and in the second exon of Hoxd1, probably originating from an ectopic transcript spanning the loxP site. However, the global transcription profile in the posterior half of the cluster closely followed wild-type patterns all along the spinal cord, indicating that the deletion of the anterior part of the cluster did not substantially alter the transcription in the remaining posterior block (Fig. 2A-C, compare with Fig. 1C-E, left). Likewise, removing the posterior part of the HoxD complex did not impact upon the transcriptional output of the anterior block, as seen in the Del(8-13) mutant spinal cord, which displayed an ‘anterior’ profile indistinguishable from the wild-type situation (Fig. 2D-F, compare with Fig. 1C-E, right). We thus concluded that the two transcription blocks are established independently from one another, each one depending on their own regulatory logic and modalities.

We next investigated the effects of deleting the transcriptionally silent DNA domains around the Hoxd8 locus, i.e. when the separation between the two blocks was impaired. To this aim, we used a set of deletions removing various central fragments of the gene cluster (Fig. 3B-E, shaded grey), while leaving both the anterior and posterior extremities unmodified. In contrast to the previous series of experiments, we observed important reallocations of transcriptional efficiencies, here visualized at brachial levels. When the broadly transcribed region ‘i’ was removed, a clear upregulation of Hoxd8 and Hoxd9 was scored, as well as increased transcription of the coding exons of Hoxd3, whereas intergenic transcripts mapping within the Hoxd4 to Hoxd3 region markedly decreased (Fig. 3B, green and red arrows, respectively). Larger deletions moving the Hoxd10 transcription unit next to either Hoxd4 or Hoxd3 led to a strong transcriptional activation of this gene at brachial levels, where it is usually silent (Fig. 3C,D). Although this ectopic activation occurred mostly at the expense of anterior, intergenic transcripts, progressively larger deletions also started to affect the transcriptional output of anterior coding sequences (Fig. 3C-E). Similar reallocations were seen at thoracic and lumbar levels (data not shown). We next investigated the expression patterns associated with these changes using WISH on open-book spinal cords. Consistently, we found transcriptional upregulation of posterior genes located centromeric to the deletion breakpoints along the AP axis of the spinal cord. The resulting ectopic activations along the AP (Hoxd12, Hoxd10) or dorsoventral (DV) axis (Hoxd8) of the spinal cord were mirrored by a decrease in steady-state level transcripts of genes lying telomeric to the breakpoints (Hoxd4, Hoxd3; Fig. 3F). Therefore, internal deletions abolishing the physical separation of the two transcription blocks led to the ectopic anterior activation of posterior Hoxd genes, at the expense of transcripts belonging to the anterior block (Fig. 3B-F).

Of the posterior Hoxd genes, Hoxd10 is instrumental in specifying lumbar LMCs (Carpenter et al., 1997) and its overexpression at anterior levels led to the mis-specification of neuronal fates (Dasen et al., 2003; Shah et al., 2004). Indeed, ectopic transcription of Hoxd10 in our mutant stocks also resulted in the production of functional protein at anterior levels, as evidenced by HOXD10 immunohistochemistry and the upregulation of its downstream target FoxP1 (Dasen et al., 2008; Rousso et al., 2008) (supplementary material Fig. S2), thus offering a potential explanation for the apparent necessity to keep the two transcriptional blocks separated, to avoid regulatory interferences leading to neuronal mis-specifications.

These reallocations of local enhancer activities between closely located Hox genes made it difficult to map regulatory sequences in vivo using deletion mutants. We designed a series of contiguous reporter transgenes corresponding exactly to those DNA intervals defined by pairs of loxP sites and further deleted by recombination in trans (see Kmita et al., 2002; Tarchini et al., 2005). Seven contiguous transgenes were constructed, covering a total genomic fragment of more than 75 kb from Hoxd3 to Hoxd11, all hooked up in frame to a lacZ reporter gene (Fig. 4A and supplementary material Fig. S3). Founder animals were screened for transgene presence, approximate copy number and integrity using PCR- and Southern blot-based genotyping strategies.

F1 animals were crossed to wild-type specimen and E12.5 embryos stained for β-gal activity (Zakany et al., 1988). At least two lines were investigated for each locus, and representative embryos are depicted in Fig. 4B-H. All lines investigated for TgNd11lac and TgNd9lac showed robust activation in the spinal cord with the expected rostral limits of expression (Fig. 4B,D, arrowheads). TgNd10lac lines, by contrast, showed variable results with weak and patchy reporter activity scored in a variety of embryonic structures in the different lines (data not shown). However, none of the lines (0 out of 5) displayed wild-type like activation in the developing spinal cord (Fig. 4C). Likewise, the TgNd8lac transgene was completely silent in the spinal cord, the only reproducible activity being found in the developing urogenital system (Fig. 4E; data not shown; 6 out of 7).

In TgNilac animals, the lacZ was inserted into an alternative upstream exon of Hoxd3. A patchy activity was recovered in the spinal cord, with an anterior limit of expression at approximately the level of the forelimb (Fig. 4F, arrowhead). However, this pattern reflected neither the rostral expression limit, nor the intensity of transcripts originating from the endogenous ‘i’ locus (compare with Fig. 1F, ig2-4). By contrast, both TgNd4lac and TgNd3lac showed strong and reproducible staining in the spinal cord. TgNd4lac displayed an expression boundary between rhombomeres 6 and 7, as for the endogenous gene, whereas the expression pattern of TgNd3lac was slightly anteriorized (Fig. 4G,H, arrowheads). Altogether, this transgene scanning approach revealed that although some of the Hoxd loci display regulatory autonomy in trans, others seem unable to elicit a transcriptional activity by themselves and hence require the context of the cluster in cis to produce their spatially correct gene expression patterns in the spinal cord.

Although the detection of transgenic β-gal activity reflects the overall regulatory potential of transgenic DNA, it remains poorly informative regarding the exact structure and expression levels of the produced mRNAs. Moreover, transcripts not including the reporter sequence will remain undetected. Because the breakpoints of our deletion stocks correspond exactly to the extremities of our transgenic fragments (Fig. 4A), we investigated their individual RNA profiles on tiling arrays by crossing transgenes over the corresponding deletions. In this way, we reconstituted the HoxD gene cluster, yet with one locus at a time in a trans configuration (e.g. Hoxd9 in Fig. 5A). Accordingly, all transcripts coming from a given endogenous Hoxd gene locus are removed (Fig. 5A, grey) and RNAs detected on the tiling array originate exclusively from the transgene (Fig. 5A, orange).

We reduced the transgene copy number to one using loxP- or FRT-dependent deletions. F1 transgenic animals were crossed to mice harbouring the corresponding gene deletion in the HoxD complex. F3 crosses generated E12.5 embryos, which were homozygous for a given deletion, but containing the related transgene. RNA was extracted from the lumbosacral region and processed as before.

In agreement with their lack of lacZ expression in the spinal cord, neither the TgNd8lac, nor the TgNd10lac transgenes showed any RNA signal (Fig. 5F,H). By contrast, the TgNd9lac construct, which showed a strong and rather faithful distribution of β-gal activity in transgenic embryos, displayed widespread and solid transcription throughout the locus in trans (Fig. 5G). Whereas the second coding exon of the transgene showed an RNA profile close to the endogenous gene, increased transcription levels were apparent at the 5′ end of the transgenic profile. The expression profile of the TgNd11lac transgene, on the other hand, displayed only weak peaks, at those locations where the strongest signals were observed in the endogenous locus. This discrepancy between the two transgenes showing apparently comparable amounts of β-gal activity can be explained both by a dilution effect, as only the posterior segment of the dissected lumbosacral piece stained positive for the TgNd11lac transgene, and by the biochemical properties of the X-gal staining (see Discussion).

When selected from within the anterior block of transcription, all transgenes showed lacZ staining and RNA signals on the tiling array, albeit with greatly varying intensities. A cluster of small peaks originating from the alternative Hoxd3 promoter where our reporter gene was integrated reflected the weak β-gal activity seen in TgNilac embryos (Fig. 5E). Likewise, the TgNd4lac profile globally matched the output of the endogenous locus, but again at greatly reduced levels (Fig. 5D). TgNd3lac, however, looked nearly indistinguishable from the endogenous profile (Fig. 5C).

This systematic analysis of TgNHoxlac in spinal cords thus uncovered important differences in the transcriptional efficiencies of the various gene loci. Whereas some loci (e.g. Hoxd9, Hoxd3) could efficiently reproduce their endogenous transcription patterns in trans, other loci remained completely silent (Hoxd8, Hoxd10). This suggested that their normal expression, as observed in wild-type animals, is controlled by regulatory elements located within the neighbouring loci.

The functional importance of Hox proteins in patterning the spinal cord along its AP axis, in particular in the determination of neuronal identities, has been well established (Dasen and Jessell, 2009). Yet, the mechanisms generating the necessary expression patterns, at the transcriptional level, are far less well documented. In many animal species, important aspects of the spatial Hox transcripts distribution are intrinsically linked to the clustered organization of these genes, a phenomenon known as spatial collinearity (Kmita and Duboule, 2003). Paradoxically, however, while the relative position of a gene within the cluster correlates with its rostral limits of expression, the presence of an intact cluster does not seem to be a strict prerequisite for this phenomenon to occur. For example, rather faithful expression patterns were observed for a number of randomly integrated Hox transgenes in the mouse (e.g. Puschel et al., 1991; Gerard et al., 1993). Also, the endogenous Hox genes in the larvacean Oikopleura are expressed with some collinear features, even though a gene cluster does not exist in this species (Seo et al., 2004). This may reflect the fact that an important part of the regulatory machinery necessary to define the spatial distribution of transcripts often resides in close proximity to the target genes, interspersed inside the Hox clusters (Whiting et al., 1991; Tschopp et al., 2009).

However, the results presented here suggest that the correct establishment of spatial collinearity in the mouse spinal cord also relies heavily upon the presence of a gene cluster at least partially organized. We find that extended stretches of the HoxD cluster are actively transcribed from multiple promoters, playing an important role in fine-tuning the overall transcriptional outcome of the different gene loci. In particular, we report the presence of two distinct transcription blocks, inside the HoxD gene cluster, that define two general expression territories. The first block involves mostly the transcription of both Hoxd3 and Hoxd4 all along the spinal cord, including both future plexii. The second block involves more posterior genes (from Hoxd9 to Hoxd11) and is restricted to a caudal part of the spinal cord, containing the future lumbosacral plexus. Therefore, these distinct regulatory modalities generate different genetic addresses at the level of the future plexii, which will subsequently innervate the musculature of tetrapod limbs. Within the two blocks, sharing of ‘enhancer mini-hubs’ may be required for efficient Hoxd transcription, a fact that establishes an inter-dependency between neighbouring gene loci and provides an additional layer of constraint to keep vertebrate Hox genes tightly clustered.

The presence of two distinct transcription blocks somewhat echoes the situation found in various dipteran species, where the ancestral Hox complex is split into two independent genomic subclusters (Duboule and Dolle, 1989; Graham et al., 1989). Although transcriptional regulation inside the two HoxD blocks relies upon shared enhancer elements, these two regulatory domains appear functionally independent from one another, defining two virtual subclusters (see Fig. 3A-F). Indeed, in the case of the spinal cord, the observed regulatory autonomy would probably allow for a Diptera-like split of the murine HoxD complex at or around the position of Hoxd8, thus separating ‘posterior’ genes (i.e. those related to Abdominal-B) strongly expressed in the lumbosacral plexus, from anterior Hox genes, heavily transcribed at both lumbar and brachial levels. Regulatory constraints imposed in other tissues, however, will prevent such a split from occurring (e.g. Tarchini and Duboule, 2006; Tschopp et al., 2009).

Each one of these two transcription blocks contains one locus that can elicit very strong expression in a transgenic context. In the anterior spinal cord, TgNd3lac clearly included a strong enhancer activity, whereas TgNd9lac was its counterpart in the posterior spine. Some loci located nearby (Hoxd4 and Hoxd11) were also able to drive transcription within the spinal cord, although with a substantially lower efficiency, an observation that was probably overlooked when contemplating previous X-gal stainings (Gerard et al., 1993; Zhang et al., 1997). We think this is because of a saturation effect, the stability of the β-gal protein and multicopy insertions in the conditions initially analyzed. Nevertheless, weak but clear signals were scored from both the Hoxd4 and Hoxd11 loci, suggesting that although these promoters can indeed respond to local enhancers elements, their full activities somehow rely on the heavy transcriptional activity of either the Hoxd3, or the Hoxd9 loci in cis. This is supported by previous observations reporting the impact of the Hoxd9 locus on the transcription of both Hoxd10 and Hoxd11 (Tarchini et al., 2005). In this view, the Hoxd3 and Hoxd9 loci may organize local ‘enhancer mini-hubs’ (Fig. 6, green halos), a situation that would explain the need for clustering within each of these blocks, to set up spatial collinearity at the proper quantitative levels. Once controlled by these mini-hubs, local and gene-specific regulatory elements may account for the slight expression differences observed within each of the two transcription blocks, along the AP and DV axes for the different Hoxd genes (see e.g. Fig. 1F).

Hox clusters display very dense transcriptional activities, owing to the presence of numerous alternative transcription start sites and enhancers intermingled within the coding regions (Mainguy et al., 2007; Rinn et al., 2007; Coulombe et al., 2010). Accordingly, we find extensive transcription outside the annotated coding sequences of the HoxD complex, which contribute to delineate the two distinct transcription blocks. When considering single-gene and larger deletions, an intricate regulatory interdependency of these transcripts and the various coding sequences is emerging.

Whereas the deletion of the Hoxd9 locus decreases the relative amounts of Hoxd10 coding transcripts (Fig. 5B,G, compare Hoxd10 with Hoxd11), removing this putative enhancer titration effect of Hoxd10 increases transcription in the 5′ region of Hoxd9 in Del(10) mutant spinal cords (Fig. 5H). A similar upregulation is seen for the TgNd9lac transgene, i.e. when the titrating influence Hoxd10 is equally removed (Fig. 5G). Moreover, perturbing this subtle transcriptional balance also leads to important changes in the spatial patterns of the genes involved (Tarchini et al., 2005), a consequence difficult to document when considering the entire lumbosacral region as in our tiling array approach.

Transgenic analyses have revealed a complex enhancer landscape within the relatively short HoxD genomic interval (Fig. 4) (Renucci et al., 1992; Gerard et al., 1993; Zhang et al., 1997; Herault et al., 1998a). Regulatory elements may act rather generically on any of the surrounding promoters, a situation illustrated by genes lacking the required enhancer elements in their vicinity that hijack the proper controls from their neighbours (e.g. Fig. 4C) (Gould et al., 1997). In this view, the presence of numerous intergenic transcripts within region ‘i’, i.e. starting from promoters located downstream of Hoxd8 (Fig. 6, arrow) may simply reflect the necessity to titrate such activities coming from the Hoxd3 ‘enhancer mini-hub’. These promoters would compete for ‘anterior’ enhancers and thereby prevent the same regulatory influence to reach the more posterior and potentially deleterious genes (Fig. 6, red cross). When aligning these DNA sequences to the three other murine Hox clusters, putative promoters sequences located around group 6 Hox genes score the highest (data not shown). The HoxD cluster has lost its group 6 gene at an early stage during gnathostome evolution (Ravi et al., 2009) and it is conceivable that some of its associated promoter elements were maintained there simply owing to their crucial role in the trapping of an anterior regulatory influence.

The appearance of paired appendages was a major innovation in early vertebrates, crucial for their successful radiation. It is believed that the molecular mechanisms underlying the development of paired fins were co-opted from median fins. This structure, which is already observed in agnathans, was not restricted to any particular AP level (Freitas et al., 2006). Therefore, the emergence of paired appendages required the concomitant development of localized and highly specialized nervous plexii, to provide the innervation for their increasingly complex target musculature. In tetrapods, the interpretation of specific combinations of Hox proteins by their target genes underlies this diversification of motor neuron pools at brachial and lumbosacral levels (Dasen et al., 2005; Dasen et al., 2008; Rousso et al., 2008).

The regulatory split of the HoxD gene cluster described in this study may give a hint as to how such differential Hox combinations were achieved at the transcriptional level, by evolving distinct and independent regulatory modules based on, and constrained by, the collinear properties of Hox gene clusters. Firstly, the two separate transcription blocks define two major rostral boundaries in the spinal cord, with Hoxd genes either expressed up to brachial and hindbrain regions or confined within lumbosacral limits. Secondly, inside the two blocks, additional enhancer elements further subdivide and fine-tune the major regulatory input from the two ‘enhancer mini-hubs’ to eventually elicit slightly differential expression patterns, thereby diversifying combinations of Hox proteins.

Alternatively, only few Hox proteins, rather than combinations, may be decisive for motor neuron specification in both plexii, a possibility difficult to demonstrate owing to the redundancy and complementarity of the system, under physiological conditions. An extreme view of this latter possibility would consider the intricate transcription profiles described in this work as being merely the by-products of a general regulatory strategy to ensure that one particular protein be present in neuronal precursors, at the right time and place. Anterior Hoxd proteins do not seem to exert a strong function in the brachial plexus, unlike HOXD10 at the lumbosacral level. Hoxd3 with its strong anterior enhancers, however, is required for the patterning of the cervical nerves (Condie and Capecchi, 1993). The observed bimodal control of the HoxD cluster in the spinal cord may thus have been selected primarily to prevent the deleterious gain-of-function effects of posterior proteins, while at the same time ensuring the appropriate expression levels of the decisive Hoxd products at the correct AP level.

We thank S. Arber, J. Dasen and C. Tabin for critical comments and suggestions, as well as for the generous gifts of antibodies. We also thank N. Fraudeau for technical assistance, S. Falk and L. Sommer for help with IHC protocols, C. Barraclough and P. Descombes for help with tiling arrays, B. Mascrez for mouse lines as well as members of the Duboule laboratories for discussions and reagents.