Temporal dynamics and developmental memory of 3D chromatin architecture at Hox gene loci

1Mammalian Hox genes encode proteins that are essential for patterning along the rostral-to-caudal body axis of the developing embryo. Mouse and human Hox genes are organized in four genomic clusters (HoxA to HoxD), where the relative position of the genes strongly impacts upon their patterns of expression. This structure-function relationship was initially described in Drosophila and further extended to vertebrates, where an additional correspondence exists between gene position and the timing of transcriptional activation (‘temporal colinearity’,; ).

2In murine embryos, transcription of Hox genes can be divided in several phases and is first detected at around embryonic day 7 (E7) at the most posterior aspect of the primitive streak region. Over time, Hox genes are sequentially activated following their chromosomal order and transcripts encoded by the last Hox group 13 genes can be detected at around E9, that is two days after the onset of activation. This transcriptional progression (the ‘Hox clock’, ) thus extends over several days. In the pre-somitic mesoderm (PSM), this sequential activation needs to be coordinated with the time-sequenced production of body segments (the ‘segmentation clock’, ), such that newly produced somites acquire distinct genetic identifiers. Next, the various states of Hox gene activity are fine-tuned and memorized, ultimately leading to domains along the rostral to caudal axis where partially overlapping sets of HOX products can be observed (‘spatial colinearity’). As a result, genes located at 3′positions (e.g., groups 3, 4) are transcribed almost along the entire embryonic axis, including the lateral plate mesoderm, paraxial mesoderm and neural tube, whereas the 5′-located group 10 or 11 are active in the posterior trunk and group 13 in the tip of the tail bud only.

3While both temporal and spatial colinear processes likely reflect one and the same organizational principle, they are nevertheless implemented with distinctive features. Spatial colinearity could be recapitulated by several single-gene transgenes (e.g.,), yet not in all instances. Indeed, a systematic analysis of modified HoxD clusters in vivo revealed that, at a late stage, the sustained transcription of these genes at the correct body level primarily relies upon local regulatory elements, which are present in transgenic constructs. In contrast, the precise timing of Hoxd gene activation depends on the integrity of the full cluster, a genomic situation observed thus far in all animals developing following a temporal rostral to caudal progressive strategy during their early development. The genomic clustering of Hox genes is thus considered as an essential feature for temporal colinearity to properly process, whereas it may not be as important for the correct distribution of HOX products along the AP-axis, at least in the late phase of spatial colinearity.

4Even though the mechanisms underlying temporal and spatial colinearities are becoming increasingly understood, many aspects of how genomic topology is translated into sequential transcriptional activation remain to be clarified. In vertebrates, two conceptual frameworks have been proposed to account for this process, the first relying on bio-molecular mechanisms (e.g., ) and the second involving biophysical forces. In embryonic stem (ES) cells, that is cells that reflect best the state of Hox genes before their activation, Hox clusters are decorated by both repressive (H3K27me3) and activating (H3K4me3) marks. Subsequently, cells that activate these genes in a time sequence resolve this bivalent chromatin state and show two opposing distributions of histone marks over the HoxD cluster: transcribed genes carry large domains of H3K4me3 marks, whereas inactive genes are covered by H3K27me3 marks only.

5The same dichotomy in chromatin marks over Hox gene clusters was observed in various parts of the E10.5 embryonic trunk, in parallel with the spatial colinear distribution of transcripts. The analysis of the 3D chromatin organization at this stage revealed a bi-modal compartmentalization, whereby active genes labeled by H3K4me3 are clustered together and physically separated from the inactive genes, labeled by H3K27me3 that are also found in a defined spatial structure. These 3D compartments, whose sizes correlate with the number of active vs inactive genes, may reinforce the proper maintenance of long-term transcriptional states at various AP levels by isolating Hox clusters from their surrounding chromatin and reducing interference between the active and inactive chromatin domains. Such distinct bimodal 3D organizations, associated with transcriptional regulation at Hox clusters, have been observed in other instances, either in the embryo or in mouse and human cultured cells.

6However, in-embryo conformation studies were reported so far only in the context of spatial colinearity, that is by comparing samples from different body levels at the same developmental stage. Consequently, a potential association between these bimodal chromatin structures and the progressive activation of transcription along the Hox gene clusters, rather than its maintenance, remained to be assessed. In this study, we describe the 3D organization of Hox gene clusters at high resolution during the implementation of temporal colinearity in the PSM and show that their stepwise activation occurs in parallel with their physical transition from a negative to a positive domain. We also show that this process is accompanied by series of long-range contacts with the flanking gene deserts, even though these contacts remain largely invariable throughout temporal colinearity, unlike what was observed during limb development. We discuss whether this stepwise transition of genes from one domain to the other may guide temporal colinearity or, in contrast, is a mere consequence of a sequential transcriptional activation.

7In order to monitor the 3D organization of Hox clusters during their sequential activation, we considered ES cells as a starting point of our time curve. These cells indeed represent early embryonic cells related to blastocyst inner cell mass cells, that is when Hox genes are all supposedly silent. We hypothesized that these cells reflect the ground state 3D architecture of the Hox clusters, which we assessed by using high-resolution 4C-seq (Circular Chromosome Conformation Capture;; ) and a variety of viewpoints within all four Hox clusters. These various baits generated similar interaction profiles with the majority of sequence reads covering the gene clusters and extending within several kilobases (kb) on either sides, as illustrated by Hoxd13, Hoxd9 and Hoxd4 (Figure 1A, Figure 1—figure supplements 1–4). Additional contacts were scored in the flanking gene deserts, though with significantly lower frequencies (see section ‘Temporal colinearity within a constitutive framework of long-range interactions’). The overall size of the strong interaction profiles exactly matched the distribution of bivalent chromatin marks in these cells, with a moderate level of H3K27me3 covering the cluster and rather weak H3K4me3 peaks labeling promoters (Figure 1A, Figure 1—figure supplements 1–4;; ). Therefore, prior to their activation, Hox clusters are already organized into 3D chromatin compartments that physically separate the chromatin decorated by bivalent marks from the genomic surroundings, even though some contacts are established at a larger scale, outside the gene cluster itself (see section ‘Temporal colinearity within a constitutive framework of long-range interactions’).

8These global 3D domains including the Hox clusters and their immediate flanking DNAs resemble the chromatin architecture found in embryonic forebrain cells, where the silent Hox clusters are covered by high levels of H3K27me3 only. A more quantitative comparison in 3D architectures between ES cells and E10.5 forebrain cells nevertheless indicated that in ES cells, Hox genes interacted more with the outside chromatin, relative to their interactions within the cluster, as compared to forebrain cells (Figure 1B,C, Figure 1—figure supplements 1–5). Therefore, despite the fact that the clusters are presumably inactive in both situations, the presence of bivalent marks in ES cells coincided with a 3D domain that has elevated relative levels of interactions with the directly surrounding regions, when compared to its counterpart in brain cells (Figure 1D, left and Figure 1—figure supplement 6A). This difference is more pronounced at the HoxB cluster (Figure 1B, right). In embryonic forebrain cells, HoxB forms a single 3D compartment, excluding the 80 kb large repeat-rich intergenic region located between Hoxb13 and Hoxb9, which loops out. In contrast, both the Hoxb13 and Hoxb9 viewpoints revealed local 3D compartments in ES cells, matching again the extent of bivalent histone marks, yet these two compartments remained separated and did not fuse. Rather, they displayed increased interactions with the nearby chromatin, as if decreased internal interactions would increase contacts outside the cluster (Figure 1B,C, Figure 1—figure supplements 2, 3 and 5). In ES cells, the HoxB cluster is thus organized in two 3D compartments, which have more interactions with their genomic surroundings than in forebrain cells (Figure 1D, right and Figure 1—figure supplement 6B).

9To have a possibly more unbiased view on how 3D compartments and the presence of H3K27me3 and H3K4me3 modifications relate to each other, in both ES and brain cells, we devised an approach to correlate 4C-seq signals with either H3K27me3 or H3K4me3 ChIP-seq signal (Table 1; ‘Materials and methods’). In both cell types, H3K27me3 marks strongly correlated with the 3D organization, suggesting a direct link between these two readouts. A considerably lower correlation was scored for HoxB, perhaps related to the absence of clustering of the two H3K27me3 marked sub-domains. In contrast, no particular correlation was observed between the 3D organization and the presence of H3K4me3 marks, in the bivalent state (Table 1), suggesting that H3K4me3 marks and/or the associated factors do not noticeably contribute to the formation of 3D compartments in ES cells.

10It was previously reported that genes covered by bivalent marks in ES cells can be transcribed at low levels, resulting in detectable spliced transcripts. We assessed whether the observed difference in the strength and homogeneity of the interaction profiles between ES cells and embryonic brain cells was associated with distinct levels of background transcription. In ES cells, RNA-seq detected transcription for most Hox genes, though generally at very low level (Figure 1—figure supplement 7A,B). RT-qPCR of a subset of transcripts confirmed that some of these low-level transcripts (particularly the Hoxd13 and Hoxb13 transcripts) constitute genuine processed transcripts (Figure 1—figure supplement 7C). In contrast, transcription of Hox genes in E10.5 forebrain cells was rarely detected, and no reliable spliced transcripts were detected (Figure 1—figure supplement 7). Therefore, when Hox genes are decorated by bivalent chromatin marks, they appear more permissive for background transcription as compared to other cell types where they are covered by H3K27me3 marks only, likely illustrating the increased resistance to transcription of the latter condition. In this context, posterior Hox genes seems to be more prone to background transcriptional activation in ES cells than more anterior Hox genes, in contrast to their subsequent dynamics of activation in future embryonic tissues where anterior genes come first. This may reflect the presence of strong enhancers in their vicinity.

11Next, we assessed whether this large 3D domain observed in ES cells is modified when Hox genes become activated in the pre-somitic mesoderm (PSM) or instead, whether the previously observed positive and negative compartments are only established at a later stage to fix and memorize particular combinations of Hox gene activities determined at earlier stages and at various body levels. For this purpose, we compared the 4C-seq profiles from ES cells with those obtained from early embryonic E8.5 PSM cells dissected out at Theiler stage 13, posterior from the approximate level of the 12th to 14th forming somite (Figure 2A, scheme; Figure 2—figure supplements 1 and 2). In the most caudal aspect of this latter cellular territory, transcriptional activation had progressed up to the Hoxd9 gene, whereas the Hoxd10 to Hoxd13 loci remained silent. This cellular population was thus composed of a mixture of cells positive and negative for Hoxd9 expression, whereas all cells were negative for Hoxd13. Conversely, the majority of cells expressed Hoxd4. The inactive Hoxd13 viewpoint interacted mostly with the domain labeled by H3K27me3, at the centromeric side of the cluster (Figure 2A, bottom left). In contrast, the active Hoxd4 gene essentially interacted with the other transcribed genes on the telomeric side of the cluster, labeled by H3K4me3 marks (Figure 2A, bottom right). The same bi-modal 3D organization was observed for all Hox gene clusters (Figure 2—figure supplements 1 and 2).

12We correlated 4C-seq signals with ChIP-seq data in both early and late embryonic samples (Figure 2B; Table 1) and Hoxd13 always strongly correlated with H3K27me3 histone marks. In contrast, in both the E8.5 tail bud and the E10.5 anterior trunk, the interactions of the active Hoxd4 gene correlated primarily with H3K4me3 marks. The contacts established by Hoxd9 correlated both with H3K27me3 and H3K4me3, either in E8.5 tailbuds, or in E10.5 anterior trunk, likely due to the presence of both expressing and non-expressing cells. At both stages where Hox clusters are partially active, the patterns of 3D compartmentalization and histone marks thus strongly correlated. Therefore, step wise Hox gene transcriptional activation, at least for the Hoxd9 to Hoxd13 genes, is accompanied by a conformational separation between active and inactive domains, which pre-figures their 3D organization at later developmental stages along the AP-axis (Figure 2; ).

13Temporal colinearity was initially defined as the sequential activation of Hox genes according to their positions in the clusters. However, studies on the global transcriptional organization of the HoxD cluster, at least in the developing spinal cord, revealed two large and regulatory-independent modules, which separate ‘posterior’ genes (the AbdB-related Hoxd9 to Hoxd13 genes) from the rest of the gene cluster. Also, in different developmental contexts such as the limbs and the cecum, groups of neighboring Hoxd genes are activated as single regulatory blocks. We thus assessed whether the transition in chromatin domains also occurred stepwise or, alternatively, if large domains consisting of multiple genes were initially organized in space, followed by sequential gene transcription within these domains.

14We first compared the 3D cluster architecture over the course of embryonic development, between the E8.5 PSM and dissected E9.5 tail buds (Figure 3). This latter sample was obtained after cutting off the most caudal part of E9.5 embryos (Theiler stage 15) right after the incipient hind limb bud, that is at ca. somite 26–27 level. Accordingly, this sample contained the tail bud proper as well as some tissue localized slightly more rostral. During this 24 hr time interval, the Hoxd10 and Hoxd11 genes become robustly activated in these cells, which are derived from a sub-population of the sample dissected at E8.5. In the E8.5 PSM, Hoxd10 and Hoxd11 are still silenced.

15The re-organization of compartmentalization occurring along with gene activation over this 24 hr period was clearly revealed by comparing the profile obtained when using Hoxd11 as a viewpoint with those obtained with either Hoxd4 or Hoxd13 (Figure 3, top ratio between E8.5 PSM and E9.5 Tail bud). When switching from an inactive to an active state, Hoxd11 re-deployed its interactions from the inactive, centromeric compartment (Figure 3, top ratio: blue shaded area) to the active telomeric compartment (Figure 3, top ratio: brown shaded area). These negative and positive compartments can be identified by the interaction profile of either Hoxd13 (Figure 3, top left) or Hoxd4 (Figure 3, top right), respectively. Accordingly, Hoxd4 shifted its interactions towards the centromeric (active) part of the cluster in both E9.5 and E10.5 samples (Figure 3; right), to contact Hoxd10, Hoxd11 and, to some extent, Hoxd12 (Figure 3, top ratio: brown shaded area). Of note, the dissected E8.5 PSM contained a mixture of cells either positive or negative for Hoxd9 transcription, which coincided with this gene showing conspicuous contacts with both extremities of the gene cluster, depending whether it was active (right) or inactive (left) (Figure 3, top, black arrows). In contrast, the E9.5 dissection contains a more homogenous cell population, strongly expressing this gene. As a consequence, in this latter sample, Hoxd9 contacts more strongly the now expressed Hoxd10 and Hoxd11 genes, whereas the interactions with Hoxd13 or Evx2 are strongly diminished (Figure 3; compare top with middle panels).

16In the E10.5 tail bud, the terminal part of the cluster containing Hoxd12 and Hoxd13 has been activated, as shown by the increased contacts established by Hoxd13 with the telomeric part of the cluster, indicating that the full HoxD array had been processed and that all genes were now (at least in part) included into a ‘positive’ compartment (Figure 3, bottom ratio: purple shading). While considerably increased, contacts of Hoxd13 with the telomeric part of the cluster were however weak (Figure 3, bottom, black arrows), most likely reflecting the restricted expression of Hoxd13 at this stage, in a small subset of the dissected cells. In the same context, contacts established both by Hoxd4 and Hoxd9 with the centromeric part of the cluster extended towards the end of the cluster along with the developmental stage, such that both reached Hoxd12 in E10.5 samples, whereas Hoxd13 was only weakly contacted, corresponding to the results obtained when using Hoxd13 as a bait (Figure 3; bottom left).

17Therefore, it appears that the colinear time sequence in Hox genes activation is paralleled by a progressive transition in the chromatin structure, with a positive domain gaining in size along with time, at the expense of the negative domain, as best seen by the extension of Hoxd4 contacts. At E8.5, these interactions extended up to Hoxd8-Hoxd9. In E9.5 samples Hoxd10 was clearly contacted, and in E10.5 Hoxd11 and Hoxd12 were also involved (Figure 3, right column). These dynamic topologies suggest a stepwise transition of the genes from the negative to the positive compartment, rather than the switch of large groups of multiple transcription units, following a discrete and global chromatin re-organization.

18During axial extension, Hox genes are activated in the most posterior aspect of the elongating embryo. It is thus possible that cells implementing this stepwise transition in chromatin domains can fix and memorize their bimodal distribution once they exit the posterior zone of activation, leading to the colinear Hox conformations observed along the AP-axis. Accordingly, one would expect cellular territories along the developing body axis to maintain the same bimodal combinations as those established at the time of their origin, during early axial extension. We looked at the similarities in bimodal profiles between posterior samples dissected at different times on the one hand, and various samples micro-dissected at different body levels, from E10.5 embryos, on the other hand (Figure 4).

19The profile obtained from E8.5 PSM (Figure 3, top), right at the onset of Hoxd9 activation globaly aligned with that observed in the ‘anterior trunk’ sample at E10.5 (Figure 4, top), that is a cellular domain with a posterior boundary positioned approximately at the Hoxd9 anterior limit of expression. In both cases, Hoxd9 clearly contacted the negative domain, as defined by the Hoxd13 contacts (Figure 4, top), whereas some weak contacts were also scored with the positive domain, as determined by Hoxd4 contacts, indicating that the posterior limit of the dissection was slightly below the Hoxd9 boundary. These contacts were somehow stronger, in proportion, in the E8.5 than in the E10.5 dissection.

20At E9.5 (Theiler stage 14), the ‘tail bud’ (i.e., from the start of the non-segmented mesoderm) was dissected from ca. somite 22 to 25 and caudally. At this stage, Hoxd9 is robustly transcribed, whereas Hoxd11 has just started transcriptional activation. The interaction profiles obtained in this tissue were most similar to those obtained when a fragment of E10.5 trunk was dissected out that grossly corresponded to the future lumbo-sacral region, at levels 22 to 28, that is the AP levels supposedly produced in the E9.5 tail bud (Figure 4, middle). At this AP level, Hoxd9 is fully activated and this was reflected by the quasi absence of contact with Hoxd13 whereas, conversely, strong interactions appeared with the active part of the cluster (Figure 4, top ratio: red and blue shading). This was controlled by using Hoxd4 as bait, since contacts were now clearly scored with Hoxd9 and Hoxd10 and, to a lesser extent, with Hoxd11 (Figure 4, top ratio: blue shading).

21In this lumbo-sacral sample, neither Hoxd12 nor Hoxd13 are as yet transcribed, which coincided with the absence of contact between Hoxd13 and the active part of the gene cluster (Figure 4, middle, left). On the other hand, Hoxd11 expectedly displayed a mixed interaction profile, contacting both the negative and positive domains, likely reflecting the presence of both expressing and non-expressing cells (Figure 4, middle). In the most caudal piece of the E10.5 mouse embryo, interactions between Hoxd12, Hoxd13 and the positive domain were finally detected, suggesting that the entire cluster falls into a single spatial domain (Figure 4, bottom ratio: purple shading). Here again, however, though the interactions were significant, they were not particularly strong, suggesting the presence of a mixed cell population. Based on these data, we propose that the bimodal distributions are frozen in those cells leaving the zone of proliferation, at the caudal aspect of the embryo where temporal colinearity is potentially processed. These 3D structures, and hence the Hox transcription programs, will thus be maintained and memorize the various AP levels from which they originate.

22In different developmental contexts, the transcriptional activity of Hoxd genes coincides with an overall remodeling of long-range chromatin interactions with the flanking gene deserts, which harbor essential enhancer elements active in these developing tissues. Colinear activation of Hoxd genes along the developing trunk is thought to primarily rely on regulatory influences intrinsic to the gene cluster itself. However, and even though their importance remains unclear, contributions of the flanking regulatory landscapes in this process have been proposed. Therefore, we assessed whether or not the reported changes in local interactions are associated with variations in long-range contacts during temporal colinearity, as was observed during limb and intestinal development.

23By using a recently developed analytical methodology, we found that all interrogated Hoxd genes displayed substantial interactions with the flanking gene deserts (Figure 5A, Figure 5—figure supplement 1A). The quantification of interactions over both the centromeric and telomeric gene deserts revealed a gene-specific interaction preference towards either one or the other desert (Figure 5B), similar to what was previously described in limb bud cells. However, in marked contrast, the dynamics of these long-range chromatin interactions were moderate, if any, and no clear modification in the contact profiles were detected between the inactive state in ES cells, and the subsequent transcriptional activation (Figure 5B,C). Hierarchical clustering of global patterns of long-range interactions revealed that the Hoxd4, Hoxd9 and Hoxd11 viewpoints systematically cluster together, whereas the Hoxd13 viewpoint always behaves as outlier (Figure 5—figure supplement 1B).

24This clustering of interactions matches with the position of a previously mapped boundary between ‘topological associated domains’ (TADs;; ). In ES cells indeed, two TADs cover approximately the gene deserts on either side and have their border at the level of the Hoxd12-Hoxd11 genes (; Figure 5A, Figure 5—figure supplement 1A). Virtual 4C from HiC data with bins that cover the Hoxd genes show highly similar priming of interactions with the surrounding TADs when compared to our 4C-seq analysis, confirming that both approaches score similar chromatin behavior (Figure 5—figure supplement 2; for analysis strategy). During limb development, genes located near this boundary (Hoxd9 to Hoxd11) change their tropism and switch their contacts from one TAD to the other, such as to interact sequentially with the appropriate enhancers. This structural re-organization is clearly illustrated in our hierarchical clustering with Hoxd11 changing its association from Hoxd4 to Hoxd13 (Figure 5—figure supplement 1C). During temporal colinearity, however, such structural re-organization is not observed and hence the stepwise transcriptional activation of Hoxd genes appears to occur within a largely constitutive framework of long-range interactions; genes up to Hoxd11 interact mostly with the telomeric domain, either before or after their activation, and Hoxd13 always interacts with the centromeric domain. This opposed tropism for Hoxd genes in ES cells, as revealed by HiC and the consequent TAD structures, is somewhat at odds with the local clustering of Hoxd genes when in a negative state, which we report here by using 4C. This paradox is discussed below.

25In this study we present the dynamics of local and long-range 3D chromatin organization during temporal colinear activation of Hox genes in vivo. Prior to their sequential activation, Hox genes are organized into local 3D chromatin compartments that encompass all bivalently marked chromatin. In ES cells, however, these 3D compartments appear less defined than the fully inactive and H3K27me3-only marked compartments observed in differentiated cells. This difference in discreteness was paralleled by the amount of observable background transcription, which was substantially higher in ES cells that in brain cells, though the biological impact of these very low level transcripts, if any, remains to be determined. This supports the view whereby a generally relaxed chromatin organization in ES cell may accompany the plasticity required for cell-faith commitment. Interestingly, in ES cells grown in the presence of two kinase-inhibitors and thought to be in a somehow more naïve developmental state, H3K27me3 marks are almost absent from Hox clusters. Considering the high correlation between the presence of H3K27me3 marks and the existence of 3D chromatin compartments, we would anticipate Hox clusters within these ES cells not to group into 3D compartments as distinct as those visible in ‘canonical’ ES cells. We therefore hypothesize that structuring into fully inactive 3D compartments is a gradual process occurring over the course of several days during embryonic development.

26We also observe that the sequential transcriptional activation of Hox genes in the PSM coincides with a gene-by-gene transfer or positioning from the inactive H3K27me3-decorated compartment to a newly formed 3D compartment containing active genes only. This indicates that the presence of Hox genes in 3D compartments of various extents, along the developed body axis is not an a posteriori mechanism used to fix and secure the long-term maintenance of various states of activity, fixed earlier by ‘classical’ transcriptional regulations acting in trans. Instead, it suggests that such spatial structures are instrumental in the precise regulation of their transcriptional timing. This is observed at least for the Hoxd9 to Hoxd13 genes and we infer that the same process occurs in the part of the cluster containing from Hoxd1 to Hoxd8. It is however not possible to assess this experimentally due to technical limitations associated with the size of the embryonic material at the corresponding developmental stages.

27Temporal colinearity was originally proposed as a mechanism to translate time into spatial coordinates, in different ontogenic contexts. While our results support this idea, the transcriptional timing associated with gene clustering may not be an absolute prerequisite to achieve the proper spatial distributions of Hox genes products, as suggested by the multiple cases where single mammalian Hox transgenes could largely recapitulate the major expression specificities along the AP-axis. In this context, it is possible that the progressive transition of Hox genes from an inactive to an active 3D compartment reflects the existence of a mechanism whose major aim would not be to precisely regulate a time sequence but instead, to protect the most ‘posterior’ Hox genes from a premature exposure to activating factors, a situation shown to block posterior elongation and hence to be detrimental to the embryo. The necessity to actively prevent the most posterior genes from premature activation is supported by their basal transcriptional activity in ES cells, where Hox clusters are less discrete than in subsequent negative tissues such as fetal brain cells. Such a basal activity, which was not scored in these latter cells, may reflect the rather generic nature of the activating signals, the general mechanism underlying temporal colinearity thus relying on de-repression.

28Studies using internal Hox cluster deletions and duplications indeed showed that the relative position of Hox genes, rather than their promoters, determines their responses to activating signals. In this view, graded signals emanating from the posterior aspect of the developing embryo would lead to a progressive de-repression of Hox clusters, implying that these clusters would display some directional sensitivity. While the nature of the activating factors is elusive, a link with the segmentation clock was proposed. Concerning the directional sensitivity, Polycomb group (Pc-G) gene products may play an important role in this process, as the distribution of H3K27me3 marks correlate with the size of the inactive 3D compartments. Recently, a somewhat graded distribution of both EZH2 and RING1B, two proteins members of the PRC2 and PRC1 complexes, respectively, was described over the HoxD cluster in ES cells, with the highest signals covering the most ‘posterior’ genes. Directionality may therefore derive from a weaker ‘anterior’ repression exerted by the Pc system. In this context, progressive alterations of the repressive system should sensitize the transcriptional threshold, while keeping on with directionality. This effect was observed in Cbx2−/− mutant embryos (a component of the PRC1 complex formerly known as M33), where the efficiency of the PRC1 complex was moderately decreased: RA treatment resulted in premature yet colinear activation of Hoxd genes. Alternatively, collinear activation may rely upon a different kind of model involving for example biophysical forces. Future experiments where the process will be witnessed at the cellular level in real time may be informative in this context.

29Hox genes are originally activated in the most posterior aspect of the gastrulating embryo. This initial wave of activation seems to involve first a poised transcriptional status, followed by an apparent anterior forward spreading, which will ultimately lead to the positioning and initiation of the expression domains in the pre-somitic mesoderm (PSM). The colinear processing of this early phase may involve preparatory modifications in the chromatin status, making the system poised for activation by factors emanating from posterior cells. In this view, the observed anterior forward spreading in expressing cells may reflect a prolonged exposure to low levels of signals diffusing from the posterior end of the primitive streak. A second (non-exclusive) possibility is that it illustrates the initial difficulty to maintain a robust boundary in Pc repression in a gene cluster where some anterior genes are fully active, with a tendency for the nearby-located genes to be de-repressed and activated.

30However, our results suggest that once the expression is finally established within the PSM, the boundary between the active and inactive compartments remain rather stable for the next couple of days, until the axial skeleton is fully determined. In this view, these chromatin domains may represent part of the machinery used to fix a given state of activation and thus translate a temporal parameter into spatial coordinates. As such, early heterochronies in Hox gene activation within the PSM will lead to subsequent re-positioning of the expression boundary, as previously observed.

31By using genetic approaches, it was previously argued that the time-sequenced activation of Hoxd genes primarily uses regulatory influences located within the gene cluster itself, with some contributions coming from more distant flanking regions. We now report that such a transcriptional activation is implemented with little-if any-differences in the interaction profiles between the target genes and their neighboring gene deserts, unlike the situation observed during limb development where new contacts appear upon gene activation. However, temporal colinearity does occur within a framework of constitutive long-range interactions, which may provide a scaffold helping the bimodal separation of active and inactive genes to take place. Further experiments with mice carrying large re-arrangements of these two gene deserts will be necessary to clearly weight the importance of flanking regions in the implementation of the Hox clock.

32Finally, while the comparison between published HiC data and our 4C datasets are generally highly consistent (e.g., Figure 5, Figure 5—figure supplement 1; ), the data reported here using ES cells raise an apparent paradox. HiC analysis in ES cells identified a boundary between topological domains positioned around the Hoxd12 to Hoxd11 gene (; Figure 5—figure supplement 2D) and such boundaries are thought to impose or reflect a physical separation between the two interaction landscapes (e.g., ). As a consequence, Hoxd13 should display more interactions with its flanking gene desert than with the other part of the HoxD gene cluster. Yet, by using several viewpoints in a 4C set-up, the HoxD cluster in ES cells appears to form a single negative compartment, despite the interspersed presence of this TAD boundary (Figure 1). In fact, a detailed analysis of the HiC dataset reveals that the HoxD cluster itself forms a ‘micro-TAD’, displaying strong internal interactions, in agreement with the 4C results reported here. As such, we consider it likely that the TAD boundary identified by HiC in ES cells represents an average description of two distinct configurations (Figure 6). For each allele, either the most posterior Hoxd13 gene forms stable interactions within the TAD on the centromeric side or, alternatively, the Hoxd11 to Hoxd1 genes interact with the TAD on the telomeric side (Figure 6). As a consequence, for each allele the entire HoxD 3D compartment becomes located towards a single TAD, on one side of the cluster with a physical separation from the other side. Molecule(s) causing these interactions are elusive and may include proteins that mediate constitutive loops between Hoxd gene promoters and their regulatory elements. The CTCF protein, which may play a role in scaffolding TADs, binds multiple sites around the Hoxd13 to Hoxd8 region. Because formaldehyde crosslinking has a very short range of action, the system generates a graded pattern of 4C and HiC interactions from the location of the actual binding sites. Therefore, while in ES cells and for each allele, TAD borders are likely located at either side of the HoxD 3D compartment (Figure 6A, black lines), our analysis of a large population of cells reflects the equilibrium that exists between these two situations (Figure 6E). At later stages, when the HoxD cluster adopts a bimodal 3D organization, the tethering of interactions, as illustrated by the existence of TADs on either side, may help implement the separation between activated and repressed Hox genes, thereby potentially reducing deleterious regulatory interferences and premature activation of the most posterior Hox genes.

33All experiments were performed in agreement with the Swiss law on animal protection (LPA). Tissue samples were isolated at the indicated time points (with maximum 6 hr delay), with day E0.5 being noon on the day of the vaginal plug. Tissue pieces for 4C-sequencing, ChIP-sequencing, RNA-sequencing and Reverse Transcriptase-qPCR were isolated in PBS and subsequently transferred to PBS supplemented with 10% Fetal Calf Serum. 4C-seq and ChIP-seq material was incubated for 45 min with 1 mg/ml collagenase (Sigma-Aldrich, St. Louis, MO), and 4C-seq material was further made single cell using a cell strainer (BD Falcon).

34Mouse ES cells were grown under feeder-free conditions on gelatinized plates in Dulbecco's modified Eagle's medium (DMEM, Life Technologies, Carlsbad, CA) supplemented with 17% fetal calf serum, 1 × non-essential amino acids (Life Technologies), 1 × Pen–Strep (Life Technologies), Sodium Pyruvate (Life Technologies), 0.1 mM β-mercaptoethanol, and 1000 U/ml LIF.

35Embryonic 4C-seq samples consisted of pooled material from multiple embryos: 129 embryos for E8.5 pre-somatic mesoderm samples, 196 embryos for E9.5 tail bud, 143 embryos for E10.5 tail bud or E10.5 lumbo-sacral trunks and around 20 embryos for each E10.5 forebrain sample. For embryonic ChIP samples, 50 μg of chromatin was cross-linked at a time, of which 10 μg was used per ChIP. To obtain 50 μg of chromatin, 750 E8.5 pre-somatic mesoderm samples or 10 E10.5 forebrains were pooled. Total RNA from E10.5 forebrain was isolated from single embryos. ES cell 4C-seq and ChIP-seq samples were prepared from samples consisting of 20 million cells. 10 μg of ES cell chromatin was used per ChIP. Total RNA was isolated from 1 million cells.

364C–seq libraries were constructed as previously described. NlaIII (New England Biolabs, Ipswich, MA) was used as the primary restriction enzyme and DpnII (New England Biolabs) was used as the secondary restriction enzyme. For each viewpoint, a total of 1 μg (E9.5 tail bud, E10.5 tail bud, E10.5 lumbo-sacral trunks, E10.5 forebrain and ES cells) or 50 ng (E8.5 pre-somatic mesoderm) of each 4C-seq library was amplified using 16 individual PCR reactions with inverse primers including Illumina Solexa adapter sequences (primer sequences in Table 2). Illumina sequencing was done on multiplexed samples, containing PCR amplified material of up to 7 viewpoints, using 100 bp Single end reads on the Illumina HiSeq system according to the manufacturer's specifications.

374C-seq reads were sorted, aligned, and translated to restriction fragments using the 4C-seq pipeline of the BBCF HTSstation (available at http://htsstation.epfl.ch;; ) according to ENSEMBL Mouse assembly NCBIM37 (mm9). 4C-seq patterns were corrected vs previously generated random 4C–seq libraries, consisting of BACs covering the mouse Hox clusters (HoxD: RP23-331E7; HoxC: RP23-430C12; HoxB: RP23-381I12 and RP23-196F5; HoxA: RP24-298M24). After random correction, three restriction fragments were removed that returned aberrant values (HoxD: chr2:74’597’000-74’597’732; chr2:74’608’796-74’609’312, HoxB: chr11:95’999’958-96’000’916) due to sequence abnormalities in the BAC template (confirmed by Sanger sequencing; not shown). Normalization and further data processing was done as previously described. Quantitative log2 ratios were calculated by dividing the quantitative fragment count between tissue samples. Unprocessed 4C-seq data is available from the Gene Expression Omnibus (GEO) repository under accession number GSE55344. Random corrected tracks are available from http://duboule-lab.epfl.ch/data.

38The directionality of long-range interactions was calculated as previously described. In Figure 5A, the smoothed 4C-seq patterns (running mean, window size 11) were obtained using the 4C-seq pipeline of the BBCF HTSstation (available at http://htsstation.epfl.ch; ). HiC data on topological associated domains (TADs) from ES cells were obtained from (http://chromosome.sdsc.edu/mouse/hi-c/database.php; ). Two TADs located centromeric and telomeric of the clusters were selected, covering genomic coordinates chr2:73400000-75960000 (discussed in ). Spearman correlation of long-range patterns was done over the region covering these TADs, with signal on the HoxD cluster itself removed (excluded region: chr2:74484971-74607492). Conventional hierarchical clustering was done to score for relationships between viewpoints.

39ChIP was performed as previously described. Cells were fixed for 5 min in a 2% formaldehyde solution at room temperature. ChIP-seq samples were fragmented to a range of 200–500 bp using tip sonication (Misonix S4000, Misonix, Farmingdale, NY), For all ChIP assays, 10 μg of cross-linked chromatin was used. Antibodies used: anti Histone H3K27me3 (#17-622; Millipore, Billerica, MA) and anti H3K4me3 (#17-614; Millipore). ChIP-seq libraries were constructed from 6 to 10 nanograms of immune-precipitated DNA according to the manufacturers instructions (Illumina, San Diego, CA). Sequencing was done using 50 or 100 bp Single end reads on the Illumina HiSeq system according to the manufacturer's specifications. ChIP-seq reads were mapped to ENSEMBL Mouse assembly NCBIM37 (mm9), and extended to 100 bp if read lengths smaller than 100 bp were used, using the ChIP-seq pipeline of the BBCF HTSstation (available at http://htsstation.epfl.ch; ). ChIP-seq data is available from the Gene Expression Omnibus (GEO) repository under accession numbers GSE55344 and GSE31570.

40Random corrected 4C-seq and ChIP-seq samples were correlated by ranking experimental values within restriction fragments (Table 1). First, to each NlaIII restriction fragment covered by the random 4C tracks within the regions visualized in Figures 1 and 2 (HoxD cluster: chr2:74454783-74622413; HoxC cluster: chr15:102715179-102909417; HoxB cluster: chr11:95992344-96244915; HoxA cluster: chr6:52058584-52234371), the average ChIP-seq signal was assigned for each condition. Restriction fragment within individual samples were ranked based on their 4C-seq or ChIP-seq value and subsequently the Spearman's rank correlation coefficient was calculated between pairs of samples.

41Total RNA from tissue samples was isolated using Trizol LS reagent (Life Technologies). Total RNA from ES cell samples was isolated using Trizol reagent (Life Technologies). For RNA-seq, the RNA was depleted from rRNAs and, subsequently, strand-specific total RNA-seq libraries were constructed according to the manufacturers instructions (Illumina). Sequencing was done using 50 bp Single end reads on the Illumina HiSeq system according to the manufacturer's specifications. RNA-seq reads were mapped to ENSEMBL Mouse assembly NCBIM37 (mm9) and translated into reads per gene (RPKM) using the RNA-seq pipeline of the BBCF HTSstation (available at http://htsstation.epfl.ch; ). RNA-seq data is available from the Gene Expression Omnibus (GEO) repository under accession numbers GSE55344. For RT-qPCR, cDNA was synthesized after DNAseI treatment (Life Technologies) using SuperScript III (Life Technologies) and oligo-dT primers (Life Technologies), using the manufacturer's instructions. For ES cells and E10.5 forebrain, 2 μg of RNA was used as input for the cDNA synthesis, for E10.5 posterior trunk 1 μg of RNA was used. Products were quantified by qPCR using EXPRESS SYBR GreenER mixes (Life Technologies) on a CFX96 PCR Detection System (BioRad, Hercules, CA). Sequences of intron-spanning primers are provided in Table 3.