Genome-Wide Views of Chromatin Structure

How Is the Stereotypical 5′ Promoter Architecture Specified?

The 5′ NFR typically contains one or more homopolymeric runs of polyA (or polyT) (10, 16). Poly-dA/dT runs are intrinsically stiff, and the bending required to wrap around the histone octamer is energetically costly for these sequences relative to random sequences (17–19), leading to decreased nucleosome incorporation on model genes in vivo, and, importantly, in vitro (16, 19–21). In addition to constituitive NFRs where nucleosomes are excluded by poly-dA/dT sequences, regulatable NFRs can be generated by the binding of certain proteins, such as Reb1 and Abf1, to their binding sites, which are not nucleosome depleted in in vitro reconstitutions but are strongly nucleosome depleted in midlog yeast cultures (21).

In contrast, AT-rich dinucleotides confer some bend to the DNA duplex, and thus spacing of these dinucleotides with 10-bp periodicity (aligning these bends in the same direction) results in DNA that is thermodynamically favored as a binding site for histones, because less energy must be expended in the bending of the DNA (22, 23). Genome-wide computational studies have recently described the analysis of pronucleosomal sequences in the yeast genome (24–27). One study utilized ~200 nucleosomal sequences to determine a dinucleotide position-specific sequence matrix (PSSM) for nucleosomes (26). Another used conservation of AA/TT periodicity across six Saccharomyces species to define a nucleosome-positioning score (NPS) (24). More recent studies have included discrimination against polyA runs (25, 27). The most notable finding of these studies has been that the first nucleosome in a typical yeast ORF (the +1 nucleosome) is often associated with strong pronucleosomal sequence elements. However, whole-genome data from in vitro reconstitution experiments demonstrate that antinucleosomal sequences explain far more of the in vivo patterning of chromatin than do pronucleosomal sequences (21, see Figure 4a).

Furthermore, while pronucleosomal sequences are most common at the +1 position, microarray studies show that the +2 and +3 nucleosomes are also generally well positioned. In 1988 Kornberg & Stryer proposed a “statistical positioning” model in which nucleosomes on random sequence could appear well positioned in population averages because of constraints imposed by packing many nucleosomes into a short region (28). A useful analogy is that of a can of tennis balls: The can imposes few constraints on the location of a single tennis ball, but three tennis balls will appear uniformly packaged thanks to the limited number of ways the small amount of free space can be distributed. This idea might well account for the surprising amount of order in yeast chromatin, since genes are fairly short (<2 kb average). Consistent with this idea, “fuzzy” or delocalized nucleosomes are typically found distal to NFRs and +1 nucleosomes, which are the two most likely candidates for the borders that constrain a given packaging unit (10, 13). The ideal test of this hypothesis—in vitro reconstitution of histone octamers into long or short shear distributions of genomic DNA, from saturating to limiting octamer: DNA ratios—remains to be reported.

3′ Nucleosome-Free Regions

Long nucleosome-depleted regions do not only occur at 5′ ends of genes—a surprising fraction of gene 3′ ends also exhibit NFRs (13, 14), which also appear to be programmed by antinucleosomal sequences. The function of these has not been explored in any detail, but it should be noted that many antisense transcripts initiate within these 3′ NFRs (13). These sites are often bound by the general transcription factor TFIIB, which has been proposed to contribute to gene looping, in which 5′ and 3′ ends of some yeast genes appear to interact in vivo (29), possibly enabling the local recycling of transcription machinery after each round of transcription.

Tata Versus Non-TATA Yeast Genes

Yeast genes can be grouped into two broad classes, based on the histone lysine acetylase (KAT) involved in Tata-binding protein recruitment to promoters (33–35). SAGA-dominated genes typically have TATA boxes, are stress responsive, are characterized by noisy, or “bursty,” expression, and are regulated by a wide range of chromatin remodeling factors. Conversely, TFIID-dominated genes lack TATA boxes, are expressed during active growth, exhibit little noise in expression levels, and are not affected by deletion of most chromatin-regulatory genes.

These two classes also exhibit differences in promoter chromatin packaging. The majority of yeast genes (~80%) are TFIID dominated, and the above descriptions of NFRs apply to this class in particular. The minority class of TATA-containing stress genes, on the other hand, exhibits more variable promoter architecture. This is true across different genes (i.e., various stress-responsive genes exhibit a range of promoter packaging states) and also appears to be true across individual cells, since these promoters often are associated with delocalized nucleosomes (12, 24, 36). Importantly, transcription factor binding sites at TATA-containing promoters are likely to be occluded by nucleosomes, although rapid exchange of nucleosomes at these promoters (see below) will allow binding sites to be accessed during transient time windows. This competition between nucleosomes and transcription factors might be expected to contribute to cell-to-cell variability (noise) in expression of downstream genes (36, 37).

Isw2

ATP-dependent chromatin remodelers utilize the energy of ATP hydrolysis to distort histone-DNA contacts, with eventual outcomes such as histone sliding, eviction, or replacement. Elegant work by Whitehouse & Tsukiyama showed that the ATPase Isw2p acts to position nucleosomes over unfavorable sequence elements at the POT1 promoter (38). Specifically, nucleosomes in isw2Δ yeast matched positions from in vitro reconstitutions, whereas in wild-type yeast the+1 nucleosome was located further 5′, narrowing the NFR and inhibiting transcription. A subsequent whole-genome study found that +1 nucleosomes or −1 nucleosomes were shifted toward the NFR (often over unfavorable poly-dA/dT tracts) at sites of Isw2 action (9). Promoters with shifted +1 nucleosomes were generally repressed, while repositioning of −1 nucleosomes at other promoters was surprisingly implicated in inhibiting antisense transcription.

Thus, we imagine one could prospectively identify locations where nucleosomes do not follow sequence cues in vivo, such as Isw2-regulated genes or PHO5, and this would reveal sites where cellular factors modulate chromatin architecture to regulate transcription or other processes. We anticipate that advances in computational predictions of thermodynamically favored nucleosome positions, coupled with identification of in vivo nucleosomes that do not match the in vitro/in silico predictions, will provide a rich source of information on the cellular machinery that regulates chromatin structure in vivo.

Histone Dynamics: Mechanism

H3 is replaced in yeast only over very highly transcribed genes, but at a given transcription rate, SAGA-dominated genes (34) exchangeH3 more rapidly than do TFIID-dominated genes (41). Why should the mechanism of regulation affect chromatin dynamics over coding regions? One possibility is that SAGA helps recruit histone-displacing factors to genes (45). Alternatively, we note that SAGA-dependent genes exhibit high levels of transcriptional noise (46), ascribed to “bursts” of transcription rather than to evenly spaced initiation events (47). When RNA polymerase moves through a histone octamer, it is believed to result in eviction of a H2A/H2B dimer (48), resulting in a hexamer that will either be repaired or evicted by collision with a second polymerase. Thus, the well-spaced polymerases at TFIID-dominated genes might be less likely to evict nucleosomes than would the bursts of closely spaced polymerases at SAGA-dominated genes: Pol2 density on genes might then account for differences between species in coding region turnover profiles.

At promoters, nucleosomes are generally rapidly replaced. While H3 replacement at several regulated promoters increased upon induction (41, 42), globally promoter H3 replacement was uncorrelated with RNA Pol2. We speculate that rapid promoter replacement could be related to the presence of “antinucleosomal” poly-dAdT sequences at many promoters. ATP-dependent chromatin remodelers are capable of sliding nucleosomes laterally, and in vitro even seemingly well-positioned nucleosomes can make short lateral excursions around their baseline position (49). Even if all nucleosomes in the genome were to make lateral excursions, those adjacent to antinucleosomal sequences could be evicted simply by being pushed onto the adjacent unfavorable sequences (50)—falling off a cliff, as it were. Thus, nucleosome eviction at promoters would result from an interaction between a location-specific intrinsic sequence tendency for histone eviction, and regulatable extrinsic factors such as level of ATP-dependent remodeler present or the extent of competition between nucleosome and transcription factor binding (36).

Histone Dynamics: Consequences

Replication-independent nucleosome replacement is likely to have functional consequences for cellular regulation of transcription, replication, etc. For example, transient DNA exposure by histone replacement may allow access of DNA-binding sites to DNA-binding factors. Indeed, it is entirely possible that nucleosome-“free” regions are in fact very transiently (or very loosely) associated with histones (44, 51), rather than statically naked—the fact that a typical NFR is about one nucleosome wide is consistent with this idea. But even nucleosomes that are highly occupied at steady state, such as the +1 nucleosome, are dynamic, leading to a picture of a nucleosome that is constrained translationally by sequence characteristics, yet which is evicted frequently and replaced rapidly, thereby transiently exposing the underlying DNA to binding factors (or allowing RNA polymerase to pass). Beyond site exposure, histone replacement shows an interesting association with two aspects of heterochromatin in yeast—heterochromatic regions are protected from histone replacement (41, 52), whereas the boundaries of these regions are associated with rapidly replaced nucleosomes (41, 51), suggesting that histone replacement serves to erase laterally spreading chromatin states (53) and thereby insulate chromatin domains from one another.

Typical Gene

Broadly speaking, two groups of modification have been mapped in yeast. One group consists of modifications that generally occur over transcribed regions, whose levels correlate with polymerase abundance, and which are possibly deposited by enzymes traveling with RNA polymerase. The other group consists of modifications that either correlate or anticorrelate with rates of replication-independent histone replacement. We hypothesize that their genomic localization patterns likely result from incorporation into the genome of histones drawn from a free pool carrying (or lacking) the modifications in this group. We elaborate on these groups below.

Transcription-Related Marks

A now-classic example of a transcription-related histone mark is H3K4 trimethylation. H3K4 is methylated by Set1p (66), which associates with the Serine 5-phosphorylated initiation form of RNA polymerase (67), and H3K4me3 is found over the 5′ ends of yeast genes at levels that correlate with transcription rate (68, 69). Similarly, H3K36me3 is found over the middle and 3′ ends of coding regions, deposited by a methylase (Set2p) associated with the Serine 2-phosphorylated elongation form of polymerase (70, 71).

A number of histone acetylation states (including H3K9ac, H3K14ac, H4K12ac, and many others) are found at the 5′ ends of coding regions. For these marks, this pattern apparently results from a combination of two effects. First, these residues are acetylated throughout the coding region during transcription. Subsequently, a histone deacetylase complex known as Rpd3S is recruited by the H3K36me3 found over the middle and 3′ end of coding regions, resulting in a shaping of the original coding region pattern to the 5′-biased pattern observed at steady state (70, 71).

Histone Modifications and Regulation of Transcription

Histone modifications often correlate with processes such as transcription, but genetic studies in yeast prove the old saw that correlation is not causation. For example, H3K4me3 occurs universally over the 5′ ends of transcribed genes, yet elimination of all H3K4me3 by deletion of SET1 is well tolerated by yeast, rather than being lethal as expected if all transcriptional initiation ceased. Furthermore, part of the phenotype of set1Δappears to result from Set1 methylation of a nonhistone substrate (72).

It is therefore crucial to emphasize what may be learned from localization studies. Specifically, localization of a mark simply identifies nucleosomes where a modifying enzyme has acted (and in some cases the process resulting in recruitment of the enzyme), not the function of the mark.

Perhaps the clearest example of this comes from studies of H3K36me3. H3K36me3 is deposited during RNA polymerase elongation, and is considered a transcriptional elongation mark, yet deletion of SET2 results in no identifiable effects on elongation per se (73). Instead, H3K36me3 deposition leads to deacetylation of lysines that were acetylated during polymerase passage. In the absence of K36me3 and resultant deacetylation, coding regions remain hyperacetylated, and so-called cryptic internal initiation sites become active (70, 71, 74). Thus, the function of this elongation mark is to reverse perturbations to a gene’s chromatin structure, which presumably aided polymerase passage through the coding region.

Replacement-Related Marks

The second group of histone modifications exhibits genomic localization patterns related to replication-independent histone replacement. Most notably, H4K16ac and H2BK16ac are found over coding regions (except over very highly expressed ORFs), but are depleted from the rapidly replaced nucleosomes flanking the NFR (68). Conversely, H3K56ac, best understood as a mark of newly synthesized histones assembled into chromatin during replication (75), is enriched in nucleosomes subject to replication-independent replacement (43, 76). The H3K56 acetylase Rtt109 can acetylate free histones, but not nucleosomal histones (77). We therefore believe that the localization of replacement-related histone marks results in part from their presence (H3K56ac), or absence (H4K16ac) in the free pool of histones, and subsequent incorporation via histone replacement.

Modulation of histone modification patterns by histone replacement has several interesting potential consequences. First, replacement of a given nucleosome would erase preexisting histone marks at that genomic location. This is of great interest, as it establishes an event horizon beyond which histone modification physiology cannot be discerned. Of course, if an enzyme remains associated with a genomic locus even after histone replacement, it will re-establish its mark—H3K4me3 is apparent at the +1 nucleosome of active genes, despite rapid turnover of +1 nucleosomes (68). However, the enrichment of H3K4me3 is greater at the +2 nucleosome than at the +1, suggesting that some fraction of nucleosomes in the population has been replaced but has not yet been remodified. In any case, marks leading directly to histone eviction might be difficult to find by their nature, requiring mutation of downstream turnover machinery or high temporal resolution kinetic studies of gene activation for their identification.

More speculatively, we note that histone replacement exhibits the potential for positive feedback. We can imagine a first histone replacement event in a given cell cycle, in which a H4K16ac, H3K56deac nucleosome is replaced with a H4K16deac, H3K56ac nucleosome. H3K56ac aids histone replacement, and this appears to be primarily an effect of enhanced eviction rather than incorporation (43, 76, 78). Thus, the first nucleosome evicted at a location should be more difficult to evict than subsequent H3K56ac nucleosomes. The loss of H4K16ac during replacement could have similar consequences via a K16deac->Bdf1->Swr1 pathway of Htz1 incorporation (55, 56, 79–81). Such local positive feedback loops could affect features ranging from gene induction kinetics to expression noise.

Another interesting feature of rapidly replaced nucleosomes is that replacement results in nucleosomes that are expected to have a relatively high affinity for one another. For example, H4K16 interacts with an acidic patch on H2A in an adjacent nucleosome in the nucleosome crystal structure (82), and K16 acetylation inhibits compaction of nucleosome arrays into 30-nm fiber in vitro (83). Thus, the pair of +1/−1 nucleosomes lacking H4K16ac at promoters might be expected to preferentially contact each other, or perhaps H4K16deac nucleosomes at other locations in the genome. Nucleosomes around 3′ NFRs also exchange relatively rapidly (41) and lack H4K16ac, so this phenomenon could contribute to the 5′ to 3′ looping of genes proposed to play a role in recycling of transcriptional machinery (29). Deciphering the influence of these factors on chromosome folding will be of great future interest.

Conserved Features of Chromatin Structure

Many features of chromatin structure are conserved from yeast to mammals. Nucleosome placement appears constrained by some of the same sequence preferences as in yeast, because patterns of dinucleotide repeats identified in yeast nucleosomes and sequences that are depleted in yeast nucleosomes can partially predict nucleosome occupancy in chicken and human chromatin (26, 27). Nucleosome-excluding sequences are widespread but not ubiquitous at transcriptional start sites (TSSs) of mammalian genes, and are enriched at ubiquitously expressed genes (87). Mammalian chromatin also exhibits nucleosome-free regions (NFRs) of approximately 200 bps centered −85 bp upstream of the transcriptional start site and surrounded by positioned nucleosomes; this has been documented indirectly by genome-wide measurement of DNAse hypersensitive sites (DHSs) (6) and by direct measurements of histone occupancy (88–90). The flanking of NFRs by well-positioned nucleosomes is consistent with the idea that NFRs themselves, or the machinery that forms them, have an instructive role in nucleosome positioning.

Next, as in yeast, histone modifications in mammals strongly reflect the anatomy of genes. For example, the start and first ~500 bps of an active gene are typically occupied by H3K4me3; H3K4me2 peaks in the body of the gene with a shallower gradient, and H3K4me1 is weakly localized distally. Also, as in yeast, H3K36me3 is associated with transcriptional elongation in human and mouse, and has been exploited to map the lengths of protein-coding and -noncoding transcripts (91). In both yeast and human cells, genome-scale mapping of asymmetrically methylated H3R2 showed that it is localized to the body of genes but absent from promoters (92, 93), although its level over genes is independent of transcription level. H3R2me2a and H3K4me3 each inhibit placement of the other mark, and this mutual inhibition likely contributes to their distinct localization at 5′ or the body of genes, respectively.

All known histone methylation and acetylation states were mapped in a series of wide-ranging studies from the Zhao group (94, 95). Figure 1b summarizes results for the methylation and acetylation states, aligned by coding regions. Active transcription in human cells is correlated with K4 and K36 methylation, along with additional histone methylations (Figure 1b). In contrast to the complex patterns of histone methylation, analysis of 18 histone acetylations showed that they are all positively correlated with transcription (90, 94–96), particularly at the 5′ ends. Furthermore, as in yeast, histone modifications tend to be found in groups of correlated marks. In human T cells, a common pattern of 17 histone modifications was associated with approximately 25% of promoters, and deviants from this pattern are all individually rare (and much of this deviation is likely to represent an artifact of using thresholding as a computational tool for analysis of genomic localization data).

Complexity in Modification Placement and Removal

We examine H3K4 methylation to illustrate this point. At H3K4, me3 and me2 are associated with transcribed genes, me1 is associated with enhancers (see below), and unmethylated H3K4 (H3K4me0) is associated with gene repression. Thus, each methylation state of H3K4 can, in principle, transduce a distinct biological signal. This fine discrimination is achieved by expansion and increased selectivity of protein lysine methyltransferases (KMT), protein lysine demethylases (KDM), and adaptor proteins that recognize these distinct modifications. In the case of H3K4, at least eight KMTs can methylate H3K4me0 all the way to H3K4me3. However, the H3K4 demethylases show distinct specificity: LSD1/KDM1 cannot demethylate H3K4me3 but will efficiently demethylate H3K4me2/1 to H3K4me0. Conversely, at least four KDMs in the KDM5 subfamily can demethylate H3K4me3/me2 to the H3K4me1 state. Some, but not all, Jumonji domain KDMs can also demethylate histone lysine me3 all the way to me0 (102). Thus, there is increased enzymatic specificity in KDMs to distinguish specific residues and individual methylation states. ChIP-chip studies of KMTs and KDMs are an efficient strategy to determine which enzyme is responsible for histone modifications at specific genes (103–107).

Furthermore, while many histone modifications common to yeast and humans are similarly distributed, notable exceptions exist. For example, H3K4me1 in mammals is found over long-distance enhancers (see below). H3K79me3, which blankets yeast coding regions, exhibits a complex pattern of localization in humans; it is found in tight peaks in the TSS of a subset of the highest expressed genes, but otherwise is observed in silent genes.

Complexity in Histone Modification Readers

Second, metazoans have seen a dramatic expansion of the number and diversity of binding domains for the various histone modifications. Distinct H3K4 methylation states are recognized by specific protein binding partners, which couple recognition ofH3K4methylation states to specific gene regulatory events. For instance, H3K4me3 may be linked to transcriptional activation, acute gene silencing, or DNA recombination by the PHD fingers of TAF3 (a basal transcription factor) (108), ING1 (a tumor suppressor) (109), or RAG2 (a recombination factor) (110, 111), respectively. In addition, the expanded “Royal superfamily” of methyl-lysine readers include the Tudor domain, chromodomain, PWWP domains, and Malignant Brain Tumor (MBT) domains (112). These examples reinforce the concept that histone modifications can transduce biological function in a manner independent of their biophysical effect on chromatin fiber; rather, in some cases, recognition of the histone modification by specific readers mediates distinct biological outcomes (63). One possible way to learn how complex chromatin modifications are decoded into distinct biological outcomes in the future is to map the occupancy patterns of modification binders genome wide (113).

Expansion of Histone Modifications

Mammalian chromatin also features more distinct histone modifications than yeast. For instance, gene silencing is enforced not just by histone hypoacetylation, or H3K9me3 (as in Schizosaccharomyces pombe), but is also signaled by H3K27me3 (mediated by the Polycomb repressive complex), by H2A monoubiquitination, or by association with the H2A variant macro-H2A. Curiously, in genomic maps, repressive marks such as H3K27me3 show only modest anticorrelation with transcription (94). Mapping of H3K27me3, H3K9me3, and DNA cytosine methylation in 12 human and mouse cell types revealed that most silent genes are associated with just one of the above modifications (114). This implies that these marks are not used redundantly, but mark silent genes of different functional categories, or genes that are slated for distinct modes of coregulation. Indeed, H3K27me3 occupancy is highly enriched over homeodomain-containing developmental regulators, whereas H3K9me3 and H4K20me3 are highly enriched over zinc finger transcription factors (94, 115), but not globally associated with silent genes. These observations support the idea that histone methylation associated with gene silencing is not merely a consequence of a lack of transcription, but rather reflects active mechanisms that enforce gene silencing toward distinct ends.

X Chromosome Inactivation

A canonical example of a large-scale, specialized chromatin domain is the X chromosome in female mammalian cells. Here, one of the two X chromosomes is transcriptionally silenced by a process termed X chromosome inactivation (XCI) (reviewed in Reference 129). The inactive X is a prototype of constitutive heterochromatin: It is transcriptionally silent, enriched in H3K27me3 and H3K9me3, heavily DNA methylated, late replicating, and cytologically condensed during interphase. Initial choice of XCI among female somatic cells is random—either X can be inactivated. The choice of XCI is dictated by the transcription of a ~15 kilobase ncRNA termed XIST from the future inactive X. XIST binds to and spreads over the inactive X, and initiates H3K27 methylation and silencing of the XCI. Ectopic expression of XIST on a human autosome is sufficient to silence a large contiguous portion, but not the entirety, of the autosome (130). The choice of which X chromosome to inactivate involves interaction between XIST and competing overlapping antisense ncRNAs, mediated by CTCF.

The mechanism of spreading of the X silencing complex is not well understood, but genomic mapping studies in Drosophila and Caenorhabditis elegans have yielded insights into this process (although dosage compensation in these organisms differs in detail from that in mammals). In Drosophila, the MSL dosage compensation complex is recruited to the X chromosome in male cells by the X-specific roX noncoding RNAs (131), where it binds actively transcribed genes and up-regulates their transcription by approximately twofold (potentially because of its activity as a H4K16 acetylase). MSL targeting to actively transcribed genes depends on the binding of subunit MSL3 to the elongation mark H3K36me3, and depletion of H3K36me3 reduces MSL binding to X (132). In C. elegans, dosage compensation occurs by halving transcription across each X chromosome in XX cells, leading to the same transcriptional output as the XO male. The worm dosage compensation complex (DCC) binds discrete foci across the X; these binding foci are distinguished by clustering of a specific DNA sequence motif, but also by proximity to upstream regions of transcriptionally active genes (133). The positive correlation between the level of DCC binding and transcription rates of the neighbor gene suggests the potential use of transcription to tune the level of DCC binding and spreading. In sum, despite the use of distinct molecular machinery, dosage compensation in several species involves the gender-selective targeting and chromatin modification-mediated discontinuous spreading over one or more sex chromosomes.

Allele-Specific Gene Expression

Mammalian somatic cells are diploid; each gene is present in two copies (alleles) on homologous chromosomes, one inherited from each parent. For a subset of genes, termed imprinted genes, only the paternal or maternal allele is active, and each allele is associated with a specialized chromatin state (134). Nonimprinted genes can also be expressed in an allele-specific manner: Specialized genes, such olfactory receptors or immunoglobulin genes, possess intricate mechanisms to ensure monoallelic exclusion, but recent evidence suggests that perhaps ~1000 human genes can demonstrate monoallelic expression (135). If single-nucleotide polymorphisms can be identified between the two relevant alleles, then either microarrays or sequencing can be used along with ChIP to identify allele-specific chromatin states. In one such analysis, 4% of RNA polymerase II occupancy sites in the genome of diploid fibroblasts (>450 genes) showed allele-specific bias (136). Although allele specificity in genome-wide chromatin maps has yet to be fully explored, we can anticipate rich information based on the known importance of chromatin modifications in imprinted genes.

The HOX Loci

All bilaterians have a segmented body plan, in which the HOX homeodomain-containing transcription factors play a major role. In mammals, 39 HOX genes are clustered on four chromosomal loci, and are expressed in a nested segmental fashion along the anterior-posterior and proximal-distal axes of the body, with each additional HOX gene being expressed in increasingly posterior or distal anatomic sites. The pattern of HOX gene expression along the anterior-posterior axis is mirrored by their physical location on the chromosomes, with 3′ HOX genes expressed more anteriorly and 5′ HOX genes more posteriorly. HOX expression is maintained through adulthood—in the skin, dermal fibroblasts from different positional identities maintain features of the embryonic patterns of HOX expression (137, 138), and this adult HOX code is required to drive the site-specific gene expression programs of these cells (139). That cells maintain their positional memory over time—in humans, over decades—predicts the existence of a robust epigenetic system to ensure the faithful transmission of transcriptional memory.

Early genetic studies of homeosis identified mutations that do not affect initial pattern formation, but are required for pattern maintenance (140). The Trithorax (Trx) family of genes are required for continued activation of HOX genes, and encode H3K4 KMT complexes and H3K4me3 binding proteins. Conversely, Polycomb group genes (PcG) are required for continued silence of HOX and other developmental genes and encode at least two main complexes. The Polycomb Repressive Complex 2 (PRC2) is an H3K27 KMT, whereas the Polycomb Repressive Complex 1 (PRC1) recognizes H3K27me3 and mediates H2A ubiquitination and nucleosome compaction. Unlike the chromatin organization at most gene loci, active HOX genes are marked by broad H3K4me3/2 that spans multipleHOX genes and their intergenic regions; these regions are also broadly bound by the Trx protein MLL1 (103). Conversely, the silent HOX genes are occupied by large continuous blocks of H3K27me3 and Polycomb group proteins (104, 105). Further, occupancy of H3K27me3 KDM UTX constitutes an additional, independent layer of regulation by targeting the beginning of HOX genes (Figure 2).

How are these long tracks of chromatin modifications established and maintained? In Drosophila, binding elements for Polycomb and Tithorax proteins have been identified (termed PREs), and PcG and Trx proteins are apparently restricted to these sites even though their cognate histone marks, H3K27me3 and H3K4me3, can spread for kilobases around them (140). In contrast, mammalian PREs have yet to be identified, and PcG and Trx proteins broadly occupy the HOX loci. The original genetic studies in flies by Lewis and Hogness identified homeotic mutations that mapped to nonprotein coding regions of the HOX loci, many of which later turned out to generate long noncoding RNAs and microRNAs (see sidebar titled Noncoding RNAs and Long-Distance Targeting of Chromatin Regulators). Drosophila PREs can also be transcribed, which may alter their accessibility and favor Trx binding over PcG binding (141–143). In humans, the four HOX loci harbor a surprisingly large number of noncoding transcripts; there are 39 HOX genes, as compared to 231 transcribed noncoding regions, many of them highly conserved (144). These noncoding RNAs are also expressed in an anatomic position-specific manner, and, importantly, the maintenance of appropriate site-specific HOX chromatin modification and gene expression require the action of HOX ncRNAs. A 2.2-kB HOX ncRNA termed HOTAIR encoded in the HOXC locus binds to the PRC2 complex and is required for PRC2 occupancy, H3K7me3 occupancy, and transcriptional silencing over dozens of kilobases of the HOXD locus on a different chromosome (144). Thus, HOTAIR RNA may guide PRC2 to the HOXD locus to mediate H3K27me3 and gene silencing. However, physical occupancy of HOTAIR on HOXD locus has not yet been shown, and indirect models of PcG positioning by RNA are also possible.