Unraveling the 3D genome: genomics tools for multi-scale exploration

Abstract

A decade of rapid method development has begun to yield exciting insights into the three-dimensional architecture of the metazoan genome and the roles it may play in regulating transcription. We review here core methods and new tools in the modern genomicist’s toolbox at three length scales, ranging from single base pair to megabase scale chromosomal domains, and discuss the emerging picture of the 3D genome that these tools have revealed. Blind spots remain, especially at intermediate length scales spanning a few nucleosomes, but thanks in part to new technologies that permit targeted alteration of chromatin states and time-resolved studies, the next decade holds great promise for hypothesis-driven research into the mechanisms that drive genome architecture and transcriptional regulation.

The physical landscape of genome biology

More than a decade after the completion of a high-quality reference sequence [1,2], we have seen substantial strides identifying elements of the genome that function in different cellular contexts. However, our understanding of the physical and spatial organization of the human genome at multiple scales remains stubbornly incomplete. The physical genome is an effectively one-dimensional object housed in a nucleus some 400,000 times shorter than its longest axis. This fundamental spatial constraint necessitates a complex hierarchical compaction of the one-dimensional genome into three-dimensional chromatin within the nuclear volume. The specifics of this three-dimensional organization set the physical landscape of genome biology – from determining which regulatory elements are accessible to transcription factors, to regulating which distal enhancer elements make contact with which genes. By analogy to protein structure, the spatial hierarchy of genomic compaction can be roughly divided into three structured length scales: primary structure, consisting of sequence elements, DNA-bound proteins and nucleosomes; secondary structure, consisting of interactions between nearby nucleosomes that shape local chromatin architecture; and tertiary structure, consisting of long-range three-dimensional features such as loops spanning a few hundreds of kilobases (kb) to chromosome domains spanning megabases (Mb) (Figure 1) [3]. Thanks to the rapidly dropping cost of sequencing, novel genome-wide methods have been developed to begin to probe chromatin structure, especially at the primary and tertiary levels. An integrative, comprehensive, multi-scale understanding of chromatin organization and a picture of how these layers of hierarchal organization bring about gene regulation is perhaps the next great challenge of genomics. In this review, we briefly describe how the three levels of genome architecture impact transcriptional regulation and highlight some of the methodological developments that have provided these insights, focusing on methods that can probe three-dimensional structure (Figure 1). Finally, we discuss the crucial blind spots in our understanding of the genomic landscape, and outline the characteristics of future technologies that are still missing from the genomicist’s toolbox.

Figure 1. Overview of chromatin structure and assays at three scales.

Chromatin structure is divided into three size scales, by analogy to protein structure. Primary structure encompasses DNA methylation (pink) and sequence features, DNA-bound factors (blue), nucleosome position and modifications (multi-colored), and DNA accessibility. Secondary structure encompasses local structures formed by nucleosome-nucleosome interactions, and although several models are shown here, the lack of methods that can probe this organizational scale of chromatin means that sequence-resolved in vivo architecture at this scale is not fully understood. Tertiary structure encompasses promoter-enhancer looping (on the order of several kb to several hundreds of kb) and megabase-scale chromosome domains. Many methods exist for assaying the primary and tertiary scales of chromatin structure for both architecture and the identity of DNA-bound trans factors, but no genomics methods directly assay secondary structure.

Transcriptional regulation in three dimensions

Eukaryotic transcription is intimately tied to chromatin structure at multiple scales. The transcription machinery assembles on the core promoter, a primary structure feature located around the transcription start site (TSS) of a gene that contains binding sites for components of the RNA polymerase pre-initiation complex [4]. On its own, the core promoter of most genes cannot activate reliable transcription because of the obstacle posed by chromatin [5]. Full transcriptional activation requires the action of transcription factors (TFs), DNA-binding proteins that also bind components of the RNA polymerase II transcription initiation complex and help to overcome this obstacle by recruiting chromatin modifying and remodeling complexes. The local secondary structure of chromatin, which may be altered by the action of these complexes, is thought to be a regulator of both transcription and TF binding by controlling the accessibility of DNA [6,7]. Indeed, both gene expression profiles and the genome-wide maps of open chromatin sites are highly cell-type specific [8].

Long-range tertiary interactions also play a role in regulating transcription. Although some TF binding sites occur in and near core promoters of highly transcribed housekeeping genes, most TF binding sites occur in enhancer regions. Enhancers are regulatory sequences that act in cis despite being as far as hundreds of kb away from their target gene(s) [9,10]. The currently dominant, if not uncontroversial, model is that enhancer sequences form three-dimensional contacts with promoters, integrating information from multiple TFs and tethering those TFs and their chromatin remodeling binding partners to target promoters [11]. A natural question thus arises: what determines which enhancers contact which promoters? 3D genomics aims to answer this question by mapping the megabase-scale chromosomal interaction domains that are thought to play a role in regulating enhancer-promoter looping [12]. The interaction domain map appears to correlate with genome-wide patterns of secondary and primary structure features like chromatin compaction and histone modifications [13], suggesting that the three length scales of chromatin architecture may feed back on each other, coordinating in their regulation of transcription [14]. Thus with the recognition over the past decade of the importance of the physical organization of DNA, there has been a growing toolbox of methods developed to study transcriptional regulation at all three length scales. The resulting data sets are slowly generating a comprehensive genome-wide picture of multiple aspects of this organization.

Primary structure

The primary structure of DNA consists, first and foremost, of DNA sequence, which can determine transcription factor binding sites, methylation sites, and in some cases, nucleosome positions, as well as identifying mutations between individuals. Primary structure also encompasses the epigenomic features that can be linearly mapped onto sequence, including histone modifications and variants and DNA accessibility.

At the smallest length scale of epi-genomics, bisulfite sequencing [15,16] and its targeted variant RRBS [17] (Figure 2A, Table 1) have revealed that most of the human genome is stably methylated on CpG dinucleotides [18] and related assays have shown that hydroxymethylation may be important in brain tissue [19,20]. Regions rich in hypomethylated CpGs coincide with some promoters and have been proposed to act as seeding regions for open chromatin, allowing early-acting pioneer transcription factors to bind DNA and locally open up the chromatin for other factors [5,21].

Figure 2. Methods that assay the primary structure of chromatin.

(A) The most widely used assays for cytosine methylation with base-pair accuracy rely on sodium bisulfite, which converts unprotected cytosines to uracils. In a sequencing reaction, methylated and hydroxymethylated sites are read as cytosines, while all other cytosines read as thymines. Tet-assisted bisulfite sequencing uses two additional enzymatic steps to selectively protect hydroxymethylcytosines from bisulfite conversion, and reduced-representation bisulfite sequencing uses a methylation-sensitive restriction enzyme to cleave near methylated CpGs, ensuring that they are read during sequencing. (B) DNA footprinting can be done with either endonucleases or transposases, which cleave unprotected DNA, and with MNase, which has hybrid endonuclease and exonuclease activity and “nibbles” free DNA until it reaches an obstacle, such as a TF or a nucleosome. DNA fragments are purified, size selected and sequenced. DNase-seq focuses on TF footprints from short fragments, whereas DNase-FLASH isolates multiple size classes to assay both nucleosome and TF footprints. (C) Fragmented chromatin can be assayed to identify the sequence of DNA bound to protein (ChIP) or RNA (ChIRP). The fragmentation methods vary, from sonication of crosslinked chromatin to MNase digestion of uncrosslinked chromatin (called native ChIP). Target proteins are pulled down with antibodies, whereas RNAs of interest are pulled down with biotinylated antisense oligos. ChIP-exo employs lambda exonuclease to digest away excess free DNA that is pulled down, increasing the resolution of the method. (D) DNA accessibility, a useful indicator of active regulatory regions, can be assayed by DNase I cleavage (DNase-seq/DHS-seq), Tn5 transposase insertion of sequencing adapters (ATAC-seq), or fragmentation by sonication (SONO-seq), or the depletion of protein-bound fragments (FAIRE-seq). A 30-nm chromatin fiber cartoon has been used to depict the contrast between open and closed chromatin in this panel.

Table 1.

Selected methods for primary chromatin structure

Method	Assay for	Description	Resolution (bp)	Required input	Notes
Whole-Genome Bisulfite sequencing (WGBS) [15,16]	Cytosine methylation and hydroxyl-methylation	Sodium bisulfite converts unprotected cyotsines to uracil. Sample is sequenced with and without bisulfite treatment.	1	20 ng gDNA	This assay cannot distinguish methylation from hydroxymethylation. Several updates have improved this protocol for small numbers of cells [15,16].
Tet-assisted bisulfite sequencing [20] and Oxidative bisulfite sequencing [19]	Cytosine hydroxyl-methylation	Hydroxymethylcytosine is glucosylated by beta-glucosyltransferase and methylcytosine is converted to carboxymethylcytosine by Tet. Upon treating these libraries with sodium bisulfite and sequencing, only the hydroxymethylcytosine reads as C [20]. Alternatively, hydroxymethylcytosine is oxidized to formylcytosine, which is converted to uracil by bisulfite treatment. [19]	1	3 μg gDNA	Recent studies have addressed the role of hydroxymethylation in brain tissue, where it is particularly enriched, and the function of endogenous Tet proteins in regulating chromatin [20].
Reduced representation bisulfite sequencing (RRBS) [17]	Cytosine methylation and hydroxyl-methylation	A methylation-insensitive restriction enzyme that cuts at CpG dinucleotides is used to generate fragments that contain at least one CpG site.	1	100–300 ng gDNA	This protocol is more cost-efficient than whole-genome bisulfite sequencing because it enriches for reads that contain potential methylation sites that are of interest.
FAIRE-seq [22].	Open chromatin	Nucleosome-bound DNA is crosslinked and removed by phenol-chloroform extraction, and remaining nucleosome-free DNA is analyzed by microarray or sequencing.	~200	106–107	This is a simpler assay for open chromatin than DNase-seq, although its resolution is somewhat lower.
SONO-seq [164]	Open chromatin	Crosslinked chromatin is sonicated as for a ChIP experiment (see below) but no immunoprecipitation is performed. This signal is compared to sonicated naked DNA.	100–350	~108	This method arose from an observation that no-IP input libraries in ChIP experiments had peaks in sonication breaks near promoters and other open chromatin sites.
DNase-seq or DHS-seq [6,8,23–25,29–31,165,166]	Open chromatin and transcription factor footprints	DNA in isolated nuclei is digested with DNase I at a concentration that must be optimized for each experiment. A library is prepared from the digested fragments by ligation of adapters and cleavage of ~20 bp sequence tags followed by size selection of a unique library molecule size or by biochemical fractionation of fragments followed by ligation of sequencing adapters	~1	107–108	In one experiment, this method reveals the genome-wide landscape of open chromatin, which correlates well with all transcription factor-bound sites. Technically difficult, involving optimization of DNase I concentration. The sequence bias of DNase I can give rise to artifactual apparent footprints on DNA [122].
DNase- FLASH [33]	Nucleosome and transcription factor footprints	This variation on DNase-seq systematically analyzes a broad range of fragment sizes to capture both nucleosome footprints and transcription factor footprints.	1	107–108	This method is particularly useful for mapping the structure of transcription start sites and simultaneously probes the positions of nucleosomes and transcription factors. regulatory regions because it
ATAC-seq [26]	Open chromatin, nucleosome positions, transcription factor footprints	Tn5 transposase is used to transpose sequencing adapter oligos into the genomic DNA of permeabilized, unfixed cells. The resulting library is then purified and sequenced.	~1	102–104	This simple, fast protocol generates very similar data to DNase-seq with lower input requirements. Tn5 transposase also exhibits sequence bias, which must be accounted for in footprinting experiments [122].
MNase-seq [36,37,167]	Nucleosome footprints (and in a variation, smaller DNA-bound proteins)	Micrococcal nuclease is used to digest chromatin in nuclei from cells lysed by cryogenic grinding or detergent. The enzyme has endonuclease activity that cuts linker DNA between nucleosomes and exonuclease activity that digests any linker not protected by the core nucleosome. Nucleosome-sized fragments are size-selected and prepared for sequencing.	~1–10	107	MNase digestion is the method of choice for global fragmentation of chromatin into nucleosome core particles. Digestion conditions used in MNase-seq experiments do not select for cleavage only in open chromatin regions, generating data genome-wide and thus requiring large numbers of reads to obtain sufficient depth. MNase is often used to fragment nucleosomes before ChIP (see below).
Chemical mapping of nucleosomes [124,125]	High-resolution nucleosome position mapping	This method requires histone mutants that contain a single cysteine in the core nucleosome near the nucleosome dyad. A thiol-reactive copper chelator is chemically linked to the unique pair of thiols on nucleosomes in fixed and permeabilized cells. Copper is added, and peroxide is used to oxidize the copper, locally generating free radicals that cleave the DNA at two stereotyped positions around the nucleosome dyad. The resulting DNA fragments are then purified and sequenced.	~1	Budding and fission yeast only	The free radical mediated cleavage of DNA eliminates enzyme-based bias, although sequence bias may still be introduced by downstream library preparation steps. This method has currently been implemented in S. cerevisiae [124] and S. pombe [125].
ChIP-seq [40]	Mapping of DNA-bound proteins (including nucleosomes)	Chromatin is fixed in cells, then fragmented by sonication or MNase digestion before enrichment for the protein epitope of interest using a specific antibody. Crosslinks are reversed using proteinase K and heat and the DNA is then prepared for analysis by sequencing, array hybridization or PCR.	~100	~107	The resolution of this method is limited by fragment size, which is determined by sonication in experiments targeted to non-nucleosome proteins, and falls in the 100–200 bp range.
Native ChIP-seq [41,42,167–169]	Mapping of DNA-bound proteins under native conditions (mainly nucleosomes)	Chromatin is digested with MNase in permeabilized but unfixed cells, and then solubilized by passing through fine-gauge needles before preparation of a sequencing library from the fragments. Mono and dinucleosome fragments are isolated by density gradient centrifugation of nucleosome particles or by size selection of DNA after adapter ligation.	~10	~103–107	This method relies on the stability of the nucleosome-DNA interaction to maintain the association of the bound DNA with the protein target of the antibody. One recent variation of this method uses low salt to stabilize transcription factor-DNA interaction for native transcription factor ChIP [42], while another has optimized the protocol for input as low as 1000 cells [169]
ChIP-exo [170]	High-resolution footprinting of nucleosomes with particular modifications and DNA-bound proteins	Crosslinked and fragmented DNA with bound proteins is digested with lambda exonuclease to remove excess unprotected DNA, before reversal of crosslinks, library preparation, and single-end sequencing.	~1	106	The development of this method dramatically increased the resolution of the long-standing crosslinked ChIP protocol.
ChIRP-seq [53,54]	Mapping of genome interactions with a target lncRNA	Biotinylated antisense oligos are designed to tile the ncRNA of interest. These oligos are then hybridized to chromatin that has been gluteraldehyde-crosslinked in intact cells, then fragmented by sonication. Oligo-bound RNA and its associated DNA are isolated with streptavidin beads, and DNA is then purified and sequenced.	~100	~107	This method permits mapping of RNA-DNA interactions far from the transcribed locus of the RNA of interest.
STARR-seq [51,52]	Functional genome-wide assay for enhancer activity	Fragments spanning the genome are cloned into a library of GFP reporter constructs between the polyadenylation site and the transcription termination site	N/A	Drosophila only	STARR-seq reveals the potential of a fragment of genomic sequence to act as an enhancer for a promoter of interest. The results therefore depend on the promoter sequence used in the assay.

At the length scale of a few to a few hundred base pairs, methods can be divided into those that probe chromatin accessibility and those that map the binding of a factor of interest on a linear DNA sequence (Figure 2D). In the first category, FAIRE-seq sequences accessible DNA by depleting nucleosome-associated DNA [22]. DNase-seq [23] uses the endonuclease DNase I to preferentially nick DNA in open chromatin, then sequences the resulting fragments [6,24,25]. A new method, ATAC-seq, uses Tn5 transposase to fragment DNA at open chromatin regions and efficiently produces sequencing libraries in a single step from low cell numbers [26]. Regardless of which method is used, open chromatin analysis is a highly effective way to study transcriptional regulation genome-wide, as nuclease sensitive sites have long been known to indicate active regulatory regions such as promoters and enhancers [27,28]. Indeed, one approach for assigning enhancers to their target promoters has been to look for correlations in open chromatin at pairs of enhancers and promoters across cell types [8].

In areas of accessible chromatin, DNase, Tn5 transposase, and micrococcal nuclease can also be used for footprinting assays, in which short fragments of DNA protected against nuclease cleavage by bound nucleosomes or transcription factors are purified and sequenced (Figure 2B) [26,29–33]. All of these enzymes can be used to map both nucleosomes and TFs, but because MNase digests all linker DNA while leaving nucleosome-bound DNA intact, this approach works well for genome-wide nucleosome footprinting [34–37], whereas DNase I is primarily used for TF footprinting because it can nick the minor groove of nucleosome-bound DNA [38]. This data must be interpreted with caution due to sequence bias in the digestion preferences of enzymes (see Outlook for 3D genomics below).

In the second category, ChIP- and ChIRP-seq have provided genome-wide maps of DNA-bound proteins and DNA-associated RNAs, respectively (Figure 2C, Table 1) [39,40]. ChIP-seq is usually performed using formaldehyde crosslinking to stabilize the association of proteins with DNA, but native ChIP using MNase to fragment chromatin is also possible for nucleosomes [41] or for TFs in low salt [42]. The quality of ChIP-seq experiments hinges on the specificity and sensitivity of the antibody. Large-scale, systematic efforts to perform ChIP with carefully validated antibodies in multiple cell lines have generated an invaluable resource for the genomics community [43,44]. Histone modification ChIP data are useful for both identification of regulatory elements [45] more generally as an input to computational strategies for determining chromatin states [46,47]. TF ChIP data sets have been fed into TF binding motif databases [48–50] and combined with open chromatin data to create genome-wide maps of presumptive TF binding in cell lines where individual TF ChIP experiments were not necessarily performed [8,30]. STARR-seq, a parallelized fluorescent reporter assay does not assay primary structure per se, has been applied to the Drosophila genome to search genome-wide for sequences that can drive transcription [51,52] and provides a complementary approach to the same problem (Table 1).

ChIRP-seq and a similar method called CHART-seq use short biotinylated oligonucleotide probes that hybridize to an RNA of interest instead of an antibody [53,54]. ChIRP-seq has already been used to show the genome-wide interactomes of lncRNAs involved in PRC2-mediated silencing, telomere function, and dosage compensation [53,54].

Secondary structure

In mammalian genomes, where most nucleosomes are not well-positioned [36,37], precise nucleosome positioning cannot account for regulation of the accessibility of TF binding sites, so chromatin compaction models have been invoked for the regulation of transcription [5]. We define secondary structure as the local nucleosome-nucleosome interactions that integrate many primary structure features such as histone modifications and DNA-bound proteins to regulate the accessibility of DNA, compaction of chromatin, and physical properties of the chromatin fiber.

The proposed 30-nm fiber model of secondary structure [55–58], built from consecutive nucleosomes that collapse into a helix is analogous to a protein alpha helix. The proposed chromatin melt model [59] suggests either zig-zags of interdigitated non-consecutive nucleosomes, analogous to the beta sheet, or a completely unstructured chromatin fiber, analogous to the collapsed random coil. A fourth model of secondary structure is beads-on-a-string uncondensed chromatin, analogous to an unstructured peptide. The first three models are proposed for compact arrangements that preclude most TF binding and transcription activity, while the fourth is believed to be the open structure that facilitates these processes [5].

Currently, no sequencing-based methods can probe the secondary structure of unperturbed chromatin and report anything beyond a description of “open” or “closed” [60]. DNA accessibility assays like DNase-seq and ATAC-seq report on an important effect of secondary structure, but they do not reveal any details about the 3D architecture of the compacted state per se or the mechanisms that maintain it. In the absence of higher throughput and higher resolution methods, the field has patched together the current understanding of chromatin secondary structure from four main approaches.

First, in vitro experiments with extracted or reconstituted chromatin have been used to study the modulation and regulation of chromatin secondary structure by cations [55,57,61], architectural proteins or complexes including linker histone H1 [62,63], HP1 [64,65], MeCP2 [66], and PRC1 [67], chromatin remodelers [68], nucleosome positioning and linker histones [69], and histone modifications, especially acetylation [70] under dilute conditions. Some functional experiments with reconstituted chromatin have also demonstrated how deacetylation and linker histone-dependent chromatin compaction blocks transcription [71]. Most of these in vitro experiments are consistent with a 30-nm fiber model of chromatin [57].

Second, many of the histone modifications and architectural proteins studied in vitro have been mapped onto the genome’s sequence using ChIP and DamID (described below) [72], as have nucleosome positions, which determine the local inter-nucleosome linker length [69]. Third, modeling based on known features of DNA elasticity and nucleosome structure have delineated the space of possible chromatin architectures[73] and shown that there is no unique minimum-energy solution, emphasizing the importance of regulatory factors and electrostatic interactions [58].

Fourth, electron microscopy and x-ray scattering methods have been used to probe gently fixed or frozen nuclei and have shown evidence for 30 nm fibers, but only in metabolically inactive cells [74–76]. FISH-based measurements have shown compaction of chromatin between two loci, but the architecture of the compacted state is unknown [77]. A lack of evidence for 30 nm fibers in the nuclei of normal somatic cells has suggested that the polymer melt model may be more appropriate [3,59,75,78,79]. Resolution of this debate will likely require an understanding of fiber polymorphism and regulation [80–82]. Discrepancies between in vitro and in vivo measurements of chromatin secondary structure likely arise from both the high density of nucleosomes and chromatin binding and remodeling factors found in the living nucleus. As new genomics methods are developed and the resolution of existing methods is increased, global maps of chromatin secondary structure should provide a framework to facilitate integration of these in vivo and in vitro data sets.

Tertiary structure

Tertiary structure in this context comprises long-range three-dimensional chromatin contacts and organization. Mapping the tertiary structure of the genome (Figure 1) is a promising approach to understanding the physical wiring diagram of the genome, linking enhancers to the genes they regulate. Many of the methods for probing tertiary structure are based on proximity ligation, a technology first developed as part of chromosome conformation capture (3C) (Figure 3, Table 2) to map the global 3D structure of a yeast chromosome [12,83,84]. The power of 3C to provide insights into transcriptional regulation was demonstrated with an investigation of looping between DNase I hypersensitive sites their a target gene at the mouse globin locus, which inspired the active chromatin hub model of transcriptional activation [85]. This hub is a pre-existing cluster of enhancers and transcriptional activation during differentiation is associated with target genes looping into the hub [86]. 3C is best used for exploring a single group of enhancers, such as the globin locus, or for validating the results of higher-throughput methods.

Figure 3. Methods that assay the tertiary structure of chromatin.

(A) DamID relies on expression of a fusion between the bacterial adenine methylase Dam and the protein of interest in cells. Dam methylates (yellow) TCGA motifs on any genomic DNA it contacts, and genomic DNA is enzymatically cleaved at methylated restriction sites, the sites are then ligated to adapter oligos (gray) that are complementary to PCR primers. Another restriction digestion cleaves non-methylated sites. The amplification step selects for densely methylated areas, in which methylated restriction sites are adjacent. (B) The chromosome conformation capture family of methods is based on a restriction digest of crosslinked chromatin followed by proximity ligation of the fragment ends in dilute solution. Proximity ligation creates junctions between fragments from genomic regions that interact in 3D. Combining this approach with different strategies for fragmentation, enrichment, and readout of the ligation junctions yields a diverse set of methods. In 3C, PCR with primers on either side of the ligation junction of interest assay a single site. In 4C, the library is cleaved with a second RE and recircularized. PCR primers oriented outward from the test (bait) locus are then used to amplify the unknown regions with which the bait interacts. In 5C, oligos are designed to bridge ligation junctions and are then hybridized to and ligated across the junctions to create “carbon copies” of the junctions that can be PCR-amplified with common primers. Capture-C involves fragmentation of a 3C library then oligo hybridization capture to enrich for all sequence fragments (junctions and non-junctions) from a subset of the genome. Hi-C is a very high-throughput version of 3C that assays all interactions between all genomic loci. Fragment ends are biotinylated (light blue squares) before proximity ligation to facilitate their enrichment for sequencing library construction. Refinements of Hi-C include proximity ligation within crosslinked nuclei rather than in dilute solution, and the use of DNase I fragmentation to replace restriction digest. Because DNase I generates a variety of end types, ends must be blunted and ligated to biotinylated linkers (pink with light blue square) to replace the biotinylation of sticky ends from RE digest. ChIA-PET assays DNA-DNA contacts involving a particular protein by incorporating a ChIP step in the workflow before proximity ligation.

Table 2.

Selected methods for assaying tertiary chromatin structure

Method	Assay type	Protocol description	Resolution	Input	Read- out	Notes and example applications
Chromosome Conformation Capture (3C) [83]	One locus to one locus	Formaldehyde crosslinking, restriction digest, ligation in dilute solution, crosslink reversal	Depends on distance from loci of interest to nearest restriction site; as high as 1 kb.	~108 yeast nuclei or ~107mammalian cells	PCR	Sequence of bait locus and interacting locus must be known. Analysis is straightforward. early work on whole-chromosome 3D conformation in yeast [83] comparing looping of enhancer regions to a promoter of interest between cell types [85] and through stages of differentiation [86]
				~104–107 cells	qPCR [171]	This is a more precise and sensitive approach than quantifying the amount of yield from PCR. measuring quantitative changes in interaction frequency between enhancers through differentiation, after genetic or knockdown disruptions of the looping machinery [172] or during induction of ectopic looping [114] high-resolution validation of interactions identified by Hi-C [94]
4C [87,88,1–73]	One “bait”/”view-point” locus to the genome	3C library is cut with second restriction enzyme, circularized, then PCR amplified with primers facing out from bait locus	Up to 2 kb sliding window (4C-seq); data is usually analyzed at multiple window sizes (“domain-ogram”).	~5 × 105 cells	Microarray or Seq., 1–2 M reads	Sequence of bait locus must be known. Provides the genomic interactome “viewpoint” from the bait locus. baits can be multiplexed for multiple viewpoints in one experiment [174]
Chromosome Conformation Capture Carbon Copy (5C) [89]	Many loci to many loci	3C with ligation-mediated amplification (LMA) of a “carbon copy” library of oligos designed across restriction fragment junctions of interest.	Up to ~1 kb	~107 cells	Microarray [89] or Seq., ~10 M reads per sample [175]	Loci to be assayed limited by complexity of LMA oligo pool high-resolution TAD mapping over a region of interest of (4.5 Mb) [93] systematic mapping of long-range looping interactions between TSSs and putative regulatory regions identified by ENCODE [175]
Capture-C [90]	Multi-plexed one to one in regions of interest	3C library is sonicated and selected loci are captured with oligo hybridization	2 kb sliding window in selected regions of interest	~107 cells	Seq. 200M reads per sample	Oligo capture targeted ~450 promoters most looping interactions happen within ~300 kb of the target promoter and within topological domains [90]. Can be used for large-scale analysis of SNPs on looping interactions.
Dilution Hi-C [91]	Whole genome to whole genome	3C restriction fragment ends are biotinylated before proximity ligation. Library is sheared and biotinylated junctions are pulled down and sequenced.	0.1–1 Mb [91], 40 kb [92], 5–10 kb [94]	~107 cells	Seq. 8.4M[9–1], 3.4 B [94] reads per sample	This is the original Hi-C method. It maps megabase-scale topological domains in mammalian genomes. Limited by sequencing depth and frequency/distribution of restriction sites in genome. Biases are due to chromatin openness, sequence composition and fragment length. mapped topologically associated domains (TADs) in mouse and human genomes [92], whose borders are conserved and occupied by CTCF binding sites and housekeeping genes showed that genome-wide, cell-type-specific looping interactions exist prior to induction of transcription and potentiate transcription after induction via TNF-alpha signaling [94] characterized the distances between putative enhancers and promoters and the numbers of looping partners observed for each [94]
DNase Hi-C [95]	Whole genome to whole genome	Crosslinked DNA bound to beads is fragmented by DNase I instead of restriction enzymes, then a biotinylated adapter is ligated before proximity ligation of molecules embedded in agarose.	1–50 kb or 1 Mb	2–5 × 106 cells	Seq., ~100–800 M reads per sample	Resolution does not depend on locations of restriction enzyme sites – only limited by sequencing depth, and thus can obtain better coverage than restriction enzyme Hi-C with deep sequencing. High-resolution data was achieved using targeted amplification to focus on lincRNA promoters. [95]
In situ Hi-C [13]	Whole genome to whole genome	The early steps of Hi-C, including restriction digest and proximity ligation, are performed in permeabilized but otherwise intact nuclei, without dissociating chromatin.	1–5 kb	~2–5 × 106 cells	Seq., 112 M – 3.6 B reads per biological replicate (primary experiment)	Higher resolution and better signal to noise than dilution Hi-C. These relatively recent experiments also include much higher sequencing depth than most previous Hi-C studies. showed that many loops are mediated by CTCF bound to pairs of motifs in a convergent orientation [13] classified genome domains defined by stable loops into chromatin states[13] haplotype-specific mapping of imprinted loci and inactivated X chromosome conformation [13] using SNPs in Hi-C reads
Single cell Hi-C [133]	Whole genome to whole genome	Same as in situ Hi-C, but single nuclei are first isolated and processed independently to add unique barcodes before amplification and sequencing	~10 Mb (single cells) – 1 Mb (pooled cells)	1–60 cells	Seq., ~10 K read pairs per cell	The first ligation-based method that can provide single cell resolution for looping interactions (which is otherwise obtainable by FISH measurements) showed that TADs are stable and reproducible, but inter-TAD and inter- chromosome contacts are stochastic [133]
ChIA-PET [103]	Whole genome to whole genome mediated by protein of interest	Chromatin is crosslinked and sheared by sonication and enriched by ChIP against a protein or histone modification of interest. Biotinylated half-linkers are ligated to ends of DNA fragments, followed by proximity ligation of fragment ends. An RE is used to cleave ~20 bp on either side of linker, then these paired-end tags are selected and sequenced.	Depends on read depth and size of genome region bound by protein of interest. Ultimately limited by size of sheared chromatin fragments. ~3kb fragments counted as self-ligation [103]	~108 cells	Seq., ~20–30 M read pairs per sample [103], 100–200 M read pairs per sample [106]	Restricts looping analysis to loops mediated by a protein of interest, such as CTCF Used to study chromatin modification patterns in CTCF-mediated loops [104], loops induced upon estrogen signaling and mediated by estrogen receptor alpha [103], loops involving RNA polymerase [105], and a combination of proteins that regulate genome architecture [106]
DamID [72,176]	Protein to whole genome	Dam is fused to a protein of interest and expressed in the cell. It methylates adenines at CATG sites in proximity to the protein of interest, which are then detected by methylation-specific PCR.	~1 kb	~106 mammalian cells (~2.5μg genomic DNA) [177]	Methylation-specific PCR and [72] microarray hybridization [177]	Can provide similar information to ChIP, but the interaction happens in live cells over a long time window. used to map the interactions of chromosome with nuclear bodies, a major one being the lamina; used to define LADs [108] also used to map DNA-protein interactions, e.g. polycomb group proteins [178] not suitable for mapping histone modifications [177]

4C, 5C and Capture-C methods (Table 2, Figure 3B) are variants of 3C that allow 3D looping interactions to be assayed in parallel at multiple sets of loci [87–90]. They have proven useful in mapping either a region of interest, such as an extended locus with multiple promoters and enhancers that have developmentally regulated interactions, or in surveying hundreds of promoter regions genome-wide. In many of these approaches, the design of PCR primers or capture oligos can also select particular SNPs, probing a single allele [90].

The highest throughput version of 3C is Hi-C, which has been used to systematically map pairwise interaction frequencies genome-wide (Table 2, Figure 3B). Its resolution is limited by the size of the restriction fragments used and by sequencing depth. Initial results were binned to a resolution of ~1 Mb in order to overcome the challenge of low genome coverage of sequenced ligation junctions [91]. At 1 Mb resolution, the human genome appears to partition into multi-megabase domains that belong to either of two effective nuclear compartments, termed A and B, which roughly correspond to active and inactive chromatin [91]. With deeper sequencing, it was possible to resolve topologically associating domains (TADs), regions spanning hundreds of kb to ~1Mb defined by the higher likelihood that two loci lying in the same TAD will interact with each other rather than with loci in a different TAD [92,93]. The resolution of Hi-C has been pushed yet further to reveal sub-TAD structure genome-wide by employing very high sequencing depth combined with careful filtering of the data [94], by replacing RE digestion with DNase I fragmentation to remove the restrictions imposed by restriction site frequency in the genome [95], and by using in situ proximity ligation that helps to reduce nonspecific ligations in combination with extremely deep sequencing [13].

The 3C family of methods has rapidly advanced our understanding of 3D genomic architecture. Interestingly, a significant portion of TAD borders are conserved between species [92] and between cell types during differentiation [93], and their boundaries are not significantly altered after stimulation of major cell signaling pathways [94], suggesting that TADs may make up a basal architecture of mammalian genomes onto which finer-grained regulatory interactions are overlaid. TAD borders contain binding sites for the ubiquitous CTCF [13,92,93], which has been associated with an insulator function that blocks communication between enhancers and promoters when placed between them and prevents the spread of heterochromatin in reporter systems [96–98]. Expression profiles of genes in the same TAD are correlated during differentiation [93] and CTCF-bounded TADs can be roughly clustered by the sets of genomic loci with which they interact [13]. TADs are also concordant with replication timing domains [99], further supporting their role as functional modules of the genome. High-resolution 5C experiments that focused on sub-TAD structure using existing ChIP data observed that long-range 3D interactions are hierarchical [100]. CTCF and cohesin-mediated long-range interactions make up the TADs that do not change appreciably as cells differentiate, whereas developmentally regulated, shorter-range contacts within TADs are associated with cohesin and the Mediator complex (a component of the transcription initiation machinery [101]). Interestingly, high-resolution Hi-C has shown that enhancer-promoter contacts are cell type specific, but at least in some cases, can be detected before transcriptional activation by a cell signaling cascade is induced, suggesting that in addition to constitutive TADs that are shared across cell types, an additional layer of cell-type-specific 3D interactions exists to potentiate the transcriptional activation of certain genes in response to exogenous signals [94,102].

The 3C family of methods focus only on DNA-DNA contacts, and although they can be used to study the protein factors that regulate the genome’s tertiary structure in combination with ChIP, more direct methods for understanding the role of particular proteins in shaping the tertiary structure of chromatin have also been developed. ChIA-PET, like 3C, involves proximity ligation, but also includes a ChIP step (Figure 3B) [103]. ChIA-PET has been applied to diverse targets including the looping interactions mediated by estrogen receptor alpha [103], CTCF-mediated loops [104], RNA polymerase II [105], and a combination of key architectural factors [106].

All of the proximity ligation methods discussed above suffer from the requirement that chromatin be formaldehyde-crosslinked, which can distort results [107]. DamID is a method for studying protein-DNA interactions through proximity methylation, and although it can be used to map primary structure, such as chromatin-bound proteins [72], it has proven particularly powerful in identifying large-scale domains associated with particular proteins that demarcate nuclear landmarks, such as the nuclear lamina [108] (Figure 3A). A major strength of DamID is that it involves proximity methylation in live cells rather than proximity ligation in fixed chromatin (Figure 3A) [72]. It has been used to identify lamin-associated domains (LADs), which consist largely of late-replicating, transcriptionally inactive chromatin and have significant overlap with inactive TADs [92,108]. TADs have been observed to switch from LAD to non-LAD status during development [93]. The level of nuclear organization probed by DamID and inter-chromosomal contacts observable by Hi-C or 4C, might be thought of as very large-scale tertiary structure, or, because they involve multiple chromosomes and chromosome territories, they might be likened to quaternary structure by analogy to multimeric proteins. Exciting work in understanding the interactions between primary and tertiary (or quaternary) structure of the genome has involved dissection of the relationship between nuclear periphery tethering, chromatin state and transcription, and is discussed in the next section.

How do the three scales of chromatin architecture interact?

Strong correlative evidence supports interactions between primary and tertiary structure. TAD boundaries are correlated with primary structure elements such as CTCF binding sites [13,92], open chromatin, housekeeping and tRNA genes, and high gene density in general [92]. The link between histone modifications and 3D interaction domains is less clear [97]. On one hand, the borders of histone modification domains often coincide with the borders of TADs and are established by some of the same factors that delineate TADs, such as insulators and highly transcribed genes [91,92,97,109]. TADs can be clustered by epigenetic state [13], and a combined Hi-C/ChIP/RNA-seq approach has also shown that both gene expression and epigenetic marks that change in response to hormones do so in a consistent manner across TADs [110]. On the other hand, TAD boundaries and chromatin modification boundaries are separable, as illustrated by H3K9me3 domains in differentiated cells: TAD boundaries are present in both pluripotent stem cells and differentiated cells, preceding the appearance of H3K9me3 domains in differentiated cells [92]. Together, these data support the idea that three-dimensional interaction domains can not only coordinate transcription, but can also partially delimit the diffusion of histone modifying factors [111].

As a complement to genome-wide correlative data, artificial tethering (reviewed more extensively in [112]) has been primarily studied at the β-globin locus. Ectopic 3D contacts are induced by insertion of insulator elements [113] or by dimerization of transcription cofactor Ldb1 at an endogenous enhancer in the Locus Control Region of the β-globin locus with an artificial zinc finger-Ldb1 fusion (or a fusion with the Ldb1 self-association domain) targeted to a chosen promoter. These forced contacts induce transcription of the corresponding beta globin gene in the absence of its endogenous activator, GATA1 [114], or even in opposition to the developmental program that normally shuts off transcription of fetal globin genes in adult tissues [115]. Interestingly, the induction of fetal genes in adult cells by forced chromatin contacts was only achieved in human cells, and not in mouse, which may be due to a difference in the permissiveness of the chromatin environment [115]. It remains to be seen how broadly results from forced contacts at the β-globin locus will translate to other enhancer-promoter interactions. A recent paper by Therizols et al. touches on all three scales, demonstrating that transcriptional activation via a synthetic TALE (transcription activator-like effector) based transcription factor repositions a gene from the nuclear periphery to the nuclear interior and leads to local chromatin decondensation [116]. By replacing the transcriptional activator domain of the TALE fusion protein with an acidic peptide previously shown to decondense chromatin in vivo, the authors functionally separate chromatin decondensation from transcription and show that although chromatin decondensation is sufficient for nuclear repositioning, it is not sufficient to shift late-replicating chromatin to an earlier replication time, which requires transcription. These results echo the separability of nuclear lamina tethering and transcriptional repression observed by Towbin et al. in C. elegans [117]. The converse experiment, involving forced tethering of reporter loci to the nuclear periphery, has shown that there is likely feedback between nuclear positioning and transcription because some (but not all) genes are repressed when relocated to the periphery [118]. Together, these data suggest that the relationships between transcriptional activation, chromatin structure, nuclear organization, and DNA replication are more complex than initially envisioned in classical models of open and closed chromatin.

Outlook for 3D genomics

Looking forward, several technological developments in the study of primary structure appear ready for exciting applications. First, sequencing technology is being pushed toward longer read lengths that can be used to map the repetitive regions of the genome and structural variation [119]. New bioinformatics methods are also being applied to model repeat structure in centromeres, one of the most fascinating examples of epigenetic modification and specialized chromosome architecture [120].

Second, the refinement of whole-genome transcription factor footprinting assays makes it possible to assay the occupancy of hundreds of TFs in parallel, assuming that the binding motifs of those TFs are known. To that end, the careful quantification of sequence bias of the nucleases used in footprinting will make this approach more feasible and reliable [121–123]. Footprinting signals due to sequence bias can range over several orders of magnitude, so novel TF binding motifs identified by footprinting should be verified by orthogonal methods, and new footprinting methodologies with lower or more manageable sequence bias should be pursued. Sequence bias of nuclease cleavage is also a problem in the currently available maps of nucleosome positions. An alternative method with great potential is chemical mapping of nucleosomes, which uses histone mutants with a unique cysteine on histone H4 to localize a free radical source at the nucleosome dyad and cleave DNA at stereotyped positions around the dyad [124,125]. Chemical mapping has base pair resolution and lacks nuclease-dependent sequence bias, but because of the requirement for a histone mutant, it has only been used in yeast. Advances in genome editing or novel methods for highly targeted coupling chemistry may make it possible to extend chemical mapping to metazoan genomes, revealing patterns of nucleosome spacing that can provide important clues about higher-order architecture [69,73,124]. Lastly, primary structure data quality from many methods will be vastly improved by precise classification and definition of cell type using single-cell RNA-seq [126] followed by chromatin profiling in well-defined cell populations, which will likely require low-input methods (reviewed in [15]). Such low-input primary structure mapping methods can also reveal the extent to which features like histone modification domains vary from cell to cell or over time. Although these measurements are extremely noisy, useful information can be gleaned from mapping the cell-cell variance in the signal (such as methylation) over the genome and in different cell types [15].

The gap in the multi-scale map of genome organization that currently exists at the secondary structure scale likely must be filled through the development of new methods. Some information may come from 3C family methods as their resolution continues to improve, through innovations like using DNase I digestion to randomly cleave fragments instead of relying on REs [95]. As sequencing costs continue to drop, the fundamental genome coverage and resolution limit of Hi-C will be dictated by the positions of restriction sites, which must be carefully accounted for in Hi-C analysis because they are not random [13,91,127]. DNase I can be used to generate almost arbitrarily short fragments for better genome coverage than in RE-based Hi-C [95]. When combined with polymer modeling, high-resolution proximity ligation methods may be able to discriminate between compaction along the chromatin fiber, as would be the case in regions that adopt a 30-nm fiber, and compaction based on interdigitation of nucleosomes from a polymer melt [59]. However, proximity ligation alone will probably not provide the base pair or near-base pair resolution that is necessary to infer the three-dimensional architecture.

As the resolution of methods for mapping tertiary chromatin interactions has improved, the challenge of analyzing, visualizing and interpreting Hi-C data has become more significant. For example, at 5 kb resolution, the contact matrix contains 20 billion pixels and custom paralleized code for graphic processing units was required to analyze these immense data sets [13]. The resulting two-dimensional contact matrices also do not fit the track-based nature of standard genome browsers. To aid in interpreting and visualizing data sets from 3C/Hi-C/ChIA-PET (see below) and related experiments, Aiden and colleagues have created a software package called Juicebox [13], adding to the pre-existing suite of R tools available from a collaboration of groups [128]. In a more generally accessible format, visualization of 3C and related data in a genome track-like format is possible thanks to new features in the WashU Epigenome Browser [129].

Interpretation of 3C and Hi-C interaction frequency maps is complicated by two primary factors. First, proximity that is captured by ligation can arise in multiple ways, including direct protein-mediated association or co-association with the same subnuclear body [130–132]. Second, most proximity ligation experiments average over the chromosome conformations of several million cells, and it is not clear whether a Hi-C map represents a series of stable 3D contacts that are present in all cells of a population, or the average of stochastic contacts that can vary dramatically between cells. The recently developed single-cell Hi-C [133] showed that single cell contacts were mostly consistent with TADs mapped in ensemble measurements, but inter-chromosome or inter-TAD contacts varied dramatically between cells. However, single cell data are by necessity sparse and at very low resolution. A different and potentially more fruitful approach was adopted by Giorgetti et al., who used Monte Carlo modeling of polymer representations of the X chromosome to simulate the possible conformations that are consistent with their 5C data, assuming that proximity ligation frequencies represent contact frequencies in a cell population for any two loci [134]. When compared with 3D fluorescence in situ hybridization (FISH, an orthogonal, single-cell method) data, the model successfully predicted the effect of deleting a region at the TAD boundary. This work provides strong evidence that TADs are fluctuating entities present as a compact domain in only a subset of any cell population, whose interactions are shaped by transient looping interactions between key loci. Concordantly, TADs mostly disappear in condensed mitotic chromosomes, but can be reestablished in the next interphase [135]. Giorgetti et al. also showed that maintenance of intra-TAD interactions helps to exclude inter-TAD interactions and, using RNA-FISH, that transcriptional activation is not always correlated in a simple way with the transient level of TAD compaction [134].

The emerging picture of the genome’s tertiary structure is dynamic, which is not surprising, considering that many trans acting factors undergo fast turnover [136,137]. A 4D view of the genome is therefore needed, with temporal information collected together with high-resolution spatial information both along the chromatin fiber and in 3D interaction space. This acute need is reflected in a recent initiative from the NIH to fund research on new methods to probe the 4D Nucleome [138]. Fluctuations can be studied by surveying variation between cells using single-cell and low-input methods [15,133,139] as well as simulations [134], but temporal information, in both ensemble and single cell experiments, is necessary to help distinguish causal relationships. Time courses are a powerful and important tool in the study of mechanisms of transcriptional activation, when combined with perturbations of chromatin state and chromatin interactions (Box 1), and they have already yielded insights into the nontrivial relationship between enhancer-promoter contacts and transcriptional activation [94,112,114]. At lower throughput but higher temporal resolution, promising approaches include high-resolution imaging of fluorescent proteins tethered to lac or tet operator arrays [140,141], the “molecular contact memory” method, which uses DamID and an inducible fluorescent fusion protein that specifically binds the methylated DNA [142], and kinetic measurements of chromatin modification and transcription using labeled antibodies and fluorescence microscopy [143].

Box 1. Chromatin Perturbation Methods.

As models are generated to explain the data emerging from chromatin structure mapping experiments performed using the methods described above, a growing number of tools are now available to test these models by actively perturbing chromatin structure and 3D genomic interactions, adding to the classical biologist’s toolbox of genetic perturbation in amenable model organisms, such as mice, flies and yeast, the now-established genome editing methods based on CRISPR/Cas9 and RNAi-based knockdowns of protein factors or RNAs.

• Advances in artificial chromosome design in both humans and yeast suggest that large-scale in vivo transcriptional regulation experiments with total control over DNA sequence may be possible in the future [154,155].

• Chromatin modifiers can be ectopically tethered to loci of interest by fusion to DNA binding domains from the bacterial lac or tet repressor or yeast GAL4, as long as their DNA binding sequences can be inserted into the locus of interest [156–158]. Reversible tethering that uses small-molecule mediated interaction between the DNA binding domain and the chromatin modifier has also been implemented to study chromatin modification dynamics [159].

• Transcriptional regulators can be targeted directly to chromatin using domains that recognize particular chromatin modifications, such as H3K27me3 [160].

• The CRISPR/Cas9 system and TAL effector domains have been developed as a platform for designing artificial transcriptional activators or repressors targeted to any sequence [161–163].

• Targeted chromatin decondensation by artificial transcriptional activation or tethering of a chromatin decondensing peptide to induced relocation of a locus from a heterochromatic to a euchromatic compartment of the nucleus shows that perturbations can be achieved at multiple architectural levels simultaneously, due to the multi-scale nature of genome regulation [116].

• Artificial zinc finger proteins have been used to impose forced looping interactions, revealing an early glimpse into the causality relationship between looping and transcriptional activation [114,115].

• The spatial organization of the nucleus can be altered by tethering of chromosome-integrated lac operon arrays to the nuclear periphery by peripheral protein-lacI fusion proteins [118].

The role of noncoding RNAs in 3D genome biology, reviewed extensively elsewhere [144,145], is an active area of investigation. For example, recent work has shown that RNAs can help to endow protein trans-acting factors with sequence specificity [146]. 3D chromosome architecture can affect ncRNA function [147] and ncRNAs conversely play a role in maintaining 3D interactions [148].

The future of genomics holds exciting opportunities to build on the existing infrastructure of chromatin modification and 3D contact maps by exploring chromatin dynamics and cell-cell or cell-type variation in genome architecture. The interface between genomics and biochemistry also appears ripe for breakthrough. Although the biochemistry community has expressed skepticism about the usefulness of large-scale epigenetic mapping efforts [149], some vindication may lie in the growing capability for hypothesis-driven methods using classical genetic or biochemical perturbations and tools described in Box 1, as well as in powerful approaches that combine genomics-based identification of factors that regulate genome architecture with in vitro characterization of their binding characteristics, such as a series of recent papers from the Lee and Cech labs that explore the interaction of the PRC2 complex and CTCF with RNAs [150–153]. We hope that such interdisciplinary, multi-modal approaches for untangling the puzzle of genomic architecture will serve as archetypes for future work.