Two papers you may want to read before inferring trees from morphological (or other) data

In this post, I'll advertise two probably undercited papers published in 2004 by Richard H. Zander regarding “Minimal values of reliability of Bootstrap and Jackknife proportions, Decay index, and Bayesian posterior probability”, and Matthew Spencer and co-workers on the “Phylogenetics of artificial manuscripts”. Two papers that should have been read by anyone trying to infer trees from morphological data including fossil taxa, or non-trivial data in general.

With the advent of molecular data (now increasingly phylogenomic data), phylogenetic inferences based just on morphological matrices have become obsolete when studying extant (i.e. living) groups of organisms. In the era of ‘Big Data’, we have access to data sets providing us with fully resolved trees, where each branch has (nearly) unambiguous support. But not so for the many groups of extinct animals and plants that lived millions of years ago, where we have nothing more than an impression/compression of an organ/individual, or a set of bones. Or, when we start to study relationships between (even within) species or the relationships between extant and extinct species. As I pointed out in several recent posts for David Morrison’s blog The Genealogical World of Phylogenetic Networks, inferring just a tree may not be an optimal choice to analyse morphological matrices including (or limited) to extinct taxa; be it because of the generally tree-unlike signal [spermatophytes][dinosaurs], convergent evolution, change over time, or ancestor-descendant relationships in general (upcoming post, 3rd October 2017). Here are two papers that I found inspiring, and which I believe should have been read by any researcher who ever tried (successfully or not) to infer trees based on morphological data matrices (or similarly complex data).

Character and split support values (Zander 2004)

The paper of Zander published in the 2nd issue of a journal called PhyloInformatics (not sure it still exists) opens interesting viewpoints on how to look at branch support values (sometimes called ‘node support’ in phylogenetic literature, which is usually a wrong term). I found this paper, when I was (desperately) searching for studies dealing with ambiguous bootstrap support patterns that I could use to support my ideas. Back then, I just got into contact with the consensus networks (Holland & Moulton 2003) because my boss, Vera Hemleben (a geneticists), brought me together with a new informatics professor in Tübingen, Daniel Huson, who just happened to have finished another update of a software named SplitsTree (Huson 1998; Huson & Bryant 2006); and we had plenty of data on beech trees and maples with complex split patterns for various reasons (Denk et al. 2002; Grimm 2003; Denk, Grimm & Hemleben 2005; Grimm et al. 2006).

Zander performed branch support analyses on “invented” binary matrices of four taxa A, B, C and D. One of the taxa, the outgroup D, is an all-zero taxon. The relationships of the three ingroup taxa A, B, and C were partly ambiguous: a varying proportion of characters supported AB vs. AC vs. BC. Zander showed that split support values (such as parsimony bootstrap, BS, and Bayesian probabilities, PP) are to be expected in such situations, and that applying a threshold of e.g. BS support > 70 makes little sense. When e.g. 50% of the characters supported AB and 50% supported AC (no support for BC) then both parsimony BS support and Bayesian PP converged to 50/0.5. As soon as one alternative had more character support than the other parsimony BS increased for that alternative (disproportionately; see also this post), whereas PP readily converged to 1.0 (e.g. when 50 characters supported AB and 42 supported AC and no BC, or the ratio was 50:40:10 for AB:AC:BC). Zander also pointed the finger to a still standing problem in phylogenetic literature: “Identifying imbalance [conflicting branch support]… in published papers is impossible (without recalculation from the actual data set)” (see e.g. my re-analysis of the Loranthaceae subset of Su et al. 2015). The take-home message of Zander's paper (which is not easy to read) can be put in a political analogy:
PP is what you get, when your parliament is manned using a winner-takes-the-seat voting procedure, BS is what you get, when your procedure uses some sort of partly proportional vote (always starting from data that supports more than a single alternative).
Although Zander didn’t discuss it, his results should be kept in mind when analysing morpho-matrices of extinct organisms. When two branching alternatives (e.g. A+B and A+C) receive ample, albeit low support this can be a direct indication that A is an ancestor of B and C, or that A is a hybrid or mix of B and C. Consequently, if A is a potential ancestor of three taxa, B, C, and D, all splits involving A would ideally receive BS or JK support of 33, and – in a perfect world – a PP of 0.33. Conversely, BS > 50 (but << 100) and PP ~ 1.0 may just reflect the signal of a majority (≥ 50%) of characters, and it remains unclear whether the minority (<50%) prefers one alternative, many, or none. 

A phylogeny of artificial manuscripts (Spencer et al. 2004)

Again, my contact with this paper was co-incidental. We had to re-but (once again) nasty anonymous peer comments why we showed networks and not trees to put forward our hypotheses (Denk & Grimm 2009), and I was searching for studies dealing with actual ancestor-descendant relationships. Fortunately, one of my colleagues in Tübingen, Markus Göker, a mycologist with good knowledge about distance-based analysis (and many other things), pointed me to the paper of Spencer et al., published in the Journal of Theoretical Biology.
Spencer et al. had several scribes duplicating an original text that they could not understand (or not fully, lacking the needed language skill). As consequence, errors were produced during the copying process, the text evolved with each copy. Some of the copies were given to other scribes to be copied again. The authors then coded the differences in the text copies and applied a range of phylogenetic inferences (trees and networks). The outcome of the inferences was then compared to the ‘true tree’, i.e. the actual evolutionary tree recognising the original copy as the common ancestor of all copies, and some copies as the ancestors of others (the copies being the direct descendants of the template). They found that the neighbour-net, a 2-dimensional (planar) meta-phylogenetic graph, best depicted the actual ancestor-descendant relationships (see also this post for an example of evolution in a 2-dimensional space and this post dealing with the evolution of flamenco music).

Really, stop relying (exclusively) on trees

When you read the papers by Spencer et al. (2004) and Zander (2004), but also the paper of Holland & Moulton (2003) introducing the consensus network approach, you may ask yourself why more than a decade after, most phylogenetic publications (morphology- or molecular-based) still just show an incomprehensive tree (or strict consensus tree), even when many branches lack unambiguous support; and why branch support below a certain threshold (currently: BS < 70, PP < 0.95) is not reported, but dropped, and the reasons for moderate or low support rarely investigated (but see postscriptum).

Why do we cling to trees, even if they lack support, are indecisive, or incomprehensive regarding the actual signal in our data? Well, trees may be simplistic, but are also simple to grasp and even simpler to interpret. Showing the comprehensive result of the analyses may alert peers to distrust your data and study (happened nearly in all papers, we eventually published between 2005–2015), and necessitates to discuss the emerging alternatives (making the paper too long). But the signal – particularly, in morphological matrices – is not trivial, but complex. This complexity may be puzzling, but is always interesting, and may point to very interesting aspects of evolution. And the tools are long available to look into these puzzling, complex patterns. So, there is no excuse for just showing a tree.

PS Another reason for seeing trees with branches of unknown support is that certain reviewers, and occasionally editors, comfortably hiding behind the Impermeable Fog (i.e. confidential review process), insist on dropping these values, regarding them as meaningless. And many authors comply rather than protest (like me). For instance, Mike Pirie, editor of Taxon, stated in his 2nd decision letter to our paper on Cynanchum (Khanum et al. 2016): “I don't see that it makes any sense to even include the low and moderate values for the PP values. I would think it would be more meaningful to just consider PP ≥ 90 as supported and anything below that as not supported.” Which is, of course, mathematically already nonsense, a PP of 0.89 indicates (ideally) a probability of 0.89 that a branch is correct, which is insignificantly (0.01) less support than a PP of 0.9. I naturally objected this and further unqualified statements by Pirie and the chief editor of Taxon regarding phylogenetic analyses, and retreated as an author to protest Taxon’s editorial policies (second time they tried to rid a paper of my beloved networks); but my co-authors were finally granted showing the meaningless values and superfluous reconstructions, i.e. BS and PP support networks. Which Pirie and his editorial colleagues considered a “detriment” to the study, and accused of helping to hide data problems “under the carpet”. Check out the paper to decide for yourself, and compare it to the standard publications in Taxon, e.g. Su et al. 2015 and my (network-based) re-analysis of part of Su et al's data (Grimm 2017).

Denk T, Grimm G, Stögerer K, Langer M, Hemleben V. 2002. The evolutionary history of Fagus in western Eurasia: Evidence from genes, morphology and the fossil record. Plant Systematics and Evolution 232:213-236.
Denk T, Grimm GW. 2005. Phylogeny and biogeography of Zelkova (Ulmaceae sensu stricto) as inferred from leaf morphology, ITS sequence data and the fossil record. Botanical Journal of the Linnéan Society 147:129-157.
Denk T, Grimm GW. 2009. The biogeographic history of beech trees. Review of Palaeobotany and Palynology 158:83–100.
Denk T, Grimm GW, Hemleben V. 2005. Patterns of molecular and morphological differentiation in Fagus: implications for phylogeny. American Journal of Botany 92:1006-1016
Grimm GW. 2003. Tracing the mode and speed of intrageneric evolution - a case study of genus Acer L. and Fagus L. D.Sc. thesis. Eberhard-Karls University.
Grimm GW, Renner SS, Stamatakis A, Hemleben V. 2006. A nuclear ribosomal DNA phylogeny of Acer inferred with maximum likelihood, splits graphs, and motif analyses of 606 sequences. Evolutionary Bioinformatics 2:279–294.
Grimm 2017. ‘Quick-and-dirty’ re-analysis of the Su et al. (2015) data of Loranthaceae and their sister groups. File S6 in the online supporting archive to: Grímsson F, Grimm GW, Zetter R. 2017. Evolution of pollen morphology in Loranthaceae. Grana DOI:10.1080/00173134.2016.1261939.
Holland B, Moulton V. 2003. Consensus networks: A method for visualising incompatibilities in collections of trees. In: Benson G, and Page R, eds. Algorithms in Bioinformatics: Third International Workshop, WABI, Budapest, Hungary Proceedings. Berlin, Heidelberg, Stuttgart: Springer Verlag, p. 165–176.
Huson DH. 1998. SplitsTree: analyzing and visualizing evolutionary data. Bioinformatics 14:68–73.
Huson DH, Bryant D. 2006. Application of phylogenetic networks in evolutionary studies. Molecular Biology and Evolution 23:254–267.
Khanum R, Surveswaran S, Meve U, Liede-Schumann S. 2016. Cynanchum (Apocynaceae: Asclepiadoideae): A pantropical Asclepiadoid genus revisited. Taxon 65:467–486.
Spencer M, Davidson EA, Barbrook AC, Howe CJ. 2004. Phylogenetics of artificial manuscripts. Journal of Theoretical Biology 227:503-511.
Su H-J, Hu J-M, Anderson FE, Der JP, Nickrent DL. 2015. Phylogenetic relationships of Santalales with insights into the origins of holoparasitic Balanophoraceae. Taxon 64:491–506.
Zander RH. 2004. Minimal values of reliability of Bootstrap and Jackknife proportions, Decay index, and Bayesian posterior probability. PhyloInformatics 2:1-13. 

No comments:

Post a Comment

Enter your comment ...