GC content of rRNA genes is comparatively unbiased

Several alphaproteobacterial groups have extremely AT-rich genomes. This has led to speculation that the topology in which the Pelagibacterales cluster with the Rickettsiales may be an artefact, due to a spurious attraction of AT-rich taxa [7]. We hypothesized that there may be less freedom for GC content to change in rRNA genes, compared with protein-coding genes, because rRNA is under structural constraints. To test this hypothesis, we analyzed the GC content of genomes, compared to rRNA genes, for single members of each species that was available in the IMG v350 database. We found that the GC content of rRNA genes is considerably less varied than that of the corresponding genomes (Figure 1). Linear regression analyses for the alphaproteobacterial groups gave a slope of 0.15 for the 16S sequences (R² = 0.63) and a slope of 0.20 for the 23S sequences (R² = 0.72), demonstrating that there is 5- to 6-fold less variation in rRNA gene GC content than in genomic GC content. There was even less bias after the highly variable sites were removed from our alignments (Figure S1 and Figure S2). These results validated our decision to use rRNA gene sequences for building trees that do not suffer from artefactual attraction of AT-rich taxa. However, the trend did not hold for mitochondrial LSU and SSU rRNA gene sequences. Linear regression yielded slopes of 0.69 (R² = 0.82) and 0.71 (R² = 0.79) for the 16S and 23S sequences in Figure 1, respectively. Therefore, the mitochondrial sequences were treated with caution and topologies were inferred both with and without them.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Relationship of rRNA gene vs. genomic GC content for Alphaproteobacteria and mitochondria.

The rRNA gene GC content was calculated from the entire sequence using a perl script, while the genomic GC content was taken from the IMG database. The “other orders” group includes the Caulobacterales, Sphingomonadales, Rhizobiales, Rhodobacterales and Parvularculales.

https://doi.org/10.1371/journal.pone.0083383.g001

Tree building strategy

We examined the phylogeny of the Alphaproteobacteria using single members of each species for which both the 16S and 23S rRNA genes were available, together with seven outgroup taxa from adjacent Proteobacteria classes (‘complete’ trees). All of the taxa used in our analyses are listed in Table S1. Using this complete dataset, trees were made from 16S and 23S rRNA genes alone to generate initial topologies, and subsequently with concatenations of the 16S and 23S alignments. We also tested the effect of reducing taxon selection across the tree by subsampling to an idealized membership based on the monophyletic groupings observed in the complete trees (‘trimmed’ dataset). All sets of taxa were aligned with both Muscle [23] and Arb-Sina [24] to test for incongruence between common rRNA gene alignment methods, and computed with maximum-likelihood (RAxML; [25]) using both GTRΓ and GTRCAT rate models. Each of these datasets was evaluated for compositional bias using RY recoding [26], as well as a novel method that we name RYMK recoding, and all were evaluated with and without mitochondrial sequences. Additionally, to understand the effect of sampling bias from the lack of any one group, we employed a clade-specific jackknifing approach in which we removed each of the monophyletic groups from the dataset and re-calculated the tree. This was done for both alignment methods, both rate categories, and plus or minus mitochondrial sequences. Finally, because many important under-examined alphaproteobacterial taxa have neither genome sequences nor 23S rRNA gene sequences available, we examined their taxonomic affiliation by aligning their 16S sequences with the concatenated 16S-23S bacteria dataset (‘combo’ dataset). This dataset was also tested under both alignment and rate model algorithms. These variations generated a total of 140 trees (summarized in Table 1), all of which are available in the Supporting Information.

The 16S and 23S rRNA gene trees

We began by creating phylogenies of the 16S and 23S rRNA genes separately (the ‘16S’ and ‘23S’ datasets in Table 1). These topologies provided a baseline to which we could compare our concatenated rRNA gene trees. The phylogenies were completed with taxa from the complete dataset, both alignment methods, both rate models, and both with and without mitochondrial taxa, for a total of eight trees each of the 16S and 23S rRNA genes (Figure S11). Whereas the 16S trees resolved the Holosporaceae within the clade including the Rhodospirillales and other orders (with low bootstrap support), the 23S trees did not consistently resolve the monophyly of Holosporaceae, nor its location (bootstrap supports are summarized in Figure S3 and Figure S4). Conversely, the 16S trees did not consistently resolve the Pelagibacterales in a specific location. Neither the 16S nor 23S phylogenies could consistently resolve the Caulobacterales, Rhizobiales and Rhodobacterales. A 16S tree that was used previously to classify the Alphaproteobacteria [10] had many similar results, but also some differences. For example, the Holosporaceae did not resolve monophyletically and was basal to the Rickettsiales (the Pelagibacterales were absent), whereas here the Holosporaceae is close to the Rhodospirillales.

The complete concatenated trees

To increase resolution we concatenated the 16S and 23S rRNA genes, and explored topological stability between different alignment (Arb-Sina, Muscle) and rate (GRTΓ, GTRCAT) methods, as well as the variable inclusion of mitochondrial sequences. The complete trees included a representative of each defined Alphaproteobacteria species present in IMG v350 and NCBI Genbank, for which 16S and 23S rRNA gene sequences were available (Table S1). Specifically, only a single strain was picked for each species, whereas in the cases of unclassified strain, all were chosen. In this set of eight trees (Figure S12), the Alphaproteobacteria is divided into three primary clades (representative final tree in Figure 2; bootstrap summary in Figure 3A). The earliest diverging clade is the Magnetococcales, represented by Magnetococcus marinus. One of the subsequent clades has the Pelagibacterales subtending the Anaplasmataceae, Midichloria mitochondrii, the Rickettsiaceae and the mitochondria (if present). The other clade has the Holosporaceae at the base, the Rhodospirillales as the next clade branching out, followed by the Sphingomonadales, then the remaining orders. The Holosporaceae is represented by Odyssella thessalonicensis and Caedibacter caryophilus, and is monophyletic in six out of the eight trees, but in two it resolves paraphyletically at the same location. The Holosporaceae family is currently classified as a member of the Rickettsiales [10,11], but this topology (with monophyletic Holosporaceae) was not seen in our final tree and it is only supported in a very small fraction of the bootstrap trees (2.0%, 2.0%, 0.3%, 0.4%, 0.4%, 6.8%, 4%, 1.3%). The support of the monophyly of the Rhodobacterales, Caulobacterales and Rhizobiales increased greatly when some species were reclassified, namely when Maricaulis maris was removed from the Hyphomonadaceae (Caulobacterales) and when the clade formed by Labrenzia and Roseibium species was moved from the Rhodobacteraceae (Rhodobacterales) to the Rhizobiales. These reclassifications are listed in Table S1, with associated changes in support shown in Figure S5. The support for the monophyly of the Rhodospirillales was not high, varying from 56% to 92% in the complete trees, and the monophyly of the Rhodospirillaceae had even less support (42–52%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Tree inferred with the Arb-sina aligned complete dataset under a GTRΓ model.

Bootstrap values (n = 1000) are indicated at the nodes. Red arrows indicate how a taxon or clade differs in the other regularly coded trees, with values in square brackets indicating in how many trees this is seen. If there are one or more differences within a family, this is indicated after the name of the family. The leaves of the phylogram are collapsed into taxonomic families and into the host phyla for mitochondria. The internal topology of the Rhodospirillales order is not the same in all primary trees, therefore it has been expanded to show all leaves (inset).

https://doi.org/10.1371/journal.pone.0083383.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Summary of bootstrap supports.

The values in the boxes are the bootstrap support percentages for the indicated bipartitions, from each of the eight inferences for each set, as described in the legend (i.e. with or without mitochondria, aligned with Arb-Sina or Muscle, GTRΓ or GTRCAT rate model). In the cases where the final tree did not agree with the proposed topology on the location of a clade, the bootstrap support for the proposed bipartition is represented in red, despite being absent in the final tree. A. Regular-coded complete datasets. B. RY-recoded complete datasets. C. MK-recoded complete datasets. D. RYMK-recoded complete datasets.

https://doi.org/10.1371/journal.pone.0083383.g003

The complete RY-, MK- and RYMK-recoded trees

Accepting the topology formed with the regular-coding complete datasets, we compared the GC content of the crown group to that of the Rickettsiales–Pelagibacterales and to that of the mitochondria, separately. As expected based on the data in Figure 1, GC content of the SSU and LSU rRNA genes exhibited much smaller variation than that of the genomes (Figure 4). Nevertheless, there were small, statistically significant differences when the GC contents of the crown group SSU and LSU sequences were compared to the Rickettsiales–Pelagibacterales sequences (Wilcoxon rank-sum p-values = 2.34 x 10^-15 and 4.70 x 10^-15, for SSU and LSU sequences respectively). This analysis indicated that using the rRNA genes mitigated GC bias substantially, but did not eliminate it completely.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Box plot of the distributions of GC content for the genomes, SSU and LSU rRNA genes for members of the primary clade headed by Caulobacter (C), the clade headed by Rickettsia (R) and the mitochondria (M).

Boxes indicate the interquartile range, with adjacent values as whiskers, outlying values as circles; median indicated by the horizontal black line. Box width is proportional to the total number of samples.

https://doi.org/10.1371/journal.pone.0083383.g004

In light of this analysis, together with previous concerns surrounding the AT richness in the Alphaproteobacteria [5-7], we tested RY-recoding of the dataset. This approach [26,27] masks GC content bias by recoding adenosine (A) and guanosine (G) as purines (R), and cytidine (C) and thymidine (T) as pyrimidines (Y). For consistency with our other analyses, we analyzed the recoded dataset with the GTR model. This approach has also been taken by others [7]. Support for most of the orders and families dropped in the RY-recoded datasets, compared to those with regular coding (bootstrap summary in Figure 3B; summary statistics in Table 1). There was a large decrease in support for the monophyly of the Rhodospirillales, Rhizobiales and Caulobacterales. However, support for the monophyly of some orders, such as the Pelagibacterales and Sphingomonadales, did not change greatly. The trees of the RY-recoded datasets that included mitochondria resolved the Pelagibacterales as a sister clade to the mitochondria–Rickettsiaceae–Midichloria mitochondrii–Acetobacteraceae clade (without the Holosporaceae). Conversely, three of the four trees without the mitochondria placed the Pelagibacterales as paraphyletic with the Rickettsiales (Figure S13).

As a consequence of the poor resolution obtained with the RY-recoded dataset, we utilized a novel variant: each RY-recoded dataset was concatenated to an MK-recoded dataset (“RYMK”). In the MK recoding, adenosine (A) and cytosine (C) were recoded as amino bases (M), and guanosine (G) and thymidine (T) as keto bases (K). Thus, the GC bias information is lost in the same way as RY-recoding, however the concatenated RYMK dataset retains the overall number of characters. This approach was designed as a tool to improve phylogenetic analyses of datasets with GC bias, rather than reflecting any biological process(es). We note that RY-recoding has been used similarly [27]. The site-independent nature of the GTR model ensured that it remained appropriate for analyzing the concatenated RYMK dataset.

Used on its own, MK-recoding suffered from similar issues to RY-recoding (Figure 3C and Figure S13). In particular, the MK-recoded datasets do not resolve the Rhodospirillales correctly; consequently the basal Holosporaceae is destabilized to the point that it clusters with the Rickettsiales. However, average support values and Robinson-Foulds metrics (Table 1) showed that the MK-recoded trees were more consistent than the RY-recoded trees, with respect to alignment method and rate model. On the other hand, in the RYMK-recoded datasets the grouping of the Pelagibacterales with the Rickettsiaceae, Midichloria mitochondrii, Anaplasmataceae and mitochondria, and the grouping of Holosporaceae with the remaining orders, are both more supported than in either the RY-recoded datasets or the MK-recoded datasets (bootstrap summary in Figure 3D; trees in Figure S14). These groupings also have more consistent topologies (7 of 8 trees; compare Figure 3B-C with Figure 3D).

The trimmed concatenated trees

We have asserted that it is advantageous to focus on LSU and SSU sequences because it maximizes the number of taxa that can be included. In contrast, genome concatenation studies rely on a smaller number of (fully sequenced) taxa. In order to test the impact of reduced taxon selection on overall topology, we constructed trimmed datasets with fewer bacterial taxa. By reducing the number of taxa on short branches, in clades that had more than 95% bootstrap support in the trees from the complete datasets, we reduced the number of Alphaproteobacteria sequences from 166 to 86, while leaving the number of mitochondrial sequences at 24. The 32 resulting trees (‘trimmed’ dataset in Table 1) are shown in Figure S15 (regular coding), Figure S16 (RY- and MK-recoded) and Figure S17 (RYMK-recoded). Despite the high inclusion threshold, the trees differed substantially from the complete phylogenies. Bootstrap summaries are provided in Figure S6 (regular coding), Figure S7 (RY- and MK-recoding) and Figure S8 (RYMK-recoding). None of the trimmed trees with regular coding resolved the Rhodospirillaceae monophyletically, the location of the Holosporaceae within the clade containing Caulobacterales varied substantially, and the Arb-Sina-aligned trees without mitochondria placed the Pelagibacterales paraphyletic with the Rickettsiales, albeit with poor bootstrap support (< 40%). The trees inferred from the RY-, MK- and RYMK-recoded datasets had even more diverse topologies (bootstrap summaries in Figures S6–S8). The Rhodospirillales were either paraphyletic with respect to the clade composed of Sphingomonadales, Rhizobiales, Rhodobacterales and Caulobacterales, or paraphyletic or polyphyletic in respect to the Holosporales. The Pelagibacterales was found basal to the remaining Alphaproteobacteria (except for Magnetococcus marinus) or basal to the clade containing the Caulobacterales in all but three trees. Despite the differences, the mitochondria resolved consistently in the trimmed and complete trees.

Trees with single clades removed

In an effort to identify the particular taxa or groups that may be critical to the observed instability of the complete phylogeny, we tested the effects of clade-specific jackknifing on topology. Several orders or families were removed, and trees were constructed with each of these single clades removed. The six orders and families removed were the Rhodospirillales (‘rhodoless’ dataset in Table 1), Magnetococcus marinus (‘magnetoless’ dataset in Table 1), the Rickettsiales including the Holosporaceae (‘exorickettsialess’ dataset), the Rickettsiales excluding the Holosporaceae (‘rickettsialess’ dataset), the Holosporaceae (‘hololess’ dataset), and the Pelagibacterales (‘pelagiless’ dataset). For each of the six datasets, eight trees were constructed (Arb-Sina and Muscle alignments; GRTΓ and GTRCAT rate models; plus and minus mitochondrial sequences). The 48 trees are shown in Figure S19. Removal of most orders had little effect on the tree. An exception was the problematic resolution of the Holosporaceae (e.g. paraphyletic in six trees out of 48, basal to the Pelagibacterales in two trees, basal to the Rhodospirillales in three trees, or nested in the Rhodospirillales in one). Removing the Rhodospirillales had the largest effects, destabilising the Holosporaceae and the Pelagibacteraceae (Holosporaceae basal to Pelagibacterales and Rickettsiales in two trees; Pelagibacterales basal to Holosporaceae and the clade of various orders in four trees).

Mitochondrial placement

Trees were made with and without mitochondrial sequences (Table 1). Despite the longer branches of the mitochondrial clade, the trees inferred from the various datasets with mitochondria better resolved the overall topology shown in Figure 2, as opposed to the trees without mitochondria. For example, the primary clade composed of Rickettsiaceae, Anaplasmataceae, Midichloria mitochondrii and Pelagibacterales – in which mitochondrial sequences also cluster – and the clade with a basal Holosporaceae and various orders are supported much more strongly in the datasets with mitochondria, especially the Arb-Sina-aligned datasets. This is true not only for the regularly coded full datasets (Figure 3A), but also for the RY-, MK- and RYMK-recoded ones (Figure 3B-D), and the jackknifed datasets. However, this is not the case for the trimmed datasets, where the changes in support with and without mitochondria vary between alignment method and inference model.

In contrast to the increase in support for the primary clade containing the mitochondria in the full trees, the monophyly of the Rhodospirillales and the Rhodospirillaceae loses support when the mitochondria are added, especially in the RYMK-recoded dataset. The internal topology of the Rhizobiales also changes. In the primary trees without mitochondria, the Hyphomicrobiaceae is basal to a clade formed by the Beijerinckiaceae, Methylocystaceae, Methylobacteriaceae and Xanthobacteraceae, while in those trees with mitochondria the family is basal to the sister subclade of the Rhizobiales, composed of the remaining families bar Parvibaculum lavamentivorans (basal to the two subclades) and with the addition of Labrenzia spp. and Roseibium sp. (Figure 2).

The datasets with and without mitochondria were separately aligned and trimmed. One hypothesis for the variations in topology between trees, with and without mitochondria, was that the trees with mitochondria were influenced by having fewer characters as a result of the Gblocks editing criteria. We tested this hypothesis by pruning mitochondrial sequences post-alignment, from all datasets that included mitochondria (‘complete mtDel’ dataset in Table 1). These new alignments were processed with both rate models, and the resulting trees (Figure S18) were compared to those of the alignments that were initially done without mitochondria. In order to compare the trees with mitochondria and the trees without, the mitochondrial leaves were pruned from the former group. Overall, the locations of the Pelagibacterales, the Holosporaceae and other groups of interest are the same in the ‘mtDel’ trees as seen in the other trees. However, the bootstrap support of all groups is lower (Figure S9), except for the support for the monophyly of the Holosporales in the trees from the Arb-Sina aligned dataset. Overall, the presence of the mitochondrial sequences adds support to the trees, but at the same time results in a smaller number of characters (due to Gblocks editing).

16S-23S dataset with several 16S-only sequences

There are a vast number of species represented in the databases solely by their 16S rRNA gene sequences. Within the Alphaproteobacteria, there are 975 validly-described type strains [28], against approximately 150 genotyped species. Several of the species for which we have only 16S rRNA gene sequences branch deeply within the tree, to the point that some are reported to be part of their own orders. As a consequence, a group of deep-branching taxa with only 16S sequences was aligned to the other 16S sequences, and the 23S rRNA gene positions left as missing data in the concatenated alignments. The result was the ‘combo’ dataset (Table 1).

In the resulting trees (summary in Figure S10, all trees in Figure S20), the deep-branching species that were added reduce the support for several nodes, indicating that this approach is not ideal. However, it does offer a snapshot into the diversity of the Alphaproteobacteria that is not covered by the genome databases. Several species classified as Rhodospirillales do not cluster with the Rhodospirillales, but instead cluster with the orders currently represented by a single genus, some of which appear to be synonyms. More specifically, in the trees, the order Kiloniellales contains Kiloniella laminariae, Kopriimonas byunsanensis, Rhodovibrio salinarum and Pelagibius litoralis. The order Kordiimonadales includes Kordiimonas gwangyangensis, Rhodothalassium salexigens and a diverse group of iodine-oxidising bacteria. The order Sneathiales contains only Sneathiella chinensis. However, in the Rhodospirillales, several species with only 16S sequences are found in basal positions. The Acetobacteraceae can be expanded to include basally Elioraea tepidiphila, Alysiosphaera europeae and Geminicoccus roseus. Furthermore, the genus Tistrella may be basal to all of the Rhodospirillales. Full genome sequences of these species may therefore increase the resolving power of future studies.

Approximately Unbiased tests

Trees from the 16S, 23S, complete with regular encoding, complete RY-recoded and complete RYMK-recoded datasets were assessed against the various alignments with the Approximately Unbiased (AU) test [29]. A p-value of 0.05 was used as the cut-off, such that trees with p-values below this number could be rejected based on an alignment. Generally, trees from a given dataset were unable to be rejected by their respective alignments from either Arb-Sina or Muscle. On the other hand, they could be rejected by most of the alignments that were based on other datasets (Table S2). For the complete datasets without mitochondria, none of the four regularly coded primary trees (GTRCAT and GTRΓ with Arb-Sina and Muscle alignments) could be rejected based on the Arb-Sina or Muscle alignments. For the complete datasets with mitochondria, of the four regularly-coded trees, two could be rejected by one of the alignments. The four trees with mitochondria agree on the topology of the orders and families and differ by a Robinson-Foulds distance of less than 22 different bipartitions (Table S2) while the four trees without mitochondria disagree in some instances regarding families, but differ by less than 18 bipartitions. Whereas all of the trees from the 16S, 23S and RY-recoded datasets could be rejected based on the two regularly-coded complete alignments, the RYMK-recoded trees could not be rejected based on 11 of the 16 comparisons with the regularly coded alignments.

GC bias

The debate over alphaproteobacterial phylogeny focuses mainly on the location of the Pelagibacterales (SAR11 clade), and particularly whether it is a sister clade to the Rickettsiales or whether they are artefactually attracted due to shared low GC content [5-7]. Despite concerns about the effects of such compositional biases on tree topologies [26], little is known about the effect of genomic GC composition on 16S and 23S sequences. Here, we found that for the Alphaproteobacteria sequences that we analyzed, the GC content of rRNA genes was 5- to 6-fold less variable than the genomic GC content (Figure 1). This validated the choice of rRNA genes for minimizing artefacts that were due to shared AT richness, although further analysis showed that the GC bias was not completely eliminated in our rRNA gene datasets (Figure 4).

Consistent with the result in Figure 1, several of our further tests argued against a GC content-driven artefact affecting the position of the Pelagibacterales. Using the jackknifing approach, we tested the hypothesis that the Pelagibacterales and the Rickettsiales attract each other. Under this hypothesis, the AT-rich Rickettsiales should mask the true phylogenetic signal by attracting the AT-rich Pelagibacterales, and therefore the removal of the Rickettsiales from the dataset should reveal the location of the Pelagibacterales independently of this attraction. However, the Pelagibacterales were placed in a similar position, both in the complete dataset (Figure 2) and when Rickettsiales were missing (Figure S19). Additional evidence against compositional bias-driven attraction comes from the recoded datasets. The RY-recoded trees were less supported on many nodes, which may be due to the presence of two instead of four character states. We introduced RYMK-recoding to help overcome the limitation caused by the reduced number of character states and we found that the resulting trees concurred with the regular dataset, regarding the clustering of the Pelagibacterales with the mitochondria, Rickettsiaceae and Anaplasmataceae. Approximate Unbiased tests confirmed that RYMK-encoding (which eliminates GC bias) was in greater agreement with regular encoding than either RY- or MK-encoding. Therefore we recommend RYMK-recoding as a superior alternative to RY-recoding for assessing artefactual attractions that may arise due to GC content biases.

Overall, in datasets of concatenated 16S and 23S sequences, we found no evidence that GC content may be contributing to a topology where the Pelagibacterales cluster with the Rickettsiales.

Effects of alignment method, rate model and taxon sampling

In general, the Muscle and Arb-Sina alignments generated trees that agreed well with each other when used in combination with Gblocks. Also, the trees generated with the two different rate models (GTRΓ and GTRCAT) did not differ substantially and had similar AU scores. Therefore it can be concluded that in the case of rRNA gene-based trees computed with RAxML, neither the choice of alignment program, nor the rate model, significantly affect the final topology.

In contrast, taxon selection greatly affected the final tree. The number of taxa present was reduced by trimming the dataset of leaves that were assumed to be contributing little to the overall tree topology. Contrary to expectations, the trees from the trimmed datasets had lower support values and were much less consistent with each other compared to the complete trees (compare Figure 3 with Figures S9-S11). Extra species, represented by 16S sequences only, were also added to yield the ‘combo’ dataset. Trees constructed with this enlarged dataset also showed reduced support (Figure S20), consistent with the reduction in characters that came from not having 23S sequences for these species. Our strategy of aligning concatenated 16S-23S sequences, using at least one representative of each species, yielded trees with the highest possible support (Figure 2).

Taxonomic observations

The divergence of the clade with Rickettsia and the clade with Caulobacter has been estimated to have occurred 1,650–2,390 million years ago [30]. Therefore, it is clear that several taxa may find themselves on long branches, which may spuriously attract [31]. To minimise the quantity of long branches, one representative of each alphaproteobacterial species, represented by both 16S and 23S sequences, was chosen in this study. Moreover, by analyzing only the 16S and 23S markers it was also straightforward to screen manually for misannotation errors and to correct or exclude problematic taxa. For example, the contig NZ_AAAP01003712 for Magnetospirillum magnetotacticum was found to be contaminated with sequence from Methylobacterium populi, from position 3,624 to 8,812 (data not shown).

In contrast to previous analyses, this study covered all of the known diversity within Pelagibacterales (i.e. subgroups I, III, IV and V, as classified by Grote et al. [9]). The monophyly of the group was confirmed with more than 95% support in all trees inferred. Therefore, our data provided no evidence that Pelagibacterales may be polyphyletic, in agreement with one recent study [9], but not another [7]. The Pelagibacterales fall basal to a clade composed of mitochondria and a Rickettsiales subclade without the Holosporaceae, in all of the full trees. This grouping has a moderate bootstrap support, but this is most likely due to the instability of the Holosporales (vide infra) and not due to an AT-attractional bias (vide supra).

In our four complete, regularly coded trees the mitochondria are a sister group to a clade formed by Anaplasmataceae and Rickettsiaceae, with high support. This result is consistent with several studies that used concatenated protein phylogenies [3,4,8], although slightly different to Georgiades et al. [8] and Thrash et al. [4], which found support for the Pelagibacteraceae as the sister clade to the mitochondria. However, the placement of the Pelagibacterales near the branch point of the mitochondria and the Rickettsiales, regardless of the order, is in greater agreement with all three of the above studies than with the results presented by Brindefalk et al. [5], Viklund et al. [6], and Rodríguez-Ezpeleta and Embley [7], where the Pelagibacterales are placed elsewhere in the Alphaproteobacteria entirely.

Magnetococcus marinus is the sole genome-sequenced representative of the Magnetococcales, a clade that is basal to the remaining Alphaproteobacteria [12,32]. At the outset of this study, its membership in the Alphaproteobacteria was unclear, although it has since been proven correct [12]. Having chosen our outgroup to include the most diverse members of the Betaproteobacteria, Gammaproteobacteria and Zetaproteobacteria, we can also confirm the membership of the Magnetococcales within Alphaproteobacteria. Its inclusion in our phylogenetic analyses greatly reduced the length of the branch leading to the remaining alphaproteobacterial orders. This enabled us to circumvent an earlier problem, in which the absence of Magnetococcus marinus meant that the choice of outgroup affected the topology of the alphaproteobacterial tree [4].

Similarly, in our trees the Holosporaceae are basal to a large clade of alphaproteobacterial orders, yet they are represented solely by Odyssella thessalonicensis and Caedibacter caryophilus (Figure 2). The former was only sequenced recently [8], while the latter is represented solely by 16S and 23S sequences. Previously, Holosporaceae has been classified as a family in the Rickettsiales [10,11]; however, in our datasets this clade does not cluster within this order. Instead, it lies basal to the clade that comprises several orders, but not the Rickettsiales and Magnetococcales. In light of our new analysis, we propose to remove the Holosporaceae from the Rickettsiales and to create a new order, the Holosporales ord. nov. Consequently, under this revised classification only the Rickettsiaceae, the Midichloriaceae and the Anaplasmataceae comprise the order Rickettsiales (sensu novo).

There are several peculiarities involving the Rhodospirillales in this work and in other studies [7,10], albeit generally reported without comment. There seems to be an instability within the Rhodospirillales clade: whereas the Acetobacteriaceae resolves with high support, the Rhodospirillaceae rarely resolves monophyletically or with high support (this study and [7,10]). In this study the deletion of the Rhodospirillales profoundly reduces the support of the location of the Holosporaceae. The Holosporaceae is located in a clade with the Rhodospirillales, Sphingomonadales, Rhizobiales, Caulobacterales and Rhodobacterales with an average support of 82% in the trees from the complete dataset, but with only 42% support (on average) in the trees without the Rhodospirillales. It can be concluded that the Rhodospirillales play a large role in supporting the topology at the internal node of the tree. Poor support, or even paraphyly [7,10], of the Rhodospirillales could be indicative of an incorrect resolution of the internal node separating the clade composed of Rickettsiales and Pelagibacterales from the clade that includes the Caulobacterales, Rhodospirillales and several other orders.

The problematic internal organisation of the Rhodospirillales is even more aggravated in the datasets that include species represented only by 16S rRNA gene sequences (Figure S20). Some species classified as Rhodospirillales actually belong to the separate orders Kiloniellales, Kordiimonadales and Sneathiellales, while several Rhodospirillales species, such Tistrella mobilis (whose genome was published too recently to be included in this study, [33]), form clades that are basal to the other Rhodospirillales. The implication is that the current family-level classification of the Rhodospirillales will need revisiting in the future.

Proposal of new subclasses

In light of the number of orders present in the Alphaproteobacteria, we propose the creation of three subclasses that are based on the tree topology found here (Figure 2), and that aid in the description of the groups. In particular, we propose to distinguish the two clades that are apical to the Magnetococcales. Therefore we propose to place: (i) the Magnetococcales in the Magnetococcidae subcl. nov.; (ii) the Rickettsiales (sensu novo), the protomitochondrion (i.e. the bacterial ancestor of the eukaryotic organelle) and Pelagibacterales ord. nov. in the Rickettsidae subcl. nov.; and (iii) the Holosporales ord. nov., Rhodospirillales, Sphingomonadales, Rhizobiales, Caulobacterales and other orders in the Caulobacteridae subcl. nov. Our proposal is summarized in Figure 5. Overall, our updated analysis supports the position of the Pelagibacterales as a sister group to the composite clade containing Rickettsiales and the mitochondrial branch. Our analysis has also provided support for taxonomic assignment of several recently sequenced species, including Odyssella thessalonicensis and Magnetococcus marinus.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Proposed subclasses of the Alphaproteobacteria.

The three proposed subdivisions are the Magnetococcidae, the Rickettsidae and the Caulobacteridae. Furthermore, the Holosporaceae should be removed from the Rickettsiales, however the identities of the family-level subdivisions of the Holosporales, such as the Holosporaceae (marked with an asterisk), are beyond the scope of this work. Under this scheme the Rickettsiales are comprised solely of the Rickettsiaceae, Anaplasmataceae and Midichloriaceae. The protomitochondrion (†) is an extinct organism that gave rise to the mitochondrial organelles of eukaryotes.

https://doi.org/10.1371/journal.pone.0083383.g005

Taxon selection

The 16S and 23S rRNA genes used to construct the concatenated datasets were obtained from IMG (v350). The perl scripts list_clustermaker.pl and fasta_acceptor.pl were used to select a sole sequence as a representative for each species present while checking whether the sequence matched that of the other paralogs. In negative cases, the most common sequence was chosen. To supplement the trees with additional Alphaproteobacteria without sequenced genomes, 16S or both 16S and 23S rRNA gene sequences were obtained from GenBank.

GC content

A perl script, GC-counter.pl, was used to determine the GC content of the rRNA gene sequences; genomic GC content was obtained from IMG. Scatter plots were created in Prism 4. The slopes of the linear regressions of both the SSU and the LSU rRNA gene GC content over genomic GC content for the Alphaproteobacteria and for the mitochondria were calculated in Excel with the slope function. Box plots were created in R using boxplot, and statistical significance was evaluated using wilcox.test.

Sequence QC

The quality of the sequences was corrected by trimming the ends according to the match from the Arb-Sina aligner [24]. The length was checked with a perl script, impostor checker.pl, and all sequences under 1200 (for 16S) or 2000 (for 23S) bases were removed. One outcome of imparting this size cut-off was to remove numerous mitochondrial sequences (listed in Table S1); however, including these mitochondrial sequences made no difference to the observed topologies (data not shown). The presence of multiple copies of rDNA in most organisms allowed sequencing errors and contaminations to be identified.

Alignment/gap-removal

The sequences were aligned either with Muscle [23] under default settings or with Sina [24] set to remove terminal unaligned bases and not to reverse complement sequences (due to issues with some mitochondrial sequences). Poorly aligned sites were curated with Gblocks [34] using the settings from [35]: -b1=(n/2)+1 -b2=(n/2)+1 -b3=n/2 -b4=2 -b5=h. All dataset variants were created by deleting the targeted taxa before the alignment step.

RY-, MK- and RYMK-recoded datasets

For the RY-recoded datasets, the completed (aligned and trimmed) datasets were recoded, by converting all A and G bases to R and all T and C bases to Y. For the MK-recoded datasets, all A and C bases were converted to M and all T and G bases were converted to K. For the RYMK-recoded datasets, the complete dataset was duplicated and one copy was RY-recoded, while the second was MK-recoded; the two differentially recoded variants were then concatenated.

Maximum likelihood trees

Maximum likelihood inferences were performed with RAxML 7.2.9 [25] with 1,000 bootstrap replicates under either a GTRΓ or a GTRCAT model. The trees were displayed in FigTree3.1 [36]. Figure 2 was further annotated with Adobe Illustrator CS4. The supplementary tree summary figures were made with Illustrator CS4, with the aid of Newick utilities [37] for data extraction. All alignment and tree files are available from the authors upon request.

Topology evaluation

The bootstrap support for various groupings of interest were obtained via the perl script descriptor.pl, using Newick utilities [37]. First the bootstrap trees were rooted to the outgroup (nw_reroot), then the leaves in these were renamed by mapping the names to the groups of interest (nw_rename), after which the clades composed of a single name were condensed (nw_condense) and finally the trees were ordered (nw_order). Once this was done, the script determined the support for monophyly of the groupings by simply counting the frequency of the names in each of the resulting trees, where a single instance would indicate monophyly.

Approximately Unbiased tests

The final trees of the full datasets were concatenated into a single file and the per site log-likelihoods were calculated with both the Arb-Sina-aligned and Muscle-aligned datasets (RAxML via option -f g), in order to be converted into a matrix (makermt --puzzle) that could be interpreted by Consel (consel and catpv). Robinson-Foulds distances were calculated between each tree in a set with HashRF (https://code.google.com/p/hashrf/).