Absence of a simple code: how transcription factors read the genome

Abstract

Transcription factors (TFs) influence cell fate by interpreting the regulatory DNA within a genome. TFs recognize DNA in a specific manner; the mechanisms underlying this specificity have been identified for many TFs, based on three-dimensional structures of protein-DNA complexes. More recently, structural views have been complemented with data from high-throughput in vitro and in vivo explorations of the DNA binding preferences of many TFs. Together, these approaches have greatly expanded our understanding of TF-DNA interactions. However, the mechanisms by which TFs select in vivo binding sites and alter gene expression remain unclear. Recent work has highlighted the many variables that influence TF-DNA binding, while demonstrating that a biophysical understanding of these many factors will be central to understanding TF function.

Questions at the interface of genomics and structural biology

After decades of research, much is now understood about how transcription factors (TFs) recognize their cognate DNA binding sites in the genome to initiate gene regulatory functions. However, potential target sites of each TF occur many times in the genome. How proteins can very precisely identify their functional binding sites in a cellular environment has not been resolved. Although closely related proteins are known to bind to distinct target sites to execute different in vivo functions, the mechanisms by which paralogous TFs select very similar, but not identical, target sites are not understood. Current knowledge on the DNA binding specificities of TFs is largely derived from research in genomics and structural biology, two fields of research that have developed along parallel lines with limited interactions and have only just begun to become integrated.

Recent studies have focused on the question of how TFs recognize a subset of putative DNA binding sites (Figure 1A) by identifying features, beyond the sequence of the core binding site, which contribute to TF-DNA binding specificity [1–4]. Several features contribute to TF-DNA readout on multiple levels (Figure 1B), including the nucleotide sequence [5–11], three-dimensional structure and flexibility of TFs and their DNA binding sites [12–15], TF-DNA binding in the presence of cofactors [1, 16], cooperative DNA binding of TFs [12, 17–19], chromatin accessibility and nucleosome occupancy [20–25], indirect cooperativity via competition with nucleosomes [26, 27], pioneer TFs that bind nucleosomal DNA [28, 29] and DNA methylation [30]. Additionally, interactions exist among all of these factors, which might alter binding in a cell type-specific manner [29, 31]. Many comprehensive reviews [8, 32–48] have discussed these different aspects of TF-DNA binding specificity, often from either a genomics or structural biology perspective. This review attempts to integrate what has been learned at the various scales from studies by these two complementary approaches, and discusses the important progress that has been made in recent years.

Figure 1. Structure-based illustration of multiple levels of TF-DNA binding specificity.

(A) The basic helix-loop-helix (bHLH) Mad-Max heterodimer (PDB ID 1nlw) binds to only a subset of putative binding sites (blue). Some TFBSs are inaccessible due to nucleosome formation (PDB ID 1kx5), while other accessible TFBSs are not selected by the TF.

(B) Higher-order determinants of TF binding include cooperativity with cofactors (e.g., Hox-Exd heterodimer; PDB ID 2r5z), multimeric binding (e.g., p53 tetramer; modeled based on PDB IDs 2ady and 1aie [228]), cooperativity through TF-TF interactions (e.g., IFN-β enhanceosome; modeled based on PDB IDs 1t2k, 2pi0, 2o6g and 2o61 [16]) and chromatin accessibility due to nucleosome formation (PDB ID 1kx5) [229].

Transcription factors recognize DNA through the interplay of base and shape readout

Structural biology has been at the forefront of the search for a protein-DNA recognition code. Cocrystal structures of protein-DNA complexes were first solved in the 1980s [49]. Since then, more than 1,600 entries of protein-DNA structures have been entered in the Protein Data Bank [50], including structures solved by nuclear magnetic resonance (NMR) spectroscopy. These structures have revealed why many TFs preferentially bind to a specific DNA sequence [39]. Namely, the preference for a given nucleotide at a specific position is mainly determined by physical interactions between the amino acid side chains of the TF and the accessible edges of the base pairs that are contacted. These contacts include direct hydrogen bonds, water-mediated hydrogen bonds and hydrophobic contacts. This form of protein-DNA recognition is known as base readout (Figure 2A). A prominent example for base readout is the formation of bidentate hydrogen bonds between arginine residues and guanine bases in the major groove of DNA [19].

Figure 2. Base and shape readout contribute to TF-DNA binding specificity.

(A) Base readout describes direct interactions between amino acids and the functional groups of the bases. Whereas the pattern of hydrogen bond acceptors (red) and donors (blue), heterocyclic hydrogen atoms (white) and the hydrophobic methyl group (yellow) is base pair-specific in the major groove, the pattern is degenerate in the minor groove.

(B) Shape readout includes any form of structural readout, based on global and local DNA shape features, including conformational flexibility and shape-dependent electrostatic potential. The IFN-β enhanceosome (PDB ID 1t2k; top) varies in minor groove shape. The human papillomavirus E2 protein (PDB ID 1jj4; bottom) binds to a binding site with intrinsic curvature.

(C) Most DNA binding proteins use interplay between the base- and shape-readout modes to recognize their DNA binding sites. However, the contribution of each mechanism to binding specificity might vary across TF families. Shape readout dominates for the minor groove-binding HMG box protein (PDB ID 2gzk; left). Base readout is a major contribution in DNA recognition by the bHLH protein Pho4 (PDB ID 1a0a; right). Both readout modes are more or less equally present in the DNA binding of a Hox-Exd heterodimer (PDB ID 2r5z; center).

TFs can also recognize the structural features of their DNA binding sites, such as sequence-dependent DNA bending [51, 52] and unwinding [53]. This phenomenon of recognizing sequence-dependent DNA structure is known as shape readout (Figure 2B). The DNA shape concept includes the static and dynamic properties of DNA structure, and the readout of enhanced negative electrostatic potential in narrow minor groove regions through arginine [13] or histidine [54] residues.

These two protein-DNA recognition mechanisms (i.e., base and shape readout, also known as direct and indirect readout [55]) were often historically presented as mutually exclusive driving forces for DNA recognition by a given protein. Only recently have structural studies [19, 56, 57] embraced the more realistic situation that most proteins use the interplay of base and shape readout to recognize their cognate binding sites. The contributions of base and shape readout, however, vary across protein families (Figure 2C; Figure 3). Recent structures of protein-DNA complexes accurately reflect the biologically correct architecture (which can affect cooperativity), revealing cofactors that bind (Figure 3A) [1] or do not contact [16] DNA, TF-DNA binding as dimers (Figure 3B–C) [58] or tetramers (Figure 3D) [19] and multiple TFs that bind DNA while forming protein-protein contacts (Figure 3E) [59].

Figure 3. Interplay of base and shape readout varies among TF families.

(A) A heterodimer of the homeodomain proteins (PDB ID 2r5z) Hox protein Sex combs reduced (Scr; cyan; top and center) and its cofactor Extradenticle (Exd; magenta; top and center) binds with its recognition helices through base readout to the major groove (blue box; bottom), whereas arginine residues of the N-terminal Scr linker read minor groove shape and electrostatic potential as a form of shape readout (beige box; bottom).

(B) A homodimer of the bHLH protein USF (PDB ID 1an4; green and pink; top and center) binds with its recognition helices through base readout to the E-box core-binding site (blue box; bottom) and recognizes flanking sequences (beige box; bottom) through extended linkers that connect the two α-helices of each USF monomer.

(C) The human papillomavirus (HPV) E2 homodimer (PDB ID 1jj4; purple and chartreuse; top and center) recognizes with its recognition helices the half-sites of its binding site through base readout (blue box; bottom), whereas the intrinsic curvature of the central spacer contributes to binding through shape readout (beige box; bottom).

(D) The four DBDs of the p53 tetramer (PDB ID 3kz8; cyan, yellow, pink, and green; top and center) bind to the major groove through base readout (blue box; bottom), whereas the Arg248 residues recognize the minor groove through shape readout (beige box; bottom).

(E) The c-Jun and ATF-2 TFs (cyan and magenta, respectively; top and center) of the IFN-β enhanceosome (PDB ID 1t2k) recognize the major groove through base readout (blue box; bottom), whereas the adjacent IRF-3 TFs (green and yellow; top and center) use their His40 residues to recognize the minor groove through shape readout (beige box; bottom).

Computational models for describing the DNA binding specificities of transcription factors

In parallel to structural biology approaches to studying protein-DNA binding specificity, sequence-based computational methods have been developed. These methods use a set of known protein-DNA binding sites to generate DNA motif models for predicting the binding specificity of any new site. Early DNA motif discovery methods [60–63] were trained and tested on: 1) small sets of aligned TF binding sites (TFBSs) collected from small-scale experiments, such as DNase I footprinting [64] or electrophoretic mobility shift assays [65], 2) simulated data, in which TFBSs were artificially inserted into background DNA [63], or 3) sets of promoter regions of coregulated genes [61]. The development of microarray- and sequencing-based assays for the high-throughput measurement of protein-DNA binding resulted in a burst of motif discovery methods; to date, hundreds of DNA motif discovery algorithms have been developed [9, 66, 67].

Most sequence-based DNA motif discovery methods use position weight matrices (PWMs) to represent the TF-DNA binding specificity [5, 8]. This type of model is simple, intuitive and can be learned from various data types: from small sets of known binding sites to high-throughput protein-DNA binding data. Traditional PWM models have the benefit of being easy to visualize as DNA motif logos [68]. However, these models are only able to describe the DNA base readout by a TF. Moreover, they implicitly assume that positions within a TFBS independently contribute to the binding affinity, an assumption that does not always hold [7, 10, 69–71]. Consequently, more complex sequence-based models of protein-DNA binding specificity have been developed (Figure 4; Table 1A) to account for positional dependencies within TFBSs, as well as other complexities in protein-DNA recognition [2, 9, 31, 72–74].

Figure 4. Timeline of genomic approaches for experimental and computational studies of TF-DNA binding specificity.

Development of experimental high-throughput DNA binding assays (above timeline axis) and computational DNA binding specificity models and algorithms (below timeline axis). Further examples of these experimental approaches and computational methods are provided in Table 1.

Table 1.

Computational models of protein-DNA binding specificity and high-throughput assays for generating the data used to train and test specificity models.

(A) Computational models of protein-DNA binding specificity
Model type	Model Description	Refs.
Position weight matrices (PWMs)	Simple probabilistic models that assume independence between positions in TF binding sites (TFBSs)	[5]
Dinucleotide weight matrices (DWMs)	Generalization of PWM models that incorporates frequencies of dinucleotides	[73, 230]
Bayesian networks	Flexible probabilistic models that can incorporate dependencies between positions in TFBSs	[63]
Hidden Markov models	Probabilistic models that can incorporate dependencies between neighboring positions in TFBSs	[70, 231]
High-order Markov models	Flexible probabilistic models that can incorporate high order dependencies between neighboring positions in TFBSs	[232]
k-mer based regression models	Probabilistic models that predict the level of TF binding based on the frequencies of mono-, di-, and tri-nucleotides	[93, 233]
Markov networks	Flexible probabilistic models that can incorporate high-order dependencies within TFBSs	[72]
Neural networks	Flexible probabilistic models that represent TF binding specificities using a system of interconnected, artificial “neurons”	[75]
Random forest models	Flexible probabilistic models that represent TF specificity using a collection of decision trees	[92]
Support vector models	Probabilistic models that can incorporate complex patterns of similarities between TFBSs	[2, 31]
Variable-order Bayesian networks	Flexible probabilistic models that can incorporate high-order dependencies within TFBSs	[234]
Thermodynamic/Energy-based models	Models that infer DNA binding affinities by fitting thermodynamic equations to experimental data	[73, 81–83, 235–237]
Atomistic/Structure-based models	Models based on known structures of TFs bound to target DNA sites	[86, 90]
Probabilistic models that incorporate structural features	Models that incorporate structural features such as groove geometries and helical parameters	[2, 79, 91, 92]
Probabilistic models that incorporate in vivo data	Models that incorporate in vivo data such as DNA accessibility, histone modifications	[238, 239]

(B) In vivo high-throughput DNA binding assays
Assay name	Assay description	Refs
ChIP-chip	Chromatin immunoprecipitation followed by microarray hybridization	[240]
ChIP-seq	Chromatin immunoprecipitation followed by high-throughput sequencing	[241]
ChIP-exo	Chromatin immunoprecipitation with exonuclease digestion followed by high-throughput sequencing	[242]
DamID	DNA adenine methyltransferase identification	[243]
DNase-seq	DNase I cleavage followed by high-throughput sequencing	[151, 244]
FAIRE-seq	Formaldehyde-assisted isolation of regulatory elements, followed by high-throughput sequencing	[149]
ATAC-seq	Assay for transposase-accessible chromatin using high-throughput	[152]

(C) In vitro high-throughput DNA binding assays
Assay name	Assay description	Refs
B1H	Bacterial one-hybrid	[102, 245]
PBM	Protein binding microarray	[94, 246]
CSI	Cognate site identifier	[247]
MITOMI	Mechanically induced trapping of molecular interactions	[101, 248]
MEGAshift	Microarray evaluation of genomic aptamers by shift	[249]
TIRF-PBM	Total internal reflectance fluorescence protein-binding microarray	[103]
Bind-n-Seq	Analysis of in vitro protein-DNA interactions using massively parallel sequencing	[250]
SELEX-seq/HT-SELEX	Systematic evolution of ligands by exponential enrichment, followed by high-throughput sequencing	[1, 82, 110]
EMSA-seq	Electrophoretic mobility shift assay followed by deep sequencing	[95]
HiTS-FLIP	High-throughput sequencing - fluorescent ligand interaction profiling	[108]
gcPBM	Genomic-context protein binding microarray	[2]

These complex models typically perform better than traditional PWMs [2, 63, 70, 73, 75], providing important insights into the DNA recognition mechanisms used by different TFs. For example, a dinucleotide-based model [73] revealed that including the nonindependent contributions between two specific positions in the DNA binding models of Hnf4a was critical for accurately predicting the genomic regions bound by Hnf4a in vivo. Another recent study [2] revealed that contributions from di- and tri-nucleotides in the DNA regions flanking TFBSs can influence TF binding specificity. Importantly, though, the flanking di- and tri-nucleotides in these models did not reflect base readout by the TFs; instead, the effect of the higher-order sequence features was exerted through local three-dimensional DNA structure (i.e., DNA shape) [13].

Interactions between adjacent base pairs are dominated by base stacking [76] and, to a lesser degree, by inter-base pair hydrogen bonds in the major groove [77]. These physical interactions give rise to DNA shape [78, 79] and explain the interdependencies between adjacent positions in a TFBS [73] and other, more complex situations. DNA shape features can be derived on a genomic scale by using a sliding pentamer window to mine Monte Carlo predictions [78]. This approach was the basis for generating a motif database of structural features of TFBSs [79], as well as for multiple studies in which hundreds of thousands of DNA sequences were analyzed in terms of DNA shape features [1, 2, 30, 79, 80].

A small but important class of sequence-based motif discovery methods represents approaches to infer DNA binding affinities by fitting thermodynamic energy-based models to experimental data (Table 1A). Similar to probabilistic models, some energy-based models assume independent contributions among positions in the TFBS [81–83], whereas others incorporate nonindependent contributions [73]. Structure-based atomistic models of DNA binding specificity have also been developed [84–90]. However, these models are not yet widely used, likely because they require knowledge of the structure of the protein (or one of its homologs) when bound to the target DNA site. Such data are not as easily available as DNA sequence data. Without having to model the complete structure of the TF-DNA complex, structural information on DNA alone can be incorporated into DNA motif discovery models. Recently, probabilistic models incorporating DNA structure-derived features [2, 79, 91, 92] were shown to perform better than models based on DNA sequence information alone. Thus, genomic and structural information are beginning to be integrated into protein-DNA binding models that account for both base- and shape-readout mechanisms.

Binding assays for probing the DNA binding specificities of transcription factors

With the emergence of new, high-throughput technologies for measuring protein-DNA binding (Figure 4; Table 1B, C), it has become more feasible to create complex models of DNA binding specificity through machine learning. However, all experimental datasets contain noise and (potentially substantial) biases, and complex models will fit the noise and biases more easily than simple PWM models. Thus, it is not surprising that in some recent studies of algorithms for training DNA binding specificity models from high-throughput data [9, 93], the models that performed best on certain in vitro datasets did not always generalize well to independent in vivo data. As more accurate datasets emerge (e.g., from genomic-context protein-binding microarrays [gcPBMs] [2, 74]), it is likely that more TFs will be better described by complex models of DNA binding specificity [43].

The rich datasets provided by high-throughput technologies have revolutionized our ability to characterize protein-DNA binding specificity. For example, the comprehensive nature of universal protein-binding microarray (PBM) data [94], which include measurements of TF binding specificity to all possible 8-bp sequences, has facilitated the characterization of low-affinity TF-DNA binding sites, which are often not captured by simple DNA binding models [95, 96]. Such sites, which are under widespread evolutionary selection [97, 98], are critical for interpreting the spatial and temporal TF gradients that arise during development [99, 100]. High-throughput datasets have revealed that closely related TFs, even when they exhibit a high degree of similarity in their DNA binding domains (DBDs; up to 67% amino acid identity), can have distinct DNA binding profiles [7, 95, 101–105]. Moreover, different TF family members can prefer different core binding sites [7, 102, 106, 107] or flanking DNA sequences [2, 108]. Thus, both base- and shape-readout mechanisms might play roles in the differential DNA binding specificity of paralogous TFs.

Perhaps the most striking finding suggested by high-throughput protein-DNA binding technologies is the large number of proteins that can bind DNA using two or more distinct modes [47]. A small number of such proteins were previously identified through structural studies [32, 39, 109]; however, recent high-throughput data suggest that this phenomenon is more common than anticipated. Variable binding modes can be classified into different categories: 1) variable spacing, in which TFs bind DNA motifs composed of two half-sites separated by different numbers of bases [7, 104]; 2) multiple DBDs, in which TFs contain multiple independent DBDs that allow them to recognize different DNA elements [7]; 3) multimeric binding, which might be more common than previously thought and can even occur in the case of TFs known to bind DNA primarily as monomers [10, 110]; and 4) alternate structural conformations, in which TFs with a single DBD can bind different DNA motifs, enabled by distinct conformations of the DBD (e.g., mouse TF SREBF1) or domains outside the DBD (e.g., yeast TF Hac1) [9, 111]. Importantly, the multiple modes of DNA binding observed in high-throughput in vitro studies are also enriched in the genomic regions bound by TFs in vivo [10, 104], suggesting that the different mechanisms of binding are biologically relevant. Further studies of TFs with multiple modes of binding are needed to understand the precise biochemical and biophysical mechanisms that allow such TFs to interact with diverse binding sites.

Studying the specificity of individual TFs via high-throughput in vitro technologies cannot provide a full picture of how these proteins achieve their diverse regulatory roles in the cell. Transcriptional regulation often involves the assembly of multiprotein complexes, which can modulate the DNA binding specificities of individual TFs [1, 16]. A complete understanding of the determinants of binding specificity in gene regulation requires the integration of all factors that affect protein-DNA binding in the cell, including cooperating or competing TFs and the local chromatin state.

From in vitro to in vivo transcription factor-DNA interactions

Transferring our knowledge of the in vitro biochemical and biophysical principles of protein-DNA interactions to an in vivo context is not straightforward. In contrast to the relatively well-defined components of a typical in vitro biochemical experiment, the cellular nucleus contains hundreds of millions of DNA base pairs (in metazoan genomes), as well as RNA, histones and countless nonhistone proteins. The overall concentration of macromolecules in the nucleus is estimated to be between 100 and 400 mg/ml [112, 113]. Within this crowded nucleoplasm [114], TFs somehow bind specific DNA sites and regulate gene expression. In addition, although the genome contains numerous potential binding sites for each TF, only some of them are actually bound in vivo, and only a fraction of the bound sites are functional. Consequently, predicting and interpreting in vivo TF-DNA binding are not trivial endeavors, even when the intrinsic sequence preferences of TFs are well characterized in vitro.

Regulatory genomic sequences targeted by TFs are primarily found in noncoding intergenic or intronic DNA, with a few exceptions [115]. The amount of noncoding genomic DNA varies from organism to organism, with metazoan genomes containing relatively large amounts of noncoding DNA (e.g., ~97% of the human genome is noncoding vs. <30% of the Saccharomyces cerevisiae genome [116]; Figure 5A). Although pioneering studies in S. cerevisiae have provided a tremendous foundation for our understanding of TF biology, the noncoding regulatory landscape in this organism is easier to parse than for metazoan eukaryotes.

Figure 5. Distinct cis-regulatory structure of unicellular and metazoan model organisms.

(A) Percentages of coding and noncoding DNA in select genomes, adapted from [116].

(B) Typical regulatory structure of a Saccharomyces cerevisiae gene, with most regulatory DNA binding sites falling within a few hundred bases of the gene’s TSS.

(C) Typical regulatory structure of a human gene, with several clusters of regulatory DNA sites (enhancers) distal to the TSS.

For (B) and (C), green dashed lines represent activating regulatory inputs, and red dashed lines represent repressive inputs.

For S. cerevisiae, most regulatory DNA sequences for a given gene fall within a few hundred base pairs of its transcription start site (TSS) (Figure 5B) [117]. In metazoans, by contrast, regulatory sequences often fall tens of kilobases or even megabases from the TSS of the target gene [118–120]. These distal elements can be upstream or downstream of the target gene, and they regularly bypass intervening genes (Figure 5C). The combination of a large search space (i.e., noncoding sequence) and the distal location of many enhancers complicates the search for regulatory DNA sequences in metazoans.

Making sense of regulatory DNA is further complicated by a lack of straightforward sequence grammar. Unlike genic coding regions, which are easily interpreted from the triplet code, noncoding regulatory elements are difficult to decode. Regulatory TFBSs are often clustered, with binding sites from different TFs in close proximity to one another. A group of TFBSs that function together to direct gene expression are referred to as a cis-regulatory module (CRM) or enhancer. The combinatorial nature of these groupings gives enhancers the ability to integrate inputs from multiple TFs, to direct the spatial and temporal patterns of gene expression. Although enhancers typically contain clusters of TFBSs and other common features (e.g., dinucleotide repeat sequences [121]), the patterns associated with these features are not sufficiently strong to permit easy discrimination between enhancers and nonregulatory DNA. In addition, sequence information is often an insufficient predictor of TF binding because in vivo TF binding preferences are influenced by additional variables, including interaction with cofactors and chromatin accessibility (discussed below). Ultimately, enhancers are difficult to decode and require further experimental work for their identification and functional characterization.

Chromatin and transcription factor-DNA binding

In the past decade, we have seen a dramatic expansion of the use of genome-wide technologies for studying in vivo TF-DNA binding and transcriptional regulation. These technologies include genome-wide chromatin immunoprecipitation (ChIP-seq) and related approaches (Figure 4; Table 1B), gene expression profiling and newer screening methods for the high-throughput identification of DNA regions with enhancer activity [122–130]. Collectively, these tools of the genomics era have facilitated the annotation of genomic regulatory regions and have served as a platform for understanding TF-DNA interactions on a global scale, informing models of how TFs achieve regulatory specificity in vivo.

One surprising finding from early genome-wide ChIP studies was that TF binding is widespread, with thousands to tens of thousands of binding events for many TFs. These numbers did not fit with existing ideas of the regulatory network structure, in which TFs were generally expected to regulate a few hundred genes, at most [131–133]. Binding is not necessarily equivalent to regulation, and it is likely that only a small fraction of all binding events will have an important impact on gene expression (Figure 6) (discussed below) [134, 135]. However, if we ignore preconceived notions regarding the expected number of direct target genes for a TF and, instead, focus only on DNA sequence, the genome-wide binding numbers begin to make sense.

Figure 6. In vitro versus in vivo transcription factor-DNA interactions.

(A) Standard and high-throughput in vitro DNA binding assays provide a motif or model representing a TF’s DNA binding preferences.

(B) Genomic DNA sequences matching a TF’s in vitro-derived motif represent potential TFBSs.

(C) Potential in vivo binding sites determined from a TF’s in vitro-derived motif far outnumber the actual number of in vivo binding sites as measured by ChIP-seq. In general, <5% of potential binding sites are identified as bound in vivo. In addition, in vivo binding strength does not always correlate with motif strength, and not all in vivo binding sites contain the expected motif. Non-DNA variables, such as nucleosomes and cofactor interactions, explain part of the difference between predicted and actual binding.

(D) Not all in vivo binding events have a regulatory impact on gene expression. Productive, functional binding must be validated experimentally using standard reporter assays or other measures of cis-regulatory function. In this hypothetical example, only Regions W and Y drive gene expression that is responsive to the TF being tested.

Considering the information content of a typical 6-bp human TF motif, one would expect matches to a motif to occur approximately once every 4 kb, with hundreds of thousands of potential binding sites genome-wide [136]. Thus, based on information theory alone, TFs actually bind far fewer regions than expected (Figure 6), due in large part to the restrictive nature of chromatinized DNA.

Nuclear DNA is associated with nucleosomes, which consist of two copies each of the histone proteins H2A, H2B, H3 and H4, or their variants. Nucleosome assembly facilitates DNA packaging in the nucleus, but also has major regulatory roles [22]. Histones are subject to extensive posttranslational modifications (PTMs) [137, 138], which can regulate chromatin compaction and affect the recruitment of certain transcriptional regulators [139, 140]. With more than 100 possible histone PTMs and a tremendous possibility for combinatorial PTM interactions, the burgeoning field of epigenomics is rapidly defining genome-wide chromatin states (i.e., distinct combinations of histone modifications and other chromatin-associated factors at a given locus) across many cellular contexts [137]. Findings from the integration of chromatin state data with TF binding data suggest that many TFs have specific histone PTM preferences that are consistent across multiple cell types [141]. Still, it is often unclear whether a specific chromatin state is simply permissive to TF binding, actively directs TF binding or is a result of TF binding. Further mechanistic elucidation of the relationships between TFs and histone PTMs will likely influence our models of TF-DNA targeting.

Aside from the regulatory potential of histone PTMs, nucleosome-DNA interactions can provide a steric impediment to TF binding and increase TF-DNA dissociation rates [142]. Consistent with this concept, most of the TFBSs identified by the Encyclopedia of DNA Elements (ENCODE) consortium fall within highly accessible (i.e., nucleosome-depleted) DNA regions [143]. Furthermore, for several TFs, simple thermodynamic models based on TF levels, DNA motif information and DNA accessibility [23, 24, 133, 144, 145] can largely explain genome-wide binding patterns. Carefully designed studies have suggested that the accessibility of TFBSs can explain most genome-wide binding patterns. However, recent studies indicate that some binding to accessible DNA regions may be a crosslinking-mediated ChIP artifact (discussed below) [146, 147], and there are factors whose binding patterns are not driven by DNA accessibility [148].

DNA accessibility in vivo is commonly measured through DNase-seq, FAIRE-seq, or, more recently, ATAC-seq (Figure 4; Table 1B) [149–152]. DNase-seq is based on the differential DNase I sensitivity of nucleosome-associated and nucleosome-free DNA. DNase I selectively cleaves DNA that is not protected by association with nucleosomes; therefore, accessible DNA regions manifest as DNase I-hypersensitive sites. TF binding to DNA protects DNA from cleavage by DNase I. Consequently, footprints of TF-DNA binding can be identified within hypersensitive regions [151]. These properties of DNase-seq data were recently exploited to characterize DNA accessibility profiles around TFBSs during a program to differentiate mouse embryonic stem cells (ESCs) into pancreatic and intestinal endoderm [153]. The data were used to quantify the impact of a given TF on DNA accessibility patterns. Ultimately, TFs were broken down into three categories: pioneers, settlers and migrants.

Pioneer TFs (Figure 7A) are characterized by their ability to bind DNA target sites, even in inaccessible regions, and, subsequently, to promote DNA accessibility. Although pioneer TF activity had been described previously [154, 155], the above DNase-seq based study expanded the catalog of TFs with pioneer activity [153]. Interestingly, TFBSs for the pioneer TF Pu.1 can be differentiated from nontargeted Pu.1 motif matches, based on DNA sequence and shape characteristics that favor nucleosome assembly [29]. True Pu.1 target sequences are highly associated with nucleosomes in cell types where Pu.1 is not expressed. This result suggests that selective pressures have favored sequences that are competent for both pioneer TF binding and nucleosome occupancy. It also highlights the importance of the interplay between these two forces in pioneer TF function.

Figure 7. Transcription factor-DNA binding strategies.

(A) Pioneer TFs can bind inaccessible, nucleosome-associated DNA sites. Pioneer factors then create an open chromatin environment that is permissive for the binding of nonpioneer factors (settler and migrant TFs).

(B) Settler TFs bind to essentially all accessible copies of their DNA target sites.

(C) Migrant TFs only bind a subset of their accessible target DNA sites.

(D) High- and low-affinity binding are driven by a TF’s specific DNA recognition properties. Nonspecific binding is driven by the electrostatic attraction between negatively charged DNA and positively charged DNA binding domains. Nonconsensus binding is driven by the attraction of TFs to repeated homo-oligomeric tracts. Indirect binding, or tethering, is driven by the interaction of TFs with another DNA binding factor (in this schematic, TF’).

In contrast, settler TFs (Figure 7B) almost always bind sites matching their DNA binding motif if these sites fall within accessible DNA; however, they do not bind inaccessible DNA sites [153]. The least defined group, migrant TFs (Figure 7C), are similar to settler TFs, although more selective [153]. Migrants only bind a subset of their target sites, even in accessible DNA; therefore, their selectivity is likely driven by interaction with additional cofactors. Although settler and migrant TFs do not evict nucleosomes like pioneer factors, TFs lacking pioneer activity can facilitate the binding of unrelated TFs by competing with nucleosomes for DNA binding; this process is termed collaborative competition or nucleosome-mediated cooperativity [26, 27]. Taken together, these data support the idea that DNA accessibility substantially contributes to the DNA binding selectivity of most TFs, with pioneer TFs being an important exception.

Functional and nonfunctional transcription factor-DNA binding

Regardless of whether one considers the widespread genomic binding of TFs to be expected or unexpected, most researchers acknowledge that a reasonable fraction of TF binding events are neutral or nonfunctional (i.e., they do not have a direct impact on target gene expression levels). ChIP-seq assays do not provide any information about regulatory function, just protein-DNA coassociation. In addition, similar to all biochemical purification assays, ChIP-seq assays must cope with false positives and false negatives (see [156] for what is necessary to confirm functional binding). Although functional binding events are certainly present within the thousands of genome-wide binding events for many TFs, neutral binding is likely to be common [135].

Thus, a major question in the TF genomics field regards how to identify functional TF binding events within the thousands of genome-wide TF-DNA interactions. What features distinguish functional from neutral binding? Can we use these distinctions to learn about TF-DNA binding strategies? The data suggest that functional binding can be identified on the basis of several distinguishing features, although these features will be influenced by the TF under study and the experimental design.

Developmentally dynamic or clustered TF peaks have been identified as enriched for functional binding events [157–162]. Functional analyses of TF targets in the Drosophila embryo suggested that the strongest ChIP peaks represent functional binding, whereas lower signal peaks do not [135]. Consistent with this model, strong ChIP peaks are more likely to be conserved across species [161, 163, 164]. However, ChIP peak strength is a less reliable indicator of function when monitoring binding in more heterogeneous tissues, likely because functional binding events only occur in a subset of cells within a tissue [162].

However, caution is needed when interpreting functionality or binding affinity from ChIP-seq signal strength. ChIP assays are usually based on the average signal across millions of cells. Thus, a medium peak might actually be a high-affinity TFBS that is only bound in 50% of cells, whereas a strong peak might be a medium-affinity TFBS that is bound in every cell. That is not to say that peak strength does not correlate with affinity or regulatory function for some TFs (as there clearly can be a strong correlation [135]); however, not all data follow this pattern. The experimental design must be considered when interpreting and building models from in vivo genome-wide TF binding data.

The implications of the many seemingly nonfunctional binding events identified by ChIP-seq should also be considered. As a point of clarification, discussions of ChIP-seq data often refer to regions of strong ChIP enrichment as TFBSs, which can be misleading. Immunopurification assays, especially those aided by crosslinking, can be rife with false positives. Indeed, recent carefully controlled ChIP-seq studies in yeast have indicated that many regions of the genome, especially those associated with highly expressed genes, are hyper-ChIPable. This situation makes it difficult to discern between functional and artifactual ChIP signals [146, 147]. The resulting high potential for artifact-based peaks in ChIP must be considered when interpreting ChIP-based studies.

The fact that potentially misleading ChIP signals are associated with highly expressed genes is interesting because highly occupied target (HOT) regions also exhibit this feature [165, 166]. HOT regions are often targeted by 10 or more unrelated TFs. They generally fall in nucleosome-depleted regions upstream of highly expressed genes. Although HOT regions can act as regulatory enhancers, many of the binding events within HOT regions are neutral (i.e., have no impact on gene expression patterns) and may result from nonspecific or indirect DNA binding [167, 168]. Interestingly, HOT region binding disappears when a modified, crosslinking-free ChIP protocol is used, suggesting that such binding could be an experimental artifact for certain TFs [169].

Nonfunctional ChIP signals may potentially be due to the capturing of transient nonspecific or indirect binding events in highly accessible DNA. Single-cell, single-molecule imaging studies of the TFs Sox2 and Oct4 demonstrated that nonspecific interactions with chromatin are central to the in vivo search for functional binding sites [170]. At physiological TF concentrations, at least for Sox2 and Oct4, these nonspecific interactions were sampled enough times to provide a measurable ChIP signal in a population of cells [170]. Nonregulatory protein-DNA interactions can be sequence-dependent, occurring via binding to spurious weak matches to the TF target sequence, as a result of the low-information motifs targeted by metazoan TFs (Figure 7D) [136]. Sequence-independent nonspecific interactions are also possible, through interactions with other chromatin-associated proteins or through the general electrostatic attraction between negatively charged DNA and positively charged DBDs (Figure 7D) [171, 172].

A recent theoretical model suggested that TFs are more attracted to repeated homo-oligomeric poly(dA:dT) and poly(dC:dG) tracts; the longer the segment, the greater the attraction (Figure 7D) [173–175]. This variation of nonspecific binding, termed nonconsensus binding, has also been observed in vitro [176]. It has the potential to shape nonfunctional and functional TF-DNA interactions [173]. Thus, although nonfunctional TF-DNA associations do not provide information about the regulatory targets of a TF, they may provide clues to the mechanisms by which TFs find their functional binding sites across the genome. To recognize their functional sites during this searching process, TFs are influenced by additional variables, including direct and indirect interactions with other TFs.

Transcription factor interactions at genomic regulatory regions

A clear theme from both classical enhancer-bashing studies and newer genomics data is that enhancers must integrate multiple TF inputs to direct precise patterns of gene expression. How, exactly, are multiple TFs assembled at enhancers? The answer to this question is likely to fall somewhere on a spectrum represented by two extremes: the enhanceosome model and the billboard model.

The enhanceosome model (Figure 8A) is based on pioneering work with the interferon-β (IFN-β) enhancer [177]. This model proposes that enhancer activity is dependent on the cooperative assembly of a set of TFs at the enhancer. Only once the cooperative unit is assembled on an enhancer will cofactor recruitment cause changes in gene expression. The cooperative assembly of an enhanceosome is dependent on protein-protein interactions and a highly constrained pattern of TF-DNA binding sites (or binding-site grammar). Enhancesome assembly does not tolerate shifts in the quality, spacing or orientation of the binding site, which can disrupt protein-protein interactions and cooperativity [59, 178].

Figure 8. Models of transcription factor assembly on enhancer DNA.

(A) Left: The enhanceosome model is characterized by cooperative TF binding and highly constrained binding site positioning. Right: Minor changes in enhancer sequence (i.e., inversion in this case, but insertions, deletions, mutations, etc., also apply) can lead to collapse of TF assembly and enhancer function.

(B) Left: The billboard model is characterized by highly flexible binding-site grammars. Although all TFs are important for enhancer function, TF binding and enhancer function are not affected by significant changes in binding site positioning or orientation.

The IFN-β enhanceosome probably represents an extreme example, as few enhancers are found under similarly stringent constraints. However, additional examples of organizationally constrained enhancers do exist [179–184]. Spatial constraints on select paired TF-TF coassociations and binding-site combinations are found in genome-wide ChIP data [148, 185, 186]. Interactions between TFs can lead to cooperative DNA binding, although this binding does not approach the extreme multifactorial cooperativity required for enhanceosome assembly. True enhanceosome and enhanceosome-like regulatory DNA elements are not common. It may be that they are only necessary under unique regulatory conditions, such as for the amplification of signals at enhancers regulated by low-abundance TFs [181] or to prevent unwanted TF synergy and ectopic enhancer activity [182].

The billboard model (Figure 8B), also known as the information display model [187, 188], hypothesizes that although individual TFBSs are essential for enhancer activity, binding-site grammar is very flexible. That is, the positioning of binding sites within an enhancer is not subject to strict spacing or orientation rules because, even though the TFs collaborate to regulate enhancer output, they do not target the enhancer as a cooperative unit. The TFs at a billboard enhancer work together in a combinatorial fashion to direct precise patterns of gene expression, but they do not depend on highly cooperative DNA binding to target the enhancer in an all-or-nothing manner. For example, the loss of a cell-specific repressive input to an enhancer will lead to ectopic target gene expression in that cell type, but will not cause the complete collapse of enhancer function. Billboard-like combinatorial binding is not uncommon in genome-wide ChIP data [189, 190]. Indeed, findings from the high-throughput dissection of mammalian enhancers suggest that the regulatory architecture of many enhancers is quite flexible [128, 191].

Another flexible enhancer architecture model—the TF collective model—was recently proposed on the basis of the genome-wide binding patterns of a panel of TFs that regulate heart development in Drosophila [192, 193]. Cardiac TFs were observed to bind their target regions in an all-or-nothing fashion, with binding driven by the collective action of many TFs, similar to cooperative binding. Similar all-or-nothing patterns of genome-wide binding have been seen in TFs that regulate mammalian hematopoiesis [194]. However, despite the similarity to cooperative binding, the binding-site grammar at targeted enhancers is flexible in the TF collective model [193].

Ultimately, the mechanisms by which multiple TFs assemble on enhancers probably fall on a continuum between the enhanceosome and billboard extremes. Distinct TF binding properties are better suited to different regulatory strategies. Noncooperative TF-DNA interactions are well suited for regulating graded gene expression, which is often necessary for homeostatic responses. Cooperative interactions are more appropriate for switch-like, on/off expression, which is often necessary in developmental cell-fate decisions [195–197]. The strategies employed by TFs and enhancers are subject to multiple evolutionary pressures. In the end, no single model can accurately describe all of the rules of transcriptional regulation.

Cellular context and transcription factor binding specificity

In multicellular organisms, gene regulatory networks are plastic, with spatial, temporal and environmental dynamics impacting gene expression patterns. Many TFs are reiteratively used in multiple cellular contexts, often directing the expression of distinct sets of genes. Characterizing the influence of cellular context on genome-wide TF-DNA binding is central to the understanding of binding specificity. Accordingly, there has recently been a dramatic increase in the number of ChIP-seq studies monitoring metazoan TF-DNA binding across multiple cell or tissue types [3, 162, 198–205] or across multiple environmental or signaling contexts [129, 206–208]. Although context-independent binding (i.e., binding events shared across multiple conditions) is common [162, 199, 204], context-specific binding is substantial in all cases, suggesting that regulatory specificity is often achieved at the level of TF-DNA binding. Importantly, DNA accessibility is dynamic, with important differences in accessibility across cell types or developmental stages within a cell type [143, 209–211]. Thus, the chromatin environment is modified by cellular context, likely through the pioneer TFs expressed in a given context, which, in turn, can impact the binding patterns of nonpioneer TFs.

Interestingly, context-independent and -dependent binding events for a given TF often represent distinct binding strategies. For example, estrogen receptor (ER) binding sites that are shared between breast and endometrial cancer cell lines are associated with high-affinity estrogen response elements (EREs), are not dependent on DNA accessibility, and tend not to colocalize with interacting TFs [204]. By contrast, cell type-specific ER binding sites are not associated with high-affinity EREs, fall within DNA that was accessible before ER activation, and colocalize with interacting TFs. Whether the colocalized TFs in the cell type-specific ER binding sites directly impact ER-DNA binding preferences, or whether they simply generate a permissive chromatin environment, remains to be tested. Nevertheless, it is clear that these binding sites represent a regulatory strategy that is distinct from that used at the cell type-independent ER binding sites.

Cell- and tissue-specific genomics data have clarified that precise patterns of gene expression result from collaboration between broadly expressed TFs and tissue-, cell-, or developmental stage-specific TFs [3, 129, 202, 212, 213]. This mechanism for refining the regulatory activity of a broadly expressed TF is not new to developmental biology. Indeed, the mechanism was evident from the findings of enhancer-bashing experiments that were performed before genomics experiments were common [214].

An interesting example of this refinement comes from two TF modules that direct the differentiation of mouse ESCs into spinal or cranial motor neurons (Figure 9A) [215]. The homeodomain TF Isl1 is an essential component of both modules. Homeodomain TFs Lhx3 and Phox2a determine whether a spinal or a cranial motor neuron, respectively, is generated. Inducible expression of these two ESC programming modules revealed that Isl1 binding is strongly influenced by context (i.e., the presence of Lhx3 or Phox2a is required for distinct Isl1-Lhx3 or Isl1-Phox2a composite binding sites, respectively). In this elegant experimental model, the programming TFs were induced concomitantly by using a polycistronic construct, in an identical cellular context (ESCs). Consequently, the observed binding differences were not due to basal differences in chromatin structure or expressed cofactors. The data suggested that Isl1 forms a complex with Lhx3 or Phox2a; the complex is then recruited to context-specific enhancers with distinct binding-site grammars to direct cranial or spinal motor neuron fate. Thus, Isl1 is necessary for both motor neuron fates, and its genome-wide DNA targeting is refined by interactions with additional cell type-specific TFs.

Figure 9. Cellular context and transcription factor-DNA binding.

(A) Isl1 is an essential factor in two separate embryonic stem cell (ESC) reprogramming modules, which generate spinal (left) and cranial (right) motor neurons, respectively. The genome-wide DNA targeting of Isl1 is markedly influenced by interaction with spinal- and cranial-specific TFs (Lhx3 and Phox2, respectively). DNA at different loci is represented in blue, red or black. DNA accessibility profiles of the reprogrammed stem cells resemble brain, not ESC, accessibility profiles, suggesting that the reprogramming TFs can induce DNA accessibility. However, this possibility remains to be functionally tested.

(B) Left column: GATA “switch” sites at the GATA2 locus remain continually bound by GATA factors through multiple stages of erythroid differentiation. GATA2 acts as an autoregulatory activator at these enhancers, and GATA1 is either repressive (red line) or neutral (gray dashed line). Right column: At the GATA1 locus, DNA methylation and, presumably, chromatin compaction prevent GATA2 from binding a “switch” enhancer in hematopoietic stem cells. As the epigenetic environment becomes permissive, GATA2 binds this enhancer and activates GATA1 expression. GATA1 then displaces GATA2 and acts as an autoregulatory activator at this enhancer.

Binding that is unaffected by a cellular context can be important and may represent the association of a TF with its canonical targets [199, 204]. For example, a variation in context-independent binding is central to the regulatory roles of the GATA transcription factors Gata1 and Gata2 (Figure 9B). These zinc finger proteins bind the DNA motif WGATAA (W = A or T) and are the primarily regulators of hematopoietic stem cell (HSC) maintenance and differentiation [216]. These factors were the subject of several recent ChIP-seq experiments covering multiple branches of hematopoietic lineage commitment. The studies identified substantial cell- and stage-specific GATA factor-DNA binding [213, 217, 218] and highlighted the key role that DNA-binding and non DNA-binding cofactors play in modifying GATA-DNA binding selectivity [194, 219–221].

GATA1 and GATA2 also act at binding sites that remain bound by GATA factors when HSCs differentiate into erythrocytes. In the GATA switch process, GATA2 (which maintains the HSC state) is displaced by GATA1 (which promotes erythroid commitment) [216]. This process is best characterized at autoregulatory enhancers targeting the GATA1 and GATA2 genes (Figure 9B), where the switch can have a neutral regulatory effect or can change the direction of an enhancer’s activity (e.g., activator to repressor) [216, 222–224]. Importantly, several ChIP-seq studies have demonstrated substantial overlap in the regions targeted by GATA1 and GATA2 at different stages, suggesting that the GATA switch might be part of a global mechanism during erythroid commitment [194, 218, 221, 225–227]. Thus, the potentially widespread GATA switch mechanism is dependent on highly similar GATA factors targeting the same DNA sequence at multiple stages of erythroid development.

Findings from the recent glut of context-specific ChIP-seq experiments demonstrate that the context-specific regulatory activity of a TF is often adjusted at the level of TF-DNA binding. A TF may bind and regulate the output of an enhancer in one cell, whereas it does not bind the same enhancer in another cell. Differential binding could be regulated via DNA accessibility or cofactor interactions; however, another mechanism is also prevalent. In many cases, a TF (or highly similar TFs, in the case of the GATA switch) targets the same enhancer across many cellular contexts. In these instances, changes in enhancer activity are likely to be regulated by changes in the coactivators or corepressors that are recruited by the bound TF, or by the action of collaborating TFs that target the same enhancer.

Selective pressures on regulatory DNA have resulted in finely tuned systems for increasing/decreasing the transcription of a given gene, although there clearly are many routes towards regulating enhancer output. It seems that the only common thread in the world of TF-DNA interactions and transcriptional regulation is that no single model is sufficient to explain all of the mechanisms used to achieve regulatory specificity.

Concluding remarks

TFs select their genomic target sites through multiple mechanisms at various levels. Some of these mechanisms are well understood; for instance, the determinants of base and shape readout are known because of the many high-resolution structures that are currently available. Models of TF-DNA binding specificity using PWMs or interdependencies between nucleotide positions in a binding site can quantitatively describe in vitro binding. Higher-order determinants of TF-DNA binding in vivo include cofactors, TF cooperativity and chromatin accessibility. However, an accurate model that integrates all of the known contributions to TF-DNA binding specificity is not yet available, because the interactions between the various factors of in vivo binding are highly complex, dynamic and dependent on many unknown parameters.

Thus, a simple recognition code does not exist between the amino acids of a TF’s DBD and the nucleotides in the TFBS. It is possible that some complex code, comprising rules from each of the different layers, contributes to TF-DNA binding; however, determining the precise rules of TF binding to the genome will require further high-quality structural and high-throughput binding data. Questions that remain to be determined include whether such a multi-rule system will ever be condensed into a single code and, if so, whether such a potential code represents the overarching principles of protein-DNA recognition or is highly specific for TF families and the cellular conditions of their activity.

Box 1. Outstanding questions.

• Will it be possible to condense the different rules that determine TF-DNA binding specificity (e.g., base and shape readout, cofactors, cooperativity and chromatin accessibility) into a simple code?

• Would such a code describe overarching principles that are valid for protein-DNA interactions in general, or would it be highly specific to a TF or a TF family?

• If a single code cannot be defined, can a set of rules that describes binding specificity at multiple levels be integrated into a complex, but unified, model?

• What kind of experimental data will be required to derive more accurate binding specificity models?

• What kind of computational methods need to be developed to derive accurate models from high-throughput genome-wide binding data?

• To what extent can higher-quality in vitro TF-DNA binding data be used to derive more accurate binding specificity models and explain in vivo TF-DNA binding?

• Beyond using cofactors to alter DNA binding preferences, how much impact do variables, such as PTMs, have on TF-DNA binding specificity?

• Considering the diverse, context-specific roles of many TFs, can a single motif ever capture a TF’s in vivo DNA binding preferences?

• Within the same cell type, how important is cell-to-cell variation in TF-DNA interactions?

• Will single-cell genomics reinforce or rewrite current models of in vivo TF-DNA binding?

• Beyond DNA accessibility, are there any instances in which the chromatin state (e.g., presence of histone modifications) acts as an epigenetic specificity determinant, or is this state primarily an effect of TF binding?

Highlights.

• TFs recognize their genomic target sites by using mechanisms at multiple levels

• Models of DNA sequence and shape can capture the in vitro TF binding specificity

• Cofactors, cooperativity, chromatin and other factors affect the in vivo TF binding

• No simple code combines all of the various determinants of TF binding specificity