Chromatin structure has long been known to affect tissue-specific transcriptional regulation [11,12]. Figure 2 illustrates the familiar organization of DNA in the genome. Briefly, the double helix wraps around histone protein octamers to form nucleosomes that then, with additional scaffold proteins, form higher-order 30 nm fibers. Transcriptionally silenced regions are generally packaged into tightly-packed heterochromatin. Actively transcribed genes typically remain in the more loosely-packed euchromatin, where DNA is more accessible to the transcriptional machinery. Within euchromatin, some stretches of DNA are less associated with histones. These "unwrapped" regions are referred to as open, accessible, or nucleosome-depleted regions. Nucleosome-depleted regions can interact with DNA-binding proteins, which can then regulate nearby chromatin structure and gene expression.

Thus, one avenue for chromatin structure to affect transcriptional regulation is through open chromatin being bound by sequence-specific transcription factors. Scientists have been studying regulatory elements using experiments that locate open chromatin since the 1970s [13]. However, despite the realized importance and interest in studying regulatory DNA, initial studies of the human genome sequence instead focused on protein-coding genes. This is partly due to difficulties with noncoding DNA: Identifying and assigning functions to regulatory elements is complicated by several issues.

What makes studying regulatory elements so difficult? First is the sheer number of elements; there are far more than genes. Our initial estimate of 30,000 to 40,000 genes in the human genome [4] has more recently been reduced to 20,000–25,000 [14]. In contrast, the number of proposed regulatory elements currently stands in the millions and continues to rise [15,16]. This complexity has given rise to the term “regulome,” symbolizing the growing set of regulatory components. The fragmentation and volume of regulatory sequence (perhaps one-third of the genome vs. 2% being coding regions) makes it difficult to study [16].

Second, they are difficult to find. Unlike protein-coding genes, which have a rigid genetic code, regulatory elements seem to lack such a rigid code and are thus more difficult to identify computationally. Where regulatory elements do follow patterns, such as conforming to transcription factor binding motifs, they are more elastic, which has led to more sequence variation in regulatory elements than in protein-coding genes. Sequence conservation yields some clues (e.g., [17]), but active functional elements may not be under detectable evolutionary constraint [18,19]. For this reason, we must currently rely mainly on experimental methods to identify regulatory elements.

Third, after identifying regulatory elements, further difficulties arise in determining their targets [20]. Regulatory elements can act both directly and indirectly to modulate transcriptional levels. For direct (or cis-) regulation, target genes are not obvious because an element does not necessarily regulate the single nearest gene; instead, they can act at large distances [21], skip genes [22], or affect multiple genes [23]. Elements can also act indirectly (in trans-) to affect genes en masse by regulating a transcription factor with many targets. These layers form the robust system required to thrive in a changing environment, but they are not easy to dissect and explain.

Fourth, regulatory elements have a variety of functions [24]. For example, there are distinctions between promoters, enhancers [25], enhancer blockers [26], insulators [27], LCRs [28], and Polycomb-bound silencers [29]. One step further, within each category there may be additional subdivisions, such as the distinction between TATA-box vs. CpG-rich promoters [30] or enhancer sub-classes [25]. Each performs a different but necessary purpose, but the sequence alone is currently unable to classify the elements by type. Computational research continues to identify subtle sequence patterns [31], but our understanding of sequence remains a key limitation.

Finally, perhaps the most important challenge to studying transcriptional regulation is that the regulome is dynamic. The human regulome varies in many dimensions, such as age, environment, developmental time, cell-cycle stage, and tissue type. In contrast, the genome is essentially constant and identical in each cell and tissue type. To identify every regulatory element in the human genome is a currently infeasible task; it would require interrogating every cell-type under every possible developmental stage in any environment and against all genetic backgrounds.

Despite these challenges, considerable recent progress has been made toward identifying and characterizing regulatory elements. Several of these difficulties have only started to become tractable in the past few years due in large part to technological improvements in sequencing and computation, which have driven new discovery in almost every biological field. In the study of transcriptional regulation, the major genome-wide findings have been primarily driven by results from chromatin accessibility and transcription factor experiments that assay regulatory elements.

Three common experimental techniques used to assay chromatin structure and identify regulatory elements are DNaseI hypersensitivity (DHS), formaldehyde assisted identification of regulatory elements (FAIRE), and chromatin immunoprecipitation (ChIP). These assays each have strengths and weaknesses.

The working hypothesis of the past several decades has been that open chromatin identifies regulatory regions. Genome-wide results now confirm this finding: Thurman et al. [16] showed that the DNaseI-seq profile recapitulates the sum of TF ChIP-seq signals. In K562 cells, with ChIP-seq results available for more than 42 factors, the correlation between the cumulative ChIP results and DNaseI-seq is very high (~0.8). Almost 95% of known ChIP-seq peaks are in regions identified as open by DNaseI assays. This result shows that open chromatin is a reasonable proxy for generic TF binding and highlights the utility of open chromatin assays.

The distribution of DNA in the genome relative to known transcribed regions is shown in Figure 3, and compared to the distribution of regulatory elements. The vast majority of regulatory elements are noncoding, with about 5% identifying known promoter elements, and the other noncoding elements almost evenly split between intergenic and intronic sequences [16]. There are some rare exonic open chromatin regions, which could regulate splicing [50], or they may overlap intronic elements. The clear emphasis on distal regulation (most elements are intergenic or intronic) supports the growing realization that distal elements are a key source of phenotypic complexity [25].

Using a sample of 126 cell types, Thurman et al. [16] reported that nearly one-third of the genome shows sensitivity to DNaseI digestion with ~15% of the genome being DNaseI hypersensitive. These percentages differ depending on the threshold used to define hypersensitivity, but they provide an idea about the overall proportion of the genome that is sensitive to DNaseI. These numbers should be considered minimums, as they are restricted to the cell-types, environments, and developmental stages assayed. Figure 4 shows how the number and percentage approaches saturation with increasing cell-types. This trend is not simply a result of false-positive DNase-seq signal, which is estimated to be less than 0.5% [16]. The numbers will grow as new cell-types and contexts are assayed. Without testing all cell-types and contexts, the exact percentage of regulatory DNA in the human genome cannot be determined, but these results demonstrate that it may be quite high. Ultimately, these regions will have to be tested to establish function.

With open chromatin data from a variety of cell-types, it is possible to classify regulatory elements as cell-type-specific, shared among multiple cell-types, or ubiquitous. This cell-type-specificity classification necessarily depends on the cell-types used. Many elements currently considered cell-type-specific may turn out to be present in several related but not yet assayed cell-types. Nevertheless, highly cell-type-specific sites are likely to retain that feature, though it may be adjusted from “cell-type-specific” to “narrow-cell-lineage-specific.”

Gaulton et al. [51] identified thousands of pancreatic-islet-specific FAIRE sites. Song et al. [35] used both FAIRE and DNaseI to show similar results in seven other cell-types. These results have illustrated that each cell-type is likely to have a unique set of specific open chromatin regions that guide specific cell fates and functions, in addition to those shared by other cell-types. Both of these studies made two additional observations in relation to cell-type-specificity: First, the cell-type-specific DHS sites were associated with cell-type-specific gene expression; and second, cell-type-specific regulatory elements tended to cluster with respect to genomic location.

Most recently, Thurman et al. [16] showed that the majority of DHS sites are found in relatively few cell-types, with the distribution depending on the genomic context of the DHS site. For example, promoter DHS sites were more likely to be ubiquitous rather than cell-type-specific.

One way a cell can affect cell-type-specific expression is by creating a cell-type-specific chromatin conformation. It is now becoming clear that a large class of regulatory elements is involved in establishing cell-type-specific chromatin structure. For example, CTCF sites have many roles affecting chromatin accessibility [52]; CTCF is the canonical insulator, but it also can create both active and repressive loops [53]. Recent studies have shown that other factors also work by altering chromatin structure. Biddie et al. [54] showed that DNaseI-hypersensitive sites, specifically with AP1 binding sites, predefine binding sites for GR receptor binding in glucocorticoid cells. Along the same lines, Shibata et al. [55] associated AP1 motifs with DNaseI signal differences across primates. Chromatin looping is also likely to be regulated by other factors, such as mediator, p300, and cohesin, which work together to establish cell-type-specific chromatin structure [56]. These results collectively illustrate the important interaction between transcription factors and chromatin structure.

A connection between transcription factor binding and chromatin structure represents one way a cell could regulate gene expression. Recently, Degner et al. [57] drew the connection between chromatin and expression in a study that used matched DNaseI-seq, genotype, and expression data from a common cell-type in 70 individuals. They identified about 9,000 sites where DNaseI signal correlated with genotype, and further showed that, in many cases, the DNaseI difference also correlates with gene expression. In a cross-species comparison, Shibata et al. [55] showed that DNaseI differences among primates are closely tied to expression differences. These results illustrate how this data may be able to inform computational models to predict gene expression, which has been an area of interest in other organisms [58,59,60]. Along these lines, Natajaran et al. [61] built a model to predict gene expression from motif analysis and DNaseI signal. They showed that performance improves when distal elements are included in the model, highlighting the relevance of distal regulatory sites. These results are beginning to unravel the complex interactions between chromatin structure and gene expression.

One of the most important ways to annotate a regulatory element after discovery is to determine what factors bind there. The primary experimental tool for such an annotation is ChIP. In the ENCODE project, data for hundreds of ChIP experiments have already been made available on the UCSC genome browser [47]. These data enable us to identify the frequency and importance of individual factors [62,63,64,65], as well as explore cooperativity among factors [66]. DNaseI footprinting has also been used to propose what factors bind an element [36,37]. These analyses require not only experimental data, but appropriate computational algorithms to extract meaningful signal.

Regulatory elements can also be divided into functional classes (i.e., silencers, promoters, enhancers). Some classes, like promoters, are relatively straightforward to positionally and functionally define: they lie just upstream of genes and operate on the adjacent gene. Others, like distal enhancers, seem to be anywhere and act on anything. In general, all functional classes share the one major property that they tend to be found in regions of open chromatin. However, they vary in other properties, such as TF binding and histone marks [67]. For example, insulators are often bound by CTCF, whereas enhancers are commonly characterized by certain histone marks (H3K4me1 and H3K27ac) or cofactors (p300) [39]. Quantifying histone marks therefore gives us a relatively straightforward way to provide an initial annotation of regulatory elements genome-wide.

The ENCODE Consortium and others have been performing ChIP-seq to identify multiple histone modifications in several human cell-types. Heintzman et al. [39] showed that enhancer histone marks are more often cell-type-specific, in contrast to promoter and insulator marks that are fairly consistent across cell-types. As histone mark data have become available from more cell-types, it has become possible to extrapolate to elements of unknown class and improve current genome annotation. On the basis of this type of data, computational researchers are designing algorithms to characterize regulatory elements. For example, Ernst and Kellis [68] designed a Hidden Markov Model to assign categories to genomic elements on the basis of their histone modification signatures. Lee et al. [31] trained a Support Vector Machine to predict mammalian enhancers from genomic sequence and validated the predictions with ChIP and DNaseI data. Classifying regulatory elements by function will be a vital step in understanding transcriptional regulation.

A complementary way to characterize regulatory elements is by cell-type specificity, rather than by experimental signature. Thurman et al. [16] used a self-organizing map (a machine learning method) to classify regulatory elements based on their open chromatin signal pattern across all cell-types. They showed that regulatory elements cluster into a limited number of similar patterns. These kinds of classifications will help us derive more meaningful annotations of regulatory elements. By grouping elements by patterns, we may be able to leverage information across open chromatin sites to better annotate how they act. For instance, we have discovered that many groups of DHS sites with similar patterns of accessibility across cell types share common transcription factor motifs suggesting similarly bound factors.

Open chromatin experiments have clearly made progress toward their primary goal to identify regulatory elements, but identifying them is only the beginning. Regulatory elements are dynamic; they drive expression specific to tissue, developmental stage, genetic background, and environment. Distal elements in particular often act in very specific contexts. To decode the human genome, regulatory elements must not only be identified, but also characterized by context specificity. We have described some initial efforts and results in these endeavors; for example, the open chromatin experiments described in this review help us characterize regulatory elements by cell-type. By combining this with ChIP for histone modifications and transcription factors, we can hypothesize the general function of each element; however, we have only begun to assign precise functions. To reveal developmental specificity of an enhancer requires a developmental experiment, such as an embryonic assay in non-human model organisms, accomplished by tethering enhancer sequences to basal promoters [69]. These assays are currently restricted to only certain classes of regulatory elements, and this is an area of active research. In addition, we have little data for environmental context and allelic background [70]. Collecting data from all these sources of variability is a challenge, but to completely understand even a single element, we must characterize it in each context.

Another key is the question of what genes a particular regulatory element affects. Neither open chromatin, nor ChIP, nor even embryonic assays directly answer this question. The current best methods to reveal gene targets are chromatin conformation capture [71,72] or ChIA-PET [73] experiments. These experiments examine the physical proximity of promoters to regulatory elements. This idea is based on the mounting evidence that physical proximity is associated with regulation [74]; however, its precise importance remains unclear [75]. In addition, experiments that assay the nuclear organization of the genome remain expensive, time-consuming, and difficult to interpret [76]. A potential alternative is to leveraging information across cell-types to correlate cell-type- and developmental-specificity of distal elements with promoter elements [16] or gene expression [77]. However, this method provides only indirect evidence for regulation. Additional research and new methods will be necessary before we will be able to accurately identify the targets of regulatory elements.

Although we still have much to learn, the past 40 years have seen phenomenal advances in our understanding of how the human genome works. These advancements have been driven by the combined effort of both experimental and computational research. Already, there are clear benefits and insights derived from our study of transcriptional regulation, as well as other levels of gene regulation, including mRNA splicing, dispersion, and decay. As we improve our annotations of regulatory elements, our goal to convert this data into information will be realized, ultimately to find the function (if present) of what we once called junk DNA.