The precise expression of genes in time and space during embryogenesis is largely influenced by communication between enhancers and promoters, which is propagated and governed by the physical proximity of these elements in the nucleus. Here, we review how chromatin domains organize the genome by guiding enhancers to their target genes thereby preventing non-specific interactions with other neighboring regions. We also discuss the dynamics of chromatin interactions between enhancers and promoters, as well as the consequent changes in gene expression, that occur in pluripotent cells and during development. Finally, we evaluate how genomic changes such as deletions, inversions and duplications affect 3D chromatin configuration overall and lead to ectopic enhancer-promoter contacts, and thus gene misexpression, which can contribute to abnormal development and disease.
The expression of developmental genes is tightly orchestrated in time and space throughout embryonic development as well as organogenesis, and transcriptional regulation is a well-recognized essential component of normal embryogenesis. Complex changes in expression are associated with, and are essential for, determining cell fate, lineage commitment, the establishment of the body plan, and organ formation (Zeller et al., 2009; Petit et al., 2017). These dynamic transcriptional processes are mainly mediated by promoters that are located near the transcription start site (TSS) and by regulatory regions termed enhancers that can be located at a considerable distance from the TSS. Once activated by specific transcription factors, enhancers produce spatially and temporally restricted patterns of activity that determine the expression of their target genes (Spitz and Furlong, 2012). They regulate gene expression by physically interacting with their target gene promoters via looping of DNA. This looping and the associated physical interaction between enhancers and promoters is propagated by proteins that facilitate this contact, such as the mediator complex (see Glossary, Box 1) (Carlsten et al., 2013). Importantly, changes in these regulatory units are thought to play a role in the evolutionary development of new or altered patterns of gene expression and the subsequent acquisition of morphological novelties (Chan et al., 2010; Long et al., 2016).
A/B compartments. Chromosomes are organized into active (A) and inactive (B) compartments that interact in a homotypic fashion (Lieberman-Aiden et al., 2009).
ATAC-seq. Assay for transposase accessible chromatin with high-throughput sequencing (ATAC-seq) that uses in vitro transposition of adaptors into open chromatin regions, followed by deep sequencing to map open regions at a genome-wide level (Buenrostro et al., 2013).
ChIP-seq. An approach that combines chromatin immunoprecipitation of the desired epitope with deep sequencing to identify genomic regions bound by a specific transcription factor or marked by a specific histone post-translational modification.
Chromocenters. Clusters of pericentromeric satellite repeats from different chromosomes that are found in specific nuclear landmark structures.
Cohesin complex. A ring-like multiprotein complex involved in the pairing of sister chromatid during cell divisions and in the production of chromatin loop between distal DNA segments.
CTCF. An 11-zinc-finger transcription factor that is enriched at boundary elements of TADs and loop domains.
DNA replication time zones. Regions of the genome that are replicated at distinct times during S phase.
DNAse-seq. A method that captures nucleosome-depleted fragments digested by the DNAse I enzyme and uses deep sequencing to map chromatin open regions at a genome-wide level (Song and Crawford, 2010).
Embryonic stem cells (ESCs). Pluripotent stem cells that are derived from the inner cell mass of a blastocyst and can differentiate into any cell type.
Induced pluripotent stem cells (iPSCs). Pluripotent stem cells that are generated by the reprogramming of differentiated cells via the introduction of several pluripotency factors.
Insulated neighborhoods. Chromatin domains defined by chromatin loops formed by a CTCF homodimer co-bound by cohesin and containing at least a gene (Hnisz et al., 2016).
Lamina-associated domains (LADs). Chromatin domains that interact with the lamina of the nucleus.
Loop domains. Domains of preferential interaction that are marked by a loop at their corner (Rao et al., 2014).
Mediator complex. A large protein complex with a variable subunit composition, which controls various processes necessary for transcription, including 3D chromatin architecture.
Polycomb or Polycomb group (PcG) proteins. Proteins that can remodel chromatin so that epigenetic repression can occur. They play key roles in stem cell identity and cell differentiation (Di Croce and Helin, 2013).
Polycomb repressive complex 1 (PRC1) and 2 (PRC2). PRC1 contributes to chromatin compaction and catalyzes the monoubiquitylation of histone H2A. PRC2 is recruited to chromatin, contributes to chromatin compaction and catalyzes the methylation of H3K27 (Margueron and Reinberg, 2011).
Topologically associating domains (TADs). Chromatin domains that exhibit high levels of internal interactions. They are separated from each other by regions of low interaction termed boundary elements (Nora et al., 2012).
Together, promoters and enhancers constitute the regulatory landscapes of genes. At developmentally active sites, these landscapes can span several hundreds of kilobases and can contain numerous enhancers (Montavon et al., 2011; Andrey et al., 2013; Marinić et al., 2013; Franke et al., 2016; Symmons et al., 2016). Such loci require an exquisitely complex regulation of their 3D structure, to assure precise and robust gene expression. In recent years, chromosome conformation capture technologies (e.g. 4C, HiC and its derivatives) have been used to quantify the frequency of chromatin contacts at specific loci or genome wide (Sati and Cavalli, 2017). These studies have shown that the genome is organized in a non-random fashion, resulting in a higher order 3D organization of entire chromosomes. On a sub-chromosomal level, chromatin domains exhibiting high levels of internal interactions have been identified using a variety of algorithms and definitions. One particular type of chromatin domain, referred to as a topologically associating domain (TAD, see Glossary, Box 1), has emerged as being particularly important. TADs are defined as chromatin domains with high internal interactions and are separated from each other by regions of low interaction called boundary elements. TADs have been shown to be a fundamental component of the 3D organization of the genome in the nuclear space, restraining and facilitating enhancer-promoter interactions mostly in a cell type-independent manner (Dixon et al., 2012; Nora et al., 2012). During embryonic development, in particular upon cell differentiation and lineage commitment, the micro-architecture of TADs becomes more structured, probably because TADs become directly involved in transcriptional control by directing enhancer-promoter contacts (Bonev and Cavalli, 2016).
Characterization of the various components that control gene regulation, i.e. enhancers, promoters and their associated 3D interactions, has become a standard approach to deciphering the mechanisms underlying transcriptional control and has thus provided new frameworks for understanding basic developmental processes, disease pathology and the evolution of organs and structures (Montavon et al., 2011; Andrey et al., 2013; Lonfat et al., 2014; Lupiáñez et al., 2015; Franke et al., 2016). Here, we review the general mechanisms of genome organization and gene regulation during vertebrate development and cell differentiation. We describe the function and organization of enhancers and discuss current concepts of 3D chromatin dynamics during development and lineage commitment, focusing on TADs and their involvement. Finally, we review the different types of mutations that can affect the integrity of chromatin structure and thus result in gene misexpression, developmental abnormalities and disease.
Enhancers and promoters form regulatory landscapes
The regulation of developmental or lineage-specific genes is tightly controlled through the activities of regulatory elements called enhancers. Enhancers were originally described as genomic regions that can drive the transcription of a reporter gene in specific reporter constructs (Banerji et al., 1981). Such a construct would classically consist of the DNA fragment to be tested, a minimal promoter (often from the β-globin gene), and a reporter gene, such as lacZ or GFP (Shlyueva et al., 2014). Once introduced into the genome of an embryo, the expression of this reporter construct is expected to reflect the regulatory activities of the tested region. Large-scale studies using this approach have identified thousands of enhancers that drive tissue- and time-specific gene expression patterns in mice or flies (Visel et al., 2007; Manning et al., 2012). Techniques that assess chromatin openness or post-translational histone modifications can also be used to localize putative enhancers (see Box 2). Indeed, as transcription factors preferentially bind to open chromatin, techniques that detect these regions, such as ATAC-seq or DNAse-seq (see Glossary, Box 1), are very useful for establishing genome-wide enhancer footprints (Song and Crawford, 2010; Buenrostro et al., 2013). Another approach utilizes the acetylation of histone 3 at lysine 27 or 122 (H3K27ac or H3K122ac), which correlates with active enhancer activities and can be probed by chromatin immunoprecipitation followed by deep sequencing (ChIP-seq; see Glossary, Box 1) (Creyghton et al., 2010; Rada-Iglesias et al., 2011; Pradeepa et al., 2016).
Enhancers feature several hallmarks that allow their identification:
Open chromatin. Enhancers are found at sites of open chromatin, which can be identified using approaches such as DNAse-seq or ATAC-seq.
Chromatin modifications. Poised and active enhancers are enriched for the chromatin modification H3K4me1. Active enhancers are further modified with H3K27ac and/or H3K122ac (Heintzman et al., 2009; Creyghton et al., 2010; Rada-Iglesias et al., 2011; Pradeepa et al., 2016). These modifications can be mapped in the appropriate cell type using ChIP-seq.
Transcription factor binding. Enhancers are bound by transcription factors, which tether chromatin remodeling and transcriptional complexes (Spitz and Furlong, 2012). ChIP-seq for transcription factors that are involved in specific biological process can be used to determine the genomic location of enhancers (Araya et al., 2014).
Conservation. Because of the evolutionary constraints imposed on enhancer sequences by the requirement to bind transcription factors, enhancers are predicted to be more conserved genomic regions (Pennacchio et al., 2006).
Other techniques and approaches that can be used to identify and characterize enhancers include:
Reporter assays. In this approach, putative enhancer sequences are placed in front of a minimal promoter and a reporter gene. By integrating such a construct into the genome, it is possible to detect, in a developing embryo, the tissue-, cell- or time-specific activity of the reporter gene, which mirrors the activity of the enhancer.
STARR-seq. In this approach, which takes advantage of the fact that enhancers can work independently of their position and orientation, the candidate enhancer region is placed downstream of the promoter. Accordingly, genome-wide libraries of fragments can be cloned and active enhancer regions that can induce their own transcription can thus be measured by the abundance of transcripts (Arnold et al., 2013).
CAGE. Cap analysis of gene expression is a method that can detect enhancer activities by measuring the presence of bidirectional capped RNA transcripts (Andersson et al., 2014).
Enhancers are bound by tissue-specific transcription factors and can thereby produce highly controlled regulatory outputs, in time, in space or in specific cell types (Spitz and Furlong, 2012). The birth of new enhancer regions, through the acquisition of transcription factor binding and subsequent novel regulator function, is thus thought to be a key component in gene neo-functionalization and a driver for evolution (Long et al., 2016). According to current concepts, complex expression patterns of developmental genes are achieved by the additive effect of multiple enhancers, with each element regulating a subset of the global expression pattern of a gene (Fig. 1). For example, in vivo studies of the α-globin gene locus, which is controlled by a cluster of enhancer elements, also referred to as a super enhancer, have shown that each enhancer seems to act independently and in an additive fashion without clear evidence of synergistic or higher-order effects (Hay et al., 2016). A recent comprehensive study of the Indian hedgehog (Ihh) locus also shows that Ihh is regulated by a multipartite enhancer ensemble consisting of at least nine enhancers with individual combinations of tissue specificities that function in an additive manner (Will et al., 2017). Furthermore, this study showed that enhancers function in a dosage-dependent manner; an increase in the copy number of enhancers results in increased gene expression, although this increase is not linear and differs in different tissues. Likewise, a significant loss of transcription of the PIM1 leukemic oncogene in a human cell line was only obtained through the combinatorial repression of several weak enhancers embedded within a super-enhancer region and not by the repression of single enhancers (Xie et al., 2017). Enhancer crosstalk is also sometimes required to control gene transcription properly. At the mouse Krox20 (also known as Egr2) locus, for example, an enhancer element is required to potentiate another enhancer that acts in an auto-regulatory fashion to sustain Krox20 expression (Thierion et al., 2017). In addition, strong enhancers at the Drosophila knirps and hunchback loci were shown to act in a sub-additive manner by buffering each other's activities thereby allowing a constant transcriptional output (Bothma et al., 2015). Thus, the regulatory landscapes of developmental genes seem to consist of complex enhancer ensembles that act in an interactive manner to achieve the precise and robust control of gene expression that is essential for normal development.
In vertebrates, enhancers often locate at a distance from the promoter they control and bypass nearby genes (de Laat and Duboule, 2013). Indeed, a genome-wide study in human cell lines has estimated that as few as 7% of distal regulatory elements control their closest promoters (Sanyal et al., 2012). The specificity of promoter elements for certain enhancers might thus play a role in directing their regulatory activities. A study of the HoxD cluster during mouse limb development, for example, determined that each of the HoxD gene promoters, despite being very similar, has a different affinity for the same set of digit enhancers (Montavon et al., 2008). Furthermore, pioneer studies in Drosophila have shown that promoters can each respond in a different fashion to particular types of enhancer (Zabidi et al., 2015; Arnold et al., 2017). Accordingly, the distance between enhancer and promoter is not the primary parameter that controls gene activation. A well-characterized case of long-range regulation occurs at the mouse sonic hedgehog (Shh) locus, where the limb-specific enhancer of Shh (called ZRS) is located more than 1 Mb away from the Shh promoter (Lettice et al., 2003; Sagai et al., 2005). Similarly, at the mouse HoxD cluster, a regulatory landscape that is distributed across a 750 kb region controls the digit-specific expression of the Hoxd13 to Hoxd10 genes (Montavon et al., 2011), and at the mouse Sox9 locus, a 1.7 Mb region controls the intricate expression pattern of this gene (Franke et al., 2016).
The insertion of regulatory sensors, which consist of a minimal promoter and a reporter gene, at multiple sites in a given genomic locus has been used to determine the extent of such regulatory landscapes (Kondo and Duboule, 1999; Ruf et al., 2011; Akhtar et al., 2013; Marinić et al., 2013; Symmons et al., 2016). In these experiments, one can observe that the regulatory inputs from enhancers are sensed over hundreds of kilobases with varying intensities. These and other studies have also revealed that regulatory boundaries exist, as demonstrated by a decrease and ultimately the disappearance of the reporter signal. The regions of activity that exist between such boundaries have subsequently been shown to correspond to TADs (discussed in detail below) (Dixon et al., 2012; Nora et al., 2012; Tsujimura et al., 2015; Lupiáñez et al., 2016). In conclusion, regulatory landscapes can extend over hundreds of kilobases, are part of or defined by the boundaries of a TAD, and characteristically consist of numerous regulatory elements that together control the expression of a target gene. These landscapes have also been referred to as regulatory archipelagos or holo enhancers (Montavon et al., 2011; Marinić et al., 2013; Hughes et al., 2014) (Fig. 1).
Chromatin domains as basic genomic regulatory units
In recent years, various technological approaches have shown that the genome is organized in a non-random fashion into higher order chromatin domains. For instance, proximity ligation technologies (see Box 3) have been used to determine the frequency at which two genomic regions can be crosslinked and are thus in close proximity to each other in the nucleus (Sati and Cavalli, 2017). Specifically, HiC and other C-technologies have revealed the existence of self-associating and insulated chromatin domains called TADs, which are separated by regions of low interactions called boundaries (Dixon et al., 2012; Nora et al., 2012). These Mb-size structures associate with multiple genomic features, such as functional chromatin modifications, DNA replication time zones, chromocenters, or lamina-associated domains (LADs; see Glossary, Box 1) (Dixon et al., 2012; Pope et al., 2014; Wijchers et al., 2015). The existence of TADs has also been confirmed by another approach that does not rely on chromatin crosslinking and ligation but on DNA sequencing from a collection of thin nuclear sections (Beagrie et al., 2017). Strikingly, TADs remain largely unchanged during lineage commitment, are stable across different cell lines, and are even conserved between species (Dixon et al., 2012, 2015).
Proximity ligation technologies are used to estimate the frequency of DNA interactions. They are based on the ability to crosslink two genomic loci that are in close physical proximity to each other in the nuclear space (Sati and Cavalli, 2017). The crosslinked chromatin is then subjected to restriction enzyme digestion, which fragments the genomic DNA. The digested chromatin is re-ligated so that regions that are crosslinked together (intra-molecular re-ligation) are attached to each other. This process produces libraries of chimeric DNA products, including adjacent restriction fragments, which are normally found at different positions along the linear genome. Several variations of this technologies exist and include:
Chromosome conformation capture (3C). This approach allows any candidate chimeric product generated from the crosslinking step described above to be quantified using quantitative PCR and specific primers amplifying the product junction (Simonis et al., 2006).
Circular chromosome conformation capture (4C). In this approach, libraries of chimeric DNA products are digested with a second restriction enzyme and further re-ligated together to decrease their molecular size. Using circular PCR, all chimeric products ligated with a desired viewpoint (i.e. a specific restriction fragment) are amplified and sequenced, generating an interaction profile measuring all DNA fragments interacting with the viewpoint (Noordermeer et al., 2011).
Carbon copy chromosome conformation capture (5C). In this approach, primers are designed over a defined genomic interval to amplify many possible ligation products. The amplified library is then sequenced or hybridized to a microarray in order to measure the interaction frequencies of all the tested fragments. Ultimately, an interaction map spanning the defined genomic interval is produced that displays a ‘many versus many’ interaction map (Dostie and Dekker, 2007).
HiC. In the HiC technique, after the initial restriction enzyme step, the ends are filled in with a biotin-marked nucleotide and subsequently re-ligated. A streptavidin pull-down step is used to enrich for the chimeric products, which are then sequenced. Sequencing of the library produces a genome-wide interaction map of ‘all versus all’ restriction fragments (Lieberman-Aiden et al., 2009; Rao et al., 2014).
Capture-C (no biotin fill in) and capture-HiC (with biotin fill in). In capture-C, the library of chimeric DNA products is sheared and hybridized to RNA baits, which consist of 100- to 150-bp-long RNA oligonucleotides attached to beads, to enrich for specific loci; the hybridized products are then sequenced in a paired-end fashion to determine the chimeric products (Hughes et al., 2014). The RNA-enriched regions can either be specific genomic viewpoints in order to produce a ‘one versus all’ 4C-like interaction track or entire loci to produce targeted ‘many versus many’ interaction maps (Franke et al., 2016; Andrey et al., 2017). In capture-HiC, specific loci or viewpoints are enriched from a HiC library, also using RNA baits (Schoenfelder et al., 2015). These methodologies allow one to either parallelize the production of thousands of interaction tracks from specific viewpoints or produce targeted HiC maps with a much lower sequencing effort compared with HiC.
The definition of a TAD and its boundaries depends on the computational algorithms used and the resolution of the HiC experiment and, as such, is somewhat arbitrary. Other types of chromatin domains that represent alternative views of genome compartmentalization have thus been described. SubTADS, for example, are regions of preferential interactions located within a TAD and denote a substructure of these domains (Phillips-Cremins et al., 2013). On a larger scale, TADs can interact with each other, organizing themselves into metaTADs if significant inter-TAD interactions occur in a given cell type or tissue (Fraser et al., 2015). Interestingly, subTADs and metaTADs display a more tissue-specific behavior than TADs (Dixon et al., 2016; Zhan et al., 2017). Loop domains, which correspond to genomic regions of preferential interactions with interaction peaks at their corners (see Glossary, Box 1), have also been described (Rao et al., 2014). Such domains and interactions occur between the boundaries of a TAD, but also within TADs, most of the time involving binding sites for the transcriptional regulator CTCF (see Glossary, Box 1) in convergent orientation. Insulated neighborhoods (see Glossary, Box 1) are specific types of loop domains that are defined by chromatin loops sustained by a CTCF-CTCF homodimer co-bound by cohesin (see Glossary, Box 1), and containing at least one gene (Dowen et al., 2014; Hnisz et al., 2016). In comparison with TADs, subTADs, metaTADs, loop domains and insulated neighborhoods are more labile across cell types and change with gene activity and regulation (Dixon et al., 2016). Accordingly, the algorithms that detect these chromatin domains are based on different concepts and thus do not compartmentalize the genome in the same way as TADs (Dixon et al., 2016). Finally, HiC interaction maps have also revealed the existence of another type of higher-order chromatin structure called A/B compartments (see Glossary, Box 1). These compartments correspond to long-range interacting territories consisting of transcriptionally active (A-compartments) and inactive (B-compartments) domains (Lieberman-Aiden et al., 2009). Recently, the 3D reconstruction of chromatin in the nuclei of single haploid embryonic stem cells (ESCs; see Glossary, Box 1) has shown that A-compartments tend to be found in a ring shape, surrounded by inactive B-compartments, which locate to the periphery and close to the nucleolus (Stevens et al., 2017).
In vertebrates, growing evidence supports the notion that TADs are shaped through the formation of stable DNA loops, called anchor loops, which connect boundary elements together (Rao et al., 2014). Accordingly, TADs correspond to a stable subset of loop domains (Fig. 2). Anchor loops were found to be strongly associated with, and dependent on, the binding of CTCF and the cohesin complex (Sanborn et al., 2015). According to the loop extrusion model, a recently proposed biophysical mechanism for TAD formation, the loading of a cohesin ring around the chromatin progressively extrudes the chromatin until the cohesin ring reaches a ‘roadblock’, which could be CTCF, to form the observed anchor loops (Sanborn et al., 2015; Fudenberg et al., 2016). According to this model, the constant extruding process is responsible for the frequent mutual interaction of all the sequences located within the TAD. The importance of CTCF and cohesin has also recently been underlined by the finding that the depletion of one of these components leads to a progressive loss of TAD structure, although other chromatin features, such as A/B compartments, are retained (Schwarzer et al., 2016 preprint; Nora et al., 2017). Moreover, two studies have shown that transcription can relocate cohesin over long distances and that its chromatin loading/unloading rate controls the extension of chromatin loops (Busslinger et al., 2017; Haarhuis et al., 2017).
TADs are believed to function as a scaffold for enhancer-promoter interactions (Fig. 2), guiding them in space and allowing them to interact with each other (de Laat and Duboule, 2013). Accordingly, it was recently proposed that TADs act as a buffer for genomic distances and allow for frequent physical interactions to occur between all of their constituent genomic elements (Symmons et al., 2016). In this latter study, the authors showed that the relative genetic distance between the Shh gene and its limb enhancer, the ZRS, has no influence on its transcriptional activity as long as both elements locate in the same TAD. In contrast, when the TAD boundary is relocated to lie in between Shh and the ZRS, Shh transcription is abolished when the genomic distance from the ZRS is large and is reduced when it is short. This result indicated that boundary elements can indeed act as chromatin insulators by reducing the frequency of chromatin interactions but are not impermeable to contact with genomic regions located outside of the TAD. This concept is supported by studies of the Sox9 locus, which showed that the duplication of boundary-containing regions results in the formation of a new TAD that is insulated from its neighbors by the duplicated boundary (Franke et al., 2016).
By combining these structural data with enhancer detection methods (see Box 2) and gene expression datasets, an accurate description of the regulatory architecture underlying particular genes can be obtained. This type of approach was recently instrumental in physically linking hundreds of genes involved in human brain development with regulatory regions under positive selective pressure as well as with non-coding variants involved in schizophrenia (Won et al., 2016). Accordingly, with the understanding of chromatin domains such as TADs, subTADs, loop domains or insulated neighborhoods in a given tissue, it is possible to refine the genomic interval important for the regulation of genes as well as the position of putative enhancer regions. It should also be noted that, aside from chromatin domains, subMb-scale chromatin interactions have been shown to structure the genome in a cell type-specific fashion (Javierre et al., 2016; Phillips-Cremins et al., 2013) and, as such, turn out to play a crucial role in gene regulation.
Functional chromatin interactions are dynamic
Despite TADs being rather invariant chromatin structures, extensive differences have been observed within them among different cell types (Dixon et al., 2015). Alterations in intra-TAD chromatin microarchitecture are best shown using viewpoint-specific, proximity ligation technologies, such as 3C, 4C and capture-C/-HiC, because all regions interacting with a specific defined region can be examined at a higher resolution than with HiC (Simonis et al., 2006; Noordermeer et al., 2011; Hughes et al., 2014; Schoenfelder et al., 2015) (Fig. 3, Box 2). Using these methods, extensive changes in the interactions between enhancers and promoters can be observed among different cell types or during embryonic development (Javierre et al., 2016; Andrey et al., 2017; Freire-Pritchett et al., 2017). The first study that used 3C to assess the chromatin architecture of the mouse β-globin locus showed that long-range interactions occur between the β-globin promoter and a cluster of enhancers called the locus control region (LCR); these interactions occur specifically in the β-globin-expressing erythroid cells, but not in brain cells that do not express it (Tolhuis et al., 2002). At other mouse genomic loci, such as at the HoxD cluster or at the Satb1 gene, extensive changes in enhancer-promoter interactions are also observed during development, allowing tissue-specific enhancer-promoter contacts to form (van de Werken et al., 2012; Andrey et al., 2013). Through the use of proximity ligation technologies, the dynamics of chromatin interactions from many specific genomic viewpoints can be simultaneously studied in different tissues or cell types (Simonis et al., 2006; Noordermeer et al., 2011; Hughes et al., 2014; Schoenfelder et al., 2015). For example, by probing the promoter interactome of 17 human primary hematopoietic cell types, one study has shown that most of the scored interactions are cell type or lineage specific and are thus highly dynamic (Javierre et al., 2016). Moreover, by studying hundreds of viewpoints during mouse limb and midbrain development, two types of chromatin interactions were identified. The first type of contact remained invariable throughout development and between tissue types. The regions involved in these interactions associate with CTCF and cohesin and are likely to play a structural role in organizing the local folding of chromatin in a tissue-independent manner. The second type of chromatin interaction was tissue and/or time point specific. The sites involved in this type of interaction show enrichment for repressive or active functional chromatin marks, indicating that they belong to regulatory regions (Andrey et al., 2017) (Fig. 3). The association between chromatin modifications and facultative chromatin interactions suggests that mechanisms involving the deposition of chromatin marks might mediate this association. Indeed, and as we describe in the section below, Polycomb-repressed regions are spatially clustered together via PRC1 and PRC2 complexes (Denholtz et al., 2013; Joshi et al., 2015; Schoenfelder et al., 2015). Conversely, at transcriptionally active regions, the mediator complex helps to recruit factors, such as the cohesin complex, that bind at promoters and enhancers and enables chromatin interactions between them (Ebmeier and Taatjes, 2010; Kagey et al., 2010; Larivière et al., 2012; Carlsten et al., 2013; Phillips-Cremins et al., 2013).
An important remaining question is whether the observed enhancer-promoter contacts are a cause or consequence of gene regulation and if looping in general is essential and sufficient to activate a gene. Most interestingly, experimentally induced ‘forced’ looping between enhancers and promoters, outside of their native cellular context, has provided evidence that looping actively controls the accuracy of the onset of gene expression (Deng et al., 2012, 2014). Specifically, the proximity with which enhancers locate to promoters increases the number of alleles that are transcribed per cell (the so-called transcriptional burst fraction), but not the number of RNA molecules produced per transcriptional burst (the burst size) (Bartman et al., 2016). However, in a naturally occurring transcriptional situation, both the burst size and burst fraction are increased, suggesting that the burst size is independent of the spatial proximity between enhancers and promoters. Accordingly, cell fate changes and lineage commitment are associated with large changes in 3D chromatin architecture, and chromatin looping appears to be a driving force in this process. To understand the ground state of this architecture, many studies have thus focused on genome structure in pluripotent cells.
3D genome architecture in pluripotent cells
Pluripotent cells (PCs) share certain features of genomic organization with differentiated cells; they maintain TAD structures as well as long-range A/B compartments that associate with regions of transcriptional activity. However, in contrast to other cell types, PCs display a less compact chromatin organization (Meshorer et al., 2006; Melcer and Meshorer, 2010; Gaspar-Maia et al., 2011). Moreover, long-range contacts between repressed regions are more non-specific in PCs than in differentiated cell types, suggesting that transcriptionally inactive genomic compartments are less strongly established in PCs than in differentiated cells (de Wit et al., 2013).
In PCs, pluripotency gene loci have a tendency to come into contact with each other in cis and in trans in the nuclear space, somehow sharing a common, cell type-specific, transcriptional machinery (de Wit et al., 2013) (Fig. 4A). Regions bound by pluripotency factors [OCT4 (POU5F1), KLF4, SOX2 and NANOG] were found to colocalize in the nucleus and to form long-range chromatin contacts with each other, indicating that pluripotency factors themselves partially control the genomic architecture of PCs (Denholtz et al., 2013; Bouwman and de Laat, 2015). A study using high-resolution imaging of SOX2-bound enhancers in mouse PCs has revealed their spatial clustering in the nucleus (Liu et al., 2014). In another experimental setting, Wei and colleagues showed that several regions bound by KLF4 are in close spatial proximity to each other in PCs and released upon differentiation or depletion of Klf4 (Wei et al., 2013). These results were further confirmed by single-cell HiC experiments in haploid mouse ESCs (Stevens et al., 2017). At the intra-TAD level, the formation of chromatin loops was also found to depend on pluripotency factors, which are lost upon differentiation (Kagey et al., 2010; Phillips-Cremins et al., 2013). In particular, several instructive enhancer-promoter interactions, within PC-specific loops, have also been found at the Sox2 locus (Li et al., 2014). Regions enriched for the binding of Polycomb proteins (see Glossary, Box 1), which are involved in the repression of developmental genes, also frequently interact with each other in PCs, highlighting a role for this protein family in genome organization during pluripotency. Accordingly, when components of the Polycomb PRC1 or PRC2 complexes (see Glossary, Box 1) are inactivated in mouse ESCs, chromatin contacts between Polycomb-repressed genes become altered (Denholtz et al., 2013; Joshi et al., 2015; Schoenfelder et al., 2015).
The establishment of PC-specific chromatin architecture has also been used as a model to study the dynamics of genome organization and the factors that influence this. At various loci, the cohesin and mediator complexes were shown to co-bind at chromatin interaction sites that were also occupied by pluripotency factors (Phillips-Cremins et al., 2013). Upon differentiation, the pluripotency factors are lost and interactions between transcription factor-bound regions are significantly decreased, suggesting that pluripotency factors tether architectural complexes such as the cohesin and mediator complexes (Levasseur et al., 2008; de Wit et al., 2013). In the reverse situation, upon the reprogramming of various cells into induced pluripotent stem cells (iPSCs; see Glossary, Box 1), the PC-specific genome topology is re-established to a large extent, and A/B domains specifically as well as enhancer-promoter interactions at pluripotency loci are very similar to those in ESCs (Beagan et al., 2016; Krijger et al., 2016). It is noteworthy that the maintenance of subMb-scale topological memory of founder cell hallmarks was observed in both studies.
In PCs, as in other cell types, two possible regulatory architectures underlie gene activity: instructive or permissive. An instructive architecture refers to tissue-specific regulatory interactions whereas a permissive one refers to tissue-independent regulatory interactions that are set independently of gene activation (de Laat and Duboule, 2013). An example of permissive architecture in PCs has recently been described at poised enhancers of early differentiation genes. These regions are enriched for H3K27me3 and interact with their putative target genes, prior to their activation, in a PRC2-dependent manner (Cruz-Molina et al., 2017). This suggests that Polycomb interactions might contribute to the formation of a permissive regulatory environment, in which the regulatory contacts of these genes are poised for activation (Fig. 4B). Similarly, several stem cell enhancers are found in tissue-invariant chromatin loops and are demarcated by the architectural proteins CTCF and cohesin. When these CTCF-binding sites are genetically disrupted using CRISPR/Cas9 (see Box 4), the misregulation of neighboring genes is observed, suggesting that this permissive environment insulates neighboring loci (Dowen et al., 2014). In summary, whether it is in an instructive or permissive state, the chromatin structure of PCs appears to be flexible and poised for transcriptional activation of early or late differentiation genes.
The use of CRISPR/Cas9 to introduce mutations in mouse embryonic stem cells (mESCs) or via pro-nuclear injection of mouse zygotes has improved the speed and efficiency with which mutations can be introduced into the mouse genome. CRISPR/Cas9 is based on the targeted double-strand break induced by the Cas9 nuclease at sites where a single guide RNA (sgRNA) precisely hybridizes. The induction of the double-strand break will trigger the double-strand break repair mechanism, which will introduce mistakes leading to the desired mutation.
Structural variants. By using sgRNAs that target two loci in cis that are up to 1.6 Mb apart, it is possible to generate deletions, inversions or duplications of the intermediate DNA region. This approach has been performed both in mESCs (CRISVar) and via zygotic pro-nuclear injection (CRISMERE) (Kraft et al., 2015; Birling et al., 2017).
Motif deletion. By using an sgRNA in ESCs or in zygotes, indels of 1 bp to several dozen of bps can be introduced at the target site. This technology is very useful for removing transcription factor-binding motifs, such as those to which CTCF binds (Wang et al., 2013; Andrey and Spielmann, 2017).
Recombination of targeting cassettes. By using an sgRNA, it is possible to increase the recombination frequency of a targeting cassette in mouse zygotes or in mouse ESCs (Wang et al., 2013; Andrey and Spielmann, 2017).
3D genome architecture during lineage commitment and development
The function and extent of dynamic chromatin interactions are particularly visible during cell differentiation and more generally during developmental processes. During lineage commitment, extensive switching of chromatin contacts between A and B compartments, which associate with active and inactive transcriptional domains, respectively, is observed. Specifically, around 36% of compartments switch from one type to another as human ESCs differentiate into various lineages (Dixon et al., 2015). Moreover, many pluripotency loci become re-positioned toward the nuclear lamina as they are shut off (Peric-Hupkes et al., 2010). Thus, higher-order genome structure undergoes a widespread remodeling during cell differentiation, which underlies the regulatory decisions associated with cell fate.
As in PCs, the 3D regulatory chromatin architecture that underlies developmental transcriptional regulation is either permissive or instructive. Accordingly, in a large-scale study of the promoter interactomes of PCs and neuroectodermal cells, around 50% of promoter-interacting regions were re-wired and 50% were retained between both cell types, arguing for extensive chromatin dynamics upon cell differentiation (Freire-Pritchett et al., 2017). Early on in differentiation, transcriptionally active pluripotent genes, such as Sox2, Oct4 and Nanog, are repressed and lose their instructive active chromatin interactions with pluripotent enhancers (Phillips-Cremins et al., 2013; Li et al., 2014). In contrast, early differentiation genes, such as early posterior neural genes, which display poised, permissive, PRC2-dependent interactions with their enhancers in PCs, maintain these pre-formed interactions and are therefore rapidly activated upon differentiation (Fig. 3) (Cruz-Molina et al., 2017).
Studies of various specific loci have also provided insights into the dynamic nature of the genome during development. RNA and DNA fluorescence in situ hybridization experiments in cells of the mouse posterior limb show that the Shh enhancer is in close physical proximity to the Shh promoter in those cells that express Shh (Amano et al., 2009). Conversely, in anterior limb cells, in which the ZRS enhancer is not active, the Shh promoter is nevertheless in close physical proximity, but no Shh transcription is observed (Amano et al., 2009; Williamson et al., 2016). Here, a permissive type of 3D structure is at play, and thus the activity of the enhancer alone dictates the expression of the Shh gene. Similarly, the transcriptional response of target genes to glucocorticoid receptor (GR) activation has been shown to occur without significant remodeling of pre-formed chromatin contacts, allowing for the rapid onset of gene transcription (Hakim et al., 2011). Interestingly, the mouse Hoxd13 promoter forms several tissue-non-specific chromatin contacts with distal enhancer regions, as well as a few tissue-specific interactions within the same genomic landscape. Specifically, Hoxd13 establishes a digit-specific interaction with the digit enhancer island-3, and a genital tubercle-specific interaction with the genital tubercle enhancer GT-2 (Montavon et al., 2011; Lonfat et al., 2014). In these cases, as most of the chromatin structure is permissive, some regulatory interactions are formed in a tissue-specific manner. Other Hoxd genes display a different regulatory architecture. The central mouse Hoxd genes, from Hoxd9 to Hoxd11, regulate limb patterning and shift their contacts between two adjacent TADs. In early and proximal limb tissues, they establish interactions with early enhancers in a telomeric TAD; during later development and in more distal limb regions, these Hoxd genes shift their interactions toward a centromeric TAD that contains digit enhancers. This complex regulatory transition allows for the formation of an intermediary zone, in which low Hoxd expression helps to pattern the wrist (Andrey et al., 2013). HOXA13 and HOXD13 then establish the chromatin structure required for digit-specific patterning (Beccari et al., 2016). At the β-globin and Satb1 loci, extensive changes in enhancer-promoter interactions occur during erythroid maturation, suggesting that these genes rely on an instructive type of transcriptional onset (Palstra et al., 2003; Vernimmen et al., 2007; van de Werken et al., 2012; Deng et al., 2014). Tissue-specific transcription factors have also been shown to mediate these newly established contacts during lineage commitment. For example, at the β-globin locus, erythroid-specific transcription factors such as EKLF (KLF1), GATA1 and FOG1 (ZFPM1) mediate the cell-specific interactions that occur between the gene locus and its associated enhancers (Drissen et al., 2004; Vakoc et al., 2005).
It is possible that dynamic chromatin interactions act as an active mechanism to control gene transcription rather than representing a passive byproduct of it. Accordingly, the chemical inhibition of transcription at the β-globin locus does not prevent chromatin interactions to occur in a tissue-specific manner with the LCRs, thereby disconnecting transcription from chromatin structure (Palstra et al., 2008). Another study has also shown that the formation of ectopic loops between the LCR and the β-globin gene in the pro-erythroblast cell line GE1, in which β-globin is normally not expressed, results in the strong overexpression of this gene (Deng et al., 2012). The diversity of regulatory architectures observed at different loci has evolved in order to produce a spatially, temporally and lineage-specific expression pattern for the involved genes. Accordingly, and as we discuss below, changes in 3D chromatin folding, in particular in the structure of TADs, induced by targeted or normal mutations have been shown to affect gene expression and to result in disease or malformation.
Altered TAD structure perturbs enhancer-promoter communication and gene expression: insights into disease
As we have highlighted above, TADs help to shape the overall architecture of the genome and thereby insulate and delineate the extent of regulatory cues. Changes that affect the insulation properties of TADs can thus allow regulatory elements to ectopically contact promoters in neighboring TADs and thereby induce potentially pathogenic gene misexpression. The disruption of TAD structures by deletions, duplications or inversions, collectively called structural variants (SVs), can lead to such effects through enhancer-adoption mechanisms (Fig. 5A-C). Using CRISPR-derived methodologies (see Box 4) to induce SVs, such mutations can be re-engineered in mice (Kraft et al., 2015). Using this approach, it has been shown that a deletion that breaks a TAD boundary and decreases the genomic distance between a set of Epha4 limb enhancers and the Pax3 gene leads to the overexpression of Pax3 in an Epha4-like pattern in the mouse limb (Lupiáñez et al., 2015). In this situation, Pax3 adopts the Epha4 enhancers and is then misexpressed, leading to a shortening of phalanges. In line with the insulating function of TAD boundaries, a similar effect was not observed when a slightly smaller deletion, leaving the TAD boundary intact, was induced. In addition, an inversion at the Wnt6/Epha4 locus re-positions the same Epha4 enhancers in the vicinity of Wnt6 and induces its ectopic expression in the developing distal limb bud, leading to a specific type of digit malformation, called F-syndrome. Finally, duplications that re-position Epha4 enhancers in front of the Ihh gene induce the overexpression of this gene in distal developing mouse limb buds, leading to polydactyly.
In cancer patients, oncogenes such as TAL1 or LMO2 have also been found to be activated as a result of boundary deletion causing the loss of insulated neighborhoods (Hnisz et al., 2016). Most of these spontaneously occurring deletions span boundary regions of several kilobases and do not precisely pinpoint the underlying set of factors that are responsible for establishing the boundary. As stated above, an important factor implicated in boundary formation is the transcription factor CTCF. Upon the genetic removal of CTCF-bound regions in mouse ESCs, which correspond to domain boundaries, ectopic interactions between genes and enhancers are observed that lead to gene misexpression (Dowen et al., 2014). This suggests that mutations that affect CTCF binding in human patients would also lead to gene misexpression and potentially to disease. Accordingly, in cancer patients with IDH gliomas, the ectopic methylation of a boundary element has been shown to associate with decreased CTCF binding, and with increased inter-TAD interactions and the misexpression of the PDGFRA oncogene (Flavahan et al., 2016). However, it is unclear if this mechanism can be applied to other loci, as CTCF was previously shown to bind independently of methylation levels and to trigger demethylase activity in mouse cells (Stadler et al., 2011).
Another type of mechanism has recently been shown to generate pathogenic effects through genomic duplications. The duplication of regions that encompass enhancers as well as a TAD boundary can result in the formation of new chromatin domains, called neo-TADs, that are isolated from the rest of the genome (Franke et al., 2016; Weischenfeldt et al., 2017). If the duplicated region does not contain genes, its regulatory activities remain restricted to the neo-TAD, and are therefore without effect. However, if a gene is included in the duplication, it will adopt the regulatory activities of the neo-TAD thereby acquiring new expression domains (Fig. 5D). This phenomenon can lead to ectopic gene expression, as shown for the Kcnj2 gene in mice, which adopts a Sox9 pattern when included in the neo-TAD that contains duplicated Sox9 enhancers (Franke et al., 2016). This de novo expression of Kcnj2 results in a congenital malformation that is characterized by nail aplasia and short digits (Franke et al., 2016). In this example, the regulation of neighboring genes remains largely unchanged. This phenomenon has also been described at the human IGF2 locus, where genomic duplications create a new TAD domain that consists of the IGF2 gene and a colorectal cancer lineage-specific super enhancer, which results in the misregulation of the oncogenic locus (Weischenfeldt et al., 2017).
It can be easily envisioned that the formation of new regulatory domains that are separated from the rest of the genome by boundaries provide an ideal setting in which to acquire new functions in an evolutionary context. These results suggest that the formation and isolation of a newly formed TAD can result in a phenotypic change in an organism that is then directly subjected to selective pressure, without affecting the parent copy of the gene. In that sense, the shifting of TADs and the recombining of regulatory activity with new target genes provides a toolbox of possibilities for how new gene functions can be acquired (Franke et al., 2016).
A crucial advance in our understanding of gene regulation came with the finding that the genome undergoes three-dimensional folding in the nucleus in a genetically determined process that directly influences gene regulation via the formation of chromatin units called TADs. These units directly influence the availability of enhancers for their target genes. The complex interplay between distal genomic sequences that contain regulatory elements and their target genes and promoters has proven to be essential for regulating genes during development and lineage commitment. In particular, the development of new technologies that measure the frequency of DNA interactions has advanced the field, enabling the analyses of multiple genomic regions at high resolution, even in single cells. Moreover, the recently developed genome architecture mapping (GAM) approach, which does not rely on crosslinking or ligation, has revealed numerous three-way interactions between chromatin regions and will increase our capacity to investigate such complex chromatin interactions (Beagrie et al., 2017). Furthermore, the development of direct visualization technologies that rely on super-resolution microscopy techniques, especially in vivo, should provide novel insights into the allele-to-allele variability of dynamic chromatin interactions.
The role of the 3D genome architecture during development or lineage commitment can be described in two ways. In the first rather structural view, the genome is divided into chromatin domains that have been described in different ways, such as TADs, loop domains or insulated neighborhoods. One interesting feature of TADs is their maintenance during the oocyte-to-zygote transition in mice or their establishment prior to genome transcriptional activation in Drosophila, which suggests that they are formed prior to the transcriptional function they control (Flyamer et al., 2017; Hug et al., 2017). Chromatin domains, and in particular TADs, are in this respect believed to act by delineating the regions scanned by enhancer and promoters, independently of tissues or cell types. The second view takes into account the extensive chromatin dynamics that occur during cell differentiation and development, within and in between domains. This is best exemplified by the recently characterized interactome footprints, which are sufficient to identify specific cell types (Javierre et al., 2016). It will be important to understand which factors induce the formation of these variable chromatin structures. Obviously, tissue-specific transcription factors play a key role in producing them but we also need to identify the mechanisms through which they act. Changes in chromatin modifications, which directly derive from the binding of transcription factors, can also be associated with the dynamics of chromatin interactions (Javierre et al., 2016; Andrey et al., 2017). Thus, transcription factors might act by recruiting other factors to the chromatin, which are linked to chromatin modifications and which themselves trigger the looping.
Future challenges will involve the dissection of the biochemical mechanisms underlying the formation of permissive or instructive chromatin loops. These findings, as well as progress in modeling technologies, will help to improve predictions of the architectural outcome and thus pathogenic effect of mutations involved in various conditions, including congenital malformations and cancers. The quantification of chromatin architecture and associated dynamics in heterogeneous tissue will benefit a lot from purer cell populations, obtained by, for example, fluorescence-activated cell sorting technologies or high-resolution microscopy. Together, these studies will also help establish the causal link between chromatin interactions and gene expression. Important studies have already paved the way to the modulation of tissue-specific loops, either by activating them prior to their normal activation or by changing their interaction partners in cell culture systems (Deng et al., 2012, 2014; Bartman et al., 2016). However, the regulatory role that these dynamic chromatin interactions play in establishing the tightly regulated expression patterns of development genes in vivo, which ultimately control the morphogenesis of organs and structures, remains to be determined.
We apologize to colleagues whose work could not be included owing to space constraints. We thank members of the Mundlos laboratory for helpful discussions.
Work in S.M.'s laboratory is funded by the Deutsche Forschungsgemeinschaft, the Berlin Institute of Health, and the Max Planck Foundation.
The authors declare no competing or financial interests.