Introduction

The chromosomes of multicellular animals have a regular and inheritable physical organization. This was first recognized in studies on the lampbrush chromosomes in amphibian oocytes and the polytene chromosomes of insects (Callan, 1963; Nora et al., 2020; Zhimulev and Belyaeva, 1991; Zhimulev et al., 1983). Subsequent research has shown that the architectural features inferred from analysis of lampbrush and polytene chromosomes are present in chromosomes throughout much of the animal kingdom. The key organizational principle is the subdivision of the chromatin fiber into a series of independent looped domains, commonly called “TADs” (Cavalheiro et al., 2021; Chetverina et al., 2017; Jerković et al., 2020; Matthews and White, 2019; Rowley and Corces, 2018). With important exceptions, the arrangement of TADs along a given chromosome tend to be invariant and are largely independent of the cell type or developmental stage. This regular and inheritable organization is a reflection of the underlying mechanism of TAD formation. TADs are separated from each other by special elements called boundaries or insulators. While these elements have been found in many different species, they have been most fully characterized in Drosophila (Cavalheiro et al., 2021; Chetverina et al., 2017). Fly boundaries span DNA sequences of 150 bp to 1.5 kb in length and contain one or more nucleosome-free nuclease-hypersensitive regions. These nuclease-hypersensitive regions are targets for a large collection of DNA binding proteins that have been implicated in boundary function. Some of the fly boundary factors are widely conserved (CTCF, BEN family proteins, GAF), while others appear to be restricted to insect linages (Su(Hw), Pita, Zw5, Pita, Mod(mdg4), BEAF) (Heger et al., 2013; Heger and Wiehe, 2014; Schoborg and Labrador, 2010).

Boundary elements in flies are not only responsible for organizing the chromatin fiber, they also have genetic activities. When interposed between enhancers or silencers and target promoters, boundary elements block regulatory interactions (Bell et al., 2001; Chetverina et al., 2014; Chetverina et al., 2017). This insulating activity provides a mechanism for delimiting units of independent gene activity: genes located between a pair of compatible boundaries are subject to regulatory interactions with enhancers/silencers present in the same chromosomal interval, while they are insulated from the effects of enhancers/silencers located beyond either boundary in adjacent regulatory neighborhoods. Genetic studies suggest that the insulating activity of boundary elements is a consequence of subdividing the chromosome into a series of topologically independent domains (Cai and Shen, 2001; Gohl et al., 2011; Muravyova et al., 2001). Organizing the chromatin fiber into looped domains enhances contacts between sequences within the loop, while it suppresses contacts with sequences outside of the loop. While it is not known whether a similar insulation mechanism is at play in mammals, the fact that mammalian chromosomes are also subdivided into TADs by boundary-like elements suggests that it might be.

A critical question in chromosome biology is the mechanism(s) responsible for determining the endpoints of the loop domains, the TADs. In mammals, a novel loop-extrusion model has been proposed to not only generate chromosomal loops but also determine the endpoints of those loops (Alipour and Marko, 2012; Fudenberg et al., 2016; Guo et al., 2012; Guo et al., 2015; Nuebler et al., 2018; Sanborn et al., 2015). In this model, a cohesin complex initiates loop formation at a loading site within the “loop-to-be” and then extrudes a chromatin loop until it bumps into CTCF-dependent roadblocks, one on each side of the extruding loop. The location of these roadblocks combined with the processive action of the cohesin complex determines the endpoints of each TAD (Davidson and Peters, 2021; Ghosh and Meyer, 2021; Mirny and Dekker, 2022; Perea-Resa et al., 2021). A key assumption of the loop-extrusion model is that mammalian boundaries are fully autonomous: they are roadblocks, and their physical presence in and of itself is sufficient to define the loop endpoint, independent of the functional properties of neighboring boundaries. In a more specific variant of this model, the relative orientation of the paired CTCF roadblocks is also important—in order to halt cohesin-mediated extrusion, the CTCF sites on each side of the loop must have a convergent orientation. In this model the CTCF roadblocks are able to block an oncoming cohesin complex in only one direction; however, their intrinsic blocking activity is independent of the orientation of other CTCF sites in their neighborhood.

While loop extrusion is widely thought to be operative in mammals, the evidence in flies is inconsistent with this mechanism. Genetics studies have shown that fly boundaries are functionally non-autonomous and that their activities in both loop formation and gene regulation depend upon their ability to engage in direct physical interactions with other boundaries (Chetverina et al., 2014; Chetverina et al., 2017). That fly boundaries might function by physical pairing was first suggested by studies which showed that the gypsy transposon su(Hw) boundary and the bithorax complex (BX-C) Mcp boundary can mediate regulatory interactions (enhancer/silencer: reporter) between transgenes inserted at sites separated by 15-32 MB or more (Muller et al., 1999; Sigrist and Pirrotta, 1997; Vazquez et al., 1993). Consistent with direct physical interactions, these distant transgenes co-localize in vivo, and in living tissue remain in contact for extended periods of time (Chen et al., 2018; Li et al., 2011; Vazquez et al., 2006).

Some of the parameters that govern pairing interactions have been defined in two different transgene assays, insulator bypass and boundary competition. In one version of the bypass assay, a set of enhancers are placed upstream of two different reporters (Cai and Shen, 2001; Kyrchanova et al., 2008a; Muravyova et al., 2001). When a single boundary is introduced between the enhancers and the closest reporter, the enhancers are unable to activate either reporter. However, if the same boundary is placed downstream of the closest reporter, bypass is observed. In this case the closest reporter, which is bracketed by the two boundaries, is still insulated from the enhancers; however, the downstream reporter is activated (Muravyova et al., 2001). Kyrchanova et al. (Kyrchanova et al., 2008a) showed that bypass activity is, in most cases, orientation-dependent. When the same boundary is placed in the upstream and downstream position, they typically need to be introduced in the opposite orientation. The reason for this is that self-pairing is head-to-head and this configuration generates a stem-loop topology, bringing the upstream enhancers in close proximity to the downstream reporter. When the two boundaries are introduced in the same orientation, the enhancers do not activate the downstream reporter because a circle-loop rather than a stem-loop is formed (Chetverina et al., 2017; Kyrchanova et al., 2008a). Head-to-head pairing seems to be a common feature of self-pairing interactions between fly boundaries and was observed for scs, scs’, iA2 and wari (Kyrchanova et al., 2008a). A preference for head-to-head interactions is also observed for heterologous pairing interactions between different boundaries in the Abdominal-B (Abd-B) region of the BX-C (Kyrchanova et al., 2011; Kyrchanova et al., 2008b). There are of course some exceptions. The BX-C Fab-7 boundary pairs with itself and with its neighbor Fab-8 in both orientations (Kyrchanova et al., 2011; Kyrchanova et al., 2008b). Bypass activity also depends upon finding the proper match. For example, when multimerized dCTCF or Zw5 binding sites are placed both upstream and downstream of the closest reporter, bypass is observed. However, there is no bypass when multimerized dCTCF sites are tested with multimerized Zw5 sites (Kyrchanova et al., 2008a). Similar results have been reported for the Zipic and Pita factors (Zolotarev et al., 2016). Boundary:boundary pairing preferences are also observed in competition assays in which different boundaries are introduced into the same transgene (Gohl et al., 2011). In these assays, the insulating activity of a boundary in the blocking position (between an enhancer and a reporter that is flanked by a 3’ boundary) is challenged by placing a heterologous boundary upstream of the enhancer. If the boundary upstream of the enhancer is a better match for the boundary located downstream of the reporter, then the insulating activity of the boundary in the blocking position can be compromised or lost altogether. These and other experiments argue that fly boundaries are functionally non-autonomous—that is, their activities depend upon their ability to physically interact with other boundaries.

However, the differences between the two models are not limited to the mechanisms for determining the endpoints of TADs. The models also differ in the possible topologies of the chromatin loops that form TADs. Only two loop topologies can be generated by the cohesin-extrusion mechanism. One is a stem-loop while the other is an unanchored loop. Depending upon the location of the road-blocks that halt cohesin extrusion and the cohesin loading sites, the chromatin fiber would be organized into a series of stem-loops. These stem-loops will be separated from each other by unanchored loops (see Fig. 1B). This is the configuration that is often illustrated in articles discussing the loop-extrusion model. If the relevant road-blocks are very closely juxtaposed, then the unanchored connecting loops will disappear and be replaced by a series of stem-loops which point in opposite orientations, as illustrated in Fig. 1C. Stem-loops are also possible in the boundary-pairing model, and they will be formed when two heterologous boundaries pair with each other head-to-tail (see Fig 1D) in cis. Like loop-extrusion, the stem-loops could be separated from each other by unanchored loops (Fig. 1E). If the boundaries in the neighborhood pair with each other head-to-tail, a series of connected stem-loops pointing in opposite orientations will be generated (Fig. 1D). In this case, the main axis of the chromosome would be a series of paired boundaries (Fig. 1F).

Figure 1.

Stem-loops are not, however, the only loop topology that can be generated in the boundary-pairing model. If the neighboring boundaries pair with each other head-to-head instead of head-to-tail, a circle-loop structure will be generated (Fig 1G). Like stem-loops, the circle-loops could be connected by unanchored loops (Fig. 1H). Alternatively, if boundaries pair with both neighbors head-to-head this will generate a series of linked circle-loops oriented in (more or less) the same direction (see Fig. 1I).

In the studies reported here we have critically evaluated these two models. We have used a combination of MicroC to analyze how TADs are organized and experimental manipulations to test the predictions of the “loop-extrusion” and the “boundary-pairing” models. For these experimental manipulations, we have used the well-characterized homie boundary from the even skipped (eve) locus. homie together with nhomie flank the 16 kb eve locus, and these two elements share many of the properties of other fly boundaries (Fujioka et al., 2013; Fujioka et al., 2009). Like the gypsy and Mcp boundaries, nhomie and homie can mediate long-distance regulatory interactions (Fujioka et al., 2016). For example, a lacZ reporter transgene containing homie (or nhomie) inserted on one homolog can be activated by an eve enhancer in a homie- (or nhomie-) containing transgene inserted on the other homolog over a distance of 2 Mb. Long-distance activation is also observed in heterologous combinations of nhomie and homie, but not in combinations with phage lambda DNA. Similarly, regulatory interactions are observed when a homie- or nhomie- containing reporter is inserted at an attP landing site in the hebe gene, 142 kb upstream of the eve locus (Chen et al., 2018; Fujioka et al., 2016; Fujioka et al., 2009). In this case, the eve enhancers drive reporter expression during development in a pattern that mimics the stage-and tissue-specific expression of the endogenous eve gene. Activation also depends upon the orientation of the boundary (5’→3’) relative to the reporter in the transgene. For homie, the reporter must be located “upstream” of the boundary (5’→3’), just like the eve enhancers (and eve gene) are “upstream” of endogenous homie (see Fig. 2B). Little or no activation is observed when the relative orientation is reversed so that the reporter is located “downstream” of homie. For nhomie, the reporter must be located “downstream”, in the same relative position as that in which the eve enhancers and eve gene are located with respect to endogenous nhomie (5’→3’: Fig. 2B).

Figure 2.

Results

The eve TAD

We used MicroC to probe the TAD organization of the eve locus and the genomic region that extends from eve to just beyond the attP site at –142 kb. For this purpose, we chose nuclear cycle 14 embryos. At this stage of development, most nuclei in our sample will have completed replication and will be arrested in G2. As sister chromosomes in flies pair with each other in register, there will be two juxtaposed copies of this region of the chromosome. The sister chromosomes should be linked to each other by cohesin in preparation for the asynchronous nuclear divisions at the end of NC14 (Collier and Nasmyth, 2022; Gligoris et al., 2014). While homolog pairing is limited in pre-cellular blastoderm embryos, there are regions of the chromosome that already show evidence of pairing (Fung et al., 1998). In fact, transvection between gypsy insulator-containing transgenes has been observed in live imaging studies of nuclear cycle 14 embryos (Lim et al., 2018). Since the NC14 nuclei are just emerging from an earlier mitosis and a round of DNA synthesis, one would imagine that other key features of chromosome organization might also be in the process of being established. Thus, it should be possible to capture evidence of intermediates like those expected to be generated as the extruding cohesin complexes move through the eve TAD and other nearby loci at this particular stage of development.

Shown in Fig. 2A is a blow-up of the TAD organization of the eve locus. As expected, the left and right boundaries of the eve TAD correspond to nhomie and homie, respectively, and sequences containing these two boundaries converge at the apex of the eve interaction triangle. All of the sequences within the TAD appear to come in contact with each other and define a high density of internal contact domain (HDIC). These contacts are expected to be generated, at least in part, by the sliding of the chromatin fiber within the TAD against itself, much like sequences in circular/supercoiled DNA can interact with each other (Fig. 1E). There are also sequences within the eve TAD that show enhanced interactions. For example, there is a darker interaction triangle connecting nhomie and the promoter region, while another triangle connects homie (and the PRE adjacent to homie) to the promoter region. These enhanced internal contacts would be expected to help insulate the eve gene from external regulatory elements and at the same time facilitate interactions between eve and its upstream and downstream enhancers.

As predicted by both the loop-extrusion and boundary-pairing models, the eve TAD is a stem-loop, not a circle-loop (Fig. 2B). This is evident from the prominent “volcano plume” at the apex of the eve TAD (volcano triangle). A volcano plume is observed because sequences flanking eve are brought into close proximity when homie and nhomie are linked to each other by either cohesin or by head-to-tail pairing (see Fig.1: see also Ke et al.). Fig. Supplemental 1 shows the hierarchy of interactions between TADs to either side of eve (L-M, K-M and J-M in Sup. Fig. 1). This conclusion is supported by previous transgene studies. As first described in the boundary bypass experiments of Cai and Shen (Cai and Shen, 2001) and Muravyova et al. (Muravyova et al., 2001), the formation of a stem-loop structure when two boundaries are linked together brings sequences immediately beyond the linked boundaries into close proximity with each other. In this assay, a stem-loop structure is required, as it enables enhancers flanking the “upstream” boundary to interact with reporters flanking the downstream boundary (Kyrchanova et al., 2008a). In contrast, a circle-loop configuration does not bring the neighbors to either side of eve into close proximity.

While a stem-loop is the topology expected for the eve TAD in both models, one feature of the TAD is inconsistent with the loop-extrusion model. According to this model, loops (TADs) are generated by a cohesin complex that forms a small bubble at a loading site somewhere in the TAD-to-be and then extrudes both strands of the growing loop at an equal rate until it bumps into the boundary road-blocks (Davidson and Peters, 2021; Ghosh and Meyer, 2021; Mirny and Dekker, 2022; Perea-Resa et al., 2021). If the loading site were located in the middle of the TAD-to-be, and the extrusion intermediates in a population of nuclei were captured by crosslinking, one should observe a crosslink-generated stripe that begins at the loading site and extends perpendicular to the chromosomal DNA until it reaches the apex that links the two TAD boundaries with imaginary lines of 45° (Fig. Supplemental 2B). This is the pattern that would be generated by a “broken zipper” when it transiently links the zipper teeth equidistant from the initial “loading site”. If the loading site is off center to the left, the crosslinking of the passing cohesin complex in a population of nuclei will generate a perpendicular stripe until cohesin come to a halt when it encounters the roadblock on the left (Fig. Supplemental 2C). Assuming that the complex continues to extrude the right strand, it will then generate a stripe at 45° degrees that extends to the apex of the TAD. If the loading site corresponds to one of the CTCF roadblocks or is located close to the roadblock, the cohesin complex will generate a 45° stripe that comes to a halt at the apex of the triangle when it encounters the neighboring CTCF roadblock (Fig. Supplemental 2D and E). If there are multiple loading sites within the TAD, there should be a series of perpendicular stripes that generate and sequentially reinforce the stripe(s) at 45° (Fig. Supplemental 2F). However, in spite of the fact that NC14 fly embryos should provide by far the best prospect of actually capturing loop extrusion intermediates, there are no vertical stripes in the eve TAD, nor are there stripes at 45° that extend to the apex of the eve TAD. Nor do we observe a population of perpendicular stripes that convert into a series of reinforced 45° stripes. There is also no evidence of stripes that are initially tilted to the left or right of a “loading site”, as might be observed if the rate of extrusion were unequal on the two strands, and then subsequently convert to a 45° stripe as cohesin comes to a halt at the first boundary encountered.

TADs in the eve environs

The attP insertion site at –142 kb is near the end of the first exon of the hebe transcription unit (Fig. 2C). It is located just within a ∼14 kb TAD, TB, that has a high density of internal contacts, and extends from the hebe promoter to the 3’ end of the neighboring dila gene. In between hebe and the eve locus there at least are ten distinct TADs, TC-TL (Fig. 2C). The endpoints of most of these TADs correspond to sequences that are associated with one or more known chromosomal architectural factors (dots in Fig. 2C).

These TADs correspond to the fundamental building blocks for organizing the 3D architecture of this chromosomal segment. Superimposed upon these TADs are regions that exhibit lower density of contacts (LDC). For example, the mef2 gene spans two TADs, TJ and TK. TK contains the mef2 distal promoter region and extends to the internal mef2 promoter, while TJ extends to near the promoter of the divergently transcribed Etf-QO gene (blue arrow). Both of these TADs are linked together by a rectangular LDC domain J-K (see Fig. Supplemental 2). Likewise, TJ is linked to its immediate neighbor TI by the LDC domain I-J. In the next level, TADs separated by a single TAD are linked together. In the region immediately above TJ, the LDC domain I-K links TI to TK (Fig. Supplemental 2). Similarly, TJ is linked to TL by J-L while TJ is also linked to the large TAD TM (which contains TER94 and the neighboring gene, pka-R2) on the other side of eve by J-M. The same pattern of a hierarchical series of LDC domains linking TAD neighbors, TAD next-door neighbors and next-next-door neighbors is observed in the DNA segment that includes the hebe gene.

--cohesin-mediated loop extrusion

Like eve, the TADs in the region between eve and the attP site in the hebe locus are defined by right angle triangles (volcanos) with a high density of internal contacts. However, unlike eve, these volcano triangles do not have a plume, and instead are surrounded by a series of LDC rectangular clouds. This is not the contact pattern that one might expect for either a series of stem-loops connected by unanchored loops or a series of stem-loops connected to each other. Moreover, there are no perpendicular or angled crosslinking stripes (broken zippers) emanating from the base of these triangles, nor are their stripes along their 45° legs. Thus, in spite of the fact that the NC14 nuclei used in this analysis are just emerging from an earlier mitosis and a round of DNA synthesis and should be in the process of assembling TADs, the crosslinking signatures that should be generated by the embrace of passing cohesin complexes are completely absent.

In the loop extrusion model, the LDC domains would arise in a subset of nuclei because the extruding cohesin complex breaks through one or more boundary roadblocks before its progression is halted (Fig. Supplemental 3A versus B, C and D) (Hsieh et al., 2020; Krietenstein et al., 2020). Breakthroughs could occur at only one roadblock on one side of the extruding loop (Fig. Supplemental 3B and C), or at multiple roadblocks on one or both sides (Fig. Supplemental 3D), giving rise to a set of overlying LDC domains like that evident in the region between hebe and eve. However, while there would have to be multiple breakthrough events to account for the hierarchical array of LDC domains observed in these NC14 embryos, there is no evidence of the 45° crosslinking stripes that would be expected to mark the legs of the LDC domains following the breakthrough. For example, the I-J LDC domain (Fig. Supplemental 2) could be generated by cohesin complexes that initiated extrusion in either TI or TJ and then failed to halt at the TI:TJ boundary (green arrowhead). In either case, there should be a 45° stripe of crosslinking that marks (at least part of) the left (TI) and/or the right (TJ) leg of the I-J LDC domain and extends to the apex of the TI-TJ LDC triangle; however, there is no evidence of such stripes. Nor is there evidence of crosslinking stripes marking one or both legs of the J-K or I-K LDC domains (Fig. Supplemental 2).

--boundary:boundary pairing

At least two different mechanisms are expected to account for the TADs and LDC domains in the region between eve and the attP insertion site at -142 kb. First, since none of the TADs in this DNA segment are topped by volcano plumes, they could correspond to circle-loops. As circle-loops are generated by head-to-head pairing, the boundaries in the region between nhomie and hebe would be expected to pair with each other head-to-head. (The TAD to TAD contact maps generated by circle-loops and stem-loops are considered further in the accompanying paper, Ke et al.)

As illustrated in Fig. Supplemental 3E-G, the coiled circle-loop TADs, though topologically independent, are expected to be in relatively close proximity to each other. This means that in addition to crosslinking events within the TADs, there will be crosslinking events between neighboring TADs. These crosslinking events could generate the LDCs seen in Fig. 2C and Fig. Supplemental 2. Since TADs next to each other (Fig. Supplemental 3E and F; Fig. Supplemental 2 I-J or J-K) would typically be expected to interact more frequently than TADs separated by one, two or more TADs (Fig. Supplemental 3G; Fig. Supplemental 2 I-K or J-L), there should be a progressive reduction in contact frequency at each step. This is generally what is seen (see Fig. Supplemental 2). While connected stem-loops generated by cohesin extrusion should also bump into one another, contacts between next-door neighbors might be expected to be less frequent than contacts between next-next-door neighbors (Fig. 1B, C and Fig. Supplemental 3A). (TAD-to-TAD crosslinking and the impact of loop topology is further examined in Ke et al., 2023.)

The second mechanism would be switching and/or combining pairing partners (Fig. Supplemental 3H-J). Though imaging studies suggest that pairing interactions can be of relatively long duration (>30 min: Chen et al., 2018; Vazquez et al., 2006)), they are not permanent, and other nearby boundaries can compete for pairing interactions (Gohl et al., 2011). If switching/combining occurs in NC14 embryos (see Fig. Supplemental 3H-J), then the precise pattern of TADs and LDC domains could also be impacted by the relative avidity of potential pairing partners in the neighborhood and also the distances separating potential partners. For example, the rectangle linking TH and TI (H-I in Fig. Supplemental 2) could represent partner switching so that the boundary between TI and TJ (green arrowhead in Fig. Supplemental 2) sometimes pairs with the boundary separating TH and TI (blue arrowhead) and sometimes with the boundary separating TG and TH (red arrowhead in Fig. Supplemental 1). Distant boundary elements could also preferentially interact with each other, generating a supra-TAD that contains multiple TAD domains. For example, at the apex of the LDC domain B-I there is an interaction dot (purple arrow, Fig. Supplemental 2) that links the TA-TB boundary to the TI-TJ boundary (green arrowhead).

Activation of a distant reporter by the eve enhancers

In previous studies we found that a minimal 367 bp homie fragment can orchestrate regulatory interactions between the eve enhancers and reporters inserted at an attP site in the first intron of the hebe transcription unit, 142 kb upstream of the eve gene (Fujioka et al., 2016). This minimal fragment lacks the single homie CTCF site, but has sites for several other generic boundary proteins including BEAF, Zipic, Su(Hw), Pita and Ibf2. In blastoderm stage embryos, lacZ reporter expression was observed in a 7-stripe pattern that coincided with the stripes of the endogenous gene. However, at this stage only a subset of eve-expressing nuclei expressed the reporter. Later in development, during mid-embryogenesis, reporter expression was observed in cells in the dorsal mesoderm, the ventral CNS, and in the anal plate region, in a pattern that recapitulated eve gene expression (Fujioka et al., 2016). At these later stages, most of the eve- expressing cells also appear to express the reporter. The eve-dependent activation of the reporter in a stripe pattern has also been visualized in live imaging of pre-cellular blastoderm-stage embryos (Chen et al., 2018). Within the 7 eve stripes, the majority of the nuclei (∼80-85%) did not express the lacZ reporter. In these nuclei, the transgene was found to be at a considerable average physical distance from the eve locus (∼700 nM). In the remaining nuclei, the transgene was located in closer proximity to the eve locus (∼350 nM, on average). Within this subset, the transgene was expressed in most of the nuclei. Moreover, in those cases in which the transgene was active, it remained in close proximity and continued to express lacZ for the duration of the experiment (∼30 min).

In the two models for TAD formation, quite different mechanisms must be invoked to account for activation of the reporter at –142 kb by the eve enhancers. In the boundary-pairing model, the transgene homie boundary at –142 kb loops over the intervening TADs and pairs with the nhomie:homie complex flanking the eve TAD (c.f., Fig. 3B and C). In the loop-extrusion model, a cohesin complex initiating loop extrusion in the eve TAD must break through the nhomie roadblock at the upstream end of the eve TAD. It must then make its way past the boundaries that separate eve from the attP site in the hebe gene, and come to a halt at the homie boundary associated with the lacZ reporter. This would generate a novel TAD, eveMammoth (eveMa), that extends from the eve homie all the way to the homie fragment at –142 kb, and encompasses both the reporter and the eve gene, including its enhancers (c.f., Fig. 3D). Of course, the eveMa TAD could also be generated by a cohesin complex that initiated in, for example, TF. However, in this case, the runaway cohesin complex would have to break through the intervening boundary roadblocks in both directions. In both the boundary-pairing and loop-extrusion models, the configuration of the chromatin fiber would lead to the activation of lacZ expression by the eve enhancers, while the reporter would be protected from the hebe enhancers by the homie boundary.

Figure 3.

To test these two models, we used a transgene that has two divergently transcribed reporters, lacZ and gfp, that are each under the control of an eve promoter (Fig. 3A). Two different versions of the transgene were generated. In the first, LhomieG, the eve-lacZ reporter is located “upstream” of homie (with respect to the 5’→3’orientation of homie in the eve locus), while eve-gfp is located downstream (Fig. 3A: top). In the second, the orientation of homie within the transgene is reversed, so that gfp is upstream of homie, while lacZ is downstream (GhomieL) (Fig. 3A: bottom). These two transgenes were then individually inserted into the attP site at –142 kb, in either the Z5 or Z3 orientation. In the Z5 orientation, the eve-lacZ reporter is on the eve side of the homie boundary, so that it is closer to the eve locus (Fig. 3A). The eve-gfp reporter is on the opposite side of the homie boundary and is farther away from the eve locus. This places the eve-gfp reporter next to a series of hebe enhancers located farther upstream, in the intron of the hebe gene, and thus it should be subject to regulation by these enhancers (Fig. 2C). In the Z3 orientation, the entire transgene is inserted in the opposite orientation so that eve-gfp is on the eve side of homie, while eve-lacZ is next to the hebe enhancers (Fig. 6A). As a control, we also generated a Z5 transgene in which lambda DNA instead of homie was inserted between the eve-gfp and eve-lacZ reporters.

LhomieG Z5 and GhomieL Z5

The results for transgenes LhomieG Z5 and the control LlambdaG Z5 are most straightforward and will be considered first. In both the loop-extrusion and boundary-pairing models, the LlambdaG Z5 control is not expected to display any regulatory interactions with the eve locus. Fig. 4C shows that this is the case: the eve-lacZ and eve-gfp reporters in LlambdaG Z5 embryos are silent at the blastoderm stage, and transcripts are only very rarely detected. Later in development, both reporters are expressed in a repeating pattern along the dorsal midline (Fig. 6E). This expression is driven by the hebe gene enhancers located just beyond the gfp-reporter (Fujioka et al., 2009).

Figure 4.

In the case of the LhomieG Z5 insert, regulatory interactions between the transgene and the eve locus will be observed if a stem-loop is established that links the transgene to eve. In the loop-extrusion model, the breakthrough cohesin complexes would generate a stem-loop, eveMammoth (eveMa), by coming to a halt at the homie boundary in the transgene and the homie boundary in eve, as illustrated in Fig. 3D. This will place the lacZ reporter in close proximity to the eve enhancers. It will also disrupt the eve TAD, as nhomie is no longer linked to the homie boundary by cohesin. In the boundary-pairing model, the lacZ reporter in the transgene is “upstream” of homie just like the eve gene and its enhancers are “upstream” of homie in the eve locus. Since homie pairs with itself head-to-head, pairing between homie in the transgene and homie in the eve locus will generate the same stem-loop as that generated by loop-extrusion. However, because homie pairs with nhomie head-to-tail, a more complex structure like that in Fig. 3B would be expected if the pairing interactions with the transgene do not disrupt the eve TAD. Consistent with a tripartite structure of the sort shown here, previous studies have shown that three Mcp transgenes on three different chromosomes can interact with each other genetically (Muller et al., 1999). Subsequent direct visualization of Mcp-mediated pairing interactions in imaginal discs by Vazquez et al. (Vazquez et al., 2006) showed that four Mcp- containing transgene separated by Mbs (at Bridges cytogenetic intervals ∼65 on the left arm of chromosome 3, and ∼83 and ∼95 on the right arm of chromosome 3) and/or located on different homologs, clustered in the same nuclear foci in 94% of the nuclei.

As predicted from the formation of a stem-loop linking the transgene to the eve locus, lacZ is expressed in blastoderm-stage embryos in 7 stripes that coincide with the 7 stripes of the endogenous eve gene. However, unlike eve, not all nuclei in the seven stripes express lacZ (compare with the eve control in Fig. 4A). Later, during mid-embryogenesis, lacZ is expressed in the mesoderm, the CNS, and the anal plate in a pattern that mimics the endogenous eve gene. At this stage almost all eve positive cells are also positive for lacZ. Unlike the LlambdaG Z5 transgene, the hebe enhancers do not drive eve-lacZ expression, as they are beyond the homie boundary (Fig. 4E). The eve-gfp reporter in the LhomieG Z5 transgene responds differently. As was observed for the LlambdaG Z5 control, the eve enhancers drive little if any gfp expression, either at the blastoderm stage or later in development (Fig. 5C, E). Instead, gfp is expressed in the midline during mid-embryogenesis under the control of the hebe enhancers (Fig. 5E), which should be in the same TAD as eve-gfp. These results would be consistent with both the loop-extrusion (eveMa) and boundary-pairing models (Fig. 3B and D).

Figure 5.

GhomieL Z5

The situation is more complicated for the GhomieL Z5 transgene. This transgene is inserted in the same chromosomal orientation as LhomieG Z5: eve-lacZ is on the eve side of the homie boundary and eve-gfp is on the hebe enhancer side of the boundary. However, the orientation of the boundary within the transgene is switched so that eve-gfp rather than eve-lacZ is “upstream” of homie. In the boundary-pairing model, flipping the orientation of homie in the transgene but keeping the same orientation of the transgene in the chromosome will have two consequences. First, as illustrated in Fig. 3C, the topology of the chromatin loop connecting the homie boundary in the transgene and the eve locus will change from a stem-loop to a circle-loop. The reason for the change in loop topology is that homie pairs with homie head-to-head, and with nhomie head-to-tail (Fujioka et al., 2016). As before, the transgene homie will pair with eve homie head-to-head to generate a simple circle-loop; however, if, as expected, it also pairs with nhomie so that the eve TAD remains intact, a more complicated loop structure would be generated (Fig. 3C). Second, the reporter that is preferentially activated by the eve enhancers will switch from eve-lacZ to eve-gfp. In order to be activated by the eve enhancer, the reporter must be “upstream” of the transgene homie. This physical relationship means that the eve-gfp reporter will be activated by the eve enhancers independent of the orientation of the transgene in the chromosome. On the other hand, the eve-lacZ reporter will not be activated by the eve enhancers. In addition, it will still be insulated from the hebe enhancers by the homie boundary.

It is not possible to form a circle-loop in the loop extrusion model. Instead, the breakthrough cohesin complex will come to a halt as before when it encounters the inverted homie boundary at –142 kb (Fig. 3D). This means that the eve-lacZ reporter will be in the same TAD (eveMa) as the eve enhancers and will be activated by the eve enhancers. In contrast, the eve-gfp reporter will be in the neighboring TAD and will not be activated by the eve enhancers. It will, however, be regulated by the nearby hebe enhancers.

Fig. 5 shows that the four predictions of the boundary-pairing model are correct: a) the eve enhancers drive eve-gfp expression, b) the hebe enhancers drive eve-gfp expression, c) eve-lacZ is not subject to regulation by the eve enhancers, and d) eve-lacZ is insulated from the hebe enhancers. This would also mean that pairing of the transgene homie in the GhomieL Z5 insert with homie (and nhomie) in the eve locus generates a circle-loop, not a stem-loop, as was the case for LhomieG Z5. In contrast, since loop-extrusion cannot generate circle-loops, the key predictions of this model, namely that eve-gfp will be silent while eve-lacZ will be regulated by the eve enhancers, are not satisfied. Instead, eve-gfp is activated by the eve enhancers, while eve-lacZ is not.

While the results described above for the GhomieL Z5 transgene are inconsistent with a “simple” loop-extrusion model, there is one potential caveat: the 5’→3’ orientation of the homie boundary is inverted so that it is “pointing” towards the eve locus rather than away from it like the endogenous homie boundary. In this orientation, it is possible that homie no longer functions as an effective roadblock, and instead the cohesin complex emanating from the eve locus (or from a loading site somewhere in between) slips past this homie fragment and continues until it is blocked by a properly oriented boundary further upstream. In this case the newly formed stem-loop, eveGargantuan (eveGa) (Fig. 3E), would include the eve-gfp reporter, and it would then be activated by the eve enhancers. However, there are two problems with postulating this novel eveGa TAD. First, eve-lacZ would be in the same eveGa TAD as eve-gfp, and it should also be activated by the eve enhancers. Second, since the homie boundary was bypassed by the cohesin complex to form the larger eveGa TAD, it should not be able to block the hebe enhancers from activating eve-lacZ. Neither of these predictions is correct.

GhomieL Z3 and LhomieG Z3

To confirm these conclusions, we examined the expression patterns of the GhomieL and LhomieG transgenes when they are inserted in the opposite orientation (Z3) in the –142 kb attP site. In the boundary-pairing model, the orientation of homie in the transgene determines which reporter will be preferentially activated, while the orientation of the transgene homie in the chromosome determines whether a stem-loop or a circle loop will be formed. In GhomieL Z3 a stem-loop will be formed, and eve-gfp will be activated by the eve enhancers; however, since eve-lacZ is adjacent to the hebe enhancers, they will activate it rather than the eve-gfp reporter (Fig. 6B). In LhomieG Z3 the topology of the loop will switch to a circle-loop, and eve-lacZ will be activated by both the eve and hebe enhancers (Fig. 6C). In the first version of the loop-extrusion model, the eve-gfp reporter in both GhomieL Z3 (Fig. 6D) and LhomieG Z3 (not shown), should be activated by the eve enhancers, since it is included in the eveMa TAD when the transgenes are inserted in the Z3 orientation, while eve-lacZ will be regulated by the hebe enhancers (Fig. 6D). In the second version of the loop-extrusion model, the functioning of the homie roadblock depends upon its orientation relative to endogenous homie. In this model, GhomieL Z3 is expected to form the eveMa TAD (Fig. 6D), while LhomieG Z3 will form the larger eveGa TAD (Fig. 6E).

Figure 6.

As would be predicted by all three models, the eve enhancers activate the eve-gfp reporter in the GhomieL Z3 insert (Fig. 6F and G). Likewise, homie blocks the hebe enhancers from activating the eve-gfp reporter, while they turn on eve-lacZ expression instead. As predicted by the boundary-pairing model, eve-dependent expression of the reporters switches from gfp to lacZ in the LhomieG Z3 insert (Fig. 6F and G). This would not be expected in the first version of the loop-extrusion model, but it is predicted in the second version. However, as indicated in the diagram in Fig. 6E, the two reporters, along with the hebe enhancers, are included in the larger eveGa TAD, and for this reason both reporters should be activated equally by both the eve and hebe enhancers. This is not the case. The hebe enhancers are blocked by the intervening homie boundary, while there is little or no eve-dependent gfp expression.

TAD formation by transgenes inserted at–142 kb

We used MicroC to examine the TAD organization of the genomic region extending from upstream of the attP site through the eve locus in 12-16 hr embryos carrying different transgene insertions. At this stage, eve is expressed in only a small number of cells in the CNS, mesoderm and anal plate, and this is also true for hebe. This means that most of the interactions detected by MicroC are in cells in which the eve gene and the two reporters are inactive. In addition, while the transgene co-localizes with eve in less than about a fifth of blastoderm stage nuclei, the frequency of physical contact is expected to be much higher in 12-16 hr embryos (c.f., Vazquez et al. Vazquez et al., 2006)). This is because at this stage of development, cell cycles are much longer, and homologs are fully paired. The key findings are summarized below.

LlambdaG Z5

While there is no evidence of contact between sequences in eve and sequences around the attP site at –142 kb in wild-type embryos, this not true for embryos carrying the dual reporter transgene. Perhaps the most unexpected finding is a weak, but clearly discernable pattern of interaction linking the LlambdaG Z5 transgene to the eve locus (Fig. 7A). However, since the LlambdaG Z5 transgene does not respond to the eve enhancers, this physical interaction is not sufficient to drive detectable expression of the reporters. Though our experiments do not allow us to unequivocally identify which sequences in the transgene are responsible for this long-distance interaction, the most likely candidates are the two eve promoters in the LlambdaG Z5 transgene. As noted above, interaction triangles are observed within the eve TAD spanning the regions between nhomie and homie (and/or the eve 3’ PRE) and the eve promoter. It seems possible that the factors supporting these intra-TAD interactions are also responsible for connecting LlambdaG Z5 to the eve locus. In this respect it is interesting to note that homie and nhomie stimulate enhancer-promoter interactions in a transgene assay (Yokoshi et al., 2020).

Figure 7.

LhomieG Z5

Unlike LlambdaG Z5, the eve-lacZ reporter in the LhomieG Z5 insert is activated by the eve enhancers, and this regulatory interaction is perfectly paralleled by the strong pattern of crosslinking between the transgene and the eve locus (Fig. 7B). On the eve side, the interactions between the transgene and the eve locus generate a heavily populated band of crosslinked sequences that span the 16 kb eve TAD. On the transgene side, there is an unequal distribution of crosslinked sequences to either side of the transgene homie boundary. As shown in the inset, the heaviest density of crosslinked sequences is on the eve-lacZ side of the transgene, consistent with the activation of this reporter by eve enhancers. To confirm that the interaction bias mimics the differences in activity of the two reporters in cells in which the reporters are expressed, we analyzed the interaction profiles from viewpoints within either the eve-lacZ or eve-gfp genes. Fig. 7D shows that eve-lacZ interacts with sequences extending across the eve TAD, while eve-gfp interactions are largely restricted to the eve homie element. While these results would be seemingly be consistent with both the boundary-pairing and loop-extrusion models, there is no evidence of cohesin breaking through multiple intervening boundaries into order to establish the novel eve(Mam) TAD. Instead, the organization of the TADs and LDC domains in the interval in between the transgene and eve (Fig. 7B) resemble that seen in wild-type NC14 embryos (see Fig. 2) (with the caveat that there are significantly fewer reads in the transgene experiment). Equally important, the same is true for the eve TAD. In order to generate the eve(Mam) TAD, the cohesin complex must break through the nhomie boundary and, in so doing, disrupt the eve TAD. However, there are no obvious alterations in either the eve volcano or plume with respect to their structure or relative contact density compared to the LlambdaG Z5 and wild type NC14 embryos. Likewise, there is no evidence of novel vertical and 45° crosslinking stripes that would reflect the boundary breakthrough events that are expected in the loop-extrusion model to accompany the formation of a physical connection between the transgene and the eve locus.

We next used the transgene homie as the viewpoint to test the predictions of the boundary-pairing and loop-extrusion models. Since homie can pair with itself (head-to-head) and with nhomie (head-to-tail) in transvection assays (Fujioka et al., 2016), a prediction of the boundary-pairing model is that there will be physical interactions between the transgene homie and both nhomie and homie in the eve locus (c.f., Vazquez et al., Vazquez et al., 2006)). Also based on what is known about how fly boundaries pair with each other (Chetverina et al., 2014; Chetverina et al., 2017; Kyrchanova et al., 2008a), self-interactions are expected to be stronger than heterologous interactions. In contrast, in the loop-extrusion model, the cohesin complex forms the eveMa TAD by linking homie in the transgene to homie in the eve locus, and this requires the cohesin complex to break through nhomie and disrupt the connections between nhomie and homie in the eve locus. In this case, transgene:homie← →eve:homie interactions should be observed, while transgene:homie← →-eve:nhomie should not. Fig. 8A shows that, as predicted by the boundary-pairing model, the transgene homie interacts with both homie and nhomie in the eve locus, and the cross-linking frequency is greater for homie:homie interactions.

Figure 8.

Since the eve TAD does not appear to be disturbed by the presence of the LhomieG Z5 transgene, these findings would suggest that the transgene homie likely pairs with both nhomie and homie in a pattern somewhat like that illustrated in Fig. 3B. While the interactions with the transgene are inconsistent with the loop-extrusion model, one could make the ad hoc assumption that runaway cohesin complexes form not one but two different novel TADs. One would be eveMa. In the other, eveElephant (eveEl), a cohesin complex that initiates in, for example, the TE TAD would come to a halt on one side at the homie boundary in the transgene, and on the other side at the nhomie boundary in the eve locus (Fig. Supplemental 4). However, this configuration would not explain how the transgene reporter is activated by eve enhancers. Since the eve enhancers in eveEl are located in a different TAD than the TADs containing eve-lacZ and eve-gfp, they would not be able to activate either reporter.

GhomieL Z5

When homie is reversed in the GhomieL Z5 transgene, the eve-gfp reporter responds to the eve enhancers instead of eve-lacZ. In the boundary-pairing model, eve-gfp is activated because the self-and heterologous-pairing interactions between nhomie and homie are orientation-dependent. As a result, the topology of the loop formed between the transgene and eve switches from a stem-loop to a circle-loop, and this brings the eve-gfp reporter in close proximity to the eve enhancers instead of the eve-lacZ reporter. Since circle-loops cannot be formed in the loop-extrusion model, activation of the distal eve-gfp reporter would require the formation of a larger eveGa TAD in order to encompass the reporters in the transgene and the eve enhancers in the same looped domain (Fig. 3E). However, as shown in Fig. 7C, we do not observe the cross-linking pattern expected for a new eveGa TAD that includes the entire transgene at –142 kb and then extends to an even more distal boundary. Nor are there any novel vertical/45° crosslinking stripes or obvious perturbations in the basic TAD organization in the region beyond –142 kb, or between –142 kb and the eve locus, as might be expected for cohesin complexes breaking through multiple intervening roadblocks. Instead, as was the case for LhomieG Z5, the GhomieL Z5 transgene is physically linked to the eve locus, and there is a prominent band of crosslinked transgene sequences that spans the eve TAD (Fig. 7C). Consistent with the expression patterns of the two reporters, the heaviest density of crosslinked sequences is on the eve-gfp side of the transgene (see inset). This is confirmed by the interaction profiles generated using either the eve-gfp or eve-lacZ gene body as the viewpoint: the most frequent contacts are between sequences within eve and the eve-gfp reporter (Fig. 7D). In addition, as was the case for LhomieG Z5, when the transgene homie is used as the viewpoint, interactions are observed between both homie and nhomie in the eve locus, with the most frequent corresponding to self-pairing interactions (Fig. 8B). Other than postulating the ad hoc eveEl TAD (Fig. Supplemental 4), this would not be expected for either version of the loop-extrusion model.

LhomieG Z3 and GhomieL Z3

To confirm the findings for transgenes inserted in the Z5 orientation, we used MicroC to analyze the TAD organization of LhomieG Z3 and GhomieL Z3. (Sup. Figs. 5 and 6). Like their Z5 counterparts, both show a strong band of interaction between the transgene and sequences spanning the eve locus. In the case of GhomieL Z3, which is predicted to form a stem-loop, bringing eve-gfp in contact with the eve enhancers, the eve-gfp sequences interact more frequently with the eve locus (see inset in Fig. Supplemental 5A). homie is in the opposite orientation in the transgene and in the chromosome in LhomieG Z3, and in the boundary pairing model is expected to generate a circle-loop. While the sample preparation for MicroC was of poorer quality than the others, the inset in Sup. Fig.6A shows that, as expected, eve-lacZ interacts more frequently with sequences in the eve locus than eve-gfp. On the other hand, there is no evidence of a larger evaGa TAD even though eve-lacZ is activated by the eve enhancers. These interactions are confirmed when either eve-lacZ or eve-GFP are used as viewpoints to plot the physical contacts between the eve TAD and the two reporters in LhomieG Z3 and GhomieL Z3 (Sup. Figs. 5B and 6B). In addition, as was the case for the two Z5 transgenes, when the transgene homie is used as the viewpoint, interactions are observed between homie in the transgene and both homie and nhomie in the eve locus (Fig. 8A and B). Again, the most frequent interactions are self-pairing interactions between homie in the transgene and the eve homie.

Loop topology and interactions with neighbors

The results in the previous sections demonstrate that the topology of the loop formed between the transgene and the eve locus depends upon the orientation of homie in the transgene and the orientation of the transgene in the chromosome. When homie in the transgene is pointing away from the eve locus, the topology is a stem-loop. When homie in the transgene is pointing towards the eve locus, the topology is a circle-loop.

Changing the topology also impacts how sequences flanking the transgene and flanking eve interact with each other. With the important caveat that the depth of our MicroC sequencing is limited, the interaction profiles reflect the topology of the transgene-induced loop. For homie pointing away from eve (LhomieG Z5 and GhomieL Z3), sequences to the TER94 side of eve and to the hebe enhancer side of the transgene are more frequently crosslinked with each other (Fig. Supplemental 7). These interactions would be expected for a stem-loop structure (Fig. 3B). For homie pointing towards eve, sequences on the TER94 side of eve interact with sequences on the eve side of the transgene. This would be expected for a circle-loop structure (Fig. 3C).

Discussion

Two different mechanisms have been proposed to explain how TADs are formed and their endpoints determined. In the first, a cohesin complex first induces a small loop at a loading site within the TAD-to-be and then progressively extrudes one or both strands until the complex encounters boundary roadblocks. If the loading site is located asymmetrically within the TAD-to-be, extrusion will continue on the unblocked strand until a roadblock is encountered. The endpoints of the loops were initially thought to be determined by the location of the first two roadblocks encountered by the cohesin complex (Alipour and Marko, 2012). In a refinement of this model, the orientation of a given roadblock relative to the direction of movement of the cohesin complex determines whether it will come to a halt (Davidson and Peters, 2021; Fudenberg et al., 2016; Guo et al., 2012; Guo et al., 2015; Sanborn et al., 2015). It has also been suggested that cohesin complexes can break through a roadblock and come to a halt at more distant roadblocks, so that in a population of cells analyzed by Hi-C or MicroC, the basic units of organization, the TADs, are overlaid by a hierarchy of LDC domains (Hsieh et al., 2020; Krietenstein et al., 2020). However, the only loop topologies that are possible in the loop-extrusion model are a stem-loop and an unanchored loop.

In the other mechanism, TAD formation is governed by boundary:boundary pairing interactions. In this case, the critical factors are the pairing and orientation preferences of potential partners, together with proximity. Since boundary pairing interactions often exhibit an orientation preference, the topology of the loop formed by paired boundaries can be either a stem-loop or a circle-loop, depending on their relative orientation in the chromosome. In this respect it differs from the loop-extrusion model in that switching the orientation of the boundary does not preclude the formation of a TAD, whereas it can in some versions of the loop-extrusion model.

In the studies reported here, we have used MicroC to analyze the TAD organization of a ∼150 kb chromosomal segment flanking the eve locus, and have experimentally tested the predictions of these two models for the mechanisms of TAD formation. Our studies do not support a loop-extrusion mechanism, and are inconsistent with this model on multiple levels.

Manipulating “TAD” formation

In our experimental paradigm, a transgene containing two divergently transcribed reporters and the eve boundary homie is inserted 142 kb upstream of the eve TAD. Depending upon the orientation of the homie boundary within the transgene, either eve-gfp or eve-lacZ is activated by developmentally regulated enhancers in the eve TAD. Since there are a series of well-defined TADs and their associated boundary elements separating the transgene from eve, a version of the loop-extrusion model in which the cohesin complex invariably arrests upon encountering nhomie and homie at the ends of the eve TAD cannot account for this long-distance regulation. Instead, the extruding cohesin complex must be able to break through the boundaries located between homie in the transgene and homie in the eve TAD. This would generate a novel TAD, eveMa, with endpoints corresponding to the two homie boundaries (Fig. 3D). In this case, the reporter on the eve side of the transgene homie would be in the same domain as the eve enhancers, and thus potentially subject to regulation. While eveMa could account for reporter expression in inserts in which the transgene homie is pointing away from eve (and also away from the reporter on the eve side of the transgene insert), it does not explain reporter expression when the transgene homie is pointing towards eve (and in this case towards the reporter on the eve side of the transgene insert). In the latter case, the reporter activated by the eve enhancers lies outside of the eveMa (cf., Fig. 5D), and thus is not included in the same TAD as the eve enhancers. In order to account for activation of the distal reporter, one must imagine that the cohesin complex breaks through the homie boundary in the transgene and continues on until extrusion is halted by an even more distal boundary. In this revision of the loop-extrusion model, both reporters would be included in the same TAD, eveGa (c.f., Fig. 5E), as the eve enhancers, and both reporters should be activated. However, this is not observed.

TAD organization: loop-extrusion versus boundary pairing

Analysis of the TAD organization in the region spanning the transgene insertion site and eve also does not fit with the expectations of the loop-extrusion model. To account for reporter activation, the cohesin complex must be able to break through the nhomie boundary. This breakthrough could be engineered by a cohesin complex that initiated loop extrusion either from within the eve TAD or from one of the TADs located between the nhomie boundary and the transgene at –142 kb. However, in the absence of the transgene, there is no indication that nhomie is particularly prone to break-through events. In fact, the eve TAD differs from most of the TADs in between eve and the transgene in that it is incorporated into larger LDC interaction triangles much less frequently. There is also no reason to imagine that the presence of a homie (or nhomie: see Fujioka et al., 2016)) transgene at –142 kb would somehow induce nhomie break-through events.

As expected from the regulatory interactions, homie-containing transgenes at –142 kb generate a strong band of crosslinking events that physically link the transgene to the eve locus. In the loop-extrusion model, this strong band of interaction should generate a novel interaction triangle (LDC or HDIC) with endpoints corresponding to the homie elements in the transgene and eve. Contrary to this expectation, a triangle of increased interaction frequencies spanning the entire region between –142 kb and homie is not observed for any of the transgene inserts. In viewpoints from the homie element in the LhomieG Z5 and GhomieL Z3 transgenes, there are two peaks in the eve TAD. The more prominent peak maps to homie, while the less prominent peak maps to nhomie. To account for these interaction peaks, the loop-extrusion model would have to assume that the presence of the LhomieG Z5 and GhomieL Z3 transgenes induces the formation of two novel TADs, the eveMa (homie:homie) TAD in Fig. 5 and another TAD, eveElephant, that links the transgene homie to the nhomie element in the eve locus (Fig. Supplemental 4). However, while the eveMa TAD could potentially explain the activation of reporters on the proximal side of the transgene homie by the eve enhancers, these TADs do not explain how the eve enhancers are able to activate reporters on the distal (hebe enhancer) side of the GhomieL Z5 and LhomieG Z3 transgene. This would require the formation of yet another TAD, eveGa, that encompasses eve, the transgene, and one or more TADs beyond the hebe gene (Fig. 3E). However, there is no indication of physical interactions between the homie boundary in eve and a boundary located beyond the hebe gene. Moreover, like eveMa, eveGa would also have to be accompanied by the formation of eveEl, since the homie element in GhomieL Z5 and LhomieG Z3 interacts with both nhomie and homie.

While these results are inconsistent with the expectations of the loop-extrusion model, they dovetail nicely with many of the predictions of the boundary-pairing model. First, in all four transgenes, homie interacts with both homie and nhomie in the eve locus, but does not interact with the multiple boundary elements located in between. This is consistent with other studies which show that boundaries often have strong partner preferences (Gohl et al., 2011; Kyrchanova et al., 2011; Kyrchanova et al., 2008b). Second, interaction frequencies are greater for self-pairing than for heterologous pairing. This fits with what is known about how partner preferences are determined (Blanton et al., 2003; Chetverina et al., 2017; Erokhin et al., 2021; Kyrchanova et al., 2008a; Zolotarev et al., 2016). Third, the pattern of reporter activation in the four transgenes is precisely as expected from previous studies on the orientation dependence of homie pairing with itself (head-to-head) and with nhomie (head-to-tail) (Chen et al., 2018; Fujioka et al., 2016). This orientation dependence means that the reporter located “upstream” of homie is the one that is activated, independent of the orientation of the transgene in the chromosome. Fourth, the pattern of physical interactions between the eve locus and the transgenes recapitulates the pattern of reporter activation. That is, when eve-lacZ sequences are preferentially crosslinked to sequences in the eve locus, then the eve-lacZ reporter is activated by the eve enhancers. Likewise, eve-gfp is preferentially crosslinked to sequences in eve when it is activated by the eve enhancers. Again this is dependent on the orientation of homie in the transgene, but not on the orientation of the transgene in the chromosome. Fifth, unlike the loop-extrusion model where only stem-loops or disconnected loops can be generated, both stem-loops and circle-loops can be, and, as we have shown here, actually are generated by boundary:boundary pairing. The topology of the loop formed between the transgene and eve depends on the orientation of the transgene homie in the chromosome. When homie in the transgene insert is pointing away from the eve locus, a stem-loop is generated. When homie in the transgene insert is pointing towards the eve locus, a circle-loop is generated. That both loop topologies can be generated by boundary:boundary pairing is confirmed by not only by the pattern of expression of the two reporters in the four transgene inserts but also by their pattern of physical interactions evident in our MicroC experiments.

Are the pairing interactions of homie and nhomie unusual?

There are two reasons to think that the properties of most fly boundaries are similar to those of homie and nhomie. First, all of the fly boundaries that have been tested in assays that are expected to require pairing do have this activity (Blanton et al., 2003; Cai and Shen, 2001; Erokhin et al., 2011; Gohl et al., 2011; Kyrchanova et al., 2008a; Kyrchanova et al., 2011; Kyrchanova et al., 2007; Kyrchanova et al., 2008b; Muravyova et al., 2001). Second, multimerized binding sites for generic polydactyl zinc finger proteins like Su(Hw), Pita, dCTCF, Zw5, and Zipic that are found associated with many fly boundaries are not only able to block enhancer/silencer action but can also mediate pairing interactions in insulator bypass and transvection assays (Erokhin et al., 2021; Kyrchanova et al., 2008a; Zolotarev et al., 2016). Moreover, as is the case for homie and nhomie, pairing interactions require the appropriate partners: multimerized sites for Zw5 do not pair with multimerized sites for dCTCF (Erokhin et al., 2021; Kyrchanova et al., 2008a; Zolotarev et al., 2016). This makes good sense, since these proteins form dimers/multimers, as do other fly boundary factors like BEAF, Mod(mdg4), the BEN-domain protein Insensitive and the Elba complex (Avva and Hart, 2016; Bonchuk et al., 2021; Bonchuk et al., 2011; Bonchuk et al., 2020; Fedotova et al., 2018; Fedotova et al., 2019; Fedotova et al., 2017; Hart et al., 1997). The ability to multimerize means that boundaries that share binding sites for the same protein, i.e., for Pita, Zipic or CTCF, can be physically linked to each other by a single protein complex.

Boundary pairing orientation and loop topology

Our analysis of the TAD organization in the eve neighborhood suggests that the predominant loop topology may be a circle-loop, not a stem-loop. Only the eve TAD (and perhaps the very small TAD, TA) is clearly a stem-loop. It is topped by a distinctive volcano plume and is not overlaid by a hierarchy of LDC domains, indicating that it must be “insulated” from crosslinkable interactions with other nearby TADs. By contrast, none of the TADs in the region spanning eve and the hebe gene have volcano plumes at their apex. Instead, all of these TADs are overlaid by a hierarchal organization of LDC domains (clouds) that diminish in crosslinking intensity as the number of in-between TADs increases. As discussed above, the contact pattern of the TADs to the left of eve are likely to arise from partner switching and/or head-to-head boundary pairing interactions that generate circle-loops. Pairing orientation versus loop topology is considered further in the accompanying paper (Ke et al.).

Orientation and long-distance regulatory interactions

Our experiments highlight an important constraint on long-distance regulatory interactions that are mediated not only by boundary elements like homie and nhomie but also by the GAF-dependent tethering and bypass elements (Batut et al., 2022; Kyrchanova et al., 2023; Kyrchanova et al., 2018; Kyrchanova et al., 2019a; Kyrchanova et al., 2019b; Levo et al., 2022; Li et al., 2023; Mohana et al., 2023), namely the orientation of the physical interactions that bring distant enhancers (or silencers) into close proximity with their gene targets. As we have shown here, the orientation of the physical interactions that bring distant sequences close together determines whether the enhancer will be able to activate its target gene. In principle the effects of orientation could be mitigated by increasing the distance between the boundary/tethering elements and either the enhancers (silencers) or the promoter. However, distance can be a critical if not a rate limiting step in transcriptional activation. For example, live imaging experiments have shown that the frequency and amplitude of transcriptional bursts depend on the distance between the enhancer and target promoter. Nearly continuous bursting and high amplitude transcription was observed when the enhancer was located close the promoter (Fukaya et al., 2016). However, at distances as little as 6 kb, transcription became discontinuous with discrete bursts, while at 9 kb, the bursts were much less frequent, and the amplitude/output was substantially reduced (Yokoshi et al., 2020). If similar distance-dependent reductions in transcriptional efficiency are common, then it would not be possible to overcome the topological constraints on regulatory interactions by simply increasing the distance between the boundary/tethering elements and the enhancers/promoters. While our studies do not directly address how chromosome architecture impacts regulatory interactions in vertebrates, it seems clear that long-distance regulatory interactions in vertebrates (c.f., Tan et al., 2023) will be subject to the same types of topological constraints as those that apply in flies.

Acknowledgements

We would like to thank the Gordon Grey for running the fly food facility at Princeton, members of the Lewis Sigler Genomics Core facility for their invaluable assistance with DNA sequencing and Sergey Ryabichko for his help with smFISH probes. We would also like to thank members of MOL431 for creative input. Special thanks to Olga Kyrchanova, Daria Chetverina, Maksim Erokhin, Pavel Georigev, Tsutomu Aoki, Girish Deshpande, Airat Ibragimov, Sergey Ryabichko, Yuri Pritykin and Alex Ostrin for stimulating discussions and sharing unpublished data.

Materials and methods

Key resources table

Plasmid construction and transgenic lines

The dual reporters construct was described previously (Fujioka et al., 2016). In short, both reporters contain the eve basal promoter (-275 to +106 bp relative to eve start site), the lacZ (eve-lacZ) or EGFP (eve-GFP) coding region, and the eve 3’ UTR (+1300 to +1525 bp). These two reporters are divergently transcribed. The 367 bp homie fragment, or a 500 bp fragment from lambda phage DNA, was placed between the two reporters. The homie fragment was placed in both orientations: in LhomieG, the eve-lacZ reporter is located upstream of the homie fragment, while in GhomieL, eve-GFP is located upstream of the homie fragment (see Fig 2A).

The -142 kb attP landing site was described previously (Fujioka et al., 2009). The -142 kb site contains two attP target sites for phiC31 recombinase-mediated cassette exchange (RMCE) (Bateman et al., 2006) and mini-white as a marker. RMCE can result in either orientation of an inserting transgene (Z5 and Z3; see text). RMCE events were identified by loss of mini-white, and the orientation of each insert was determined by PCR.

smFISH probe preparation

The sequences of target genes were obtained from Flybase (flybase.org)(Gramates et al., 2022). To design probes, the target gene sequences were submitted to the Biosearch Technologies Stellaris RNA FISH probe designer tool (biosearchtech.com), with parameters set to 20nt probes, 48 probes per gene, and mask level 5 for Drosophila melanogaster. The sequence of the probes used are listed in Supplemental Data S1. The probes with 3’ MdC(TEG-Amino) modifications were then ordered from Biosearch Technologies in 96-well plate format, cartridge purification and salt free conditions and in 5 nmol scale. To prepare the fluorophore for probe labeling, the Sigma Atto NHS ester fluorophore (Atto 565 NHS or Atto 633 NHS) were dissolved in dimethylformamide at 5mg/mL as the stock solution. For probe labeling, 1 nmol of each probe was combined and mixed with Atto NHS ester fluorophore at 1mg/mL in 500μL of 0.1M sodium bicarbonate (pH 8.0-8.5) for 2 hours, and the probes were then precipitated with 0.5M potassium acetate (pH 5.2) in 75% EtOH. After that, the probes were dried and resuspend in 50μL of DePC-treated H2O. From there, the probes were injected into HPLC in which the percentage 0.1M triethylammonium acetate varied from 7% to 30% with a flow rate of 1mL/min. The detector to monitor coupled probes during the HPLC purification were set to 260nm for DNA and the absorption wavelength of the coupled fluorophore. The coupled probes were collected when both 260nm and the absorption wavelength channel were detected. The coupled probes were dried and diluted into appropriate concentration in DePC-treated H2O for smFISH experiments.

smFISH

smFISH methods were adopted and modified from previous studies (Little and Gregor, 2018; Trcek et al., 2017). 100-200 flies were placed in a cage with an apple juice plate at the bottom of the cage. Early stages of embryos were for 7hours, while for later stages embryos, collection was overnight. Embryos from each plates were washed into collection mesh and dechorionated in bleach for 2min, then fixed in 5mL of 4% paraformaldehyde in 1X PBS and 5mL of heptane for 15min with horizontal shaking. The paraformaldehyde were then removed and replaced with 5mL methanol. The embryos were then devitellinized by vortexing for 30s, and washed in 1mL of methanol twice. Methanol were then removed and replaced by PTw(1X PBS with 0.1% Tween-20) through serial dilution as 7:3, 1:1, and 3:7 methanol:PTw. The embryos were washed twice in 1mL of PTw and then twice in 1mL smFISH wash buffer (4X SSC, 35% formamide and 0.1% Tween-20). Embryos were incubated with ∼5nM coupled smFISH probes in hybridization buffer (0.1g/mL dextran sulfate, 0.1mg/mL salmon sperm ssDNA, 2mM ribonucleoside vanadyl complex, 20μg/mL RNAse-free BSA, 4X SSC, 1% Tween-20 and 35% formamide) for 12-16h. Embryos were then washed twice for 2 hours in 1mL smFISH wash buffer, followed by 4X washing in 1mL PTw. For DAPI/Hoechst staining, the embryos were stained with 1ug/mL DAPI or Hoechst in PTw, and washed 3X with 1mL PTw. Finally, the embryos were mounted on microscope slides with Aqua PolyMount and #1.5 coverslip for imaging.

Imaging, image analysis and statistics

Embryos from smFISH were imaged by using a Nikon A1 confocal microscope system, Plan Apo 20X/0.75 DIC objective. Z-stack images were taken with interval of 2μm, 4X average, 1024X1024 resolution, and the appropriate laser power and gain were set for 405, 561, and 640 channels to avoid overexposure. Images were processed by ImageJ and the maximum projection was applied to each of the stack images. To measure the stripe intensity of early embryos, multi-channel images were first split into single channels and the stripe signal was highlighted and selected by the MaxEntropy thresholding method. For APR intensity measurement, the ROI tool was used to crop out the APR region of late-stage embryos. The cells with APR signal were also highlighted and selected by MaxEntropy thresholding. The particle measurement tool was used to measure the average intensity of all cells that have the signal. At the same time, the background signal (average intensity) was measured on cells without a signal in the same embryo. The relative intensity (signal to background) for each embryo was calculated by using the stripe signal and the background from the same embryo. To make comparisons between independent biological replicates, the average background signal of all embryos from each replicate was calculated. The relative intensity of each embryo from each replicate were normalized based on the average background signal of all embryos from same replicates. GraphPad Prism was used for data visualization and statistical analysis. To compare intensity from embryos in different groups, different signals from the same embryo (e.g., lacZ and GFP) were paired and paired two-tailed t-tests were used to calculate p-values. All raw measurements and normalized data are included in Supplemental Data S2.

MicroC library construction

Embryos were collected on yeasted apple juice plates in population cages for 4 hours, incubated for 12 hours at 25℃, then subjected to fixation as follows. Embryos were dechorionated for 2 mins in 3% sodium hypochlorite, rinsed with deionized water, and transferred to glass vials containing 5 mL PBST (0.1% Triton-X100 in PBS), 7.5 mL n-heptane, and 1.5mL fresh 16% formaldehyde. Crosslinking was carried out at room temperature for exactly 15 mins on an orbital shaker at 250rpm, followed by addition of 3.7 mL 2M Tris-HCl pH7.5 and shaking for 5 mins to quench the reaction. Embryos were washed twice with 15 mL PBST and subjected to secondary crosslinking. Secondary crosslinking was done in 10mL of freshly prepared 3mM final DSG and ESG in PBST for 45 mins at room temperature with passive mixing. The reaction was quenched by addition of 3.7mL 2M Tris-HCl pH7.5 for 5 mins, washed twice with PBST, snap-frozen, and stored at -80℃ until library construction. Micro-C libraries were prepared as previously described (Batut et al., 2022) with the following modifications: 50uL of 12-16hr embryos were used for each biological replicate. 60U of MNase was used for each reaction to digest chromatin to a mononucleosome:dinucleosome ratio of 4. Libraries were barcoded, pooled and subjected to paired-end sequencing on an Illumina Novaseq S1 100 nt Flowcell (read length 50 bases per mate, 6-base index read).

Micro-C data processing

MicroC data for D. melanogaster were aligned to custom genomes edited from the Berkeley Drosophila Genome Project (BDGP) Release 6 reference assembly (dos Santos et al., 2015) with BWA-MEM (Li and Durbin, 2009) using parameters -S -P -5 -M. Briefly, the custom genomes are simply insertions of the transgenic sequence into the –142kb integration site, as predicted from perfect integration. These events were confirmed using PCR post-integration. The resultant BAM files were parsed, sorted, de-duplicated, filtered, and split with Pairtools (https://github.com/mirnylab/pairtools). We removed pairs where only half of the pair could be mapped, or where the MAPQ score was less than three. The resultant files were indexed with Pairix (https://github.com/4dn-dcic/pairix). The files from replicates were merged with Pairtools before generating 100bp contact matrices using Cooler (Abdennur and Mirny, 2020). Finally, balancing and Mcool file generation was performed with Cooler’s Zoomify tool.

Virtual 4C profiles were extracted from individual replicates using FAN-C (Kruse et al., 2020) at 400bp resolution. The values were summed across replicates and smoothed across three bins (1.2kb). Viewpoints were determined based on the most informative region for interpretation. Ultimately we decided to use 800bp regions in the gene body of either GFP or lacZ, moving downstream from the eve promoter. However, similar results were obtained using viewpoints between transgenic homie and the promoters of either gene. For homie viewpoints, due to transgenic homie being masked, some loss of signal is expected at the viewpoint (Figure 6).

Supplemental Figures

Figure Sup. 1. TAD organization of the even-skipped locus and neighboring sequences MicroC contact pattern for the chromosomal region spanning the attP site at –142 kb and the eve locus. Like the eve volcano, this contact pattern was generated using aggregated previously published NC14 data (Batut et al., 2022; Levo et al., 2022). Arrow indicates the –142kb locus where transgenes are integrated into the genome. LDC domains are labeled according to the interacting TADs: for example, H-I indicates crosslinking events generated by physical interactions between sequences in two neighboring TADs, TH and TI, while I-J indicates interactions between sequences in TI and TJ. Likewise, G-I indicates crosslinking events generated by physical interactions between sequences TADs TG and TI, which are separated from each other by TAD TH. Note that unlike the TADs to the left of eve, the eve TAD does not interact frequently with its neighbors. Purple arrow indicates potential interaction dot between the boundary marking the border of TA and TB and the boundary marking the border of TI and TJ. Red, blue and green arrowheads mark boundaries that could engage in partner switching: see text.

Figure Sup. 2. Trouble ahead, trouble behind. A) Illustration of a broken zipper—zipping in front but not closing behind. B) Crosslinking pattern expected for cohesin loading at a site very close to the middle of the TAD-to-be: The vertical line of linked DNA is generated when the two chromatin fibers in the TAD-to-be are transiently linked together as cohesin passes. When cohesin bumps into the CTCF roadblocks (BE) at the boundaries of the TAD, the vertical line stops. C) Cohesin loading site is offset from center of TAD-to-be to the left. The crosslinking initiates at the loading site and extends vertically until the cohesin complex bumps into the CTCF roadblock on the left. At that point extrusion of the left chromatin fiber halts, while it continues to extrude the chromatin fiber on the right. The line of crosslinked sequences then extends at a 45° angle until cohesin comes to a halt at the CTCF roadblock on the right. D and E) The loading site is at or close to the boundary on the left (D) or the right (E). In this case extrusion will be unidirectional and the DNAs crosslinked as cohesin passes will form a line that extends at a 45° angle until cohesin comes to a halt at the CTCF roadblock. F). Cohesin is able to load at multiple sites in the TAD-to-be. In this case there will be a series of vertical lines whose intensity will be correlated with the relative frequency that a give loading site is used. The vertical lines will extend in each case until cohesin comes to a halt at the closest CTCF roadblock. At that point, the line of crosslinked DNA will extend at a 45° angle until that particular cohesin complex comes to halt at the CTCF roadblock that is farther away. The 45° line of crosslinked DNA will become progressive more intense as it approaches the apex of the triangle.

Figure Sup. 3. Models for the formation of LDC domains. Several different mechanisms could explain the LDC domains. A-D). In the loop extrusion model, the LDC domains arise because cohesin sometimes breaks through a CTCF roadblock. In A) there are three primary TADs, TAD1, TAD2, and TAD3. B and C) Occasional breakthrough events can generate a secondary TAD, LDC1-2 (TAD1+TAD2) or LDC2-3 (TAD2+TAD3). D) Even less frequent breakthrough events can generate a tripartite TAD (TAD1+TAD2+TAD3) that will correspond to LDC1-3. E-G). LDC domains arise because TADs sometimes bump into their neighbors. In E, TAD1 bumps into TAD2 to give LDC1-2, while in F TAD2 bumps into TAD3 to give LDC2:3. G) Less frequently TAD1 bumps into TAD3 to give LDC1:3. H-J) LDC domains arise because boundaries can switch partners. H and I) In H boundary A (dark green) skips boundary B (red) and pairs with boundary C (tan). This generates a new TAD, TAD1+TAD2 which corresponds to LDC domain 1-2. In I, boundary B (red) skips boundary C (tan) and pairs with D (blue). This generates a new TAD, TAD2+TAD3 which corresponds to LDC domain 2-3. J) In J, boundary A (dark green) skips boundary B (red) and boundary C (tan) and pairs with boundary D (blue). This generates a new TAD, TAD1+TAD2+TAD3 which corresponds to LDC domain 1-3.

Figure Sup. 4. Loop extrusion: eveElephant. A) Loop extrusion orientation dependent model for LhomieG Z5 transgene as the sketched in Figure 2E. This configuration generates eveMa. B) Instead running through nhomie, the cohesin complex comes to a halt at the nhomie boundary, generating eveElephant. A second cohesin complex generates the eve TAD by loading at a site within the eve locus. Both eveElephant and the eve TAD are stem-loops. However, while this double stem-loop configuration would explain why the transgene homie makes contact with both nhomie and homie, the eve enhancers are isolated in their own loop and would not make contact with the two reporters in the transgene.

Figure Sup. 5. MicroC of GhomieL Z3. A) MicroC map of GhomieL Z3. Gray lines are masked regions due to repeated sequences. In this case, focusing on the zoom-in map from top to bottom, most likely P-element remnants 3’ of the lacZ gene, transgenic homie, and P-element remnants 3’ of the GFP gene. Compare directly to GhomieL Z5 for changes in interaction preference due to change of homie orientation (Figure 4C). B) “Virtual 4C” map from either lacZ or GFP gene body viewpoints.

Figure Sup. 6. Micro C of LhomieG Z3. A) MicroC map of LhomieG Z3. Gray lines are masked regions due to repeated sequences. In this case, most possibly P-element remnants 3’ of the lacZ gene. Compare directly to GhomieL Z3 for changes in interaction preference due to change of homie orientation (Figure 4B). B) “Virtual 4C” map from either lacZ or GFP gene body viewpoints.

Figure Sup. 7. Blowup of interactions generated by the transgene with sequences neighboring –142 kb and the eve TAD. The interaction pattern indicates that the loop topology changes as a result of the orientation of transgenic homie. Blow-up of off-diagonal interactions, focused on the interaction between the endogenous eve locus and the –142kb transgenic insert with A) LhomieG Z3, B) LhomieG Z5, C) GhomieL Z3, or D) GhomieL Z5. Note the switch in preferential interaction between regions adjacent to transgenic homie and endogenous homie (indicated by arrow brackets). In cassettes where transgenic homie is in the same orientation in the genome as endogenous homie (A and D), regions upstream of transgenic homie preferentially interact with regions downstream of endogenous homie. In contrast, when transgenic homie is in the opposite orientation in the genome as endogenous homie (B and C), regions downstream of transgenic homie preferentially interact with regions downstream of endogenous homie. This suggests a different loop topology when the orientation of the transgene homie element in the chromosome is flipped.