The Planctomycetes-Verrucomicrobia-Chlamydiae (PVC) superphylum is an assemblage of bacterial phyla which is consistently recovered as a monophyletic group in trees using a range of different phylogeny estimation methods and data (Wagner and Horn, 2006; Pol et al., 2007; Hou et al., 2008; Pilhofer et al., 2008; Kamneva et al., 2010). It includes the Planctomycetes (Fuerst and Sagulenko, 2011), the Verrucomicrobia, the Chlamydiae, the Poribacteria, the Lentisphaerae, and the OP3 candidate phyla (consisting only of uncultured organisms), along with several other groups. A range of characters that were previously either considered absent or rare amongst bacteria, but which are common or ubiquitous in archaea or eukaryotes, have recently been identified in some PVC members (Table 1, Appendix). These include, for example, the presence of membrane coat (MC)-like proteins and condensed DNA. A list of such features has been described elsewhere (Devos and Reynaud, 2010; Reynaud and Devos, 2011). In this article we review several analyses of the taxonomic distribution of those characters.

The taxonomic distribution of a heritable character can be shaped by many different genetic events. These include vertical or horizontal inheritance of genes responsible for the phenotype, loss or duplication of these genes, and independent or convergent acquisition of a phenotype. Phylogenetic trees estimated from molecular data, i.e., from multiple sequence alignments (MSAs), play a key role in our attempts to estimate the contribution of those mechanisms to the evolution of any character. To use MSAs in this way, we need to know (some of) the genes or proteins responsible for expressing the character of interest, and to be able to estimate reasonably accurate MSAs for the relevant sequences. The precision of the phylogenies estimated from these MSAs is assessed by branch support values such as non-parametric bootstrap values; note, however, that such values represent the precision (i.e. sampling error) with which the topology of these phylogenetic trees are estimated, not the accuracy of these estimates. The resulting gene or protein tree is then compared to the corresponding organism tree. The phylogenetic tree describing the evolution of a gene family which does not experience any gene loss, duplication, incomplete lineage sorting, intra-gene recombination, or horizontal gene transfer (HGT), and which only evolves via a process of point mutation, will have the same topology as that of the corresponding organism (Figure 1). In contrast, where the evolution of a gene family does involve gene loss, duplication, incomplete lineage sorting, intra-gene recombination, or HGT, the gene tree topology may no longer correspond to the organism tree topology. For this reason, the observation of differences between the topologies of gene and organism trees has been used to identify gene families that may have experienced gene loss, duplication, or HGT during their evolution.

Clearly, the accuracy of such approaches depends on the accuracy of the gene and species trees used in the analysis; if the topology of either (or of both) the gene and organism trees are inaccurately estimated, then the evolutionary history of the gene family may be wrongly inferred.

Assuming gene and species trees have been correctly inferred and rooted, and that no incomplete lineage sorting or intra-locus recombination has occurred (Degnan and Rosenberg, 2009; Stolzer et al., 2012), any difference between gene and organism trees can be explained by either (i) only gene loss and duplication events, (ii) only HGT and gene loss events, or (iii) a mixture of gene loss, gene duplication, and HGT (Figure 1). Note that there are many possible causes of error that could lead to inaccurate estimation of gene or species tree topologies. Such errors could lead to inaccurate conclusions concerning the frequency of gene loss, duplication, and HGT events within a gene family. These include errors in sequencing or gene prediction (Prosdocimi et al., 2012), alignment (Löytynoja and Goldman, 2008), or differences between the substitution model used to estimate the phylogeny and the true process of point mutation experienced by sequences during their evolution (Huelsenbeck and Rannala, 2004; Mar et al., 2005; Kolaczkowski and Thornton, 2008; Roure and Philippe, 2011). Furthermore, using a comparison of gene and organism trees to analyze the frequency of gene loss, duplication, and HGT events requires rooted phylogenetic trees; even with correctly estimated unrooted topologies for both gene and organism trees, errors in the inference of the position of the root of these trees will also lead to inaccurate inference of the frequencies of gene loss, duplication, and HGT (Swofford et al., 1996; Huelsenbeck et al., 2002). The inference of the position of the root of the Tree of Life (ToL) offers additional challenges due to the lack of outgroup organisms to use for comparison (Bapteste and Brochier, 2004; Lake et al., 2009).

To estimate the relative contributions of gene loss, duplication, and HGT to the taxonomic distribution and phylogenetic tree topology of a gene family requires a model of the processes, including the relative frequencies, of gene loss, gene duplication, and HGT. The accuracy with which the parameters of evolutionary models (such as phylogeny topologies, but also relative frequencies of different kinds of evolutionary change such as point substitutions between different nucleotides, but presumably also relative frequencies of gene loss, duplication, and HGT) are estimated is reduced as the evolutionary time-scale increases. Thus, we should be very cautious about the inference of ancient HGT events that may have occurred close to the origin of the eukaryotes, as is the case for the characters discussed in this review; alternative scenarios involving only vertical transmission of genetic material should also be carefully considered.

Currently, relatively few characters with unexpected taxonomic distributions in PVC can be analyzed in this way, i.e., characters for which one or more of the gene families associated with the character has been identified, and for which MSAs can be built that give reasonably precise estimates of the topology of the phylogenetic tree of the gene or protein families. An example of a character that cannot be studied in this way is the presence of MC proteins in PVCs (MCs are also found also in all eukaryotes; Santarella-Mellwig et al., 2010); we have identified several relevant gene families, but the sequences of these families are so different from each other that it is impossible to confidently estimate an MSA for them, despite the clear similarity of their predicted structural features and domains (Devos, 2012). Thus, so far only three characters have been analyzed in this way: the presence and absence of (i) sterol synthesis, (ii) tubulin, and (iii) c1 transfer genes. The results of these analyses are summarized in Table 1 and in the Appendix.

An analysis of the taxonomic distribution of sterol synthesis genes in eukaryotes, planctomycetes, and other bacteria illustrates many of the issues discussed here (Pearson et al., 2003). Sterol synthesis is found in almost all eukaryotes, but in only a few bacteria. Pearson et al. present phylogenetic trees of the only two genes involved in sterol synthesis in PVCs (Appendix). Both trees include several (more than 10) eukaryotic sequences, at least one non-PVC bacterial sequence, and one PVC sequence. Both trees contain an internal branch that partitions all eukaryotic sequences from all bacterial sequences (Figure 2).

Because the estimated gene tree corresponded to the canonical “16S rRNA phylogeny” (with three monophyletic domains of eukaryotes, bacteria, and archaea), the authors concluded that “a recent lateral gene transfer from higher-order eukaryotes to bacteria is unlikely. The bacterial sterol biosynthesis genes are not closely related to genes found in any extant group of eukaryotes, and the bacterial gene trees are parsimonious with respect to 16S rRNA phylogeny.” Alternatives to the 16S rRNA phylogeny for the ToL were not considered.

However, due to the absence of members of these gene families from most bacterial genomes, the authors concluded that the observed taxonomic distribution of these families is best explained by an ancient HGT between bacteria and eukaryotes. Alternative explanations involving no HGT, i.e., only vertical genetic transfer and gene loss, were considered less likely than the HGT scenario. This is because, in the implicit model of gene family evolution used by the authors, one HGT is considered more likely than the large (but unspecified) number of gene losses inferred by alternative scenarios.

Another reason why non-HGT scenarios are considered unlikely in this, and other similar, analyses, is that they require the inference of a change in the rate of gene loss along different lineages. More specifically, they require the inference of an initial period in which gene loss is high, followed by a long period in which the gene families are retained by a small number of organisms, i.e., lineages in which the rate of gene loss is much reduced. Thus, one (or a few) HGT events are considered more likely than a relatively more “complex” scenario in which the rate of gene loss (and other events) varies (i.e., is time-heterogeneous) across the phylogeny.

Similar conclusions were reached, for similar reasons, by many analyses of PVC characters with unexpected character distributions (Table 1 and Appendix.).

In many cases, these analyses have concluded that the most likely explanation for the phylogenetic distribution of these gene or protein families involves one or more ancient HGT events. In this article, we review these previous studies, focusing in particular on two aspects of their results and conclusions. Firstly, we highlight, as reported by the authors, that the divergence of PVC and eukaryotic/archaeal members of these families are ancient events. Secondly, most of these studies conclude that one or more HGT events are likely to have occurred during the evolution of these families; we discuss three issues that make us cautious about making such inferences, i.e., that (i) it is difficult to obtain accurate inference of phylogenetic tree topologies for single gene families that diverged over the time-scales involved in these analyses (ii) inference of HGT also requires an estimate of the species tree for the organisms involved in the analysis, which is also difficult, and (iii) the need for increased use and further development of probabilistic models of the different kinds of genetic events that can shape the taxonomic distribution of gene families (i.e., gene duplication, vertical inheritance, gene loss, and HGT).

Therefore, we encourage caution in the inference of ancient HGT events to explain unexpected character distributions, and look forward to the further development of probabilistic models for comparing HGT and non-HGT-based models. Additionally, we think that it is important for such analyses to consider alternatives to the three-domain 16sRNA-based ToL when assessing the evidence for HGT in general. In the case of the PVC characters, alternatives include the possibility that the engulfment of a thaumarchaeon by a PVC bacterium was involved in the origin of the eukaryotes (Forterre, 2010), a stepwise vertical evolution of archaea and eukaryotes from a PVC common ancestor (Devos and Reynaud, 2010; Reynaud and Devos, 2011), or the reductive evolution of Planctomycetes from a complex proto-eukaryote-like last universal common ancestor (Fuerst and Sagulenko, 2011, 2012). Many of these issues are generally relevant to inference of HGT events, not just in the context of unexpected character distributions involving PVCs.

Estimating patterns of genetic information transfer, i.e., phylogenetic tree topologies, that include ancient lineage divergences is difficult (Gribaldo and Philippe, 2002; Delsuc et al., 2005). In particular, over large time-scales, we expect variation in the (nucleotide or amino acid) substitution processes acting on different branches of the phylogeny, while most phylogenetic inference uses homogeneous substitution models that assume the same process acts on all branches of the tree (Roure and Philippe, 2011). Model misspecification of this kind has been found in a range of different sequence datasets, and has been shown to potentially lead to systematic error in phylogeny reconstruction (Shavit Grievink et al., 2010). Such errors can lead to the estimation of branches with high support values (i.e., with low sampling error) that are not present in the true phylogeny.

The size of the dataset (i.e., the number of MSA alignment columns) used to estimate phylogenies for individual gene families is clearly restricted by the size of the gene being analyzed. Organism phylogenies are typically estimated from much larger data sets obtained by combining data from many different genes. Thus, phylogenies estimated for individual genes typically experience increased sampling error compared to organism trees (Jeffroy et al., 2006; Rokas and Carroll, 2006; Castresana, 2007). The phylogenetic trees used to investigate the role of HGT in establishing unexpected taxonomic distributions of characters in PVCs all involve ancient lineage divergences, typically around the time of the origin of the eukaryotes or earlier. Thus, we should be aware that there could be significant undetected errors in both the gene and organism phylogenies used in these analyses. In particular, some of the gene trees used in these analyses include many long branches clustered together within the tree (Pearson et al., 2003; Bauer et al., 2004; Chistoserdova et al., 2004), a feature that could be the result of systematic errors often referred to as “long branch attraction” (Huelsenbeck, 1997; Anderson and Swofford, 2004; Bergsten, 2005).

Many of the analyses reviewed in this paper assume that the three-domain ToL estimated from early 16S rRNA analyses and other later studies (Woese et al., 1990) is correct, and do not take into account any alternative ToLs. However, the organism phylogenies used to infer HGT in the evolutionary history of PVC gene members may contain potentially major topological errors. One way of highlighting this is to consider current disagreements concerning the ToL. In recent years, several alternatives to the classical three-domain ToL have been proposed, for example the “eocyte” hypothesis in which the archaea are not monophyletic (Cox et al., 2008; Foster et al., 2009), and many alternatives have been proposed for other features of the ToL, such as the interrelationship between eukaryotes and archaea, and the relationships between the major eukaryotic groups (Burki et al., 2007; Hampl et al., 2009; Roger and Simpson, 2009; Desmond et al., 2010). In addition to this disagreement and discussion of fundamental features of the ToL, it has also been shown that the phylogenetic signal present in “universal” proteins (i.e., which are found in the majority of cellular organisms) are not sufficient to resolve with confidence the topology of the ToL (Desmond et al., 2010). A recent analysis of the origin of land plants (Laurin-Lemay et al., 2012) nicely illustrates many factors that can make it difficult or impossible to estimate a true organismal phylogeny, including closely spaced speciation events, incomplete lineage sorting, gene duplications, and HGTs. Of particular relevance to analyses involving genes found in the PVC superphylum is the ambiguity of the phylogenetic position of the group (Stackebrandt et al., 1984; Janssen et al., 1997; Ward et al., 2000; Jenkins and Fuerst, 2001; Brochier and Philippe, 2002; Di Giulio, 2003; Fieseler et al., 2004; Teeling et al., 2004), further highlighting the uncertainty associated with the topology of the ToL.

Hopefully, in the future, improved taxon sampling and the use of more sophisticated models of character evolution may lead to a more accurate estimate of the ToL (Philippe et al., 2011).

Thus, given the range of alternatives proposed for many different parts of the ToL, it is clearly important to be cautious when using it to investigate possible HGT events. In particular, where alternative hypotheses for the topology of regions of the ToL may lead to different conclusions about the occurrence of HGT, then such analyses should be carried out using all plausible alternative ToLs.

A key component of HGT inference using the comparison of gene and species trees is a model, whether implicit or explicit, of the processes of vertical and horizontal genetic transmission, as well as of the relative probability of gene duplication and loss.

The use of explicit probabilistic models of these processes enables the use of formal statistical tests of whether the observed data (gene and organism trees) better fit a scenario with or without HGT. However, relatively little attention has been given to models of this kind, and in most cases, the inference of HGT is considered outside such an explicit model-testing framework. Part of the reason for this is perhaps that researchers are cautious about explicitly building models for a process (HGT) which is relatively poorly understood, in terms of factors likely to influence variation in the rate at which it occurs such as gene size, features, and degree of divergence of accepting and donating genomes, and other factors (Sorek et al., 2007; Boto, 2010). Similar uncertainty surrounds the dynamics of the processes of gene duplication and gene loss. For example, there is strong evidence that some gene families in parasitic bacteria experience an increased rate of gene loss, analogous to the heterotachy observed in the process of point mutations in sequence evolution (Kolaczkowski and Thornton, 2008), i.e., where the rate of an evolutionary process varies along different lineages of a phylogeny.

However, despite the problems associated with the use of explicit models of such processes, we feel that it would be good if assertions of HGT (or non-HGT) were carried out in the context of exploration of some of the existing models of this kind (Thiergart et al., 2012). Explicitly declaring the sets of assumptions being made when HGT is (or is not) being inferred, i.e., by using explicit probabilistic models of this kind, would ease the process of identifying and discussing the central assumptions lying with the inferences. Hopefully, in the future, a better understanding of the processes of gene family evolution will aid the development of more sophisticated and accurate models of these processes, leading to a wider application of methods of this kind.

As described above, given:

it is important to be cautious about inferring the occurrence of ancient HGT to account for unexpected distributions of characters and gene families. Ideally, we feel such inferences should be made in the context of testing whether or not HGT is supported using a range of different explicit models, while also taking into account the uncertainty and proposed alternatives of the trees (both organism and gene trees). Thus the HGT origin of those PVC features is still not established. This uncertainty highlights the importance of taking into account new and alternative hypotheses and ideas in analyses of this kind (Devos and Reynaud, 2010; Forterre, 2010; Reynaud and Devos, 2011; Fuerst and Sagulenko, 2012). In this context, discussions about the evolutionary origins of some of the characters found in PVCs echo the initial reaction to the platypus by European scientists at the beginning of the nineteenth century; with its mosaic of characters not previously seen together in the same organism (including, amongst others, a beak like a duck, eggs similar to those of reptiles or birds, together with the fur and milk production found in other of mammals) it was initially assumed to be a hoax. In contrast to this initial reaction, however, the discovery of the platypus provided a rich source of ideas and understanding about the history and process of both mammalian and non-mammalian evolution, along with better understanding of the connections between these taxonomic groups. In a similar way, with its surprising and unexpected combination of characters, the PVC superphylum might help reveal new and unexpected links and similarities between bacteria and other cell types, including our own.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.