Although the sequence similarity between eukaryotic and prokaryotic rhomboid family proteins is relatively low (around 10-15% identity in the conserved region), the entire superfamily could be retrieved from the protein sequence databases within three iterations of the PSI-BLAST program with a high statistical significance and without any false positives. The conserved core of the rhomboid family consists of six conserved TMHs (Figure 1). The predicted catalytic serine is located in TMH5, whereas the predicted catalytic histidine is in TMH7; TMH3 contains two additional histidines and an asparagine, which are conserved in the great majority of the rhomboid-family proteins (Figure 1). The roles of these conserved residues are not known, but, given the remarkable evolutionary conservation, it seems likely that they also contribute to catalysis; indeed, it has been shown that the conserved asparagine is required for the cleavage of Spitz by RHO 7.

When examining the multiple alignment of the rhomboid superfamily proteins, we noticed that several eukaryotic members appear to be inactivated proteases, as indicated by the loss of the predicted catalytic serine or histidine (Figure 1, and data not shown); these inactivated forms could be regulators of active rhomboid proteases. Several other proteins lack one or more of the conserved residues in TMH3; it remains unclear whether or not these are active proteases.

Bacterial and archaeal members of the rhomboid superfamily contain six TMH, whereas the eukaryotic members typically have an additional seventh TMH, which may be attached to the core either from the amino terminus or from the carboxyl terminus as discussed below.

The phyletic distribution pattern of the rhomboid family shows that this intramembrane protease is extremely common in all three kingdoms of life, but is not necessarily essential for cell function. Rhomboids are missing in the microsporidian Encephalitozoon cuniculi, a eukaryotic intracellular parasite with a highly degraded genome, the archaea Methanothermobacter thermoautotrophicus and Thermoplasma volcanium, and several bacterial species, primarily parasites with small genomes but also species with moderately sized genomes, such as Xylella fastidiosum (see COG0705 at 22). In two instances, a representative of the rhomboid family is present in only one of a pair of relatively close genomes (present in T. acidophilum but missing in T. volcanium; present in the spirochete Treponema pallidum but missing in the related bacterium Borrelia burgdorferi), which suggests relatively recent, repeated losses of this gene. Most of the prokaryotic species have a single gene coding for a rhomboid-family protein, although some have two or three paralogs (see COG0705 22); in contrast, eukaryotes show expansion of the rhomboid family, with seven members in Drosophila, and as many as 13 in Arabidopsis.

The multiple alignment of the 6-TMH core of the rhomboid family (Figure 1) was employed to construct a phylogenetic tree using the least-squares algorithm with subsequent optimization using the maximum likelihood (ML) method (see Materials and methods). Only the conserved regions including the TMH and short adjacent stretches shown in Figure 1 were used as the input for tree building, whereas the poorly conserved intervening regions were omitted to avoid noise from potentially misaligned residues (except for the Bayesian analysis, which used the complete alignment; see Materials and methods). The alignment used for phylogenetic reconstructions included 87 sequences and 149 aligned sites. The phylogenetic tree of the rhomboid family presents a complex and unexpected picture (Figure 2). Neither the eukaryotic nor the archaeal subsets of the family appear to form monophyletic clades. Instead, the eukaryotic rhomboids are split between two major subfamilies, which are positioned in the midst of different prokaryotic branches (Figure 2). The first subfamily, which includes six of the seven Drosophila rhomboids, clusters with a distinct prokaryotic assemblage, consisting primarily of Gram-positive bacteria as well as a subset of archaea; this clade is strongly supported by bootstrap analysis (Figure 2). The proteins in this group of eukaryotic rhomboids, which we designated the RHO subfamily, typically have an extra TMH added carboxy-terminally to the 6-TMH core; some of these proteins also contain EF-hand calcium-binding domains amino-terminally of the core (Figure 2).

The second eukaryotic subfamily, which we designated the PARL subfamily, after PARL, the human ortholog of Drosophila RHO7 6, resides within a large, heterogeneous prokaryotic cluster (Figure 2). Within this subfamily, PARL and its orthologs from other animals and from fungi have distinct domain architecture, with an extra TMH added to the amino terminus of the core, whereas the rest have only the core (a carboxy-terminal TMH and a ubiquitin-associated domain are appended in one Arabidopsis protein; Figure 2). Thus, the existence of two distinct subfamilies of eukaryotic rhomboids is supported by features of domain architectures that appear to comprise shared derived characters. Within these two major eukaryotic subfamilies, evolution apparently proceeded by both ancient and more recent duplications. Several lineage-specific expansions of paralogs 23 are noticeable, in insects, mammals and plants (Figure 2).

Archaeal rhomboids are scattered over the phylogenetic tree, with two major clusters and, in addition, three isolated proteins joining different bacterial branches (Figure 2). There is no indication of an affinity between any of the archaeal and eukaryotic rhomboids. Although many of the bacterial rhomboids form phylogenetically coherent clusters corresponding to the established bacterial lineages, there are also several clusters that have an odd composition, such as the grouping of proteobacterial and Gram-positive species; some of these clusters are well supported by bootstrap (see clusters 1-4 in Figure 2).

Unexpected tree topologies often emerge due to artifacts of phylogenetic analysis methods. This concern is particularly serious for highly divergent families of membrane proteins, such as the rhomboids, in which parallel amino-acid substitutions are likely. Therefore we investigated the phylogeny of the rhomboid family in greater detail using several independent phylogenetic methods and the corresponding statistical tests. First, we assessed the robustness of the topology of the tree shown in Figure 2 using the Kishino-Hasegawa (KH) test whereby the clade of interest is forced into various positions on the tree and the likelihoods of the resulting topologies are estimated. Specifically, the KH test was used to evaluate two alternative topologies, in which the RHO and PARL subfamilies formed a clade, and two topologies, in which the RHO subfamily formed a clade with archaeal rhomboids (Figure 2 and Table 1). Each of these alternative topologies had a significantly lower likelihood than the original topology shown in Figure 2 (see Table 1).

In addition, a tree of the rhomboid family was constructed using the Bayesian inference method, which has recently become a practical alternative to the more traditional methods of phylogenetic analysis 2425. The tree produced using the MRBAYES package 26 showed the same major clades as the tree in Figure 2 (data not shown); moreover, clustering of the RHO and PARL subfamilies of eukaryotic rhomboids with the respective prokaryotic clades was supported by high posterior probabilities (Figure 2).

We also attempted to construct a phylogenetic tree of the rhomboid family by using the maximum parsimony method 27. The resulting tree contained the same major clades as the trees constructed using ML and MRBAYES; however, the number of parsimony-informative sites was insufficient to obtain high bootstrap support with this approach (data not shown).

We also tested alternative phylogenies using neighbor-joining search with constraint trees 27. The alternative phylogenies reflected two distinct hypotheses: first, clustering of the RHO and PARL subfamilies of eukaryotic rhomboids with the prokaryotic rhomboid families as suggested by the tree topology in Figure 2; and second, monophyly of the eukaryotic rhomboids (Figure 3). The phylogenies corresponding to these alternative hypotheses were compared to the best phylogeny using three statistical tests (Table 2). The hypothesis 1 tree was not significantly different from the best tree under any of these tests whereas the hypothesis 2 tree was significantly (p < 0.05) worse than the best tree according to each of the tests (Table 2).

The concordance of the results obtained with several independent methods for phylogenetic tree construction and statistical analysis specifically aimed at testing the alternative hypothesis of monophyletic origin of eukaryotic rhomboids shows strong support for the major aspects of the tree topology in Figure 2 and, in particular, for the polyphyly of eukaryotic rhomboids.

The phylogenetic tree of the rhomboid family shown in Figure 2 and supported by the additional tests described above follows neither the 'standard model' scenario 2829, with the major split between the archaeo-eukaryotic and bacterial lineages nor the 'mitochondrial' scenario, which postulates acquisition of a gene by eukaryotes from the pro-mitochondrial endosymbiont. Neither can this tree be explained by postulating a small number of lineage-specific gene losses. The parsimonious interpretation of the rhomboid family tree seems to be that the evolutionary history of this family had been replete with horizontal gene transfer (HGT) and lineage-specific gene loss events. In particular, in spite of the presence of rhomboids in the majority of modern life forms from all three primary superkingdoms, phylogenetic analysis suggests that this family has not been inherited from the last universal common ancestor (LUCA). Instead, the tree topology seems to indicate that this family emerged in some bacterial lineage and afterwards had been widely disseminated by HGT, and then lost in some lineages. Both archaea and eukaryotes seem to have acquired rhomboids on several independent occasions. In particular, at least two HGT events seem to have contributed to the origin of eukaryotic rhomboids, one of them yielding the RHO subfamily and the other one the PARL subfamily, with a possible additional HGT in plants (Figures 2,3).

Given the broad phyletic representation of both subfamilies of eukaryotic rhomboids, both the RHO subfamily and the PARL subfamily must have been acquired through HGT at an early stage of eukaryotic evolution, definitely before the divergence of the major crown-group lineages. This early epoch in eukaryotic evolution is thought to have been dominated by HGT from multiple bacterial symbionts 3031.

An alternative to this multiple-HGT scenario is that LUCA already had multiple, paralogous rhomboids, which evolved by a series of ancient gene duplications, and the odd topology of the phylogenetic tree is due primarily to differential loss of these ancient paralogs. Although this cannot be ruled out formally, this hypothesis implies the existence of an elaborate signaling system in LUCA and, accordingly, suggests that LUCA was a complex organism, which might have had as many genes as modern bacteria. Theoretical analysis of evolutionary scenarios constructed on the basis of the phyletic patterns of COGs by applying the parsimony principle shows that the complexity of the inferred gene set of LUCA critically depends on the relative rates of gene loss and HGT at the early stages of evolution 32. A complex LUCA with around 2,000 genes is predicted only when one assumes that the rate of gene loss is an order of magnitude greater than the rate of HGT. However, explicit reconstruction of the gene set of LUCA under the assumption of equal rates of gene loss and HGT leads to a hypothetical genome that consists of only around 600 genes but appears to be 'compatible with life', that is, it includes genes responsible for most, if not all, essential cellular functions 32. We currently believe that this is the most realistic, albeit inevitably imprecise, reconstruction of LUCA's gene set. With respect to the rhomboid family and other families whose phylogenetic trees show similar patterns, this makes the multiple-HGT interpretation the scenario of choice. Further theoretical, comparative-genomic and experimental analyses aimed at determining relative rates of gene loss and HGT will help in a more objective assessment of the validity of this argument.

The multiple-HGT interpretation of the evolutionary history of the rhomboid family, while supported by the above argument, seems, at least at first glance, distinctly counter-intuitive, given that this family is nearly ubiquitous among extant life forms. Indeed, when attempts are made to construct parsimonious evolutionary scenarios on the basis of phyletic patterns alone 313233, there is no chance that such a widespread family is not assigned to LUCA. It should be realized, however, that these approaches are inherently probabilistic, and extensive HGT can fool them 34. For the rhomboid family, the multiple-HGT mode of evolution seems to be particularly plausible. It seems likely that the ultimate ancestor of the rhomboid family evolved from a nonenzymatic integral membrane protein, probably a transporter that might have been involved in an early primitive form of export of signaling peptides in bacteria. The protease active center might have evolved in such a transporter by chance emergence of the suitable catalytic amino acids within two or three of the TMHs (Figure 4). This would enable the transition from simple transport to the RIP mode of controlled export of signaling molecules. Emergence of RIP could have conferred a major selective advantage on the respective bacteria and might have resulted in an evolutionary sweep whereby the gene carrying this trait was repeatedly fixed, rather than eliminated, after HGT. In terms of the evolution of sequence itself, the requirements for the conservation of the protease activity apparently 'locked' the rhomboid family in a regime of relatively slow evolution, which ensures significant sequence similarity between all family members (Figure 1). The scenario of origin from non-catalytic transporters might potentially apply to other integral membrane enzymes, including intramembrane proteases involved in RIP, such as presenilins and their homologs 1415 and the archaeo-eukaryotic signal peptide peptidase 35.