How not to make a phylogeographic study

A citation alert pointed me to the paper of Zhang et al. (2017) to be published in Tree Genetics & Genomes, a failed attempt to make a biogeographic study on a small Ulmaceae genus: Zelkova. The severe concerns raised by at least one peer (not me) were largely ignored by the authors and the editor, providing us with a paper that managed to combine the most important pitfalls in (plant) biogeographic studies.

Background—What we know about Zelkova

Zelkova is a small Ulmaceae (elms and relatives) genus with five or six species that show a disjunct distribution pattern: in western Eurasia, we have the geographically very restricted species Z. sicula on Sicily, Z. abelicea on Crete, and Z. carpinifolia in scattered in Georgia (Transcaucasia) to northern Iran. In East Asia, we have one or two species that survived only in China proper (Z. schneideriana and Z. sinica), and one widespread species in China, the Korean peninsula, and Japan: Z. serrata. The biogeographic pattern of the genus appears to be trivial: the western Eurasian and East Asian species each form a lineage. In the western Eurasian lineage, the Mediterranean species Z. abelicea and Z. sicula are close sisters (Denk & Grimm 2005; Christe et al. 2014). In the East Asian lineage, molecular data so far failed to separate Z. schneideriana (and Z. sinica, if the sparse data are genuine) from the widespread Z. serrata. However, ITS patterns shared between Z. carpinifolia and the East Asian species Z. serrata, are indicative for a recent split, or a longer contact between Z. carpinifolia and Z. serrata. This may have left a persisting imprint in the biparentally inherited nuclear gene regions (such as ITS) but cannot be captured by the maternally inherited plastomes, primarily sorted by geography. But whereas the genetic differentiation in the western Eurasian species is well studied (Christe et al. 2014), the data for East Asia are still scarce. Being not feasible, neither we (Denk & Grimm 2005) nor Christe et al. provided an explicit biogeographic-historical reconstruction, a so-called ‘ancestral area analysis’ (AAR). Where the paper of Zhang et al. aimed to step-in.

Zhang et al.'s study—according to their abstract

Zhang et al. “… sequenced and combined ITS and trnL-trnF, psbA-trnH [correct order: trnH-psbA], and rbcL to reconstruct a phylogenetic tree.” Using Bayesian dating, “the age of Zelkova was traced to Cretaceous ca. 70 Ma, the crown divergence age of the Ulmaceae. Generic diversification was started at Paleocene ca. 53 Ma.” Not a result, but an assumption: the respective node heights, ages, were accordingly constrained. They applied several AAR methods: “Based on the reconstructed ancestral area, northeastern China was speculated to be the place of origin, from where species migrated westward to western Asia–southern Europe and dispersed to Japan and Korea within East Asia.” and further “presumed” that “the Chinese subtropical region” [i.e. today the south of China] is the “diversity center” [trivial: the only region with three species; however, the AARs only scored for “China” and “China + Japan”], and that the genus’ distribution pattern relates to the “East Asian monsoon onset at about 22 Ma as well as the Qinghai–Tibetan Plateau uplift” [always a good guess]. The last sentence of the abstract states that: “Species extinction in northeastern China perhaps coupled with the climate cooling event, the glacial epoch during the Quaternary.” Let alone the somewhat awkward phrasing (the native English-speaking co-author died 7 months before the paper was submitted), everything stated in the abstract or conclusions that is not trivial, is without any basis.

Zhang et al.’s prime error: poor data basis

No georeferenced samples…

Fig. 1 in Zhang et al. shows a distribution map of Zelkova species, but none of these points were actually sampled by the authors. Instead they sequenced the four said gene regions for 12 individuals, including nine from botanical gardens (one only sequenced for rbcL): the two Z. carpinifolia, the Georgian-Iranian species, and the rest representing the East Asian species Z. serrata, Z. sinica (possibly a mis-determination), and Z. schneideriana growing in China in the wild (Z. sinica may indeed be extinct). The remaining three were from herbarium material: one Z. serrata from Sichuan (W. China), and each one of the two remaining western Eurasian species.
Relying exclusively or mostly on arboretum material is problematic for two reasons: 

1. Being on cultivation for possibly more than one generation, arboretum trees may be (unnatural) hybrids.
2. The provenance of the original seed/seedling maybe unknown. 

Hybrids can produce misleading/unrepresentative data; and all material with unknown provenance is effectively useless for detailed, species-level biogeographic studies (unless the original distribution area of the species is very small). In all cases studied, extra-tropical tree genera, including Zelkova (Christe et al. 2014), showed substantial (plastid) haplotype variation within species, which can exceed interspecies divergence. In Zelkova, Z. sicula falls within the variation of Z. abelicea, and Z. schneideriana and Z. sinica seem to be nested in Z. serrata.

… and likely unsuitable set of markers

The authors combined the nuclear ITS region with three plastid regions, two non-coding spacers (trnL-trnF, slow-evolving; trnH-psbA, fast-evolving) and a highly conserved gene, the rbcL gene. Signal in the ITS is nonetheless complex, and one has to deal with intraindividual variation (Denk & Grimm 2005). The patterns found by Christe et al. (2014) indicate a principal correlation between ITS divergence and combined trnL/trnH-psbA haplotypes (above species level), but they also relied on our ITS data for the East Asian species pointing to non-trivial differentiation patterns and a closer relationship between Z. carpinifolia and Z. serrata than seen in plastid geneaologies.
Zhang et al. write in the first sentence of the results: “Bayesian inference of the ITS dataset and cpDNA datasets showed no [in]congruence [a mistyper; see mail by the authors from 1/9/2017] at main topological nodes of two phylogenetic trees [possible translation: no (highly) supported conflict between ITS and cpDNA inferences]; thus, both datasets can be combined into one.” Naturally, no documentation/proof is provided for this claim. The supplement EPS-files (a Bayesian inference and ML bootstrap cladogram) show unambiguous support for the genera and the long-resolved relationships within Zelkova (e.g. our study or Christe et al. 2014), but not so in the East Asian lineage critical to the conclusions of the study (Fig. 1); which may be due to incongruence between ITS and plastid data.

Fig. 1 The East Asian subtree of Zelkova, extracted from the Bayesian majority rule consensus tree (left) and ML tree (right) provided in the "complimentary material fig. S1 and fig. S2" of Zhang et al. Note the low posterior probabilites (PP, branch labels, left) for four out of six (potential) branches, which can be lower than ML bootstrap support (branch labels, right); a direct indication of problematic signal. The position of "Z. serrata 5" may be a missing data artefact (rbcL likely invariable within this lineage). No phylograms are provided, the data are still confidential, so the actual branch-lengths are unknown.

The trnL-trnF spacer and rbcL gene regions are low-divergent to invariable and often uninformative below the genus level in many northern-hemispheric extratropical tree genera. In fact, branches with lower Bayesian PP than ML BS support in the East Asian subtree of Zhang et al. may be an indication for a lack of discriminating signal in all combined gene regions.
[Zhang et al. have so far not released or shared their data, the provided GenBank accession numbers are still confidential and a data matrix has not been provided. Once their data are released, I’ll provide a full re-investigation of all available data on the genus.]

Zhang et al.’s second major error: bad root

A chronogram, a dated tree, is a rooted ultrametric phylogenetic tree. Thus, its age estimates depend on a well-informed root and reasonably estimated branch-lengths. Zhang et al. included also Ulmus (5) and Hemiptelea (2 samples) of the Ulmaceae, and Celtis (1) and Pteroceltis (2) of one of their sister families, the Cannabaceae. As node constraints, they relied on an obscure, partly misinterpreted mix of secondary evidence (ages estimated by others) to constrain the minimum root age of their tree to 70 Ma. They considered 70 Ma a valid estimated for the “crown age of Ulmaceae”, i.e. the minimum age of the ‘most recent common ancestor’ (MRCA) of Hemiptelea, Ulmus and Zelkova. In addition, they fixed the stem (root) ages (probably, their description on p. 4 is a bit confused) of Ulmus and the East Asian Zelkova lineage to 50 Ma based on “early fossil[s] found in northeastern China” (p. 4). Unfortunately, they forgot to properly root their tree before the dating. In their tree, Zelkova is sister to the remaining Ulmaceae, which form a grade, and the Cannabaceae are placed as sister to Hemiptelea (possibly long branch attraction). And as consequence the constrained nodes are closer to the tree root than they should be (Fig. 2); and the 70 Ma are used for the MRCA of Ulmaceae and Cannabaceae, not only Ulmaceae.

Fig. 2 The chronogram used by Zhang et al., and its properly rooted counterpart. Annotations to the original figure (left) in blue, red font indicates first-order errors. Two of the three used age priors (constraints) inform the minimum age of a wrong node, hence lead to methodologically wrong estimates. With respect to the few tree leaves and the fossil record of Zelkova accross the entire Northern Hemisphere (including North America), a fossilised-birth-death dating would probably lead to more useful estimates (given a meaningful taxon and gene sample). 

Age estimates using relaxed or strict clock models are further the product of branch-lengths (relative or absolute), which brings us back to the data basis: the combined regions show likely different divergence levels, and this is particularly problematic when comparing ingroup diversity (Zelkova) with the diversity between genera (Ulmus, Hemiptelea) and families (Ulmaceae vs. Cannabaceae).

Zhang et al.’s final nails in the coffin:

Poor ancestral area coding …

The idea of the study was to produce an AAR (ancestral area reconstruction) for the genus. The authors hypothesise a north-eastern Chinese origin of the genus. For the geographically highly restricted western Eurasian species geographic scoring is trivial: the Mediterranean sister species Z. abelicea (Crete) and Z. sicula (Sicily) were scored as “southern Europe (D)” (labelled “North America” [?!] in their fig. 3), and the Transcaucasian-Iranian Z. carpinifolia as “West Asia (C)” (labelled “Europe”[!] in their fig. 3). Zhang et al. scored their Z. sinica and Z. schneideriana as “China (A)”; both species are – according to their own tree – a subset of the widespread Z. serrata, all of which were scored as “China + Japan (and Korea) (AB)” lacking proper provenance information. Meaningful would have been, regarding the authors’ hypothesis, to score only southern, southwestern (subtropical) Chinese material as ‘A’, and north-eastern Chinese-Korean-Japanese material as ‘B’. The authors speculate and discuss that the genus originated in north-eastern China, which is closer to Korea and as distant from Japan as it is from the southern and southwestern Zelkova populations in China. Biogeographic scoring should be done adapted to the main question of a study, and not based on political boundaries (or other artificial standards).
OTUs (‘operational taxonomic units’; tip taxa) should also not be scored as ambiguous. If a species is found in two (or more) of the used biogeographic regions, one needs material from each region. If only material from one region is available, only that region must be scored. For Zhang et al., a meaningful coding was probably impossible due to the lack of georeferenced material of Z. serrata. Thus, they lacked exactly the material crucial for what they wanted to study.

… and mis-interpretation of AAR results

Using two tools with different AAR algorithms Zhang et al. (p. 5) found: “The ancestral area of Zelkova in S-DIVA [method 1] was shown as AC [=West Asia+China] or AD [geographically impossible], since A was the intersection of both areas AC and AD; therefore, A [=China] may be elected as the ancestral area of Zelkova. Similarly, the DEC result [method 2] showed that the ancestral area of Zelkova was A, as the intersection of A and ABCD [= no idea].” Left aside the awkward phrasing, this is nonsense. Naïvely applied biogeographic inference methods ignoring the fossil record are not only biased by the current-day situation (in Zhang et al.’s case 8 of the 12 included OTUs where scored for exclusively or possibly China), when their result is ambiguous then all results must be discussed or the analysis discarded as useless.
The latter applies in Zhang et al.'s case. One does not need any sophisticated inference with ambiguous results to put forward the authors' hypothesis. Common sense would suffice. Just based on the current-day situation (see also Fig. 2), one could hypothesise the genus originated in China and migrated west and north; although it may be not that easy, when we include evidence from the fossil record, morphology, and ITS mutation pattersn (see discussion in Denk & Grimm 2005).

Zhang et al.—A highly valuable bad example

The study of Zhang et al., and equally bad studies, expose the inefficiency of the widely applied non-transparent (“confidential”) peer review system to prevent the publication of scientifically obscure and badly crafted papers (not rarely, but also not exclusively, the products of tireless Chinese researchers in desperate need of publications in international peer-reviewed journals who team up with a retired, more-or-less accomplished U.S. scientist to increase their chances). Tree Genetics & Genomes is (well, was, its Impact Factor falters) a proper mid-tier journal with focus on population genetics; and I can only speculate why the editor, Pär Ingvarsson, accepted (“communicated”) the paper (with respect to the decreasing Impact Factor, 2.4 in 2014 to 1.6 in 2016, one reason may be to get a hold on the booming Chinese science market).

Aside from that, it’s a very fine example of how not to do a biogeographic inference. Errors done by Zhang et al. can be found in many other plant biogeographic papers, but maybe not so easy to see. Here, you have them all together on 10 printed pages, which is a nice service for university teachers. You can print out the paper, hand it over to your (under)graduate or Ph.D. students and have them make a list, what went wrong here. And, if you find the time, you can engage in a discussion on the confidential peer review system afterwards.

Linked material

My correspondance with the authors and editors of Zhang et al.; and the related post

References

Christe C, Kozlowski G, Frey D, Bétrisey S, Maharramova E, Garfi G, Pirintsos S, Naciri Y. 2014. Footprints of past intensive diversification and structuring in the genus Zelkova (Ulmaceae) in south-western Eurasia. Journal of Biogeography 41:1081–1093.
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.
Zhang M-L, Wang L, Lei Y, Sanderson SC. 2017. Cenozoic evolutionary history of Zelkova (Ulmaceae), evidenced from ITS, trnL-trnF, psbA-trnH, and rbcL. Tree Genetics and Genomes 13:111 [e-paper]. https://link.springer.com/article/10.1007/s11295-017-1182-4; see here for an annotated version.