A phylogenomic data-driven exploration of viral origins and evolution

Response to Harish et al.: No ‘small genome attraction’ artifact
Arshan Nasir, Kyung Mo Kim and Gustavo Caetano-Anollés

In their eLetter, Harish, Abroi, Gough and Kurland criticize our structural phylogenomic methods, which are supported by large-scale structural and functional data and well-established comparative genomics, phylogenomics, and multidimensional scaling approaches (1). Harish et al. would like to see the origin of Eukarya at the base of the Tree of Life (ToL) (2). So despite invalidating critique (3), they go on to challenge the early cellular origin of viruses. Their claims include a warning to “nonspecialists” that the rooting of our trees is distorted by what they dub is “a small genome attraction artifact of genome content-based phylogenetic analysis”. Here we examine their reasoning, which mingles with misunderstandings and misinterpretations of cladistic methodology (Please refer to Table 1).

1. Their claim that our rooting approach uses outgroup taxa is incorrect. We do not “use a hypothetical pseudo-outgroup, … an artificial all-zero taxon … to root the ToL”. No outgroup taxon (presumably extant or artificial) was ever used or defined in our study (1). Their confusion of outgroups with ancestors showcases cladistics misunderstanding (Table 1). Contrary to their claims, our rooting method is grounded in early and well-established cladistic formalizations (4, 5) and is direct because it polarizes character transformations with information solely present in the taxa being studied (the ingroup), distinguishing ancestral from derived character states and rooting the trees a posteriori. Our rooted trees also comply with Weston’s generality criterion (6, 7), which states that as long as ancestral characters are preponderantly retained in descendants, ancestral character states will always be more general than their derivatives given their nested hierarchical distribution in rooted phylogenies. Biologically, structural domains spread by recruitment in evolution when genes duplicate and diversify, genomes rearrange, and genetic information is exchanged. This is a process of accumulation and retention of iterative homologies, such as serial homologues and paralogous genes (7), which is global, universal and largely unaffected by proteome size. The Lundberg method (5), which does not attach outgroup taxa to the ingroup as they claim, simply enables rooting by the generality criterion (8).

2. Their confusion of a priori and a posteriori character polarization questions their understanding of cladistic methodology. Their claim that we use a pseudo-outgroup to polarize character state changes a priori is inconsistent with our methodology of first reconstructing an undirected phyletic network and then polarizing character transformations with a direct method (1). They miss the fact that rooting is not a neutral procedure. While the length of the most parsimonious trees is unaffected by the position of the root, making a priori polarization unnecessary (4), rooting impacts the homology statements of the undirected networks (4, 5). “The length of a tree is unaffected by the position of the root but is certainly not unaffected by the inclusion of a root” (9).

3. Their claim that small genome size affects rooting and induces attraction artifacts is conceptually and empirically debunked. Their recitation that organisms and viruses with small genomes (irrespective of their taxonomic affiliation) would be attracted to basal branches of our trees is incorrect. During searches of tree space and prior to rooting, we optimize individual character change in unrooted networks. This allows unrestricted gains and losses of domain occurrence or abundance throughout their branches. Thus, rooting plays no role in defining unrooted tree topology and cannot be distorted by genome size, which is a property of taxa (not individual characters changing in trees). This was already described in our supplementary text (1) and made explicit in a recent study (10). Polarization is only applied a posteriori by: (a) considering character spread in nested branches while accounting unproblematically for homoplasy (Weston’s rule), (b) searching for the most parsimonious solutions with Lundberg while treating homologies as taxic hypotheses, and (c) allowing gradual build-up of evolutionary innovation (including loss and punctuation) that complies with the principle of spatiotemporal continuity (PC). These three mutually supportive technical and biological axiomatic criteria were confirmed experimentally by Venn group distributions in ToDs and by visualizing clouds of proteomes in temporal space [see Figs. 5 and 8 of (1)]. Even Felsenstein’s suggestion of inverse polarization (11) of our ordered (Wagner) characters, which can be polarized in only two directions, produced suboptimal trees [see Figs. 3 and 4 of (3)]. Fiction is even debunked empirically. Plotting the node distance (nd) for each terminal node (i.e. taxa) from the root node of a ToL – on a scale from 0 (most basal) to 1 (most recent) – against the genome sizes of taxa showed substantial scatter (rho = 0.80), poor lineal fits (several peaks and troughs with 10 iterations of the LOWESS fitting method and a smoothing of q = 0.05) shallow monotonic increases (flat lines in Archaea and Bacteria), and no distortions/mixing of taxa among the four supergroups, viruses, Archaea, Bacteria and Eukarya. Genomes of similar sizes were scattered throughout the nd axis suggesting that genome size was not a significant determinant of taxa position in our trees. Similarly, genome size scatter for individual nd increased towards the base of the trees, including scatter for supergroups (rho values of 0.55, 0.63, 0.75 and 0.85 for 266 viruses, 30 Archaea, 31 Bacteria and 28 Eukarya respectively), debunking the alleged basal attraction artifact. The order of appearance of supergroups matched their proteomic complexity, from simple to complex, which also matched scaling patterns of use and reuse of structural domains in proteomes (3). This emerging property of trees supports evolution’s PC. Labeling the phylogenetic positions of the smallest proteomes in our trees showed that the smallest genomes were neither attracted towards the root nor caused distortions in 3-supergroup or 4-supergroup views of life. Their claim that “the rooting in viral lineages is an inevitable consequence of pre-specifying ‘0’ or ‘absence’ as the ancestral state” is therefore conceptually and empirically false.

4. Confusion of characters and taxa bootstraps their preconceptions. In rushing their unsupported claim that the rooting of trees of structural domains (ToDs) is also unreliable and affected by small genome size, they wrongly considered ToDs as being “uninterpretable in terms of the definition of the (domain) superfamilies which it comprises”, because homology “within” superfamilies “can be ascertained based on similarity of sequence, structure and function”. But superfamilies are the taxa and proteomes the characters, and definitions of taxa (superfamily hidden Markov models) do not need to follow either definitions of characters (superfamily growth in proteomes) or statements of homology tested in ToDs. What is ‘uninterpretable’ however is the putative effect of genome size on ToDs, since each proteome embodies a character, which by definition (Kluge’s Auxiliary Principle) is independent of others. So fiction bootstraps their preconceptions, including the idea that domains, the evolutionary units of proteins, do not evolve. Are 1,200 structural folds fortuitous findings or the makings of intelligent cause? Where does significant evolutionary signal of the ToDs, including a match to the geological record (12), come from? Even an exploration of the mapping of functions in genotype space shows the centrality of structure in defining evolutionary constraints (13).

5. They fictionalized comparative genomic research. They state that our Venn diagrams and summary statistics of domain distributions in supergroups are unconvincing in light of other comparative genomic analyses (14, 15). However, Abroi and Gough (14) argue that viruses may be a source of new protein fold architectures, a conclusion strongly supported by our Venn analysis, and Abroi (15) showcases the distribution and sharing of superfamilies between viral replicon types and cells, which is largely consistent with our analysis (3). There are no irreconcilable differences between these studies. Instead, our phylogenomic analysis dissects the alternative evolutionary scenarios that can be posed with the comparative method (1, 15).

Conclusions. It is ironic that attempts to controvert our direct methods of character polarization come from authors that are themselves proponents of the use of polarized characters, but with arbitrary transformation costs carefully engineered a priori to attract large eukaryotic genomes to the base of their trees (2). These characters violate the ‘triangle inequality’ (16), a fundamental property conferring metricity to phylogenetic distances. Its violation invalidates phylogenetic reconstruction (17). Their asymmetric step matrices require that they be solely optimized on a rooted tree, making them even prone to genome size attraction artifacts. In addition to this self-inconsistency, transformation costs also violate genomic scaling and processes responsible for scale-free behavior of proteins, challenging evolution’s PC and artificially forcing biological innovations to the base of universal trees (3).

Arshan Nasir1
Kyung Mo Kim2
Gustavo Caetano-Anollés3*

1Department of Biosciences, COMSATS Institute of Information Technology, Islamabad 45550, Pakistan
2Microbial Resource Center, Korea Research Institute of Bioscience and Biotechnology, Daejeon, South Korea
3Evolutionary Bioinformatics Laboratory, Department of Crop Sciences, University of Illinois at Urbana-Champaign, Urbana, IL 61820, USA

References
1. A. Nasir, G. Caetano-Anolles, A phylogenomic data-driven exploration of viral origins and evolution. Sci. Adv. 1, e1500527 (2015).
2. A. Harish, A. Tunlid, C. G. Kurland, Rooted phylogeny of the three superkingdoms. Biochimie 95, 1593-1604 (2013).
3. K. M. Kim, A. Nasir, G. Caetano-Anollés, The importance of using realistic evolutionary models for retrodicting proteomes. Biochimie 99, 129-137 (2014).
4. J. S. Farris, Methods for computing Wagner trees. Syst. Zool. 19, 83-92 (1970).
5. J. Lundberg, Wagner networks and ancestors. Syst. Zool. 18, 1-32 (1972).
6. P. H. Weston, Indirect and direct methods in systematics, inn Ontogeny and Systematics, C, J. Humphries, Ed. (Columbia University Press: New York, 1988), pp. 27–56.
7. P. H. Weston, Methods for rooting cladistic trees, in Models in Phylogeny Reconstruction, Systematics Association Special Volume No. 52, D. J. Siebert, R. W. Scotland, D. M. Williams, Eds. (Clarendon Press, Oxford, 1994), pp. 125–155.
8. H. N. Bryant, Hypothetical ancestors and rooting in cladistic analysis. Cladistics 13, 337-348 (1997).
9. A. V. Brower, M. C. de Pinna, Homology and errors. Cladistics 28, 529-538 (2012).
10. A. Nasir, K. M. Kim, G. Caetano-Anollés, Global patterns of protein domain gain and loss in superkingdoms. PLoS Comput. Biol. 10, e1003452 (2014).
11. J. Felsenstein, The statistical approach to inferring phylogeny and what it tells us about parsimony and character compatibility, in Cladistics: Perspectives on the Reconstruction of Evolutionary History, T. Duncan, T. F. Stuessy, Eds. (Columbia University Press, New York, 1984), pp. 169-191.
12. M. Wang, Y.-Y., Jiang, K.M. Kim, G. Qu, H.-F. Ji, H.-Y. Zhang, G. Caetano-Anollés, A molecular clock of protein folds and its power in tracing the early history of aerobic metabolism and planet oxygenation. Mol. Biol. Evol. 28, 567-582 (2011).
13. E. Ferrada, A. Wagner, Evolutionary innovations and the organization of protein functions in genotype space. PLoS ONE 5, e14172 (2010).
14. A. Abroi, J. Gough, Are viruses a source of new protein folds for organisms? - Virosphere structure space and evolution. Bioessays 33, 626-635 (2011).
15. A. Abroi, A protein domain-based view of the virosphere–host relationship. Biochimie 119, 231-243 (2015).
16. W. C. Wheeler, Systematics: a course of lectures (John Wiley & Sons, Hoboken, New Jersey, 2012).
17. W. C. Wheeler, The triangle inequality and character analysis. Mol. Biol. Evol. 10, 707–712 (1993).

Table 1. Fact-checking the narrative of Harish et al.

Fiction: We root trees with the “outgroup comparison method” that uses an “hypothetical pseudo-outgroup”, i.e. “an artificial all-zero taxon”
Fact: We do not use indirect character polarization methods of rooting that rely on outgroup taxa. Instead, we root trees with direct methods, which do not use outgroups or pseudo-outgroups and avoid auxiliary and ad hoc assumptions (1). We note that the root branch serves as a basal character state vector for the time arrow (9). Inclusion of outgroups or pseudo-outgroups results in an optimized state-set derived from a more inclusive set of taxon terminals, which changes tree topology, tests of character homology, and the results of phylogenetic analyses. See section 1 in main text

Fiction: Our rooting approach uses “a hypothetical pseudo-outgroup: an artificial construct based on presumed ancestral states, … an ‘all-zero’ or ‘all-absent’ hypothetical ancestor. The Lundberg method involves estimating an unrooted tree for the ingroup taxa only, and then attaching an outgroup … to the tree a posteriori to determine the position of the root node”.
Fact: We do not combine outgroups and ancestors during tree reconstruction. “(The) use of one hypothetical ancestor that combines inferences based on outgroup comparison with those based on other methods of polarizing character transformations to root a cladogram is invalid (8).” Inferences regarding ancestral states apply to the outgroup or ingroup nodes, respectively, and “cannot be combined into a single hypothetical construct (8)”. Section 1

Fiction: “The state '0' or 'absence' of a superfamily is assumed to be ‘ancestral’ a priori”.
Fact: We first optimize character change in unrooted networks using the Wagner algorithm and then polarize character transformations a posteriori, most parsimoniously, and complying with Weston’s generality criterion. Lundberg simply enables the task. Section 2

Fiction: “We can expect the ‘all-zero’ pseudo-outgroup to cluster with proteomes described by the largest number of ‘0s’ in the data matrix”.
Fact: During searches of tree space, optimal trees are retained by minimization of character state changes in branches of competing trees (4). In absence of computational optimization, the topology or rooted trees cannot be predicted from patterns in character state vectors of ingroup or outgroup taxa. Section 3

Fiction: “The Lundberg method distorts phylogenetic analysis” in ToLs and ToDs. “The rooting in viral lineages is an inevitable consequence of pre-specifying ‘0’ or ‘absence’ as the ancestral state”.
Fact: Character polarization plays no role in defining unrooted tree topology, which cannot be distorted by genome size. Their claim is “erroneous since polarization also involves spread in the nested lineages of the uToLs and is only applied a posteriori, allowing gains and losses throughout branches of the tree” [supplementary text in (1)]. If this were to be true then we should observe a comb-like pattern of organism appearance in the ToL, which is clearly not the case. Section 3

Fiction: The tree of structural domains (ToD) is “uninterpretable in terms of the definition of the superfamilies which it comprises” and its rooting is “unreliable”.
Fact: Harish et al. confuse the meaning of characters and taxa and extend their unsubstantiated claims to ToDs. To build ToDs, domain homologies (defined with robust hidden Markov models) define families and superfamilies that are used as taxa, while their proteomic abundances are used as characters. ToDs contain significant phylogenetic signal that follows a molecular clock linking the molecular and geological records (12). Section 4

Fiction: “Inferences based on statistical distributions of superfamilies alone are unconvincing, especially in light of other recent analyses”
Fact: The other recent analyses that are mentioned (14,15) failed to couple the comparative genomic studies to phylogenomic reconstructions, which in our case (1) helped weed out evolutionary scenarios that were non-significant or incompatible with reconstructed history. Section 5