Reply to Holmes and Duchêne, “Can Sequence Phylogenies Safely Infer the Origin of the Global Virome?”: Deep Phylogenetic Analysis of RNA Viruses Is Highly Challenging but Not Meaningless

摘要

REPLY In their Letter to the Editor of mBio , written in response to our recent article on evolution of the global RNA virome ( 1 ), Holmes and Duchêne submit that the extreme sequence divergence between the RNA-dependent RNA polymerases (RdRps) makes it impossible to infer deep relationships between RNA viruses from any type of sequence analysis. We certainly agree with Holmes and Duchêne that extreme caution is due in the analysis and interpretation of deep phylogenies, and in particular, that alignment quality is central to our ability to resolve long-distance evolutionary relationships. If the alignment is largely wrong (i.e., does not align homologous protein sites) or noninformative (i.e., cannot be used to distinguish between alternative histories), it is of no utility for phylogenetic reconstruction. Moreover, even a correct and informative alignment does not guarantee correct phylogenetic reconstruction due to the technical limitations of the software, systematic biases of the available evolutionary models, and the fundamentally random nature of sequence divergence. Therefore, formal phylogenetic analysis should be accompanied by careful consideration of the associated biological data and examined in terms of the implications of the respective evolutionary scenarios. Where exactly lies the boundary between an alignment that is suitable for phylogenetic reconstruction and one that is “highly unlikely to be accurate” is far from being an easy question. In the ideal situation (high sequence similarity, random homoplasy), one might need as little as O (log k ) informative sites to resolve a tree of k sequences ( 2 ). With real-life data, it is critical that sequence similarity, even if extremely low between the most distant sequences, changes according to a pattern, consistent with the tree structure. Fortunately, the structure of proteins with their nearly invariant functional sites, strongly conserved structural core, variable bulk, and extremely fluid interface surfaces naturally provides such an essential pattern, with each class of sites in the scale of evolutionary rates allowing for good resolution at the appropriate range of distances. Furthermore, the very definition of a “random” site is not a trivial matter. An alignment site that contains all 20 amino acids might appear completely random, but in fact, its validity and utility greatly depend on the amino acid distribution pattern. An obvious hypothetical example is a site where 110 sequences have one amino acid, 110 sequences have another amino acid, and the remaining 18 sequences each have different amino acids. Such a site would contain a strong bipartition signal, and if the other positions of the discordant sequences show affinity with one of these two groups, it would be highly informative for tree reconstruction. The alignment of 228 RNA-dependent RNA polymerases (RdRps) from RNA viruses and 10 reverse transcriptases (RTs) that was employed in our work ( 1 ) to construct the global tree of RNA viruses does indeed push the envelope of usable sequence similarity. As Holmes and Duchêne note, there are no invariant sites, no sites without gaps, more than 96% of the alignment columns contain more than 50% of gaps, and where sites are aligned, the similarity is low (the median distance between RTs and RdRps is 5.0 substitutions per site as estimated by PhyML). However, some of these metrics, although correctly calculated, do not give the full picture of the alignment properties. Although as indicated above, only 441 sites contain less than 50% of gaps in an alignment of the total length of 12,200, the median length of the RdRp core is 497 amino acids, so that actually, 89% of a typical sequence is part of a reasonable alignment. The plot of the conservation (alignment column homogeneity) and gap content shows multiple, sharp peaks of relatively high conservation and low gap content. Moreover, these regions correspond to well-known motifs that are conserved among the RdRps, across the evolutionary distance of more than five substitutions per site, on average ( Fig.?1 ). Although this level of conservation might appear insufficient to capture the deepest relationships between the RNA viruses, one should keep in mind that, at the deepest level, there are few major clades to resolve (according to our analysis, the RT and five branches of RdRps). The alignment statistics rapidly improve at the shallower levels: even within each major branch, the clade-specific conservation is readily apparent ( Table?1 ). FIG?1 Sequence conservation profile along the core alignment of the RdRps and RTs. The homogeneity metric is based on the BLOSUM62 scores between the consensus amino acid and the actual amino acids in the alignment column and are scaled from 1 (all residues are the same) to 0 (the score is not different from the random expectation). The fraction of gaps is computed using sequence weights ( 6 ). The amino acids conserved in five prominent motifs