Introduction

Quartet analysis was first described by Strimmer and Von Haeseler (1997) utilizing posterior probabilities to analyze the phylogenic relationships among four closely related species. They used simulations to differentiate two different modes of evolution: star phylogeny, in which four taxa evolved simultaneously (i.e., a polytomy) versus bifurcation, in which taxa came about by splitting lineages in two. Strimmer and Von Haeseler (1997) demonstrated that quartet analysis is a powerful technique that provides information that standard phylogenetic techniques could not. Within a genera, the computational creation of standard phylogenies for the comparison of all gene pairs, even with relatively small genome sizes are prohibitive. Furthermore, quartet analysis can also be useful as a technique for the verification of standard rDNA sequencing and phylogenic reconstruction, insuring that the majority of other genes support specific relationships in rDNA trees.

Since the original paper by Strimmer and Von Haeseler, other authors have utilized quartet analysis to demonstrate genomic mosaicism and horizontal gene transfer (HGT) by observing phylogenetic discrepancies in different genes within the same group (Zhaxybayeva and Gogarten 2002, 2003). This technique has also been improved by using more conservative branch support measures such as bootstrap values or by including additional taxa in tree reconstructions before mapping to reduce the influence of long-branch attraction (Zhaxybayeva and Gogarten 2002, 2003).

Crenarchaeota is a phylum in Archaea well known for its extreme thermophilic members, although recent studies have shown them to be ubiquitous in the marine environment (DeLong 1992, Fuhrman 1992, Barns et al. 1996). Crenarchaeota are also known to have high levels of HGT (Barns et al. 1996; Ribeiro and Golding 1998; Young 2001; Kunin et al. 2005; Reno et al. 2009; Nelson et al. 1999). However, a comprehensive analysis of HGT within the phylum has not been performed; there is a need to develop methods that are comprehensive, quantitative, and statistical (Ragan 2001).

In the literature, putative HGTs provide only indeterminate evidence when they are not put into a phylogenetic or systematic framework. Reciprocal BLAST and other distance based methods (Reno et al. 2009; Garcia-Vallvé et al. 2000) are not sufficient to eliminate other possibilities, such as gene loss and long-branch attraction. Garcia-Vallvé et al. (2000), therefore, suggested that their findings should represent a “first approximation” of the extent of HGT within the taxa studied. In their critical review of HGT studies, Kurland et al. (2003) discussed how distance based methods can lead to false positives. They suggested that the widespread use of BLAST as a means of determining HGT has systematically inflated the extent and importance of HGT. Salzberg et al. (2001) demonstrated just how dangerous distance based methods can be when providing erroneous results. They reviewed published data that purported HGT from bacteria to human based on protein BLAST searches between genomes of different organisms. They concluded that the results were premature and demonstrated that gene loss and consequently long-branch attraction were much more likely the cause of the genomic mosaicism that linked human and bacteria genes.

The workflow of quartet analysis in Fig. 1 is illustrated as a means to analyze genomic mosaicism and HGT. A quartet of orthologous proteins (QuartOP) from four taxa has only three possible topologies with one branch support value per topology, and can be graphed in two dimensions on an equilateral triangle called barycentric coordinates (Strimmer and Von Haeseler 1997; Zhaxybayeva and Gogarten 2002, 2003). Each corner of the triangle is mutually exclusive and represents one possible tree topology. The closer a point is to a corner, the more support it lends to that topology. A set of four taxa that has strongly supported relationships will have most or all of its data points within one particular corner.

Fig. 1
figure 1

Workflow of quartet analysis (details are described within the text). Four taxa were chosen (1.1); QuartOPs were found using BLAST (1.2); QuartOPs were bootstrapped and were used to create maximum likelihood phylogenies (1.3) and were then graphed in barycentric coordinates (1.4)

Full size image

At the time of analysis, 35 annotated Crenarchaeota genomes were downloaded from the UCSC Archaea Genome Database (archaea.ucsc.edu). It was hypothesized that this workflow could be used to distinctly demonstrate HGT within these 35 genomes without a priori knowledge of biogeography, molecular physiology, or other such information. HGT can be determined through purely analytical approaches.

Four taxa were chosen for quartet analysis. The results suggested horizontal transfer of a ferredoxin-related gene. At the end of this analysis, we discuss how the workflow could be used to analyze intra-phylum HGT more comprehensively. We also discuss the strengths and limitations of this method. The four taxa chosen were: Metallosphaera cuprina Ar-4T, Acidianus hospitalis W1T, Vulcanisaeta moutnovskia 768-28T, and Pyrobaculum islandicum DSM 4184T.

Methods

Preliminary Analysis of 16S/26S Phylogeny

From the UCSC Archaea Genome Database (archaea.ucsc.edu), 35 annotated Crenarchaeota genomes were downloaded. A distant outgroup within a different archaeal phylum, Archaeoglobus profundus DSM 5631T, was also used for analysis. Using a Perl script, 16S and 26S sequences were extracted from the genome annotations and assembled into a fasta file, which was subsequently aligned with ClustalW using standard settings (Gap Opening Penalty: 15, Gap Extension Penalty: 6.66, Weight Matrix: IUB). These alignments were used to build neighbor joining and maximum likelihood trees (Supplemental files A) in MEGA 5.05 (for maximum likelihood: general time reversible model G+I, partial deletion of gaps, nearest neighbor interchange for heuristic search, and 500 bootstraps; for neighbor-joining: Tamura-Nei model, partial deletion of gaps, and 500 bootstraps) (Tamura et al. 2011).

Choice of Quartet

When selecting the four species for quartet analysis, the following need to be considered for selection of the two sets of two paired species: (1) each paired species should have a short branch length between them in the 16S phylogeny in order to reduce the chances for long-branch attractions and (2) the distance between the two separate pairs is large relative to the branch lengths between species in each pair, so that the QuartOP topologies as well as 16S topology would be readily apparent and clearly supported. The two pairs were M. cuprina Ar-4T and A. hospitalis W1T (with a branch distance between them of 0.09 in the maximum likelihood phylogeny; V. moutnovskia 768-28T, and P. islandicum DSM 4184T with a branch distance of 0.06. The major node separating the two groups had a strong bootstrap branch support of 95 and the branch distance between the two groups was 0.112.

Generating QuartOPs

The annotated protein sequences from each species in the quartet were then extracted using BioPerl’s searchIO function, and these sequences were used to create a BLAST database which included c.a. 1,000 proteins per genome using the makeblastdb command from NCBI’s command line BLAST suite. The size of the database for BLASTP was determined by the total described proteins in the database for the compared species. The QuartOPs were determined with the BLASTP function using the coding sequences from one of the quartet species, in this examination the M. cuprina genome as query sequences. The query species used should be selected based upon the largest available genome. There are 243 ORFs in the M. cuprina genome that do not occur in its closest neighbor, Metallosphaera sedula (Liu et al. 2011). The other species of the quartet were also used as queries and no significant differences were observed. The expect value (E) was set at 1e-6 (http://www.ncbi.nlm.nih.gov/blast/Blast.cgi?CMD=Web&PAGE_TYPE=BlastDocs&DOC_TYPE=FAQ#expect). The lower expect value used results in a better statistically significant match. These generated top hits (highest bit score) from each of the compared sequences together with the query sequence from M. cuprina formed individual QuartOPs, that were placed into a fasta file and used as the quartets in the maximum likelihood reconstruction. A considerable reduction in quartOPs was observed during our examination, when using bidirectional BLAST. For the current analysis a bidirectional BLAST was not performed as it results in the potential exclusion of true orthologs (Wolf and Koonin 2012).

Alignment, Tree Building, and Graphing

Each QuartOP was aligned with ClustalW using default settings and used to build bootstraps of four species maximum likelihood trees using RAxML (matrix: PROTGAMMABLOSUM62F, bootstraps: 500) (Stamatakis 2006). The generated bootstrap support values were then graphed in barycentric coordinates using a modified excel template developed by W. Vaughan (wvaughan.org/ternaryplots.html). These processes were automated using a Perl script to transfer output from one program to the next.

Genomic Mosaicism: HGT Versus Gene Loss

From eight putative genes with bootstrap values greater than 95 and observed in the barycentric plot (Fig. 2, corner B), one QuartOP (a ferredoxin iron-sulfur binding domain protein) disagreed strongly with the majority topology having the greatest bootstrap support and was chosen for further analysis. The values in the barycentric plot are mutually exclusive as a quartOP cannot occupy more than one corner. The other 7 putative genes were not analyzed in this study. Orthologs from 35 genomes were found by blastp, and a maximum likelihood tree was built using MEGA. Compared with RAxML, MEGA provides better graphic support and was useful for confirmation of RAxML results. This tree was then compared with the maximum likelihood 16S phylogeny.

Fig. 2
figure 2

Plot of branch support values in barycentric coordinates of the chosen quartet. The topology (A) had 879 QuartOPs that agreed and 811 that are strongly supported (>95 bootstrap). (B) Had 29 QuartOPs that agreed and 8 that are strongly supported. (C) Had 35 QuartOPs that agreed and 19 that are strongly supported. This topology for (A) was in opposition with the 93 % majority. The QuartOP for (B) was chosen for further analysis as it strongly supported the topology of (A. hospitalis, V. moutnovskia) and (M. cuprina, P. islandicum)

Full size image

Results

16S and 26S Sequence Analysis for Quartet Selection

When performing sequence analysis for the selection of quartets, the phylogenies with the best basal branch support should be utilized for comparison. In comparison with 16S, the 26S phylogeny had less branch support at the basal branches in both maximum likelihood and neighbor joining trees, yet did not improve branch support at the apical branches (Supplemental file A). The outgroup also fell between Crenarchaeota species in both 26S phylogenies indicating long-branch attraction. The 16S maximum likelihood and neighbor joining phylogenies produced very similar topologies, the basal outgroup branched earlier than other members within the phylum. The 16S phylogenies had a better basal branch support in comparison with 26S and was less susceptible to long-branch attraction. The weaker support of the 26S results compared with 16S may be due to a high rate of substitution determined lambda for the 26S sequence of these species. Therefore, 16S phylogenies were used as the basis for choosing the quartet.

Incongruent QuartOPs

The sequences with the highest bit score from each of the compared sequences together with the query sequence from M. cuprina were used to form a QuartOP. After alignment with ClustalW and bootstrap construction using RaxML a barycentric graph was constructed. The barycentric graph is useful for graphically showing phylogenic grouping. The graph showed strong support for the topological grouping of M. cuprina with A. hospitalis and P. islandicum with V.moutnovskia. This topology garnered support from 93 % of the QuartOPs and very strong support from 85 % (bootstrap branch support greater than 95). However, there were 27 QuartOPs that strongly supported other topologies (Fig. 2) which could indicate HGT. While any of these QuartOPs could be further analyzed for HGT, for the purposes of this investigation a QuartOP with strongest support was utilized for further analysis.

HGT of a Ferredoxin Gene

The QuartOP that was chosen strongly supported the topology of (A. hospitalis, V. moutnovskia) and (M. cuprina, P. islandicum) (B in Fig. 2). This topology was in opposition with the 93 % majority (A in Fig. 2), as well as 16S (Fig. 3). Assuming that the support for a conflicting topology represents a real case of genomic mosaicism, this could be due to either HGT or gene loss (Fig. 4). The two cases can be differentiated by comparing the topologies of the ferredoxin phylogeny. It would be expected that in the case of HGT, the organism that putatively received the transferred gene would cluster closely with the other group. The ferredoxin phylogeny includes all 35 taxa and the 16S phylogeny provides a good representation of the overall organism phylogeny. If gene loss in M. cuprina occurred, the M. cuprina ortholog in the QuartOP would be an unrecognized paralog of some distant gene. This would place M. cuprina on a long-branch, which could have led it to group with either V. moutnovskia or P. islandicum due to long-branch attraction. Alternatively, if the gene was transferred from V. moutnovskia to A. hospitalis, A. hospitalis would fall with Vulcaniseata after V. moutnovskia branched off from P. islandicum.

Fig. 3
figure 3

Comparison of ferredoxin-related gene phylogeny (upper tree) with that of 16S (lower tree). A. hospitalis falls in with V. moutnovskia, suggesting HGT. The suspected HGT event is represented visually by a trace line on the 16S phylogeny

Full size image
Fig. 4
figure 4

Schematic concept of genomic mosaicism caused by HGT or gene loss. The transfer or loss of a suspected gene can lead to phylogenies (bottom) that differ from the “true” or assumed organism phylogeny (top). The phylogeny of orthologous genes with loss from A2 (at low right) would differ from that if instead A2 received a gene from B1 (lower left). HGT caused A2 grouping with B1 after splitting from B2

Full size image

Comparing the ferredoxin gene phylogeny with the 16S phylogeny, M. cuprina maintains its relative place, while A. hospitalis groups with V. moutnovskia (Fig. 3). This comparison suggests that HGT is more likely than gene loss. A. hospitalis is not on a long-branch and no additional taxa are included between V. moutnovskia and P. islandicum. These results suggest HGT from V. moutnovskia to A. hospitalis. If gene loss were to have occured, we would have expected A. hospitalis to be separated from V. moutnovskia and P. islandicum with other species interleaved. Based on grouping, the relative chronological order of the transfer event can also be inferred. Since the ferredoxin in A. hospitalis grouped with V. moutnovskia and not P. islandicum, the gene was transferred after the V. moutnovskia lineage diverged from P. islandicum. Since the ferredoxin was transferred to A. hospitalis and neither of the two Metallosphaera, the ferredoxin gene was received after A. hospitalis diverged from the Metallosphaera lineage. MEGA was used to reconstruct the full 35 species phylogenies due to the stronger graphing support and to corroborate the findings of RAxML.

Discussion

HGT occurs more often in microorganisms that reside in similar environments (Zhaxybayeva and Gogarten 2003). Even between the Bacteria and Archaea domains, HGT can occur with high frequency (Nelson et al. 1999; Koonin et al. 2001). Nelson et al. (1999) found that the novel bacterial species Thermotoga maritima had approximately 24 % of genes more similar to Archaea than to its closest bacterial relatives.

Vulcanisaeta moutnovskia and A. hospitalis are acidophilic thermophiles and are found in hot springs around the world (You et al. 2011; Gumerov et al. 2011; Mavromatis et al. 2010). It is not unreasonable that a HGT event was observed between the V. moutnovskia and A. hospitalis lineages, as high rates of HGT has been reported in Crenarchaeota and many are extremophiles that reside in the same environment (DeLong 1992; Fuhrman 1992; Barns et al. 1996). However, there are still open questions, for example, which genes are more frequently horizontally transferred, and to what extent HGT takes place within the phylum or with other species outside the phylum.

From single gene sequences extracted from the genome annotation files, we have observed 27 homologous sequences that are putative HGTs (B and C in Fig. 2) and the results have demonstrated the phylogenetic relationships among four taxa, given the support of the majority of QuartOPs. The use of bootstrap for the branch support values, rather than, for example, maximum likelihood posterior probability, leads to a conservative estimation of the extent of HGT (Zhaxybayeva and Gogarten 2003). Manual inspection of the alignment and incorporation of one specific QuartOP (a ferrodoxin related gene) into a larger phylogeny has confirmed that this gene was indeed horizontally transferred from the V. moutnovskia lineage to A. hospitalis.

The results have shown that quartet analysis can be used in screening homologous sequences for putative HGTs and is useful in visually describing genomic mosaicism and HGT within four taxa. However, each putative HGT must be incorporated into a larger phylogeny to differentiate HGT from unrecognized paralogy, composition bias, or phylogenetic reconstruction artifacts. A putative HGT may also be due to composition bias (a type of convergent evolution), providing a false positive, as pointed out by Jermiin et al. (2004).

Some proteins may be excluded due to lack of orthologs in all four species in a quartet. Because an ortholog must be found within all four species of a quartet to become a QuartOP, the analysis often loses half or more of a given taxon’s genes. The analysis done by Zhaxybayeva and Gogarten (2002) demonstrated this problem, finding taxa quartets that had as few as 82 genes in QuartOPs. This is a potentially larger problem than a first order estimate based purely on percentages, since variable presence of a gene in various taxa is indicative and characteristic of HGT (Gogarten and Townsend 2005).

Compared with standard phylogenetic techniques, quartet analysis is better suited for finding putative HGT events. Creating standard phylogenies for comparative analysis using total genomes are computationally demanding. Zhaxybayeva et al. (2006) and Kubatko and Degnan (2007) discuss the common technique of concatenating sequences to help resolve phylogenetic discrepancies and the use of the resulting phylogenies as a reference to which to base the level of HGT. Kubatko and Degnan (2007) analyze the concatenation approach by simulation and show that it performs poorly in predicting and supporting true topologies. Zhaxybayeva et al. (2006) point out that concatenation requires the assumption that the concatenated genes have a uniform phylogenetic history. No such assumptions are required when using quartet analysis.

Zhaxybayeva et al. (2004) discusses a related technique, the bipartition method, in which they used “QuintOps” instead of “QuartOps.” Here, they took a similar approach of finding orthologs between genomes, by performing BLAST of ORFs against the genomes of the other species. They used this method to help determine tree topologies that were left unresolved by standard phylogenetic techniques, by looking at “plurality” (i.e., majority) of supported quintOps.

Quartet analysis can show relationships among only four species. It cannot say anything about the other 31 Crenarchaeota. Limited computational power potentially plays a large role in this type of analysis for large groups since the number of possible quartets grows exponentially as more species are added. It may simply not always be practical or even useful to comprehensively analyze large data sets. Furthermore, not all quartets are equal. A quartet where all species are very closely related may not provide any additional information of interest; there may not be enough divergence between the species in the quartet to resolve phylogenetic topology. Furthermore, if one species in a quartet is substituted for another that is a very close neighbor, this new quartet might not provide any additional information. It is noteworthy that substituting V. moutnovskia for V. distributa would have found the same ferredoxin HGT event (data not shown). Other quartets may give faulty information, if each of the taxa is on long-branches that separate them from the other three species.

Conclusions

Analysis by quartets has not yet been performed on any group in the phylum Crenarchaeota. The results have shown that quartet analysis can be used to screen homologous sequences for putative HGTs and is useful in visually describing genomic mosaicism and HGT within Crenarchaeota species and taxa in general. It should be possible to create a mechanical algorithm to choose informative quartets based on a global phylogenetic topology, such as one based on 16S, and use those quartets as comprehensive representatives of the genomic mosaicism of the whole group. Such an algorithm does not yet exist, but it seems inevitably possible.