Low-Affinity Binding Sites and the Transcription Factor Specificity Paradox in Eukaryotes

Abstract

Eukaryotic transcription factors (TFs) from the same structural family tend to bind similar DNA sequences, despite the ability of these TFs to execute distinct functions in vivo. The cell partly resolves this specificity paradox through combinatorial strategies and the use of low-affinity binding sites, which are better able to distinguish between similar TFs. However, because these sites have low affinity, it is challenging to understand how TFs recognize them in vivo. Here, we summarize recent findings and technological advancements that allow for the quantification and mechanistic interpretation of TF recognition across a wide range of affinities. We propose a model that integrates insights from the fields of genetics and cell biology to provide further conceptual understanding of TF binding specificity. We argue that in eukaryotes, target specificity is driven by an inhomogeneous 3D nuclear distribution of TFs and by variation in DNA binding affinity such that locally elevated TF concentration allows low-affinity binding sites to be functional.

INTRODUCTION

The study of gene regulation by transcription factors (TFs) dates back more than half a century, to the pioneering work by Jacob & Monod (1961), the isolation of Escherichia coli and phage repressors (Gilbert & Muller-Hill 1966, Ptashne 1967a), and the subsequent realization that these proteins bind DNA at specific sites (Gilbert & Maxam 1973, Ptashne 1967b). This last discovery raised a new fundamental question: How do these repressors recognize and bind to a specific stretch of DNA from among the millions of possible sites in the E. coli genome? Early X-ray crystallography structures for the Lac and λ repressors (Anderson et al. 1981, Ohlendorf et al. 1982, Steitz et al. 1982) suggested that TFs primarily recognize DNA sequences by forming hydrogen bonds with base pairs in the DNA major groove. Initially, it seemed that simple rules might predict which DNA sequences are bound by any TF (Pabo & Sauer 1984). However, with more structures solved, it became apparent that TFs use a variety of structural mechanisms to recognize DNA (Garvie & Wolberger 2001, Luscombe et al. 2001) and that a simple DNA recognition code might not exist (Pabo & Sauer 1992, Slattery et al. 2014).

TFs can be grouped into distinct structural families on the basis of their DNA binding domains (DBDs). In eukaryotes, the largest of these are the zinc finger (ZF), homeodomain (HD), basic leucine zipper (bZIP), and basic helix-loop-helix (bHLH) families (Lambert et al. 2018). With the exception of ZF proteins, TF paralogs from the same DBD family tend to recognize similar DNA sequences in vitro (Jolma et al. 2013, Weirauch et al. 2014) (Figure 1a). Early genetic experiments that swapped closely related homeobox DBDs resulted in severe homeotic transformations (reviewed in Mann et al. 2009, Merabet & Mann 2016), indicating that TFs with very similar DBDs nevertheless execute distinct functions in vivo. Moreover, many TFs have crucial functions as master regulators of cell fate (Vierbuchen & Wernig 2012) and perform their functions reliably, despite highly similar sequence recognition. For instance, family members of the bHLH class of TFs control essential processes as different as myocyte differentiation [bHLH MyoD (Tapscott et al. 1988)], regulation of the circadian clock [bHLH Clock and BMAL1 (Dierickx et al. 2018)], and the decision to proliferate or differentiate [bHLH Max (Carroll et al. 2018)], despite recognizing very similar binding sites (Figure 1b).

Figure 1.

Shared binding mechanisms and sequence recognition among transcription factor (TF) paralogs with distinct in vivo functions. (a) Crystal structures, binding motifs, and broad in vivo functions of three examples from the Drosophila melanogaster homeodomain TF family: All three TFs—Aristaless (PDB ID: 3LNQ), Engrailed (PDB ID: 1DU0), and Antennapedia (PDB ID: 4XID)—share a conserved structural fold and recognize similar binding motifs, represented as information content logos (motif logos derived from Noyes et al. 2008). Despite the similarities in DNA recognition and sequence readout, Aristaless, Engrailed, and Antennapedia have distinct in vivo functions: distal leg development (Campbell & Tomlinson 1998), posterior compartment identity along the anterior-posterior (AP) axis (Nusslein-Volhard & Wieschaus 1980), and thoracic fate specification (Struhl 1982), respectively. (b) Crystal structures, binding motifs, and broad in vivo functions of three homo- and heterodimeric examples from the Mus musculus and Homo sapiens basic helix-loop-helix (bHLH) TF family: MyoD:MyoD (PDB ID: 1MDY), Clock:Bmal1 (PDB ID: 4H10), and Max:Max (PDB ID: 5EYO). Each of these TF dimers recognizes a highly similar, reverse-complement-symmetric motif [consensus = CACGTG; motifs taken from Jaspar (Khan et al. 2018)], yet has a distinct in vivo function: MyoD is a master regulator of myocyte fate specification (Tapscott et al. 1988); Clock:Bmal1 are central players in establishing a functional circadian clock (Dierickx et al. 2018); and Max is involved in either cell proliferation or differentiation, depending on its binding partners (e.g., Myc or TFs from the MXD class of bHLH TFs) (Carroll et al. 2018).

To solve this specificity paradox, it is necessary to develop a thorough understanding of the DNA binding mechanisms that TFs employ. In contrast to initial expectations, the majority (approximately two-thirds) of contacts between TFs and DNA are van der Waals (VdW) interactions as opposed to direct hydrogen bonding to bases; the latter occurs as frequently as water-mediated bonds (i.e., in approximately one-sixth of contacts) (Luscombe et al. 2001). This finding implies that sequence specificity emerges not only from specific hydrogen bonding patterns but also from DNA shape readout, which—through VdW, water-mediated, and DNA backbone bonds—contributes significantly to the total free energy of binding (Rohs et al. 2009, 2010).

Multiple studies have demonstrated the importance of DNA shape readout in conferring both binding affinity and DNA binding specificity (reviewed by Rohs et al. 2010, Slattery et al. 2014). For instance, analysis of the crystal structure and binding preferences of the Drosophila TALE (three-amino-acid loop extension) TF Extradenticle (Exd) in complex with the Hox HD factor Sex combs reduced (Scr) revealed that positively charged amino acids within the N-terminal arm of the Hox HD insert into the DNA minor groove, sensing its electrostatic potential (Abe et al. 2015, Joshi et al. 2007). Intrinsic geometric and physical properties of DNA, such as helical twist and electrostatic potential, can be readily predicted from the DNA sequence (Li et al. 2017, Yang et al. 2017) and used to analyze DNA recognition mechanisms (Rube et al. 2018, Zhou et al. 2015). Deviations from classical Watson-Crick base pairing (Nikolova et al. 2011) can also be critical for TF recognition, as exemplified by the presence of Hoogsteen base pairs in the structures of Matα (Aishima et al. 2002) and the p53 tetramer bound to DNA (Kitayner et al. 2010). These examples underscore the idea that, despite being constrained by particular DBD structures, TFs exploit a wide range of mechanisms to bind to a particular DNA sequence.

The notion that any one TF can bind different DNA sequences also dates to early work in prokaryotes. For instance, sequence degeneracies were identified within the λ operator of the E. coli promoter −10 element (Maniatis et al. 1975), and the binding affinity of the λ phage Cro and λ repressors was found to directly depend on the operator sequence (Hochschild et al. 1986). These observations were initially summarized using consensus sequences, which specify degenerate positions with IUPAC ambiguity symbols, and were later summarized using more refined position weight matrices (PWMs) (Stormo et al. 1982), which tabulate nucleotide base frequencies for each position within the TF-DNA interface. A theoretical framework in which competition between mutation and selection defines the information content of the PWM relative to a random genomic background (Berg & von Hippel 1987) in turn led to a widely used visualization known as DNA sequence logos (Schneider & Stephens 1990), in which letter height corresponds to the information gain measured in bits (Figure 1).

Today, TF binding preferences are best determined by performing high-throughput in vitro binding assays (see sidebar titled (Experimental Methods for Detecting Low-Affinity Binding Sites). The two most popular technologies are (a) methods such as protein binding microarrays (PBMs) (Badis et al. 2009, Berger et al. 2008, Bulyk 2007, Mukherjee et al. 2004, Weirauch et al. 2014) and cognate site identifier (Rodriguez-Martinez et al. 2017, Warren et al. 2006), in which binding is quantified for tens of thousands of immobilized DNA probes in parallel, and (b) methods such as high-throughput systematic evolution of ligands by exponential enrichment (HT-SELEX) (Jolma et al. 2010, Zhao et al. 2009), SELEX-seq (Slattery et al. 2011), Spec-seq (Stormo et al. 2015), SMiLE-seq (Isakova et al. 2017), and DAP-seq (Bartlett et al. 2017), in which high-complexity libraries consisting of either random DNA or genomic fragments are subjected to one or more rounds of selection for TF binding followed by deep sequencing. PWMs derived from these data have been collected in databases such as CIS-BP (Weirauch et al. 2014), JASPAR (Khan et al. 2018), HOCOMOCO (Kulakovskiy et al. 2018), and UniPROBE (Hume et al. 2015).

EXPERIMENTAL METHODS FOR DETECTING LOW-AFFINITY BINDING SITES.

In recent years, high-throughput in vitro binding assays have been modified to improve the detection of low-affinity binding sites, providing an alternative strategy for the laborious trial-and-error process of identifying them in vivo; the latter typically requires enhancer bashing and a series of binding assays or mutational analyses to narrow down a small enhancer fragment (Crocker et al. 2015, Zandvakili et al. 2018). For instance, capturing specificity contributions from regions flanking the TF-DNA interface as observed in crystal structures (the so-called core binding site) can help resolve subtle but important binding preference differences in TF paralogs and TF complexes (Gordan et al. 2013, Shen et al. 2018) but requires targeted DNA probe designs. Assays such as BET-seq (Le et al. 2018), HiP-FA (Jung et al. 2018), MITOMI (Maerkl & Quake 2007), and Spec-seq (Stormo et al. 2015) yield quantitative affinity measurements that also cover the low-affinity range but require prior knowledge about a TF binding site to design the sequence pool. Complementary deep-sequencing-based assays such as HT-SELEX (Jolma et al. 2010), SELEX-seq (Slattery et al. 2011), and SMiLE-seq (Isakova et al. 2017) can measure a relative DNA binding affinity (the binding affinity of a transcription factor for any particular DNA sequence relative to the optimal sequence for the same transcription factor) and show new promise, given the recent emergence of highly accurate modeling approaches (for further details, see sidebar titled Sequence-Based Models of Low-Affinity Binding).

Direct comparison between in vitro DNA binding specificities and genome-wide in vivo TF binding profiles, as assayed using high-throughput methods such as ChIP-seq (Barski et al. 2007, Johnson et al. 2007, Mikkelsen et al. 2007) and related methods (Cheetham et al. 2018, He et al. 2015, Rhee & Pugh 2011, Skene & Henikoff 2017, Southall et al. 2013, van Steensel et al. 2001, Wang et al. 2007), has revealed that a majority of in vivo binding events are not accompanied by an obvious match to the corresponding PWM (Wang et al. 2012, Yang et al. 2006). This observation underscores the TF specificity paradox and suggests that current PWM models are missing a crucial aspect of TF–DNA sequence recognition. Binding sites alternatively referred to as low affinity (Crocker et al. 2015, 2016), suboptimal (Farley et al. 2015), or submaximal (Bhimsaria et al. 2018), which are poorly predicted by most PWM models, have been suggested as an important potential solution to this problem.

Here, we review evidence for the role played by low-affinity binding sites in eukaryotic gene regulation. We discuss how these sites are recognized by TFs and how recently developed computational tools can be used to distinguish them from nonspecific binding sites (see sidebar titled Sequence-Based Models of Low-Affinity Binding). We argue that low-affinity sites can help solve the problem of TF specificity but that this solution requires TF concentrations to be locally boosted to a sufficiently high level for low-affinity sites to be effectively bound. We suggest that eukaryotic nuclear architecture and transcriptional hubs may have coevolved with functional low-affinity binding sites to solve the eukaryotic TF specificity paradox.

SEQUENCE-BASED MODELS OF LOW-AFFINITY BINDING.

Despite the availability of high-quality data, the computational identification of relevant low-affinity sites from in vitro binding data has proven challenging. For example, researchers were unable to identify biologically functional low-affinity binding sites (Crocker et al. 2015) from oligomer enrichment tables derived from SELEX-seq data (Slattery et al. 2011). To be successful, computational approaches require an optimal statistical representation of the data generation process and a flexible representation of the features across the full TF-DNA binding interface that determine binding free energy.

A systematic comparison of methods for analyzing PBM data (Weirauch et al. 2013) revealed that algorithms that fit biophysically motivated models that use DNA sequence features as predictors (Riley et al. 2015, Zhao & Stormo 2011) achieve superior quantification of binding affinities. The resulting models can be represented as position-specific affinity matrices (Bussemaker et al. 2007) and visualized as energy logos (Foat et al. 2006), in which letter heights correspond directly to binding energy differences in bases at each position (Figure 4b, below).

Although technically more challenging due to the discrete nature of probe counts, analogous direct fitting of biophysically motivated feature-based models of protein-DNA interaction to SELEX data recently became possible and will help overcome the limitations of methods based on the oligomer enrichment tables initially used to analyze SELEX data. These feature-based algorithms include (a) an application of convolutional neural networks that treats binding site prediction as a classification problem in terms of bound and unbound probes (Alipanahi et al. 2015) and (b) two distinct maximum-likelihood algorithms, BEESEM (Ruan et al. 2017) and NRLB (Rastogi et al. 2018).

Of these, NRLB is currently the only algorithm capable of learning feature-based models of arbitrary length that accurately capture the entire range of binding affinity. When tested on Drosophila enhancers, NRLB not only successfully identified functional low-affinity Hox binding sites experimentally demonstrated to be >100-fold weaker than the best site in the fly genome, but also accurately predicted the quantitative loss of enhancer activity in vivo as individual sites were sequentially mutated (Rastogi et al. 2018).

If these algorithms can be further improved and leveraged to create a comprehensive resource of accurate sequence-to-affinity models for all TFs, this approach has the potential to transform the way in which enhancers are found and analyzed.

THE UNIQUE CHALLENGES OF EUKARYOTIC GENE REGULATION

Most early examples of TFs and their DNA binding sites came from prokaryotes such as E. coli and its phage. Consequently, theories regarding gene regulation were developed in the context of these organisms, which differ significantly from multicellular eukaryotes. Most prokaryotic TFs consist of only a DBD and an effector domain that mediates direct interactions with RNA polymerase (Perez-Rueda et al. 2018). The TFs of prokaryotes with larger genome sizes—and thus more genes to regulate—have additional DBD architectures, resulting in novel TF families and DNA recognition mechanisms (Minezaki et al. 2005). Consequently, a typical prokaryotic DBD binds with sufficient specificity to almost uniquely specify its target sites in the genome (Wunderlich & Mirny 2009). Use of the same strategy in eukaryotes would require DBDs with much larger binding sites and greater specificity to deal with much larger genomes. Surprisingly, however, the opposite phenomenon is observed: On average, eukaryotic DNA binding sites tend to be small, degenerate, and insufficiently specific to uniquely define a small number of genomic locations (Wunderlich & Mirny 2009).

To begin to reconcile this discrepancy, it is informative to ask whether there are differences in TF composition in eukaryotes and prokaryotes that might play a role in diversifying DNA recognition beyond single, short motifs. Indeed, a systematic comparison of TFs across the kingdoms of life (Charoensawan et al. 2010) noted that the number of non-DBDs, the total number of unique domain architectures, the average number of DBDs per TF, and the average length of TFs increased in eukaryotes and, to an even greater extent, in metazoans (Figure 2a). In contrast, however, both the average length of DBDs and the number of DBD families remained largely constant (roughly 60 amino acids per DBD and approximately 60–70 DBD families per organism) (Figure 2a). These observations suggest that structural constraints may place an upper bound on family diversity and DBD size and, consequently, on the achievable specificity of TF binding. However, instead of increasing DBD diversity, eukaryotes dramatically increased the number of paralogs within individual TF families (Figure 2a). Most notably, the human TF repertoire is dominated by two major families: the C2H2 ZF proteins, with >700 members, and HD proteins, with roughly 200 members (Lambert et al. 2018). The existence of so many structurally related members of the same DBD family makes it even more challenging to understand how individual specificities and TF functions are achieved.

Figure 2.

Prokaryotic versus eukaryotic transcription factors (TFs): design principles and binding mechanisms (Charoensawan et al. 2010).(a) Comparison between prokaryotic and eukaryotic TF design principles. (i) Simplified cellular structure of prokaryotic and eukaryotic cells. (ii) Similarities in TFs. The average amino acid (aa) length of the DNA binding domain (DBD) (~60 aa) and the total number of distinct structural families per species are conserved across animal kingdoms. (iii) Differences in TFs. Compared to prokaryotes, eukaryotic and metazoan TFs are on average longer (in terms of total aa length) and have a higher number of unique domain architectures (i.e., combinations of DBDs and non-DBDs). In addition, eukaryotes have a much larger number of TF paralogs per structural family of DBDs. (b) Mechanisms of binding site diversification in eukaryotes beyond adding new DBDs. Prokaryotic binding sites are tailored to family-specific DBDs, with high information content and long binding sites. Following paralog expansion, eukaryotes diversified their binding site preferences by adding (left) tandem arrays of DBDs to specify longer or more complex sequence footprints and (right) new interaction domains that allowed for combinatorial binding logic. Other abbreviations: bHLH, basic helix-loop-helix; bZIP, basic leucine zipper; POU, domain shared by Pit-1, Oct-1, and Unc-86.

EUKARYOTES EXPLOIT COMPLEX DOMAIN STRUCTURES TO ENHANCE TRANSCRIPTION FACTOR DNA BINDING SPECIFICITY

Eukaryotes use several distinct combinatorial strategies to increase genomic target specificity. One strategy equips individual TFs with more than one DBD of the same family (Figure 2b) and, less frequently, with DBDs from different families (Lambert et al. 2018). In principle, this can lead to longer protein-DNA footprints with higher binding specificity, approaching that of prokaryotes. For example, human C2H2 ZF TFs can contain as many as 40 ZF domains, each recognizing 3 to 4 bp (Fedotova et al. 2017, Lambert et al. 2018). However, most ZF TFs have fewer fingers or use only subsets of fingers to recognize particular binding sites (Barazandeh et al. 2018, Fedotova et al. 2017, Najafabadi et al. 2015). Nevertheless, some ZF protein binding sites can be as long as 60 bp; an example is CTCF, which helps orchestrate chromosome architecture (Nakahashi et al. 2013). In a variant of this strategy, the human tumor suppressor p53 combines a DBD with a tetramerization domain (Wang et al. 1994, Weinberg et al. 2004), creating a four-subunit complex that recognizes an unusually long (>20-bp) binding site (Funk et al. 1992, Rastogi et al. 2018). It is noteworthy that both ZF TFs and p53 tetramers often exert highly specialized functions in the cell that are not restricted to specific cell types, thus allowing for a prokaryotic-type approach to recognize genomic target sites.

An alternative strategy combines DBDs with domains that mediate homo- or heterotypic interactions with other proteins (Figure 2b). This type of interaction is seen in some of the heavily expanded TF families in the human genome. For instance, bZIP proteins form a large variety of homo- and heterodimers (Miller 2009, Rodriguez-Martinez et al. 2017, Vinson et al. 2002). Cooperative TF complexes can also be formed by exploiting novel architectures beyond the DBD, such as the use of short linear motifs to mediate protein-protein interactions (Davey et al. 2012, Jolma et al. 2015, Merabet & Mann 2016, Slattery et al. 2011). Although new sequence specificities are created and footprints are extended by combining two TFs in this manner, this mechanism is typically not sufficient to bind a unique set of sites in vivo.

A few known examples of complexes consist of three or more cooperatively interacting TFs. The most prominent example is the IFN-β enhanceosome, in which eight factors are assembled on a DNA sequence with very stringent constraints on binding site position and orientation (Thanos & Maniatis 1995); given the paucity of protein-protein contacts seen in IFN-β crystal structures (Panne 2008, Panne et al. 2007), the assembly of the enhanceosome appears to be driven by a highly optimized DNA sequence. However, large-scale assemblies with strict constraints on binding site architecture seem to be the exception rather than the norm: While there are many examples of cooperatively bound TF pairs within enhancers (Jolma et al. 2015, Merabet & Mann 2016), most eukaryotic enhancers tolerate a much looser binding site arrangement (Arnosti & Kulkarni 2005, Panne 2008, Spitz & Furlong 2012).

Complex formation by pairs of TFs can reveal latent specificities (Slattery et al. 2011), whereby the binding preferences of the complex cannot simply be explained as a combination of the individual TF specificities (Ansari & Peterson-Kaufman 2011, Jolma et al. 2015, Mann & Chan 1996, Slattery et al. 2011). For instance, the Drosophila melanogaster Hox protein Scr is sensitive to variations in the DNA minor groove width only when bound in complex with its HD cofactor Exd (Abe et al. 2015, Zhou et al. 2015); Exd positions the N-terminal arm of Scr near the minor groove through a direct protein-protein interaction, which explains why this sensitivity is not observed when Scr binds as a monomer (Joshi et al. 2007). While latent specificity extends the classical concept of cooperative binding (Oehler et al. 1990) and thus contributes to the diversification of TF specificities, this gain in specificity is not sufficient to define a small subset of specific binding sites in eukaryotic genomes.

Recently, a direct link was made between DNA shape and TF affinity: By solving the X-ray crystal structures of the same Exd-Hox heterodimer bound to four different DNA binding sites of different affinity, a relationship between affinity and intrinsic DNA shape was revealed (Zeiske et al. 2018). All four ternary complexes had very similar protein and DNA structures, regardless of binding site affinity. However, although the predicted structures of unbound high-affinity binding sites were similar to the bound DNA shape, the shapes of lower-affinity sites were different, making binding energetically less favorable (Zeiske et al. 2018). Thus, in addition to intrinsic DNA shape playing a role in paralog binding specificity, differences in intrinsic DNA shape can also impact TF affinity.

Epigenetic DNA modifications also have the potential to modulate TF binding affinity and specificity. The most extensively studied case is the effect of CpG cytosine methylation on TF binding. Several studies have shown that methylation can influence TF binding both in vivo (Domcke et al. 2015) and in vitro (Kribelbauer et al. 2017, Mann et al. 2013, Yin et al. 2017, Zhu et al. 2016, Zuo et al. 2017). Depending on the TF, CpG methylation can both increase and decrease DNA binding affinity. Moreover, CpG methylation can alter DNA binding affinity in a paralog-specific manner, creating new low-affinity binding sites that are preferentially bound by specific members of the same TF family (Kribelbauer et al. 2017, Yin et al. 2017).

SPECIFICITY-AFFINITY TRADE-OFF: LOW-AFFINITY BINDING FACILITATES TRANSCRIPTION FACTOR PARALOG SPECIFICITY

The strategies outlined above help us understand how eukaryotes have been able to exploit cooperative binding to increase specificity from the perspective of an individual TF complex that binds across the genome. However, such strategies do not resolve the specificity problem from the point of view of an individual binding site that must recruit a particular TF. This paradox is especially striking for the eukaryote-specific family of HD TFs (Burglin & Affolter 2016, Charoensawan et al. 2010): Two large-scale studies comparing HD binding specificities found that most HD TFs bind relatively short (6–8-bp) sequences with a high degree of overlap between paralogs (Berger et al. 2008, Noyes et al. 2008). Given this degeneracy, the optimal site for a given HD will typically also be a high-affinity binding site for many other family members (Crocker et al. 2015). In addition, unless the available set of paralogs is restricted by cell type–specific TF expression, high-affinity binding sites appear to be useful only when neither paralog specificity nor tissue specificity is required (Mann et al. 2009). Therefore, the question is, how can a binding site be created that has high enough affinity for its target TF but also favors that TF over its close paralogs?

Recent work has demonstrated the importance of low-affinity sites in paralog-specific gene regulation. Two studies showed that replacing low-affinity sites with high-affinity ones resulted in ectopic gene activation (Crocker et al. 2015, Farley et al. 2015), suggesting that paralogous TFs in other tissues ignore the endogenous low-affinity site, yet occupy the artificial high-affinity one. In other work, the modification of Paired HD binding sites upstream of Drosophila Rhodopsin gene promoters resulted in altered spatial expression patterns, implying that subtle differences in affinity can cause a switch in sequence selectivity from one HD TF to another (Rister et al. 2015). More recently, low-affinity sites were shown to be important for distinguishing between the binding of and regulation by two different K50 HD proteins in Drosophila: Bicoid and Orthodenticle (Datta et al. 2018).

Low-affinity binding sites were also required for the correct interpretation of Hedgehog (Hh) signaling: Multiple low-affinity Cubitus interruptus (Ci) sites in Hh-regulated enhancers were required for activation in cells receiving low Hh signaling, whereas enhancers with artificially generated, high-affinity sites caused Ci to behave as a repressor (Ramos & Barolo 2013). Other studies have demonstrated the ability of low-affinity sites to distinguish between repression (by Senseless) and activation (by Pax2) in the Drosophila peripheral nervous system (Zandvakili et al. 2018) and to properly time the expression of important developmental genes (Gaudet & Mango 2002).

Another notable example is a Drosophila mesodermal enhancer controlled by the Ets domain paralogs Yan and Pnt. High-affinity Yan sites are required for repression, while low-affinity Pnt sites are required for activation (Boisclair Lachance et al. 2018). Because of Yan’s relative preference for the high-affinity sites, low levels of this TF are sufficient to prevent activation by Pnt, again suggesting that differences in binding site affinity of different TF paralogs can be exploited to fine-tune enhancer activities.

Taken together, these findings indicate that functional low-affinity binding sites (a) are common in eukaryotes, (b) contribute to paralog specificity, and (c) are capable of fine-tuning gene expression patterns both spatially and temporally. Low-affinity binding sites also have the advantage of requiring less stringent sequence conservation, as it is easier to recreate multiple low-affinity sites than to maintain a single high-affinity site. Consistent with this notion, low-affinity sites in functionally conserved enhancers can exhibit rapid turnover in closely related species (Crocker et al. 2015, 2016). More generally, weak cooperative interactions have been proposed to contribute to the large increase in the number of processes that eukaryotic cells need to execute (Gao et al. 2018). However, given that genomes contain many more low-affinity sites than high-affinity ones, it is still difficult to fathom how TFs can bind to their desired targets in a manner that is not only paralog specific but also locus specific.

THE CRUCIAL ROLE OF LOCAL TRANSCRIPTION FACTOR CONCENTRATION

To understand how eukaryotic cells achieve significant binding at low-affinity binding sites, it is helpful to consider two parameters that are equally important in determining the average TF occupancy at a particular DNA binding site (Figure 3a). The first of these, [TF]_free, is the local concentration of unbound, or free, DNA binding domains; the second, K_d, is the dissociation constant, which quantifies the strength of the protein-DNA interaction and is inversely proportional to affinity. The relationship between [TF]_free and [TF]_total is highly context dependent and difficult to predict. However, as long as thermodynamic equilibrium conditions hold on the relevant spatial and temporal scales, average TF occupancy depends only on the ratio of [TF]_free to K_d (i.e., is proportional to both [TF]_free and the affinity of the binding site). When [TF]_free equals K_d, the bound state and the unbound state are equally probable, resulting in an average occupancy of 50%. At low TF concentrations, the occupancy is proportional to [TF]_free; at high concentrations, occupancy saturates at the maximum value of 100%. To maintain significant occupancy at binding sites of lower affinity (higher K_d), the TF concentration needs to increase.

Figure 3.

Transcription factor (TF) occupancy is a function of both binding site affinity and TF concentration. (a) TF binding to a specific site depends on both the free TF concentration ([TF]_free) and the K_d (the inverse of affinity) of the binding site. TF occupancy as a function of free TF concentration is shown for six different binding site affinities (with K_d from 1 nM to 100 μM). Depending on the concentration and K_d, TFs are either unbound ([TF]_free ≪ K_d) in a responsive regime, where occupancy varies with concentration ([TF]_free ~ K_d), or saturated ([TF]_free ≫ K_d). (b) The influence of nuclear compartment size on TF concentration and on the ability of sites of different affinity to be bound by the TF. Three different concentration regimes (low, intermediate, and high) are highlighted. On the low-concentration end, having only a single TF molecule inside a nucleus with a diameter of 2–10 μm (green) yields [TF]_total ~ 10 pM, which represents an upper bound on [TF]_free. At such a low concentration, a weak binding site (K_d ~ 1 μM) would essentially be unbound; to achieve significant occupancy, 100,000 TF molecules would be required, resulting in a nuclear concentration of ~1 μM, similar to the K_d. Eukaryotic nuclear architecture brings groups of regulatory elements into proximity in 3D, forming transcriptional hubs of ~100 nm in diameter (orange). Restricting a single TF molecule to a subnuclear compartment of this size also corresponds to a TF concentration of 1 μM. Recruiting a larger number of free TF molecules to the hub would allow even ultralow-affinity sites with K_ds of up to 100 μM to be occupied, while saturation would no longer be limited to just the near-optimal sites (K_d ~ 10 nM).

It is informative to consider how many molecules of a particular TF would be required to reach ~50% occupancy for a given K_d range (Figure 3b). Having one TF molecule in a typical eukaryotic nucleus with a diameter of ~6 μm corresponds to a total nuclear TF concentration of ~10 pM, which is an upper bound on [TF]_free. This value is far lower than the typical K_d values (~10 nM) for optimal TF binding sites, which would therefore rarely be bound under these conditions. Having 100,000 TF molecules in the nucleus would correspond to a concentration of ~1 μM, at which even low-affinity sites with a K_d of ~1 μM (i.e., 100-fold weaker than a typical high-affinity site) are significantly bound. However, if the distribution of TF molecules within the nucleus is uniform, all binding sites will experience the same [TF]_free, implying that binding at all high-affinity sites will be saturated and that therefore these sites will no longer be responsive to variation in TF concentration. That eukaryotes would choose to have their TFs constitutively occupy a large number of sites in the genome seems implausible, and indeed TF concentration levels may be tuned to the K_d values of their top target sites (Brewster et al. 2014, Gerland et al. 2002). However, functionally relevant low-affinity binding sites will not be sufficiently occupied to impact gene expression at TF concentrations at which saturated binding is avoided for the strongest binding sites in the genome.

PHASE SEPARATION, TRANSCRIPTIONAL HUBS, AND SUBNUCLEAR COMPARTMENTS

In any given cell type, only a portion of the genome is accessible for TF binding (Guertin & Lis 2013), as shown in both DNase-seq (Hesselberth et al. 2009) and ATAC-seq (Buenrostro et al. 2013) datasets. While nonuniform chromatin accessibility may appear to address the binding site selection problem, there are still far too many binding sites within accessible regions for most TFs. Moreover, accessible regions in chromatin are themselves a consequence of prior TF binding and activity, which further raises the question of how these regions are selected to begin with.

A more complete answer to the target specificity conundrum may lie in the highly compartmentalized structure of eukaryotic nuclei (Dixon et al. 2012, Furlong & Levine 2018, Nora et al. 2012). During the past decade, chromosome conformation capture (3C) and derived high-throughput assays (Dekker et al. 2002, Lieberman-Aiden et al. 2009) as well as DNA imaging studies (Boettiger et al. 2016, Chen et al. 2014, Chen et al. 2015, Fraser et al. 2015, Giorgetti & Heard 2016) have revolutionized our understanding of 3D genome architecture. Imaging studies with tagged TFs have shown that they are nonrandomly distributed within both fixed and unfixed nuclei and form clusters with high protein and polymerase content (Cisse et al. 2013, Liu et al. 2014, Mir et al. 2018, Tsai et al. 2017; see Furlong & Levine 2018 and Liu & Tjian 2018 for recent reviews).

Importantly, these studies suggest the existence of spatial clusters of regulatory elements that are potentially coregulated by the same set of TFs. An intriguing advantage of such subnuclear compartments, or transcriptional hubs, is that TF concentration is likely to differ greatly from the nuclear average. For example, if one restricts a single TF molecule to a compartment with a diameter 10% that of the nucleus, its concentration increases 1,000-fold. A single TF molecule in a typical nuclear hub with a diameter of ~100 nm (Furlong & Levine 2018) corresponds to a local [TF]_free of ~1 μM, equivalent to having 100,000 TF molecules distributed uniformly throughout the nucleus (Figure 3b).

The 3D compartmentalization of the nucleus thus provides a powerful mechanism for increasing local [TF]_free for specific regulatory elements inside a hub. It also immediately provides a rationale for the pervasiveness of functional low-affinity binding sites: Any binding site with a K_d below 1 μM would be fully saturated inside a hub, a quality not desirable for tunable gene expression. In contrast, the K_ds of low-affinity targets are likely within a tunable range, even at the high TF concentrations occurring within hubs. Control of specific enhancers inside a hub (Figure 4) can be further refined using latent specificity (Slattery et al. 2011), TF cooperativity (Jolma et al. 2015), binding site syntax (i.e., the relative order, orientation, and spacing of binding sites) (Farley et al. 2016), and clustering of low-affinity sites (also referred to as homotypic clusters of TF binding sites) (Ezer et al. 2014, Gotea et al. 2010) (Figure 4a). Moreover, the high TF concentration within transcriptional hubs allows low-affinity sites to be bound, which in turn provides the flexibility for TF paralogs to functionally diverge (Figure 4b,c).

Figure 4.

Low-affinity binding sites and their importance in guiding transcription factors (TFs) to their cognate target sites. (a) The role of clusters of low-affinity sites in guiding TF binding within transcriptional hubs. Since local TF concentration inside a hub is >1 μM (see Figure 3b), the preferred K_d for responsive TF binding sites is also 1 μM or higher, which is low relative to optimal sites. Since many different DNA sequences can be used to realize low-affinity binding, each enhancer inside the hub has a high probability of harboring at least one low-affinity binding site, resulting in ambiguity. (Left) An extreme case in which each of the four enhancers has one low-affinity binding site within the appropriate K_d regime. Averaged over time or a population of cells, the TF will occupy each enhancer equally, resulting in a uniform distribution with no specificity for the target enhancer. (Right) Clusters of low-affinity binding sites confer target enhancer specificity by increasing its share in TF binding. On average, the TF will now predominantly occupy its target enhancer, which has the largest cumulative binding affinity. (b,c) The importance of low-affinity sites in paralog-specific binding. (b) Sequence specificities of the closely related Drosophila melanogaster TF paralogs Labial (Lab) and Ultrabithorax (Ubx), represented as energy logos (Foat et al. 2006). The optimal, high-affinity site for Lab is strongly bound by both Lab and Ubx (left logo) due to their shared DNA recognition strategy. The sequences that maximize specificity for Lab and Ubx, respectively (middle and right logos), are low affinity (~10,000-fold weaker than the optimal site) and, unlike high-affinity sites, have equal contributions across nucleotide positions within the binding interface. (c) Replacing a low-affinity, yet highly specific, binding site within the target enhancer for one paralog (TF1, purple) with a high-affinity site will cause loss of paralog-specific binding. TF2 (green) will no longer predominantly occupy its target enhancer; instead, both TF1 and TF2 will be directed to the same enhancer.

While nuclear hub assembly provides a plausible explanation for how low-affinity binding sites can effectively be occupied in cells, it raises the question of how these microenvironments form in the first place. Low-complexity protein domains, also termed intrinsically disordered regions (IDRs), are present in many nuclear proteins (Shin & Brangwynne 2017, Zhu & Brangwynne 2015) and TFs (Chong et al. 2018; see Alberti et al. 2019 for a recent review). Recent studies suggest that IDRs can promote the creation of distinct microenvironments via a process referred to as phase separation (Banani et al. 2016, Erdel & Rippe 2018, Hyman et al. 2014, Wheeler & Hyman 2018). Phase separation is thought to underlie the formation of heterochromatin via the biophysical properties of HP1 (Strom et al. 2017) and the 3D segregation of superenhancers into hubs, driven by specific combinations of TFs and mediator proteins (Hnisz et al. 2017, Sabari et al. 2018). Another notable example is the regulation of olfactory receptor (OR) gene expression, in which phase separation has been proposed to play a role in mediating long-range interchromosomal interactions that silence all but one of >1,000 OR genes in olfactory sensory neurons (Monahan et al. 2019). Finally, activation domains of several TFs phase-separate together with transcriptional coactivator proteins, and the formation of these condensates is involved in gene activation (Boija et al. 2018).

Despite these examples, how phase-separated TF hubs are formed is currently unknown. Their formation may emerge from a combination of multiple weak protein-protein and protein-DNA interactions (Figure 5). Regardless of the mechanism, we suggest that the ability of phase-separated hubs to form in higher eukaryotes enabled closely related TFs to regulate distinct target genes by taking advantage of paralog-specific low-affinity binding sites (Figure 5).

Figure 5.

Weak protein-protein interactions as a means to concentrate transcription factors (TFs) and assemble nuclear hubs. (a) The inhomogeneous 3D nuclear distribution of TFs provides a regulatory layer that complements the information encoded in the DNA sequence, by limiting TFs to distinct nuclear compartments and boosting their local concentration. (b) The assembly of transient, phase-separated protein condensates is mediated by weak protein-protein interactions between intrinsically disordered regions (IDRs) of both TFs and mediator (Med) proteins. (c) These microenvironments allow TFs to bind to paralog-specific low-affinity binding sites. A combination of weak protein-protein and protein-DNA interactions likely drive hub formation. Whether TF-DNA binding precedes phase separation, or vice versa, and what role chromatin plays remain to be determined.

CONCLUDING REMARKS

We discuss above how the occupancy by a particular TF along the genome may be shaped as much by the nonuniform 3D distribution of TF molecules within the nucleus as by the 1D landscape of DNA binding affinity (Figure 5). A key remaining challenge is to find ways to characterize and represent transcriptional hubs and the resulting profile of local TF concentration along the genome to enable quantitative predictions of TF occupancy. A convergence between genetics and cell biology, two fields that are historically distinct, will likely be required to achieve this goal.

SUMMARY POINTS.

1. There is a trade-off between a DNA sequence (a) having a high binding affinity for a particular TF and (b) being able to preferentially recruit that TF in favor of closely related, paralogous TFs from the same structural family.

2. The emergence of large TF families in metazoans through duplication and divergence has made it necessary for the cell to find ways to make use of low-affinity binding sites.

3. To what degree a particular binding site in the genome is occupied by a TF molecule depends on the ratio of [TF] to K_d, i.e., between the (effective, local, free) TF concentration [TF] and the dissociation constant K_d that quantifies the protein-TF interaction.

4. When a TF is restricted to a subnuclear volume with a diameter 10% that of the entire nucleus, its concentration goes up 1,000-fold, and thus the formation of phase-separated subnuclear compartments, or hubs, is a powerful strategy for making use of low-affinity sites.

5. Recent advances in high-throughput genomics and statistical learning are starting to make it feasible to build highly accurate models of protein-DNA interaction that can reliably detect low-affinity binding sites in genomic DNA.

6. To predict which promoters and enhancers are regulated by a particular TF, it is not sufficient to know its full genomic affinity profile; information about the TF’s 3D distribution within the cell nucleus is also important.

FUTURE ISSUES.

1. New methods are required for estimating local TF concentration, either by direct experimental observation or by computational inference.

2. Better strategies are needed to predict which TFs from among a set of close paralogs will preferentially bind to a given DNA sequence.

3. Methods are required to quantify the nonuniform distribution of TF hubs within eukaryotic nuclei.

4. We need a better understanding of how hubs emerge from a combination of weak protein-protein and protein-DNA interactions.

Glossary

Transcription factor (TF)

a protein containing at least one DNA binding domain along with domains that mediate interactions with cofactors and transcriptional machinery

Paralogs

related transcription factors in the same organism that belong to the same family; paralogs typically have similar DNA binding domains, yet different biological functions

DNA binding specificity

a quantification of the relative binding affinity of a transcription factor (or transcription factor complex) across all possible DNA sequences

Consensus sequence

a DNA sequence pattern specifying the frequently observed nucleotides at each position of a transcription factor binding site

IUPAC

International Union of Pure and Applied Chemistry (https://iupac.org/)

Position weight matrix (PWM)

a mathematical model that predicts binding affinity by quantifying the effect of base substitution at each position in the binding site

Protein binding microarray (PBM)

a method to characterize the DNA binding specificity of a transcription factor by using <105 immobilized DNA probes

Systematic evolution of ligands by exponential enrichment (SELEX)

a method for characterizing the DNA binding specificity of a transcription factor by using a random DNA library

Low-affinity binding site

a DNA site bound up to 1,000-fold more weakly than the optimal DNA sequence, but still more strongly than the immediately surrounding sequence

Transcriptional hub

a type of subnuclear compartment characterized by a high concentration of RNA polymerase and the site of active transcription

Intrinsic DNA shape

the structure of a DNA binding site prior to binding of a transcription factor; this shape may be different from the DNA shape in the transcription factor–bound complex

Transcription factor paralog specificity

the ability of a DNA ligand to recruit a particular paralog among multiple available transcription factors from the same structural famil

Transcription factor locus specificity

the ability of a specific genomic site to preferentially bind a transcription factor as opposed to other potential sites in the genome

Binding saturation

a situation that arises whenever the local free transcription factor concentration is much larger than the dissociation constant of the DNA binding site

Subnuclear compartment

a three-dimensional membraneless volume within a nucleus with distinct molecular composition; this compartment is transiently created by a complex network of protein-protein and protein-DNA interactions