Homeodomain proteins: an update

Here, we provide an update of our review on homeobox genes that we wrote together with Walter Gehring in 1994. Since then, comprehensive surveys of homeobox genes have become possible due to genome sequencing projects. Using the 103 Drosophila homeobox genes as example, we present an updated classification. In animals, there are 16 major classes, ANTP, PRD, PRD-LIKE, POU, HNF, CUT (with four subclasses: ONECUT, CUX, SATB, and CMP), LIM, ZF, CERS, PROS, SIX/SO, plus the TALE superclass with the classes IRO, MKX, TGIF, PBC, and MEIS. In plants, there are 11 major classes, i.e., HD-ZIP (with four subclasses: I to IV), WOX, NDX, PHD, PLINC, LD, DDT, SAWADEE, PINTOX, and the two TALE classes KNOX and BEL. Most of these classes encode additional domains apart from the homeodomain. Numerous insights have been obtained in the last two decades into how homeodomain proteins bind to DNA and increase their specificity by interacting with other proteins to regulate cell- and tissue-specific gene expression. Not only protein-DNA base pair contacts are important for proper target selection; recent experiments also reveal that the shape of the DNA plays a role in specificity. Using selected examples, we highlight different mechanisms of homeodomain protein-DNA interaction. The PRD class of homeobox genes was of special interest to Walter Gehring in the last two decades. The PRD class comprises six families in Bilateria, and tinkers with four different motifs, i.e., the PAIRED domain, the Groucho-interacting motif EH1 (aka Octapeptide or TN), the homeodomain, and the OAR motif. Homologs of the co-repressor protein Groucho are also present in plants (TOPLESS), where they have been shown to interact with small amphipathic motives (EAR), and in yeast (TUP1), where we find an EH1-like motif in MATα2.


Introduction
In 1994, we wrote a review on homeodomain (HD) proteins for Annual Review of Biochemistry together with Walter Gehring that became a standard reference . Sadly, Walter passed away in May 2014 as a consequence of a tragic traffic accident (Affolter and Müller 2014;Affolter and Wüthrich 2014;Levine 2014;Mlodzik and Halder 2014a;Mlodzik and Halder 2014b;Schier 2014;Wieschaus and Nüsslein-Volhard 2014). In his honor, we provide here an update of this review, although we will barely be able to scratch the surface, given that over 17,000 publications containing Bhomeobox,^Bhomeodomain,^or BHox^in the title or abstract have appeared since the first publications. We will highlight a few of the novel findings of the past two decades, with special emphasis on topics that were of particular importance to Walter Gehring.
Our understanding of the homeobox gene family has expanded substantially in the last 20 years, not least because the numerous completed genome sequences allow comprehensive analyses. While many findings and the basic framework from 1994 are still valid, numerous revisions and refinements have This article is dedicated to the memory of Walter Gehring.
Electronic supplementary material The online version of this article (doi:10.1007/s00412-015-0543-8) contains supplementary material, which is available to authorized users. been made since then with regard to classification and homeobox gene numbers per individual genome. More structural data has also become available, and the characterization of the molecular roles of HD proteins has tremendously advanced.
The homeobox was originally discovered as a shared sequence element of about 180 bp in homeotic genes in Drosophila melanogaster, which gave rise to its name (McGinnis et al. 1984a;Scott and Weiner 1984) (for review, see Bürglin 2013a;Bürglin 2013b;Pick 2015). Soon, it was realized that this motif was also conserved in vertebrates (McGinnis et al. 1984b), and the first vertebrate homeobox gene was cloned from Xenopus laevis by Andrés Carrasco and colleagues in the laboratory of Eddy De Robertis (Carrasco et al. 1984), across the hall from the laboratory of Walter Gehring at the Biozentrum in Basel. In a sad coincidence, Andrés Carrasco also passed away in May 2014 (Blumberg 2014).
The homeobox sequence encodes the HD, a globular domain of about 60 amino acids that normally functions as a DNA-binding domain. We now know that in animals, there are usually around 100 homeobox genes in protostome species, e.g., 103 in Caenorhabditis elegans (Hench et al. 2015), 103 in Drosophila melanogaster (Sup. Fig. S1), 121 in the sea snail Lottia gigantea (Simakov et al. 2013), 111 in the polychaete worm Capitella teleta (Simakov et al. 2013), and at least 92 in the oyster Pinctada fucata (Morino et al. 2013). In the leech Helobdella robusta, an expansion has taken place, resulting in 181 homeobox genes (Simakov et al. 2013). In the deuterostome branch, 96 homeobox genes were found in the sea urchin Strongylocentrotus purpuratus (Howard-Ashby et al. 2006), and 133 in Amphioxus branchiostoma . Most vertebrates have about 250 homeobox genes, due to two extra rounds of genome duplication and subsequent loss of paralogs (Holland 2013). In teleost fish, one additional round of genome duplication followed by gene loss increased the number to over 300 (Holland 2013). Overall, about 15-30 % of all transcription factors in animals are HD proteins (de Mendoza et al. 2013), which represents about 0.5-1.25 % of all proteins in a given species. In plants, similar numbers of homeobox genes can be found, e.g., 110 in Arabidopsis thaliana (Mukherjee et al. 2009). In fungi and single-cell organisms, the number tends to be small, usually less than a dozen (Derelle et al. 2007), but in Acanthamoeba, the homeobox family has expanded to 25 (Clarke et al. 2013). In a number of unicellular eukaryotes, homeobox genes seem to have been lost entirely (de Mendoza et al. 2013;Derelle et al. 2007), though in some cases, e.g., Paramecium, they were subsequently found (de Mendoza et al. 2013) (Sup. Fig. S1).
HD transcription factors fulfill a plethora of biological functions. There is probably no tissue in plants or animals that does not require them to function properly. In animals, they act from the earliest stages of development onward (Driever and Nüsslein-Volhard 1988;Töhönen et al. 2015), and they are essential in embryonic stem cells (Young 2011). They play crucial roles in patterning, in particular the Hox genes (Capellini et al. 2011;Kmita and Duboule 2003;Maeda and Karch 2015;Pearson et al. 2005;Rezsohazy et al. 2015;Seifert et al. 2015;Zakany and Duboule 2007). Many are involved in nervous system development (Schulte and Frank 2014;Vollmer and Clerc 1998;Zagozewski et al. 2014), and not surprisingly, disruption of homeobox genes leads to various genetic disorders and diseases (Kumar 2009;Liu et al. 2015;Purkayastha and Roy 2015;Quinonez and Innis 2014;Wang et al. 2014a). Also in plants, homeobox genes regulate numerous aspects of development, e.g., stem cell maintenance, lateral outgrowth, stress response, or light response (Brandt et al. 2014;Costanzo et al. 2014;Hay and Tsiantis 2010;Ratcliffe and Riechmann 2002;Tsuda and Hake 2015). While much progress has been made in understanding the function of many HD proteins, even in the model system Drosophila 12 homeobox genes have not yet been subject to intensive study, and therefore, lack a descriptive name associated with their function (Sup. Fig. S1).

HD sequence and classification
The HD sequence The Btypical^HD is 60 amino acids long. The originally described consensus sequence was biased toward animal homeobox genes, particular of the ANTENNAPEDIA (ANTP) class (Bürglin 1994b;Gehring et al. 1994). As more genomes were sequenced, more divergent HDs were encountered. Here, we created a new profile of the conserved residues (a protein logo) of the HD sequences from a single animal, Drosophila melanogaster, including a few selected other HDs (Sup. Fig. S1, Fig. 1). This provides a less biased profile, although still 46 % of the HD sequences are of the ANTP class (Table 1). Compared to the previous profile (Bürglin 1994b;Bürglin 1995;Gehring et al. 1994), the overall pattern of amino acid conservation stays essentially the same. Due to more divergent sequences, individual positions show now more variability than evident in the previous profile (Fig. 1). In C. elegans (Hench et al. 2015) and plants (Mukherjee et al. 2009), the HD profiles show even more variability due to higher numbers of divergent HDs.
Some examples of this additional variability can be found in key residues important for the hydrophobic core of the HD, which constitute the signature of the HD and seemed to be essentially invariant, i.e., leucine (L,16), phenylalanine (F, 20), tryptophan (W, 48), and phenylalanine (F, 49). However, they can be substituted with amino acids of similar properties. For example, instead of the core signature sequence WF (pos. 48, 49), position 48 can be, e.g., phenylalanine (F), or tyrosine (Y), while position 49 can be occupied by tyrosine (Y), tryptophan (W), or small hydrophobic residues such as methionine (M), isoleucine (I), or leucine (L). Another important residue is the basic residue arginine (R) at position 5 of the HD (Sup. Fig. S1), which is found in 99 of the 106 Drosophila HDs.

Updates to HD classes
The principles underlying the classification of HDs have been outlined previously (Bürglin 2005;Bürglin 2011;Holland 2013). Briefly, in the case of animals, orthologous homeobox genes that can be traced at least to the urbilaterian split are placed into families. Families with similar features (e.g., a particular additional domain) are grouped together into classes. Classes may be merged into a superclass or subdivided into subclasses. Such a simplistic system, however, does not fully reflect the real evolutionary complexity. New homeobox genes can arise by duplication and may diverge substantially from the precursor in a relatively short time frame, giving rise to new families that may be restricted to a single taxonomic class. The most important aspect of a good and consistently used classification is that orthologous genes in different species are properly identified and not confused with paralogs. The updated classification presented here is based on two criteria: on the one hand, the sequence similarity of HDs to each other, which is used to generate phylogenetic trees (Fig. 2); on the other hand, flanking conserved domains and motifs in the HD proteins can be used as classifiers (Fig. 3). Here, we suggest retaining only one superclass, TALE, because of its deep evolutionary conservation (Figs. 2 and 3).
The HD is embedded in proteins that can differ substantially in size. Some proteins are barely larger than the HD itself, e.g., mouse Hopx (73 amino acids) (Kook et al. 2006) or C. elegans CEH-7 (84 amino acids) (Kagoshima et al. 1999), while some are large and contain many other domains, e.g., Arabidopsis Ringlet 1 (AT1G28420, 1705 amino acids), or human ZFHX3 (ATBF1, 3703 amino acids, Sup. Fig. S3). Figure 3 shows the different domains and some of the smaller motifs found in various classes of homeobox genes. In addition to the major domains, smaller motifs or regions can be conserved within families, or even between related families, many of which are not shown in Fig. 3 Fig. S1) using LogoBar (Pérez-Bercoff Å and Bürglin 2010;Pérez-Bercoff et al. 2006). The higher the bar, the stronger a position is conserved. Letters inside the bars indicate amino acid residues. Open bars indicate gap regions that were introduced to accommodate longer atypical HDs. Numbers underneath are based on the standard numbering for HDs with 60 residues, and Babc^marks the positions of the three extra residues in loop 1 of TALE HDs. The three alpha helices are indicated with shaded boxes. At the bottom, the consensus sequence (most frequent residue) is shown; residues underneath each position are listed in decreasing order of frequency of occurrence 1988). In Drosophila, it was demonstrated that the core of this region with the residues BSSYF^is an activation domain in Ubx and Scr proteins in Drosophila (Tour et al. 2005).
Many classes of homeobox genes in animals were already discovered in 1994 (Bürglin 1994b;Gehring et al. 1994). However, a number of new classes have appeared, and some refinement and re-evaluation has taken place. Furthermore, most of the plant homeobox gene classes were discovered only after 1994. We will discuss these new findings below.
Overall, given that numerous complete genome sequences are now available, the classification of HD proteins for bilaterians and vascular plants is probably quite complete. More information and analysis of Blower^eukaryotes may reveal novel types of HD proteins that evolved in specific branches. E.g., in Dictyostelium a new group of double homeobox genes has evolved (Clarke et al. 2013). Also, within particular phyla, new types of HD proteins may evolve. For example, in the genus Caenorhabditis a novel, highly divergent type of double HD emerged, termed HOCHOB (Hench et al. 2015).
Using the model system Drosophila melanogaster as an example to summarize the complement of homeobox genes in a single organism (Fig. 2, Sup. Fig. S1) we find that a large fraction is made up of ANTP (46 %) and PRD-LIKE (18 %) homeobox genes (Table 1). The remaining fraction (36 %) is shared by the remaining classes, which all encode large domains flanking the HD. Some homeobox genes were lost in the evolutionary lineage leading to Drosophila, e.g., the HNF class and the Prep family (MEIS class) are missing (Chi et al. 2002;Mukherjee and Bürglin 2007).

Animal HD classes and motifs
Most of the animal HD classes emerged early in metazoan evolution. A number are already present in sponges, while most exist in Placozoa, Cnidaria, and Ctenophora (Ryan et al. 2006;Ryan et al. 2007;Ryan et al. 2010;Srivastava et al. 2008;Srivastava et al. 2010b). In this section, we give brief updates regarding the different animal HD classes, except for PRD, SIX/SO, and POU, which are dealt with further below.

ANTENNAPEDIA (ANTP) class
The ANTP class can be broadly divided into two subclasses. The HOXL (Hox-like) group encompasses genes most similar to the Hox genes ( Fig. 2), i.e., those genes found in the Hox cluster. According to established guidelines, only homeobox genes in the Hox cluster should be named Hox genes (or genes derived from the Hox cluster, if this has secondarily broken up). Many of the HOXL genes have a short motif, the Hexapeptide (HEX, aka YPWM), upstream of the HD (Fig. 3). In the AbdB family this motif has diverged, although a tryptophan (W) is still present. The NKL (NK-like) group is comprised of the NK type homeobox genes, a number of which are found in the NK cluster (see below). A few NKL genes also encode a HEX motif (e.g., Tlx). In many NKL families, an EH1 motif is found toward the N-terminus (see below). Some genes, such as Drosophila engrailed (en), cannot be easily assigned to either subclass. The distinction into HOXL and NKL is not always clear-cut, due to the fact that HOXL genes are probably derived from NKL genes (see below).
The Octapeptide/Hep/EH1/TN/GEH motif, a Groucho interaction motif The Octapeptide was first discovered in PAIRED domaincontaining proteins of fly and humans as a short, conserved sequence motif between the PAIRED domain and the HD The updated classification scheme that we suggest for future use retains only the TALE superclass, which is conserved from animals to plants. Number (Nr.) of proteins within a class/subclass are given. Residues found at position 5 (Pos. 5) and position 50 (Pos. 50) of the HD are indicated in the columns; numbers before residues indicate the number of less frequently found residues a The PRD class contains three additional proteins that lost their HD (i.e., Poxn, Poxm, Sv), which is indicated in brackets and not counted as HD protein proper b Residues of the human sequences HNF1A and SATB1, respectively (Burri et al. 1989;Noll 1993). Subsequently, variants of this motif were discovered in other HD proteins, and given different names, i.e. Hep (Allen et al. 1991), EH1 (in En, Hemmati-Brivanlou et al. 1991;Hui et al. 1992;Joyner and Hanks 1991;Logan et al. 1992), TN (in NK/Tinman proteins, Bodmer 1995;Lints et al. 1993), and GEH (Goriely et al. 1996). The common similarity was not always noted, and also its significance was unclear, due to the shortness of the motif. However, as more sequences became available, the motif was better defined (Harvey 1996;Smith and Jaynes 1996), and eventually was found to occur in many PRD-LIKE, PRD, and NKL class HD proteins, as well as other transcription factors, such as Fox, Ets, and T-Box (Copley 2005;Shimeld 1997;Yaklichkin et al. 2007). We refer to the motif from here on as EH1, the most commonly used name (Copley 2005;Yaklichkin et al. 2007). The core of the motif spans seven residues with only a few conserved positions. Even when only the EH1 motifs encoded by the PRD class proteins are compared, the limited conservation of the motif can be noted (Fig. 4a, Sup. Fig. S5). Since the motif is so small, convergent evolution cannot be excluded. In the PRD-LIKE HD protein UNC-4 of C. elegans, the EH1 motif is found Cterminal to the HD, which implies either a duplication event, or a de novo origin of the motif at this position (Winnier et al. 1999). Nevertheless, the fact that this motif has been well conserved in many homeobox families across the bilaterian  Fig. 2 Classification of HD proteins. A phylogenetic tree of the HD sequences in Sup. Fig. S1 was created using neighbor joining in Clustal X (Larkin et al. 2007). Classification of the HDs is marked on the right. Genes belonging to genomic clusters are marked in different colors in italics. The three PRD class families with HDs are indicated. Due to the limited number of sequences, the tree should only be taken as simple guide to illustrate how similar or divergent HD sequences are when compared to each other. Hence, not all genes within a class fall within the same clade divide, and even in Cnidaria, suggests that this motif has been subject to strong evolutionary constraint. We would like to suggest that the common ancestor of ANTP, PRD, and PRD-LIKE homeobox genes already encoded this motif.
Drosophila En is a transcriptional repressor protein. Functional mapping experiments revealed that an important motif that conveys the repressor activity is EH1 (Smith and Jaynes 1996). Further experiments revealed that EH1 binds to the Groucho (Gro) co-repressor protein to exert its repressor function (Fisher and Caudy 1998;Jiménez et al. 1997;Jiménez et al. 1999;Muhr et al. 2001;Papizan et al. 2011;Winnier et al. 1999). EH1 in other proteins was also confirmed to interact with Gro as in the case for the zinc-finger factor Odd-skipped (Goldstein et al. 2005). Gro is a co-repressor that works together with many developmental transcription factors (Jennings and Ish-Horowicz 2008;Mannervik 2014). Gro and its human orthologs TLE (Transducin-like Enhancer of split) are characterized by an N-terminal glutamine-rich domain and a conserved WD-repeat. Gro/TLE not only recognizes the EH1 motif, but also another motif, termed WRPW based on its sequence. The WRPW motif is found in factors such as Hairy or Runt (Chen and Courey 2000;Cinnamon and Paroush 2008;Turki-Judeh and Courey 2012). Structural studies have shown that the EH1 and the WRPW motif bind into the central groove of the beta propeller structure of the WD repeats of TLE1 (Jennings et al. 2006). The two motifs bind on top of the mouth of the central proteins, the lower part shows classes of plant HD proteins. Not all domains and motifs are shown, which is mostly the case for motives conserved only within families channel and EH1 forms a short amphipathic alpha helix that binds to this hydrophobic recess. In yeast, the Gro homologous factor is the TUP1 protein that also functions as a transcriptional repressor (Smith and Johnson 2000). In a genetic screen to identify mutations of yeast MATα2 defective in repression, a number of point mutations were isolated, many of which mapped to the N-terminus of MATα2 (positions 4, 9, 10, Fig. 4b) (Komachi et al. 1994). We observed that the N-terminus is well conserved in many fungal MATα2 proteins, in particular in those positions that, when mutated, relieve repression ( Fig. 4b, Sup. Fig. 2S). These positions are characterized by small hydrophobic residues with the same spacing as in the EH1 motif. The only exception is that an isoleucine (I) residue is present instead of a F/Y/H residue at the first hydrophobic position. Thus, also in yeast, a slightly modified version of the EH1 motif exists that interacts with a Gro family molecule, i.e., TUP1.
Gro/TLE has been proposed to function via its interaction with histone deacetylases (HDAC) and subsequent chromatin modification, and exert long-range repression through oligomerization (Turki-Judeh and Courey 2012). More recent evidence suggests alternative pathways. For example, in yeast, it has been found that rapid depletion of the Cyc8-Tup1 co-repressor results in de-repression of target genes, while reassociation of Tup1 leads to rapid repression before any repressive chromatin structure can be formed (Wong and Struhl 2011). Co-activators such as Swi/Snf, SAGA, and mediator complex can be rapidly recruited upon TUP1 depletion. Thus, it is thought that TUP1 repression acts primarily through masking of activation domains in transcription factors. In Drosophila, Gro is found at transcription start sites containing hypoacetylated histones H3 and H4, and at sites that exhibit strong RNA polymerase pausing. Activation and repression responses can be very rapid in vivo, also suggesting that Gro/ TLE modulates transcription, rather than cause general chromatin repression (Kaul et al. 2014). Recently, it has been demonstrated that the Mediator subunit Med19 can bind directly to the HD of Hox proteins (Boube et al. 2014). Thus, perhaps, and similar to yeast, Gro/TLE blocks recruitment of mediator to the HD by binding to the EH1 motif.

PAIRED-LIKE (PRD-LIKE) class
We consider the PAIRED class (see below) separately from the PRD-LIKE class, while others group them together (e.g., Zhong and Holland 2011a). The PRD-LIKE class genes encode only the HD as the major conserved domain. Like the NKL genes, about 10 PRD-LIKE families encode an EH1 motif toward the N-terminus (Vorobyov and Horst 2006). A second small motif found in over half of the PRD-LIKE proteins is the OAR motif, first identified in otp, al, and rax (Furukawa et al. 1997). The OAR motif is encoded near the C-terminus of ten families of PRD-LIKE homeobox genes (Galliot et al. 1999;Vorobyov and Horst 2006). The OAR motif is thought to play a role in transcriptional activation (Vorobyov and Horst 2006).

TALE superclass
The typical HD has 60 residues that fold into a globular structure with three alpha helices connected by two short loops Gro and TUP1 interaction motifs. a Protein logo of the EH1/ Octapeptide Gro interaction motif derived from PRD class proteins. The EH1 motif from the PRD class protein alignment (Sup. Fig. S5) was taken and a logo created using LogoBar (Pérez-Bercoff et al. 2006). Underneath the logo the most common residues in descending order are shown. b Conserved N-terminal region of fungal MATα2 HD proteins (for the complete alignment see Sup. Fig. S2). Asterisks mark positions were mutations in MATα2 were isolated in a TUP1 interaction screen (Komachi et al. 1994). Note the matching pattern of hydrophobic residues to the EH1 motif (Bürglin 1994b;Gehring et al. 1994). HDs with deviations from this length have been characterized as Batypical^and usually accommodate extra residues either in loop 1 between helices 1 and 2, and/or in loop 2, between helices 2 and 3 (Bürglin 2005;Bürglin 2011). BAtypical^proved not to be a useful classification characteristic, since the insertion (or deletion) of extra residues has occurred multiple times independently in evolution. For example, even in the well-conserved SIX/SO class, the C. elegans gene unc-39 (ceh-35) encodes an extra residue in loop 1 (Dozier et al. 2001) (Sup. Fig. S1). However, one special group of HD proteins, the TALE superclass, is characterized by a HD with 63 residues, where three extra residues are inserted in loop 1 (Bertolino et al. 1995;Bürglin 1995;Bürglin 1997) (Sup. Fig. S1). The TALE HD proteins are highly conserved in evolution and are present in single-cell eukaryotes, in plants, and in animals in parallel with typical homeobox genes, and therefore represent an ancient split into two types of HD proteins (Bharathan et al. 1997;Bürglin 1997;Bürglin 1998;Derelle et al. 2007). In plants, the TALE proteins can be divided into two classes, BEL and KNOX (Mukherjee et al. 2009). The KNOX and BEL factors have been shown to heterodimerize (Bellaoui et al. 2001;Lee et al. 2008). In animals, the TALE group has split into five classes (PBC, MEIS, IRO, MKX, TGIF) with different domain configurations (Fig. 3). One of these classes, MEIS, is further subdivided into two families (MEIS and PREP) (Bürglin 1997;Mukherjee and Bürglin 2007). Both PBC and MEIS HD proteins, including upstream conserved domains, are already present in Acanthamoeba, which does indicate an ancient role for these TALE HD proteins (Clarke et al. 2013).

CUT class
The CUT class has been divided into several subclasses based on the associated domains, i.e., CUX (comprising the Drosophila cut gene), ONECUT, SATB, and COMPASS (CMP) (Bürglin and Cassata 2002;. Proteins of the CUX, ONECUT, and SATB classes encode one to three copies of the about 80 residue-long CUT domain (Fig. 3). The crystal structure of the CUT domain has been determined and found to comprise five main alpha helices, with helix 3 binding in the major groove of the DNA, and it showed structural similarity to the POU-specific domain (Iyaguchi et al. 2007;Yamasaki et al. 2007). In the Cux family, the N-terminal region (though not in Drosophila) is part of an alternative spliced product called CASP (CDP/ CUX alternatively spliced cDNA), which is found in yeast and plants as a distinct, separate protein (Bürglin and Cassata 2002;Gillingham et al. 2002). CASP localizes to the Golgi and its N-terminal region is predicted to adopt a coiled-coil structure (Gillingham et al. 2002;Malsam et al. 2005), which might be used in the Cux family for protein-protein interaction.
The CMP genes do not encode CUT domains. Instead, their association with CUT class genes is solely based on the N-terminal COMPASS domain that is found both in the CMP proteins (e.g., Drosophila Dve) and in vertebrate SATB proteins. Analysis of the crystal structure of the COMPASS domain showed that it has an ubiquitin-like structure and can form tetramers (Wang et al. 2012;Wang et al. 2014b). Oligomerization is essential for SATB proteins to exert their function when binding to matrix attachment regions (MARs). In SATB proteins, a DNA-binding CUT-LIKE domain follows the COMPASS domain (Wang et al. 2014b).

HNF class
The mammalian HNF1A (LFB1) transcription factor was initially described as an atypical HD protein with a long insert in loop 2 of the HD (Finney 1990;Frain et al. 1989) (Sup. Fig. S1). Subsequently, orthologs in invertebrate species were discovered and a conserved domain (HNF domain) of about 90 amino acids was found upstream of the HD. Structural analysis of HNF1A revealed that the HNF domain is comprised of five alpha helices, with helices 2 to 4 showing structural similarity to the POU domain (Chi et al. 2002). It is thought that the HNF class diverged from the POU class of homeobox genes.

LIM class
LIM class HD proteins encode two LIM domains upstream of the HD (Bürglin 1994b). The LIM domain is about 50-60 residues long and is comprised of a double zinc finger motif with a predominant consensus of CX 2 CX 16-23 HX 2 CX 2 CX 2 CX 16-21 CX 2 (C/H/D) (Kadrmas and Beckerle 2004). The LIM domain is involved primarily in protein-protein interaction (Kadrmas and Beckerle 2004;Zheng and Zhao 2007). Six families have been defined in bilaterians (Hobert and Westphal 2000;Srivastava et al. 2010a). A closely related family, LMO (aka rhombotin), encodes only two LIM domains, possibly having lost the HD secondarily (Boehm et al. 1991;Srivastava et al. 2010a). Unlike many other HD-associated motifs (e.g., PRD, POU, CUT), the LIM domain occurs in many other protein classes, varying in number between 1 to 6 copies, and associating with numerous other domains and motifs; overall, 14 LIM classes have been defined, many of which are involved in cytoskeletal function (Kadrmas and Beckerle 2004;Koch et al. 2012;Te Velthuis et al. 2007).

ZF class
Zinc finger (ZF) class homeobox genes encode C 2 H 2 and C 2 H 2 -like zinc fingers (ZF) in addition to the HD. C 2 H 2 zinc fingers are typically involved in DNA binding (Najafabadi et al. 2015). The number of zinc fingers (2-23) as well as the number of the HDs (1-6) can vary substantially, e.g., human ZFHX3 (aka ATBF1) has 23 ZFs and 4 HDs (Sup. Fig. S3). In vertebrates, five families were defined (Adnp, Tshz, Zeb, Zfhx, Zhx) (Holland et al. 2007). Three families, Zeb (Drosophila zfh-1), Zfhx (Drosophila zfh-2), and Tshz are conserved across the bilaterian divide. The vertebrate Tshz family members encode a divergent HD, but the two Tshz paralogs in Drosophila (teashirt and tiptop) lack the HD (Koebernick et al. 2006;Santos et al. 2010); this probably represents a secondary loss of the HD. The Adnp and Zhx families seem to be vertebrate specific. The Homez gene is derived from the Zhx family (Bayarsaihan et al. 2003), even though it does not encode ZFs, which most likely were lost secondarily. In fungi, C 2 H 2 -HD proteins have also been identified, although their relationship to the Metazoan ZF-HD has not been systematically investigated (Xiong et al. 2015).

PROSPERO (PROS) class
The PROS homeobox genes encode a highly divergent HD with extra residues in loop 2 (Sup. Fig. S1). They also lack the usual basic residues at the N-terminus of the HD. C-terminal to the HD is the 100 amino acids long PROSPERO domain (Bürglin 1994a). X-ray structure analyses revealed a continuity between the HD and the PROSPERO domain; the third alpha helix of the HD is extended, and, together with three further alpha helices, a four-helix bundle is formed that could contribute to DNA binding (Ryter et al. 2002).

CERS (aka LASS) class
In most studied cases, HD proteins act as transcription factors. However, in the CERS (aka longevity assurance (LASS)) class, the HD is embedded in a protein carrying multiple transmembrane (TM) regions, where the TM regions following the HD constitute a TLC (TRAM, LAG1, CLN8) domain (Mesika et al. 2007;Mizutani et al. 2005;Pewzner-Jung et al. 2006). The CERS genes encode ceramide synthases, and the HD does not appear to be essential for this function. Not all CERS genes encode a HD, suggesting that the HD may have been acquired in a translocation event at some point in evolution. Experimentally, virtually the complete HD could be deleted without affecting function, although residues at the very end of the HD and in the linker between the HD and the second TM region are required for function (Mesika et al. 2007). In a 1-hybrid system, the isolated HD was shown to be able to bind DNA, suggesting that it did not lose its DNAbinding capacity (Noyes et al. 2008).

Plant HD classes
In plants, the HD proteins can be divided into 11 classes (HD-ZIP, BEL, KNOX, WOX, DDT, PLINC, PHD, NDX, SAWADEE, PINTOX, and LD) based on the associated domains (Mukherjee et al. 2009;Viola and Gonzalez 2015) (Fig. 3). One of these classes, HD-ZIP, is divided into four related subclasses (HD-ZIP I to IV), since their members all encode leucine zippers following the HD. HD-ZIP III and IV both contain a START and a HD-SAD (HD-START associated domain) (Mukherjee et al. 2009;Schrick et al. 2004); the START domain has been implicated in lipid/sterol binding (Alpy and Tomasetto 2014;Schrick et al. 2014). The HD-ZIP III class, in addition, has a MEKHLA domain at the C-terminus. This domain is related to PAS domains and regulates dimerization, and thereby transcriptional activity, of the HD protein, via some cell intrinsic signal or mechanism (Duclercq et al. 2011;Magnani and Barton 2011;Mukherjee and Bürglin 2006). Two subclasses, HD-ZIP II and HD-ZIP IV, encode a CxxC motif downstream of, or within the leucine zipper (hence aka zipper-loop-zipper (ZLZ) motif), respectively (Ciarbelli et al. 2008;Nakamura et al. 2006). It has been suggested that intracellular redox state can influence the activity of these factors via these cysteine motifs (Tron et al. 2002).
The DDT class is characterized by a DDT domain, which is found in numerous other factors involved in chromatin regulation, and binds to the SLIDE domain in ISWI chromatin remodeling factors (Doerks et al. 2001;Dong et al. 2013). DDT class HD proteins contain additional conserved domains named D-TOXA to D-TOXH, and WSD, named because of the sequence conservation to BAZ/Williams syndrome transcription factor (WSTF) chromodomain proteins (Mukherjee et al. 2009). The bipartite WSD domain was already described in the BAZ proteins as BAZ1 and BAZ2 motifs (Jones et al. 2000). Recently, D-TOXC was characterized as a winged helix-turn-helix domain, named HARE-HTH, which is found in eukaryotic and prokaryotic proteins involved in DNA binding or modification (Aravind and Iyer 2012). Further, D-TOXD and WSD (BAZ1 and BAZ2) were named WHIM1, WHIM2, and WHIM3 and were also implicated in the interaction with ISWI factors (Aravind and Iyer 2012).
Several classes have distinctive domains with conserved cysteine/histidine residues, that are putative or confirmed zinc fingers, i.e., the PLINC (plant zinc finger) Bdouble finger^(two motifs C-X 3 -H-X 9 -D-X-C and C-X 2 -C-X-C-H-X 3 -H) (Hu et al. 2008), the D-TOX ZF Bfinger^in the DDTclass (Mukherjee et al. 2009), and the SAWADEE domain (Mukherjee et al. 2009), which has been shown to be a novel chromatin-binding module that probes the methylation state of the histone H3 tail (Law et al. 2013). Further, the PHD finger is a well-characterized zinc finger resembling a RING domain and also plays a role in binding to methylated histones (Pena et al. 2006).

EAR and WUS repressor motifs
A number of plant homeobox genes also function as transcriptional repressors. A small repressor motif with conserved leucine residues (core consensus sequence: LxLxL), named ERFassociated amphiphilic repression (EAR), was first identified in ERF transcription factors (ethylene-responsive element binding factors) (Ohta et al. 2001). Putative EAR motives were subsequently also described in the N-terminus of HD-ZIP II HD proteins (Ciarbelli et al. 2008), in the C-terminus of several WOX HD proteins (Ikeda et al. 2009;van der Graaff et al. 2009), as well as in the N-terminus and C-terminus of BEL class proteins (therein named ZIBEL motif) (Mukherjee et al. 2009) (Fig. 3). More comprehensive searches discovered similar motives in many plant transcription factors (Causier et al. 2012;Kagale et al. 2010;Kagale and Rozwadowski 2011). Functional studies have implicated the EAR motif in repression in WOX HD proteins (Ikeda et al. 2009). WOX homeobox genes also encode a WUS box (consensus sequence TLxLFP) that has been demonstrated to be involved in repression (Ikeda et al. 2009;Lin et al. 2013). The EAR motif interacts with TOPLESS (TPL) and TPL-related proteins, which are one family of plant homologs of the animal Gro/TLE family of proteins (Causier et al. 2012;Liu and Karmarkar 2008). Thus, the interaction of WD repeat proteins of the Gro/TUP1/TPL group with transcription factors, including HD proteins, is an ancient feature of the eukaryotic transcription machinery.

Genomic clusters of homeobox genes
A number of homeobox genes are organized into clusters and the various genome sequencing projects have uncovered more clusters since 1994. The best known clusters are the four paralogous mammalian Hox clusters with 39 Hox genes, which correspond to the Drosophila Antennapedia complex and the Bithorax complex (Bürglin 1994b;Bürglin 2011;Deutsch 2010;Duboule 2007;Gehring et al. 1994;Lonfat and Duboule 2015;Lonfat et al. 2014;Pick 2015;Rezsohazy et al. 2015). The Hox cluster is well conserved in tetrapods; however, as more Hox clusters have been isolated from different animals, more variation has been noted (Ikuta 2011). In teleost fish, due to the extra round of genome duplication, the Hox cluster organization can be quite capricious, with many losses (Kuraku and Meyer 2009;Martin and Holland 2014). In the genus Drosophila, rearrangements have occurred in the Hox cluster and it has split into two subclusters several times independently (Negre and Ruiz 2007), while in sea urchins and tunicates the organization is very disordered or split multiple times (Deutsch 2010;Duboule 2007); also, in nematodes, the cluster has substantially degenerated (Aboobaker and Blaxter 2003b). In butterflies and moths, extra Hox genes have been inserted in the cluster via duplication (Ferguson et al. 2014). Cnidaria have only few Hox genes (Ryan et al. 2007), and it appears that the full expansion of the Hox cluster occurred only in bilaterians (Deutsch 2010).
The Hox cluster is by no means the only homeobox gene cluster. Tandem duplication is one of the most common mechanisms in eukaryotes to increase gene diversity (Fan et al. 2008). A smaller cluster with three genes, called the ParaHox cluster, was originally found in amphioxus (Gsx, Xlox [mammalian Pdx], Cdx) (Brooke et al. 1998;Bürglin 2011). In Drosophila, it is disrupted, and only two genes (ind [Gsx] and cad) are present (Fig. 2). The ParaHox and Hox genes probably have arisen via duplication from each other. Members are present in Placozoa and Cnidaria, with the beginnings of the Hox and ParaHox clusters emerging in Cnidaria (Chourrout et al. 2006;Garstang and Ferrier 2013;Holland 2013;Hui et al. 2008;Srivastava et al. 2008). Together with a few additional Hox-related genes, ParaHox genes are grouped into the HOXL genes (Fig. 2).
The NK cluster (aka tinman complex) was initially discovered in Drosophila (Jagla et al. 2001). As more sequence information became available, further genes could be incorporated into the cluster (Bürglin 2005;Bürglin 2011;Cande et al. 2009). Beginnings of an NK cluster with several NK genes were found in the sponge Amphimedon queenslandica, with no evidence of HOXL genes (Larroux et al. 2007). More recently, it has emerged that at least one ParaHox/Hox-like gene already existed in sponges and that some Hox genes may have been lost in some sponges (Fortunato et al. 2014;Mendivil Ramos et al. 2012). The NK cluster homeobox genes and a number of disperse NK homeobox genes are grouped into the NKL subclass based on their HD (Fig. 2). However, since HOXL genes are likely to be derived NKL genes that have arisen later in evolution, the NKL subclass is paraphyletic with respect to the HOXL subclass genes, and therefore some classifications have abandoned this distinction (Holland 2013).
A number of other clusters exist, not counting tandem duplicated genes. The PRD-LIKE homeobox genes are usually dispersed in the genome. However, a small cluster with three genes exists, named HRO (homeobrain, rx and orthopedia) that is already found in Placozoa and Cnidaria (Mazza et al. 2010). This also demonstrates the ancientness of the PRD-LIKE homeobox genes. Also, in Placozoa, a cluster of two LIM homeobox and two LIM-only genes was found (Srivastava et al. 2010a).
Newly evolving clusters have also been discovered. In mice, the PRD-LIKE Rhox genes have expanded into a large cluster with 33 genes, while in humans, only three genes are present in the equivalent chromosomal location (Maclean et al. 2005; MacLean and Wilkinson 2010). Another family that has expanded in mice is the Obox gene family (Rajkovic et al. 2002). The PRD-LIKE Dux genes have also been subject to rapid evolution and duplication in mammals and primates (Leidenroth et al. 2012;Leidenroth and Hewitt 2010).

HD structure and function
Structure of the HD The basic structure of the HD as a globular domain with three alpha helices had already been determined by 1994 (Bürglin 1994b;Gehring et al. 1994). Since then, numerous additional structures of HDs have been determined, often in complexes with DNA, or with additional flanking domains, or with cofactors (see Sup. Table S1). A key residue for sequence-specific DNA binding, position 50, was already defined by then, and this residue also allows one to distinguish some HD classes (Table 1). The importance of water molecules in the interface of the HD and the DNA was already noted (Bürglin 1994b;Gehring et al. 1994). More refined X-ray crystallography and modeling gave further insights into how these water molecules contribute to DNA contacts (e.g., Billeter et al. 1996;Li et al. 1995). While most of the specificity of the HD-DNA interaction resides in the major groove contacts, the minor groove contacts of the N-terminal arm contribute to the strength of the binding as well. Considerable variation can be found in the sequence of the Nterminal arm between different HD families, yet often basic residues are present in these positions. Arginine (R) is particularly favored at position 5 of the HD (Table 1, Sup. Fig. S1). Arginine residues are preferentially found in narrow minor grooves, which tend to be A-tracts (Rohs et al. 2009).
The HD contains a helix-turn-helix motif, and thus similarities to bacterial DNA-binding proteins exist (Laughon and Scott 1984). For example, the similarity to the Hin recombinase family was pointed out early on (Affolter et al. 1991). Structural analysis of the Hin recombinase shows that the DNA-binding domain is indeed composed of three alpha helices in a similar arrangement as the HD (Feng et al. 1994).

DNA binding of the HD
The basic DNA-binding properties of the HD proteins were also known in 1994 ). Since then, many binding sites where determined with Selex and other methods, more recently using high-throughput approaches. This resulted in databases where binding site preferences for individual transcription factors, or their DNA-binding domain, can be looked up (Affolter et al. 2008;Berger et al. 2008;Jolma et al. 2013;Noyes et al. 2008), for example, in Jaspar (http:// jaspar.genereg.net) . Mutational analysis of the En HD explored the potential binding space of the HD further; novel En HD mutant variants were uncovered that displayed substantially altered binding preferences .
Although this information will certainly be of great value to better understand gene regulation by HD proteins, recent evidence suggesting that clusters of low affinity sites also play important roles in gene regulation by HD proteins (Crocker et al. 2015) suggests that in silico searches using high-affinity sequences might miss many functionally relevant sites (see also below).
It is clear that the limited sequence specificity of the HD itself, comprising TA-rich sequences of hardly more than four base pairs, is not sufficient to explain how genes can be activated in a selective manner in vivo ). Several mechanisms are exploited by HD proteins to increase DNAbinding specificity, involving either flanking domains or cofactors. In a number of cases (Fig. 3), additional domains provide extra DNA-binding capacity (e.g., PAIRED, POU, PROS, ZF, CUT). In a few cases, multiple HDs occur in a given protein (e.g., ZF class, DVE family) (Sup. Fig. S3), the most extreme case to date is found in C. elegans, with CEH-100 having 12 HDs (Hench et al. 2015).
In the case of cofactors, the HD or flanking regions provide protein-protein interaction interfaces that allow either other DNA-binding cofactors to bind together with the HD transcription factor (e.g., the HEX motif), allow oligomerization (e.g., COMPASS domain, or leucine zipper in HD-ZIP genes), or provide other types of protein interactions with components of the transcription machinery. Further, DNA shape plays a role in adding specificity.
Below, we discuss a few examples of the mechanisms by which HD proteins can interact with partners to exert their task as transcriptional regulators.

MATα2-MATa1 interaction
The yeast mating type locus contains two homeobox genes, the TALE homeobox gene MATα2 and the typical homeobox gene MATa1. MATα2 can regulate different subsets of genes by forming heterodimers with either MCM1 or with MATa1. MATα2 and MATa1 individually lack strong DNA-binding affinity, but together, they bind strongly to their binding sites (Li et al. 1995). In the complexes the two HDs bind in tandem to their binding site, and MATα2 contacts the MATa1 HD via its Cterminal tail downstream of the HD (Fig. 5a). This structural complex also illustrates well, how water molecules are positioned in the interface between helix 3 and the DNA (Sup. Fig. S4).

PBC-Hox interaction
Hox proteins usually do not bind alone to enhancers or promoter regions. Proteins encoded by two TALE classes, PBC (Exd in Drosophila), and MEIS (Hth in Drosophila) have been shown to be important cofactors for Hox function. These TALE proteins are usually expressed rather broadly, so their segment/tissue specificity resides mostly in the Hox cofactors (Mann and Affolter 1998;Mann et al. 2009;Rezsohazy et al. 2015). This interaction is evolutionarily ancient and is also found in the sea anemone Nematostella (Ferrier 2014;Hudry et al. 2014;Merabet and Galliot 2015). Notwithstanding, PBC/MEIS proteins function also independently of Hox proteins (Laurent et al. 2008;Schulte and Frank 2014).
Crystal structures of several PBC and Hox HDs when bound to DNA have been determined, such as, for example, Exd-Scr, shown in Fig. 5b (Joshi et al. 2007;Mann et al. 2009). In all these structures, it was found that the HEX motif upstream of the Hox HD interacts with the PBC HD. In Abd-B proteins, the conserved tryptophane (W) plays this role in interaction. It is also interesting to note that the EH2 conserved region in En proteins, with its rudimentary similarity to the HEX motif, confers interaction with Pbx1 (Peltenburg and Murre 1996). Interaction between PBC and Hox proteins is, however, not confined to the HEX motif; in fact, this particular interaction may even be dispensable in some contexts. For example, regions at the C-terminus of the Ubx HD (named UbdA) can provide additional interaction interfaces (Foos et al. 2015).
MATα2/MATa1 and PBC/Hox interactions may have a common ancient evolutionary origin (Bürglin 1998). However, the arrangement of the two factors when bound to DNA differs between yeast (MATα2-MATa1, Fig. 5a) and metazoa (Exd-Scr, Fig. 5b), in that the two protein types exchanged positions (TALE-typical → typical-TALE, note arrows in Fig. 5c). Perhaps, this is a consequence of a secondary loss of an upstream MEIS or PBC domain in MATα2, necessitating changes in the dimer interactions. However, MEIS proteins may also bind on the other side of the Hox protein as in the class 3 interactions shown in Merabet and Lohmann (2015) (Fig. 5c).
The current structural studies are limited by the fact that neither is the highly conserved upstream PBC domain included, nor are structures of complexes available that include the MEIS protein. MEIS/Hth interacts with PBC through the Nterminal subdomain (PBC-A) (Fig. 5c) and MEIS/Hth can also interact with the Hox proteins (Amin et al. 2015;Mann and Affolter 1998;Mann et al. 2009;Merabet and Hudry 2013;Merabet and Lohmann 2015). In such multimeric complexes, DNA specificity would of course be further increased.

DNA shape plays a role in DNA-binding specificity
Even though TALE cofactors increase the DNA specificity of the Hox proteins, the HDs of Hox proteins themselves are still very similar in sequence and bind to similar sequences. Thus, the conundrum how individual Hox proteins can exert their highly specific functions in different tissues in vivo is still not sufficiently resolved. Recent insights into how PBC/Hox complexes bind DNA may aid in resolving this puzzle: DNA shape, i.e., structural features such as minor groove width, roll, and twist, also play an important role in specificity Dror et al. 2014;Rohs et al. 2009;Slattery et al. 2011;Yang et al. 2014). Exd/Hox dimers display different binding specificities (Rohs et al. 2009). DNA shape contributes to these differences, and DNA shape predictions revealed that anterior and posterior Hox proteins prefer sequences with distinct minor groove topographies (width minima) Yang et al. 2014). A key residue for minor groove contacts is arginine at position 5 of the HD. In the case of Exd/Scr binding to the fkh250 site, two additional residues, arginine 3 and histidine -12, also insert into the minor groove (Joshi et al. 2007) (Fig. 5b), and they are important for the binding preferences of Scr, since they select DNA sequences with a narrow minor groove at the Hox half-site . Future in silico prediction of DNAbinding specificities will certainly benefit by taking such DNA structural features into account .

Additional protein-protein interactions provide specificity
Protein-protein interaction is not restricted to PBC/Hox interactions via the HEX motif. Experiments with Drosophila Scr have shown that in salivary glands, Scr but not Antp can form homodimers (Papadopoulos et al. 2012). Glutamine at position 19 in helix 1, which is found in many Hox proteins (Sup. Fig. S1), is critically important for this dimerization. However, Antp fails to dimerize because of short regions N-and Cterminal of the HD (Papadopoulos et al. 2012). Such short linear motifs (SLiMs), an example of which is the HEX motif, are becoming the focus of further studies, since they can contribute to differential, specific protein-protein interaction in different tissues (Merabet and Galliot 2015;Merabet et al. 2009;Sivanantharajah and Percival-Smith 2015). In an in vivo screen, many novel protein-protein interactions that occur in different cellular contexts were identified (Baëza et al. 2015). These studies showed that the HEX and the UbdA motives play key roles in providing specificity and that mutations in these motives can shift the interaction profile. A further mode of selective protein-protein interaction has been demonstrated for the mediator complex. It can bind to the HD of some Hox proteins, but not to other HDs, such as those in PBC proteins (Boube et al. 2014).

Low-affinity binding sites
Protein-protein interactions as well as DNA shape can address some of the conundrum of how TALE-Hox proteins discriminate between different promoters/enhancers. Somewhat contradictory is the observation of low-affinity binding sites with reduced specificity . In recent experiments, Crocker et al. showed that Ubx together with the Exd and Hth cofactors binds to low-affinity sites in the promoter of the shavenbaby (svb) gene (Crocker et al. 2015;Merabet and Lohmann 2015). Multiple low-affinity binding sites were required to achieve robust expression of this promoter. Mutation of these binding sites to high-affinity consensus sites decreased tissue specificity, allowing other Hox factors to bind and broaden the expression domain. Thus, while low-affinity sites might be able to better discriminate between different TALE-Hox complexes, multiple sites are necessary to compensate for such low-affinity sites.

PAIRED (PRD) class
The PRD class of homeobox genes was of prime interest to Walter, ever since the discovery that the Drosophila gene eyeless is homologous to vertebrate PAX-6 genes, and that these genes are involved in eye development (Small eye mutations in mouse, Aniridia mutations in human) (Quiring et al. 1994). This triggered a fruitful line of research into the origin and evolution of eyes in his laboratory resulting in highlights such as the spectacular finding that Drosophila eyeless and mammalian PAX-6 can induce ectopic eyes in tissues such as legs and wings in Drosophila (Gehring 2005;Gehring 2012;Gehring 2014;Halder et al. 1995;Hayakawa et al. 2015).

PRD class classification
PRD class homeobox genes encode a PAIRED domain and a HD with most often a serine residue at position 50 of the HD. In addition, PRD proteins may contain an EH1/Octapeptide motif and an OAR motif (Fig. 6a). In many species, PRD genes are called Pax genes, and in several instances, they have lost their homeobox secondarily (see below).
The PAIRED domain was originally discovered by Daniel Bopp and colleagues in the laboratory of Markus Noll at the Biozentrum (Bopp et al. 1986;Noll 1993). Structural studies show that the PAIRED domain, which is about 128 amino acids long, is composed of two subdomains (Xu et al. 1995), which were named PAI and RED (Jun and Desplan 1996). In sponges and Placozoa, only a single Pax gene exists, while in Cnidaria, four types of genes are found (PaxA, PaxB, PaxC, PaxD) (Hill et al. 2010;Suga et al. 2010). In bilaterians, five distinct PRD families have been described (Miller et al. 2000;Underhill 2012).
Only recently has it become apparent that there are actually six families in bilaterians (Fig. 6c), since the Eyegone (Eyg) family is also found in hemichordates and sea urchins (Friedrich and Caravas 2011). Our own phylogenetic analysis  (Li et al. 1995). N and C indicate N-terminal and C-terminal ends of the fragments, respectively. The DNA used as binding site is shown underneath and is color coded to show the binding regions of the two HDs. Arrows indicate the direction of the third alpha helix. b Crystal structure of the Drosophila Scr (green) and Exd (orange) HDs co-crystalized on the fkh250 binding site (PDB ID: 2R5Z) (Joshi et al. 2007). The N-terminal arm of the HD and the linker to the HEX motif make multiple contacts in the minor groove. c fkh250 sequence used in the Scr/Exd/DNA crystal complex (b), and schematic view of a hypothetical complex of the full-length proteins together with Drosophila Hth (Homothorax, ortholog of mammalian MEIS proteins). The two HD fragments used in the crystal are in black outlines, and arrows indicate the direction of the third alpha helix confirmed this finding, both when the PAIRED domain was used (Fig. 6c), and when the HD was used (Fig. 2). Two families, Eyg and Poxn, have been lost in vertebrates.
The PAIRED domain has significant sequence similarity to Tc1-like transposases (Ivics et al. 1996), suggesting that it was derived from a transposase (Breitling and Gerber 2000). A PRD domain-like protein has also been found in the protozoan Giardia lamblia (Wang et al. 2010), though its relationship to the transposases and the PAIRED domain has not been elucidated yet. The HD sequences of PRD class proteins are most similar to those of PRD-LIKE HDs (Fig. 2), suggesting that a Paired box gene merged with a PRD-LIKE homebox gene in early metazoan evolution prior to the emergence of sponges and Placozoa (Galliot et al. 1999;Underhill 2012). Some of the PRD class homeobox genes encode two additional motifs, i.e., the EH1 and the OAR motif. The OAR motif is present only in Pax7 proteins (but not Pax3) and is also found in PAX3/7 homologs of oyster (Sup. Fig. S5), demonstrating that this motif has been conserved across the bilaterian divide. Given that these motifs also exist in several PRD-LIKE families, the most parsimonious explanation is that the original PRD class gene that captured a PRD domain also had an EH1 and an OAR motif (Underhill 2012;Vorobyov and Horst 2006).

The PAIRED domain structure
Several 3D structures of the PAIRED domain have been determined (Sup . Table S1), and the DNA-binding specificity of several has been investigated (Mayran et al. 2015). The structure consists of two subdomains (Fig. 6b). The N-terminal PAI subdomain is characterized by a short beta motif and a domain with three alpha helices that fold in a helix-turn-helix fashion similar to the HD. The C-terminal RED subdomain also contains three alpha helices that fold in a HD-like fashion (Xu et al. 1999). The two subdomains are joined by a linker region of eight amino acids. The first structure of the PAIRED domain of the Drosophila Paired protein bound to DNA revealed that the main DNA contacts were made by the PAI subdomain (Xu et al. 1995). The subsequent X-ray structure of Pax-6 showed that the RED domain also can contact the DNA, and that the linker region of Pax-6 makes extensive contacts with the minor groove of the DNA (Xu et al. 1999).

Molecular tinkering in the PRD class
The PRD class genes represent an interesting case of molecular tinkering (Jacob 1977), an idea which Walter was particularly fond of. While the original PRD gene most certainly encoded a complete PAIRED domain (PAI and RED), an EH1 motif, a HD, and an OAR motif, we find that through loss of motifs, a wide variation of combinations has been created (see Fig. 6a, Sup. Fig. S5). Only a subset of the Pax3/7/Prd family (Pax7 and some invertebrate genes) has retained all motifs, while OAR was lost from most of the other genes (Vorobyov and Horst 2006). The Pax4/6/Ey family lost the EH1 motif, and Pax1/9/Poxm as well as Poxn lost the HD. Interestingly, the Pax2/5/8/Sv family lost only the last half of the HD in vertebrates, although in Drosophila Sv this HD remainder completely diverged. The Eyg family evolved the PAI subdomain rapidly, losing at least the N-terminal beta strands (Friedrich and Caravas 2011), and in the most extreme case in nematodes, the whole PAI subdomain was lost (Hobert and Ruvkun 1999). Conversely, nematodes also code for proteins that only retained the PAI domain (Hobert and Ruvkun 1999), and two of these also contain the EH1 motif (NPAX-1/ NPAX-4, Fig. 6a, Sup. Fig. S5). Finally, in vertebrates, a Pax10 protein (Pax3/6/Prd family) exists that lost the PAIRED domain; this protein itself was lost in mammals (Feiner et al. 2014;Ravi et al. 2013). The Pax3/6/Prd family lost the EH1 motif, although another conserved motif is present in a similar location (named PAX6 in Sup. Fig. S5). It has been suggested that this motif is Octapeptide-like (Keller et al. 2010). However, this motif does not have the characteristic conserved pattern of hydrophobic resides.
The variation of domains and motifs is also replicated to some extent within individual genes through alternative splicing, which can alter DNA-binding specificity (Underhill 2012). For example, the C. elegans Pax-6 gene vab-3 has an alternative splice form (mab-18) that lacks the Paired box (Zhang and Emmons 1995). Another example is the alternative splicing of Pax-3 in olive flounder, which can produce transcripts encoding a disrupted PAIRED domain, and/or lacking a HD (Jiao et al. 2015). Experimentally, the functional separation of the PAIRED domain and the HD has also been demonstrated: a construct of Ey lacking the HD is able to rescue the ey2 mutant phenotype (Punzo et al. 2001).

SIX/SO (aka SINE) class
The SIX/SO class of HD proteins is characterized by a 120 amino acids long SIX/SO domain upstream of the HD (Fig. 3). The HD itself is also noteworthy, since basic residues in the N-terminal region of the HD are absent (Sup. Fig. S1, Table 1), which suggests that the N-terminal arm may not interact with the minor groove of the DNA. Two SIX/SO class genes, Optix (Six3 family) and sine oculis (so, Six1 family), play a role in eye development like several of the PRD homeobox genes. Drosophila Eyes absent (Eya), a special protein tyrosine phosphatase, has been shown to be a cofactor of SIX/SO proteins. Molecular and structural analysis revealed that human SIX1 interacts directly with human EYA2 (Patrick et al. 2013). The SIX/SO domain is a globular domain comprised of six alpha helices that has no obvious similarity to other structures (e.g., helix-turn-helix motifs, etc., Fig. 7). Modeling and mutational analyses suggest that the sixth helix binds in the major groove of the DNA, and together with the third helix of the HD, the two domains provide specific DNA binding. EYA2 does not bind DNA, instead it interacts with helix 1 of the SIX/SO domain and provides thus the co-activator role for SIX1 (Patrick et al. 2013). DNA-binding studies suggest that the SIX/SO domain modifies the DNA-binding properties of the HD. Berger et al. (2008) produced a DNA-binding profile for the SIX1 HD (core TATC, Fig. 7) in their high-throughput study (Berger et al. 2008), which differed from the profile identified when full-length SIX1 was used (TT[t/a]C) (Liu et al. 2012) (Fig. 7). In addition, an additional conserved pair of residues, TC, was revealed, which we suggest is bound by helix 6 of the SIX/SO domain. The latter motif matches the binding site MEF3 (consensus: TCAGGTTTC) (Patrick et al. 2013). Overall, this illustrates how through the action of a tethered flanking DNAbinding domain, the specificity of the HD can be altered.

POU class
The POU class of transcription factors is characterized by a conserved POU-specific domain of about 70 resides that is located upstream of a HD that usually contains a serine residue at position 50 of the HD (Herr et al. 1988) ( Table 1). The POU domain has so far not been found independently of the HD, unlike the PAIRED domain discussed above. X-ray structures have shown that the POU domain is composed of a compact, globular DNA-binding domain with four alpha helices that have a similar fold to bacteriophage repressor molecules (Assa-Munt et al. 1993). Helices 2 and 3 form a helix-turnhelix motif and base-specific contacts are made with helix 3 in the major groove of the DNA.
POU proteins can bind as homodimers or heterodimers to DNA. The human POU protein OCT1 (POU2F1) exemplifies another way of increasing as well as modulating DNA- On the right side, the PRD class families are indicated, together with their typical structural organization. Deuterostome branches are highlighted in pink, protostome branches in yellow. Branch lengths in the Eyg family were reduced where indicated (Ce, Crem, CG). Note that the branching of the six families should not be taken as evidence for how they evolved from an ancestral precursor. Drosophila proteins: Sv: Shaven; Poxn: Pox-neuro; Poxm: Pox-meso. The phylogenetic tree was created using PhyML as implemented in SeaView (Gouy et al. 2010). About 100 residues containing the C-terminal region of PAI and the complete RED subdomain from the multiple sequence alignment in Sup. Fig. S5 were used for tree generation. For species codes, see Sup. Fig. S5 binding specificity. OCT1 binds as a dimer, and strikingly, can bind in two different conformations (Reményi et al. 2001). In one conformation, it binds to a DNA sequence termed PORE (Fig. 8a), which is not palindromic. In this configuration, the binding site is longer (15 bp), and the two molecules sit further apart so that their respective binding sites are adjacent to each other. In the second configuration, the binding to the palindromic MORE DNA sequence is symmetric and the site is shorter (12 bp) (Fig. 8b). Here, each dimer is oriented more longitudinally along the DNA so that the two POU domains bind the DNA in a nested fashion. The protein-protein contact interfaces between these two configurations are very different; in the case of MORE, the Cterminus of the HD contacts the POU domain at the Nterminus of helix 1 and the loop between helices 3 and 4, while in the PORE configuration, the N-terminus of the HD contacts the POU domain in the helix 1 to helix 2 region (Reményi et al. 2001). Thus, two different configurations of how the protein can bind to DNAyield different sequence specificities. These provide also different interfaces for other proteins to interact with OCT1 so that cofactors, e.g., OBF1 (POU2AF1), only recognize one conformation (Reményi et al. 2004).

Conclusion
HD proteins predominantly function as transcription factors that activate or repress gene expression. We have seen that a HD by itself is not sufficiently specific to bind to targets in gene promoters. On the one hand, additional flanking domains, or cofactors are used to add extra specificity. These extra protein domains cannot only add specificity, but they can also alter the specificity of the HD itself. How the necessary specificity is achieved in vivo is still poorly understood. DNA shape appears to play an important role in addition to the base pair sequence, and contributes to specificity. On the other hand, even low-affinity sites are used. This is compensated to   (Patrick et al. 2013). Note that the MBP fusion protein, which is part of the crystal, is hidden in this view. The HD is in green, the SIX/SO domain is in orange, and EYA2 is in cyan. Helix 6 of the SIX/SO domain and helix 3 of the HD are arranged such that they could fit in the major groove of the DNA. Underneath, the DNA logo of the binding site determined for mouse Six1 is shown, visualized using Weblogo (Crooks et al. 2004) with the sequences from Sup. Table 2 from (Liu et al. 2012). The binding site logo determined in the high-throughput study by Berger et al. (2008) using the mouse Six1 HD only (Berger et al. 2008) is shown at the bottom, retrieved and visualized using Jaspar  some degree through the use of multiple copies of such binding sites to achieve specificity. Other interacting cofactors may also contribute towards stabilizing low-affinity sites. Overall, the most common theme is that the combinatorial interaction of multiple factors is required in the regulatory region of a specific gene for proper regulation. Most HD proteins can interact with multiple partners, either other HD proteins (both homo-or hetero-interactions are possible), or other types of transcription factors. Many of these interactions can be mediated by SLiMs. Some of the motifs are required for coupling to the transcription machinery, e.g., EH1, which can mediate a repression state. The discovery of the homeobox genes was seminal for our improved understanding of developmental and evolutionary biology. Initially, as exemplified by the homeotic genes, it demonstrated that sequence-related proteins can play similar yet different functions, i.e., that duplicated and subsequently divergent paralogous genes allowed specialization in different body segments. Likely, the expansion of the homeobox genes contributed to the Cambrian explosion (Holland 2015).
The realization that homeobox genes are well conserved from flies to vertebrates provided a fundamental technological revolution: in a flurry of activity many different types of developmental control genes were isolated based on sequence similarity. A new area of reverse genetics was born that allowed breakthroughs in mammalian developmental biology. Furthermore, it demonstrated for the first time that the fundamental molecular mechanisms underlying metazoan development are evolutionarily conserved. Another consequence of these findings, though it may seem obvious nowadays, was the realization that transcription factors play a key role in decoding the genetic blueprint and converting it through a cascade of events into cell fate decisions and cell differentiation that ultimately gives rise to a complex multicellular organism.
Further insights into development and evolution of organisms stem from the knowledge gained over the last decades about key regulators of developmental processes, whether they are transcription factors, signaling molecules, or regulatory RNAs. Homeobox genes represent only a subset of all these regulators, yet their analysis has provided many important insights into the evolutionary events that have taken place over hundreds of million years.
For example, how can major changes in body morphology evolve? We now know that a mutation in a HD transcription factor can lead to drastic altered body shapes, since a whole cascade of downstream target genes is affected (e.g., Ronshaugen et al. 2002). Thus, evolutionary events need not only occur in small, gradual steps, but larger jumps are also possible, although they may not be as frequent.
Another example is gene loss. It is perhaps self-evident that, as the multicellular complexity of an organism grows, the number of regulatory factors has to increase. This is well exemplified by the increase in the number of homeobox genes when going from single-celled eukaryotes to multicellular plants or animals. Conversely, one might perhaps expect that losing key developmental regulators such as homeobox genes, once acquired, would be a taboo. Yet, we observe again and again that homeobox genes were lost in evolution. For example, the Hox cluster in C. elegans is very degenerate and several genes were lost (Aboobaker and Blaxter 2003a;Aboobaker and Blaxter 2003b). Similarly, the PRD class shows that two of its families, one comprising the apparently  (Reményi et al. 2001). The HD is in green, and the POU-specific domain is in magenta. Underneath each panel are the respective binding sites used in the X-ray studies as well as schematic views of OCT1 DNA-binding domains. a OCT1 dimer binding to the PORE DNA sequence. The two PAX6 monomers are distinguished by different color intensity (PDB ID: 1HF0). b OCT1 bound to the MORE DNA sequence (PDB ID: 1E30). Note that only half of the dimer is shown, due to the complete symmetry in conjunction with the palindromic binding site.
Bimportant^gene eyegone, were lost early in the chordate lineage (Fig. 6). Most impressive is the loss of 34 homeobox families in parasitic tapeworms (Tsai et al. 2013). A last example is evolutionary innovation. The Drosophila homeobox gene bcd is essential for early embryogenesis, where its protein forms a gradient through the embryo (Driever and Nüsslein-Volhard 1988). Yet, bcd is an evolutionary novelty, existing only in Cyclorrhaphan flies, and the gene itself was derived from a Hox3 cluster gene (Stauber et al. 1999;Stauber et al. 2002). A related observation is that 14 homeobox genes in C. elegans lack obvious orthologs in other Caenorhabditis species (Hench et al. 2015). Clearly, they have emerged only recently, and seem to be subject to rapid evolutionary change.
The last 20 years were certainly very exciting for Walter. He was still scientifically active until the last moment, and alas, he will not see the results of his latest experiments. Nonetheless, he had the satisfaction of seeing many, though by no means all, of his scientific predictions and hypotheses confirmed.