HGT turbulence

Abstract

Horizontal gene transfer (HGT) often leads to phylogenetic incongruence. When “duplicative HGT” introduces a second copy of a pre-existing gene, the two copies may then engage in gene conversion, leading to phylogenetically mosiac genes. When duplicative HGT is followed by differential gene conversion among descendant lineages, as under the DH-DC model, phylogenetic analysis is further complicated. To explore the effects of DH-DC on phylogeny reconstruction, we analyzed two sets of sequences: (1) an augmented set of plant mitochondrial atp1 sequences for which we recently published evidence of DH-DC; and (2) a set of simulated sequences for which we varied the extent of chimerism, the number of chimeric genes and nucleotide substitution rates. We show that the phylogenetic behavior of evolutionarily chimeric genes is highly volatile and depends on both the degree of chimerism and the number of differentially chimeric genes present in the analysis. Furthermore, we show that the presence of chimeric genes in gene trees can spuriously affect the phylogenetic position of purely native sequences, especially by attracting these sequences toward basal positions in trees. We propose the term “HGT turbulence” to describe these complex effects of evolutionarily chimeric genes on phylogenetic results.

Introduction

Horizontal gene transfer (HGT) is very common and of great importance in bacterial evolution^1,2 and is also relatively common in certain eukaryotic lineages.^3,4 One consequence of HGT is incongruence between phylogenies reconstructed from genes with different histories of transfer, including of course no transfer at all. Finding and examining phylogenetically atypical genes has become a standard task in HGT studies.^5-7 In cases where an entire gene has been transferred, without the complication of recombination/conversion with a native homolog, the donor lineage can in principle be identified as the nearest phylogenetic neighbor.^7-9 Gene conversion enters the picture either during the act of gene transfer, when transiently-present foreign DNA directly converts (replaces) part of a native locus^10,11or, after duplicative HGT, via potentially ongoing gene conversion between co-existing native and foreign copies.¹² Furthermore, gene conversion can occur in either a continuous or discontinuous manner.^13,14 Overall, then, gene conversion can lead to a potentially complex and diverse set of patchwork recombinant sequences, especially if it occurs repeatedly, and differentially, over the course of speciation.¹² Each recombinant gene, if analyzed as a whole, might or might not reflect the true evolutionary history of either or both parental sequences. When parental sequences contribute differentially to the number of informative characters in a recombinant sequence, this sequence will tend to resemble the parental sequence that contributes more informative characters (e.g., refs. 15 and 16), whereas when parental sequences contribute similar numbers of informative characters, the recombinant will potentially be quite different from both parental sequences, depending of course on the degree of divergence of the two parental sequences from each other.¹⁷

When properly recognized and dealt with, recombination poses few problems for phylogenetic analysis and interpretation. In practice, however, recombination detection is challenging and often subject to failure. First, it is well established that recombination detection programs perform poorly when sequence divergence is low.^18-21 Unfortunately, plant mitochondrial genomes, which collectively constitute a premiere model system for eukaryotic HGT studies,^4,22 usually have very low rates of nucleotide substitution.²³ Second, gene conversion often involves very short tracts of DNA,^13,14 which can make recombination detection very difficult. For instance, ten previously published recombinant regions between plant mitochondrial and chloroplast genes range in length from only 14 to 79 nucleotides.^24,25 Third, existing recombination detection programs are generally designed to identify a single or only a small number of recombination breakpoints.^26,27 Intricate gene conversion during the process of duplicative HGT and differential gene conversion (DH-DC) can, however, lead to mosaic gene structures, with multiple foreign regions interspersed with native regions on a fine scale.

We recently reported such mosaicism in mitochondrial atp1 and matR genes belonging to different groups of flowering plants (angiosperms). We show that these mosaic genes largely escaped detection by recombination-detection programs and were recognizable only by direct visual inspection of DNA sequence alignments.¹² In this report, we explore the effects of chimeric sequences on phylogeny reconstruction by conducting phylogenetic analyses on simulated sequences and on an augmented set of the atp1 sequences analyzed in reference 12.

Results

Phylogenetic analysis of naturally-occurring mosaic genes

In a recent study,¹² we reported the presence of three differentially mosaic types of mitochondrial atp1 genes in the angiosperm genus Ternstroemia (Pentaphylaceae, Ericales) and concluded that they arose via DH-DC, with the blueberry genus Vaccinium (Ericaceae) the best candidate to be the donor group in the initiating HGT event. As shown in Figure 1 (adapted from Fig. 2A of ref. 12), each of the three major clades within Ternstroemia possesses a differentially mosaic atp1 gene, each with multiple (four to five) foreign regions interspersed with native regions. In the one atp1 phylogeny presented in reference 12, the mosaic Ternstroemia genes were all placed within the Eriaceae, in a paraphyletic relationship with respect to Vaccinium.

Figure 1.

Three types of mosaic mitochondrial atp1 genes in Ternstroemia (adapted from ref. 12). The multi-colored boxes represent atp1 genes of the three subclades within Ternstroemia. Black vertical lines represent the 38 nucleotide positions inferred12 to have differed between donor and recipient atp1 genes at the time of atp1 transfer from Vaccinium to a common ancestor of Ternstroemia. Lines at the top of the boxes and red shading indicate sites and regions, respectively, of putatively foreign, Vaccinium ancestry, while bottom lines and blue shading represent native sites and regions. White lines centered within the boxes represent the only two sites that otherwise differ within the Ternstroemia clade. “T-ips” refers to the Ternstroemia subclade containing T. impressa, T. peduncularis and T. stahlii) and “T-glj” to the subclade containing T. gymnanthera, T. longipes and T. japonica (see also Fig. 2A).

To better understand how mosaic genes affect phylogeny reconstruction, we sequenced the mitochondrial atp1 gene of Chamaedaphne calyculata, a close relative of Vaccinium,²⁸ using the same set of primers and method as in reference 12, and employed it together with varying subsets of previously sequenced Ericales atp1 genes in maximum likelihood phylogenetic analyses. The Chamaedaphne atp1 sequence was deposited in GenBank under accession number JN808446. Consistent with our recent study,¹² and in contrast to organismal phylogeny (Fig. 2A), in an analysis that included all relevant genes, the three types of mosaic atp1 genes in Ternstroemia formed a paraphyletic assemblage, with T. fragrans sister to the Vaccinium/Chamaedaphne clade, the T-glj clade the most distant from the Vaccinium/Chamaedaphne clade and the T-ips clade in an intermediate position (Fig. 2B). Remarkably, when only one type of mosaic atp1 gene was included in a given analysis, each of the three types fell in a different phylogenetic position (Fig. 2C–E). This shows that the phylogenetic position of a mosaic gene can vary depending on the inclusion of additional, related mosaic genes and emphasizes that this position provides little or no reliable information on the nature of the gene’s parental sequences.

Figure 2. Phylogenetic analysis of mosaic mitochondrial atp1 genes. (A) Chronogram showing organismal relationships and divergence times of relevant taxa belonging to the Ericales. As described in reference 12, the chronogram was constructed by the BEAST program⁴³ using a Eurya reference fossil calibration of 86 Myr ago.⁴⁴ (B-F) Maximum likelihood phylogenies of mitochondrial atp1 genes from the taxa shown in (A), with these analyses varying as to which members of Ternstroemia (shown in red), whose atp1 gene is differentially mosaic, were included. RAxML³² version 7.0.4 was used to construct all phylogenies with a GTR+Γ+I substitution model. A total of 1000 bootstrap iterations were performed, with all bootstrap values ≥ 50% shown on the trees. Phylogenies were rooted using Fouquieria, Marcgravia and Pentamerista as unshown outgroups (hence the stub branch at the base of each gene tree).

The only other topological difference among all five atp1 trees (Fig. 2B–F) involves Chamaedaphne, which was weakly placed (41% bootstrap support) within Vaccinium when all mosaic genes were included (Fig. 2B), but was placed as sister (with 68–92% support) to a monophyletic Vaccinium in the other four gene trees. This result raises the possibility that the inclusion of mosaic sequences in phylogenetic analyses can affect not only the placement of related mosaic sequences, but also the placement of apparently native sequences.

Phylogenetic analysis of simulated chimeric sequences

We used simulation studies (conducted using Seq-Gen²⁹) to further explore the effects of chimeric sequences on phylogeny reconstruction. The following simulation parameters were chosen to be the same as those used in the above analysis of atp1 sequences: (1) sequence length, 1200 nucleotides; (2) substitution model, GTR; (3) gamma shape parameter, 0.218; (4) proportion of invariant sites: 0.371; (5) nucleotide frequencies, 0.271 (A), 0.207 (C), 0.261 (G) and 0.261 (T); and (6) GTR relative rate parameters, A < - > C = 0.818, A < - > G = 1.938, A < - > T = 0.244, C < - > G = 0.884, C < - > T = 2.219 and G < - > T = 1.000. All but the first of these parameters are based on PhyML³⁰ estimates derived from the mitochondrial atp1 alignment shown in Figure S1, except with the mosaic Ternstroemia genes excluded (as in Fig. 2F) because recombinant genes have been shown to alter the estimation of substitution rate heterogeneity.³¹ The remaining simulation parameters were independent of the atp1 data and include the number of sequences, their topology, relative branch lengths and absolute amount of divergence (Fig. 3A). For simplicity, only one recombination breakpoint was allowed in each chimeric sequence, with these chimerics constructed to contain varying proportions (from 50:50 to 10:90) of two of the 16 “parental” sequences generated by the simulations. The 16 purely-simulated sequences together with the artificially constructed chimeric sequence were used in phylogenetic analyses performed using RAxML version 7.0.4.³²

Figure 3. Variable phylogenetic placement of chimeric genes demonstrated by simulations. Artificial sequences were simulated as described in the text. Chimeric sequences were generated by combining the 5′ and 3′ portions of sequences 1 and 9, respectively (both circled in red), with different length ratios (10:90, 30:70 or 50:50) of the two parental sequences. One thousand tree-building iterations were performed. The tree shown in each panel is based (for computational ease) on the concatenated sequences of the first 100 iterations, while the bootstrap support values are from all 1000 trees. All bootstrap values are ≥ 95% except for those shown.

Two sets of simulation analyses were performed. In the first set (Fig. 3), we varied the parental proportions that comprise the chimeric sequence(s) and the number of chimeric sequences included in a given analysis. In analyses with a single chimeric sequence, this sequence grouped with 100% bootstrap support with its majority parental sequence when the parental ratio was 10:90 (i.e., when the chimeric sequence consisted of 10% of sequence 9 and 90% of sequence 1; Fig. 3B). The same topology was obtained with the 30:70 chimera (Fig. 3C), but the bootstrap value dropped to 92. When the chimera was 50:50 (Fig. 3D), it went to the base of the tree, in between the two main clades of simulated sequences, and with bootstrap support reduced along the branches leading to both parental sequences. These results thus show that the phylogenetic position of chimeric genes can vary substantially as the proportion of parental sequences that comprise the chimerics varies.

The phylogenetic position of the 50:50 chimeric sequence also varied substantially depending on whether it was the only chimeric sequence in the analysis (Fig. 3D) or whether the 30:70 and/or 10:90 sequences were also included (Fig. 3E–G). Inclusion of the 30:70 sequence, either with (Fig. 3E) or without (Fig. 3F) the 10:90, “pulled” the 50:50 from the base of the tree to its periphery, together with the 30:70 and parental sequence 1 (and with the 10:90 when included) and with strong support (96% and 92%, respectively). This peripheral attraction was more subdued when the 50:50 was paired with the 10:90 (Fig. 3G) as opposed to the 30:70 (Fig. 3F), presumably because of the greater proportion of sequence length shared by the 30:70 and 50:50 (80%) relative to the 10:90 and 50:50 (60%). Also, there is evidence for a mutual attraction between the 10:90 and 50:50 (but not between the 30:70 and 50:50), in that the bootstrap support for the 10:90 grouping with parental sequence 1 was reduced from 100% (Fig. 3B) to only 54% (Fig. 3G). These simulation results are consistent with the empirical results (Fig. 2) in showing that the phylogenetic position of a chimeric gene can vary substantially when different additional chimeric genes are sampled.

We also found that chimeric genes can perturb phylogenetic analysis by producing branch length bias. First, certain chimeric genes were directly associated with elongated branch lengths. In the single-chimeric analyses, the 10:90 chimera had a notably longer branch length than sequence 1 (Fig. 3B), while the 30:70 branch length was even longer (Fig. 3C), with the (homoplasious) substitutions contributed by the 30% of the chimeric sequence originating from sequence 9 presumably responsible for this greatly extended branch length. Similar results were obtained in Figure 3E–H, where multiple chimeric sequences were included in each analysis. Second, the presence of the 50:50 chimeric in the single-chimeric analysis of Figure 3D resulted in the loss of a molecular clock among the non-recombinant sequences in this analysis, with the branch leading to the (1–4) clade only 60% of the length leading to the (5–8) clade and the branch leading to the (9–12) clade only 58% of the length leading to the (13–16) clade. There are consistently two additional—albeit much less pronounced—sets of branch length differences in the six trees (Fig. 3B, C and E–H) in which one or more chimeric sequences are sister to sequence 1.

The second set of simulations showed that chimeric sequences can also confound phylogenetic analysis by altering the placement of non-recombinant (native) sequences. In these simulations (Fig. 4), we varied the absolute amount of sequence divergence across the tree, with the 50:50 sequence the only chimeric sequence in each analysis. Figure 4A shows the same tree as Figure 3D, while Figure 4B–D has expanded divergence by a factor of 2 ×, 5 × and 10 × , respectively. As the simulated sequences become more divergent, the parental sequences show an increased and pronounced, tendency to be attracted toward the base of the tree by the chimeric gene. Also, the branch leading to the chimeric sequence becomes increasingly short, approaching zero in Figure 4CandD. It is important to note that this effect is not simply the result of there being increasingly more informative characters from Figure 4A–D. In analyses with 5 × or 10 × divergence (data not shown), but only 1/10th the sequence length (120 nucleotides rather than the 1200 in the simulations shown in Figure 4), the topology and bootstrap values were essentially identical with those in Figure 4CandD, respectively, with these being substantially different from those in Figure 4AandB. The different topologies shown in Figure 4 must therefore result from deterministic effects arising from the varying levels of divergence in these trees. These simulations thus show that inclusion of chimeric sequences can distort the branching pattern of nonchimeric, native sequences and that the extent of this distortion varies directly with the absolute level of sequence divergence.

Figure 4. Chimeric sequences increasingly disrupt native-sequence topologies as divergence increases. (A) The same tree as shown in Figure 3D. (B–D) Maximum likelihood analysis of the same simulated sequences, but with increasing branch-length scales across the set of sequences: (B) 2 × the scale in (A); (C) 5 × ; and (D) 10 × . One thousand tree-building iterations were performed. The tree shown in each panel is based (for computational ease) on the concatenated sequences of the first 100 iterations, while the bootstrap support values are from all 1,000 trees. All bootstrap values are ≥ 95% except for those shown.

Discussion

The analyses reported in this study show that the inclusion in phylogenetic analyses of chimeric sequences arising from HGT and gene conversion can produce a variety of spurious phylogenetic results. These include the misplacement of both chimeric and native sequences, as well as branch length distortions. Accordingly, we introduce the term “HGT turbulence” as a general moniker for this category of phylogenetic artifacts. Mutually reinforcing evidence for these types of HGT turbulence was apparent in the phylogenetic analyses of both simulated sequences and a naturally occurring set of native and chimeric mitochondrial sequences, with the simulations providing greater opportunity to crystallize and illustrate specific sets of sequence interactions and consequences. It is also important to realize that while these simulations were framed and presented in the context of HGT, their results apply equally well to conversion between paralogs arising from internal gene duplication as to xenologs arising from HGT.

HGT turbulence is probably relatively common in bacteria,^1,2 given the prevalence of HGT and recombination during bacterial genome evolution. For example, such diverse bacteria as Neisseria meningitidis,^33,34 Streptococcus pneumoniae,^35,36 Helicobacter pylori³⁷ and Wolbachia³⁸ have been found to be so recombinogenic that scientists have resorted to using multiple loci (e.g., multilocus sequence typing³⁹) as opposed to a single locus to identify clones. Surprisingly, however, the phenomenon of HGT turbulence has never been explicitly addressed in the bacterial literature. One reason for this is that these studies have mainly focused on minimizing the effect of recombination and thereby inferring accurate evolutionary relationships, or on quantifying the number of recombinant genes, rather than actually exploring the topological alterations caused by HGT turbulence. Also, the precise origins of horizontally transferred genes (or gene fragments) in bacterial genomes can be extremely difficult to recover, especially when transfer and/or subsequent recombination have occurred on a fine scale.

Several other studies^17,31,40-42 have used recombinant sequences in phylogenetic simulations, but most of these have focused on the issue of whether recombinants are detectable and/or how to detect them. Perhaps most relevant to our study is the 2002 study by Posada and Crandall,¹⁷ which also explored the relationship between the location of recombination breakpoints and the phylogenetic placement of recombinant sequences, reaching similar conclusion to ours on this point. However, none of these studies explored the effects of combining multiple, related chimeric sequences in the same analysis, nor did they show that chimeric sequences can, under certain conditions, substantially alter the phylogenetic behavior of native sequences. Further simulation studies on the effects of recombination on phylogenetic inference would therefore appear to be called for.

Recognition of the remarkable frequency and extent of horizontal gene transfer, and its often great evolutionary importance, is arguably the greatest accomplishment of the past 15 years of comparative genomics research. Because HGT is so common and important, recognizing and properly dealing with HGT turbulence is likewise important. This is so not only because of the obvious need for obtaining accurate estimates of gene and species phylogeny, but also because otherwise too many cases of chimeric HGT, including complex situations involving DH-DC, will continue to go overlooked.

Supplementary Material

Additional material

Supplementary PDF file supplied by authors.