For comparing phylogenetic diversity among communities, go ahead and use synthesis phylogenies

Should we build our own phylogenetic trees based on gene sequence data, or can we simply use available synthesis phylogenies? This is a fundamental question that any study involving a phylogenetic framework must face at the beginning of the project. Building a phylogeny from gene sequence data (purpose-built phylogeny) requires more effort and expertise than subsetting an already available phylogeny (synthesis-based phylogeny). If phylogenetic diversity estimates based on these two types of phylogenies are highly correlated, using readily available synthesis-based phylogenies is justified for comparing phylogenetic diversity among communities. However, a comparison of how these two approaches to building phylogenetic trees influence the calculation of phylogenetic diversity has not been explicitly tested. We generated threepurpose-built phylogenies and their corresponding synthesis-based trees (two from Phylomatic and one from the Open Tree of Life). We then used a simulation approach to generate 1000 communities with a fixed number of species per site and compared the effects of different trees on estimates of phylogenetic alpha and beta diversity using Spearman’s rank-based correlation and linear mixed models. Synthesis-based phylogenies generally over-estimated phylogenetic diversity when compared to purpose-built ones. However, their resulting measures of phylogenetic diversity were highly correlated (Spearman’s r > 0.8 in most cases). Mean pairwise distance (both alpha and beta) is the index that is most robust to the differences in tree construction that we tested. Measures of phylogenetic diversity based on the Open Tree of Life showed the highest correlation with measures based on the purpose-built phylogenies. For comparing phylogenetic diversity among communities, our results justify taking advantage of recently developed and continuously improving synthesis trees such as the Open Tree of Life.


Introduction
Phylogenies describe the evolutionary history of species and provide important tools to study ecological and evolutionary questions (Baum and Smith ). Recently, phylogenies have been used to better understand patterns of community assembly. The phylogenetic structure of ecological communities can lend insight into the processes by which local communities assemble from regional species pools (Webb et al. ). For example, if closely related species are more likely to co-occur in the same habitats, we might suspect that these species share traits that allow them to have a positive growth rate under the environmental conditions in these habitats. To test whether closely related species are more or less likely to co-occur, one common approach is to calculate the phylogenetic diversity of communities and then compare the observed phylogenetic diversity with those expected by chance through di erent null models. There is a growing body of literature using this community phylogenetic approach, documenting the phylogenetic structure of ecological communities across taxa and scales (Webb et al. , Cavender-Bares et al. , Helmus et al. , Vamosi et al. , Cardillo , Smith et al. , Li et al. , Marx et al. ).
As an important facet of biodiversity, phylogenetic diversity (Faith ) also plays a crucial role in conservation biology by complementing more traditional taxonomic measures of biodiversity (e.g., species richness). For example, two communities can have the same number of species but di er drastically in their phylogenetic diversity depending on relatedness of the constituent species. The community with higher phylogenetic diversity, representing taxa more distantly related to each other, is expected to be more stable and productive given its greater evolutionary potential to adapt to changing environmental conditions (Forest et al. , Maherali and Klironomos , Lavergne et al. ). Therefore, all else being equal, a community with higher phylogenetic diversity should have higher conservation priority.
The information gained from phylogenetic diversity analyses are only as good as the species composition data and the phylogenies from which they are generated. In this manuscript, we explore how tree generation a ects these phylogenetic diversity metrics. Generally, ecologists and evolutionary biologists use two common approaches to build phylogenies for community phylogenetic analyses. The rst approach is for a researcher to generate his/her own phylogenies for a set of target species based on gene sequence data. We refer to such phylogenies as purpose-built phylogenies. The second approach is to derive phylogenies based on available synthesis trees, such as the Open Tree of Life , or classi cations, such as the Angiosperm Phylogeny Group (APG IV et al. ), by pruning or sampling, respectively, from the resource so that the phylogeny contains only the target species. We refer to such phylogenies as synthesis-based phylogenies. To a certain extent, one can argue that a synthesis tree could be a purpose-built tree for a larger set of species, but the sources for deriving the synthesis-based trees vary in scope, methodology, assumptions, and content (see Materials and Methods for further description of source trees for synthesis-based phylogenies). From a researcher perspective, a purpose-built phylogeny is a major undertaking but o ers potential to utilize taxonomic and phylogenetic expertise often needed in order to successfully construct trees. Synthesis trees, as compilations of peer-reviewed phylogenetic hypotheses, o er an immediately available, but typically less customizable output to researchers. We thus use these two terms (purpose-built and synthesis-based) to categorize the underlying methods and researcher cost-bene ts to obtain phylogenies.
Generating a purpose-built tree requires more e ort and expertise than subsetting a well-developed phylogeny or sampling from a classi cation. Generally, purpose-built trees are constructed by using newly generated sequence data and then combining those data with data already available on GenBank; although in many cases the researcher may simply use what is in phylogenies. Therefore, they provided little practical advice about selecting between purpose-built and synthesis-based phylogenies for ecological studies. In this study, we compared phylogenetic diversity metrics calculated from purpose-built phylogenies and corresponding phylogenies derived from three commonly used sources. It is important to note that we do not treat the purpose-built phylogenies as a gold standard and we recognize that sampling bias of both taxa and genes, combined with variation introduced through the tree-building process (e.g., tree reconstruction methods, assessment of support, etc.), can compromise the accuracy of purpose-built phylogenies. However, these issues -and others -apply also to the source trees used for synthesis-based phylogenies, although perhaps at di erent scales. Our aim here is to quantify the in uence of the two tree construction techniques on measures of phylogenetic diversity that are commonly employed in the rapidly growing eld of community phylogenetics.

Purpose-built phylogenies
We collected three "purpose-built" phylogenies from published and unpublished sources.

Commonly available phylogenies
For each of the three purpose-built phylogenies, we generated four phylogenies based on di erent sources with which to compare phylogenetic alpha and beta diversity. The rst two were generated using Phylomatic v . (Webb and Donoghue ) using two di erent backbone trees: R20120829 (APG III ) and zanne2014 (Zanne et al. The fourth phylogeny was a random coalescent phylogeny generated using the rcoal function from the R package ape (Paradis et al. ). The random tree was then scaled to have a root age that was the average root age of tree_apg, tree_zanne, and tree_otl. Results based on the random phylogeny should not correlate with those based on other phylogenies.
Not every species from the purpose-built phylogenies was found in all of the synthesis phylogenies.
For the pine rockland phylogeny, out of species ( . %) were found in all phylogenies. For the alpine plant phylogeny, out of species ( . %) were found in all phylogenies. For the Florida ora phylogeny, out of species ( . %) were found in all phylogenies. Therefore, we pruned the purpose-built phylogenies to have the same species as their corresponding synthesis tree. In practice, one could insert species that were missing from the derived phylogeny as polytomies in the same genus, so that all species could be included in the analysis.

Generation of community assemblages
For each purpose-built phylogeny, we simulated presence/absence site-by-species matrices.
Each matrix has sites, with species within each site randomly selected from the phylogeny tips representing the species pool. We xed species richness of each site to be to remove any e ects of species richness on the phylogenetic diversity measures. Without setting all sites to have the same number of species, results based on di erent phylogenies will correlate with each other. For example, it is likely that results from tree_random will be highly correlated with results from other phylogenies (Appendix Fig. A ). This is because most phylogenetic diversity metrics correlate with species richness, which, in turn, will lead to correlations among them and confound the comparisons of e ects of phylogeny per se on the measurement of phylogenetic diversity.
Removing the constraint of using the same species richness does not a ect our results and conclusions (Appendix Fig. A , A ). In our current setting, the maximum total number of species across sites is × = , which is similar to the number of tips in the largest purpose-built phylogeny in our study. We selected species from the species pool randomly because previous studies demonstrated that di erent approaches to species selection give similar results (Swenson ). For phylogenetic beta diversity, we applied UniFrac (Unif), inter-assemblage MPD (MPD_beta), inter-assemblage MNTD (MNTD_beta), and phylogenetic community dissimilarity (PCD) to all possible unique combinations of assemblage pairs. Unif is derived from the Jaccard dissimilarity index and calculates the total branch length unique to each assemblage relative to the total branch length of all species in a pair of assemblages (Lozupone and Knight ). Therefore, it measures the fraction of evolutionary history unique to each assemblage. MPD_beta and MNTD_beta were derived from MPD and MNTD, respectively, but instead of comparing species within the same assemblage, they compare species from two di erent assemblages (Webb et al. ). PCD measures pairwise phylogenetic dissimilarity between assemblages by asking how much of the variance of values of a hypothetical trait among species in one assemblage can be predicted by the values of species from another. PCD is independent of species richness of the pair of assemblages and has relatively higher statistical power than other common metrics (Ives and Helmus ).

Phylogenetic diversity measurements
As PD and MNTD are both correlated with species richness (Miller et al. ), null models that retain species composition while randomly shu ing tips of the phylogeny are commonly used to standardize phylogenetic diversity results. Despite the fact that MPD is independent of species richness, its variance changes relative to species richness (Miller et al. Unif, MNTD_beta, and PCD are available. However, the pairwise beta diversity metrics share the same core formula with their corresponding alpha diversity metrics. We thus expect that the results based on SES of these beta diversity metrics will be the same as those based on the observed diversity values in our simulations. Given the similarity in results between raw and standardized phylogenetic alpha diversity measures and the large computational burden of calculating SES for phylogenetic beta diversity metrics, we did not include the results for SES in this study.

Statistical analyses
We have two primary goals. First, we want to test the correlation between phylogenetic diversity values calculated from purpose-built phylogenies and those calculated from synthesis phylogenies.
Second, we want to investigate whether phylogenetic diversity calculated from synthesis phylogenies over-or under-estimates phylogenetic diversity when compared to purpose-built phylogenies. For the rst goal, we calculated the average Spearman's rank-based measure of the correlation between phylogenetic diversity values from all phylogenies across the simulations. We used rank-based correlation because it is the relative phylogenetic diversity, not the absolute one, that we are interested in. For the second goal, we used Linear Mixed Models

Alpha diversity
Phylogenetic alpha diversity (PD, MPD, and MNTD) values calculated with di erent phylogenies (tree_purpose, tree_apg, tree_zanne, and tree_otl) were highly correlated. The median Spearman's correlation of the simulations was larger than . across all comparisons (p < . for all simulations and comparisons; Fig. ). In most cases, the median Spearman's correlation was larger than . , especially for PD and MPD. Therefore, PD and MPD were more robust to varying the source of the phylogeny than MNTD. Across all comparisons, diversity values based on tree_otl showed the highest correlations with those based on tree_purpose, with an average correlation across all comparisons of . . As expected, diversity values based on the random phylogeny tree_random were not correlated with diversity values based on other phylogenies, with median Spearman's correlations close to zero (Fig. ).
The slopes of linear mixed models (LMM) were all less than one (Table ), suggesting that diversity values based on synthesis phylogenies generally over-estimated the diversity values based on the purpose-built phylogenies. The PD metrics based on the Open Tree of Life phylogeny (tree_otl) had estimates closest to those calculated from the purpose-built phylogenies (Table ).

Beta diversity
The phylogenetic beta diversity results (Un , MPD_beta, MNTD_beta, and PCD) show a similar pattern to the alpha diversity results. Beta diversity of community pairs based on di erent phylogenies was also highly correlated, with the median Spearman's correlation from the simulations greater than . across all comparisons ( Fig. ). Overall, phylogenetic beta diversity is more sensitive to the source of the phylogeny than alpha diversity. MPD_beta is the most robust beta diversity metric to the source of the phylogeny, followed by MNTD_beta, Unif, and PCD.  Table : Slopes based on linear mixed models (LMMs). Within the model, the response variable is the phylogenetic alpha diversity values based on the purpose-built phylogeny; the predictor is the phylogenetic alpha diversity values based on one of the synthesis phylogenies (tree_apg, tree_zanne, tree_otl, and tree_random). Therefore, slopes less than one indicate overestimations. Numbers within parentheses are the % con dence intervals for the slopes. The slopes of LMMs were generally less than one (Table ), suggesting over-estimates of beta diversity from the synthesis-based phylogenies compared with the purpose-built phylogenies.
However, slopes for MPD_beta values based on tree_otl were all greater than one, suggesting that beta PD metrics were under-estimated when compared to those calculated from the purpose-built trees. Metrics based on tree_zanne for the ora of Florida dataset were also under-estimated (Table ). For the other beta diversity metrics (i.e., Unif, MNTD_beta, and PCD), tree_otl generally gave results closer to those based on the purpose-built trees than did the other synthesis-based phylogenies.

Discussion
We examined how di erent phylogenies, purpose-built and synthesis-based, in uenced phylogenetic alpha and beta diversity measures commonly used in community phylogenetic  Table : Slopes based on linear mixed models (LMMs). Within the model, the response variable is the phylogenetic beta diversity values based on the purpose-built phylogeny; the predictor is the phylogenetic beta diversity values based on one of the synthesis phylogenies (tree_apg, tree_zanne, tree_otl, and tree_random). Therefore, slopes less than one indicate overestimations, and slopes greater than one are underestimates. Numbers within parentheses are the % con dence intervals for the slopes. analyses. We found two main results. First, the synthesis phylogenies generally over-estimated phylogenetic diversity compared with purpose-built phylogenies. This is not surprising because synthesis phylogenies generally have higher proportions of polytomies than purpose-built ones, which, in turn, leads to larger distances between species within these polytomies. This result agrees with Boyle and Adamowicz ( ) and Qian and Zhang ( ) but contradicts Swenson ( ), who found that phylogenies with more polytomies under-estimated phylogenetic diversity.
Second, phylogenetic diversity values calculated from synthesis trees were highly correlated with those based on purpose-built phylogenies, even if they were over-estimated. These results hold for both alpha and beta diversity and for phylogenies with di erent numbers of tips. While our study focuses on plants, we expect that our results will generalize to any taxonomic group. Therefore, phylogenies derived from synthesis trees can provide similar results to purpose-built phylogenies while saving e ort and time when quantifying and comparing phylogenetic diversity of communities.
One main reason for this conclusion is that, as ecologists and conservation biologists, we mostly care about the relative diversity among communities instead of their absolute diversity. For example, for a set of communities within one region, we may be interested in which communities have the highest/lowest phylogenetic diversity. The absolute phylogenetic diversity of each community does not mean much without comparing it to other communities. Because phylogenetic values based on di erent phylogenies are highly correlated with each other, the information available for community phylogenetic questions does not di er much between approaches. Even though such synthesis phylogenies may over-estimate absolute phylogenetic diversity for communities, the relative phylogenetic diversity among communities will be similar to those calculated from typically better resolved but less accessible phylogenies. Based on the information provided by relative values of phylogenetic diversity, the potential improved resolution of purpose-built trees for calculating the absolute PD may not be worth the e ort for community phylogenetic questions.
Our nding that phylogenetic diversity metrics are relatively insensitive to the phylogenies from which they are derived has been supported by other recent studies.  Donoghue ). These studies, however, only focused on alpha diversity. Our study extends the literature by also examining the e ects of phylogenies on beta diversity. We found the same pattern for beta diversity and alpha diversity. Taken together, a general pattern emerges: community phylogenetic alpha and beta diversity metrics are robust to reasonably good modern phylogenies.
Why are phylogenetic diversity values from purpose-built and synthesis phylogenies highly correlated? There are two possible reasons. First, both purpose-built and synthesis phylogenies likely share a similar systematic backbone and empirical resources such as genes, taxonomies, and expert knowledge. This guarantees that phylogenetic diversity based on these phylogenies will not be dramatically di erent. Second, phylogenetic diversity metrics aggregate (by summing or averaging) all information into one value for each site, which could help bu er most uncertainty and further mask most of the di erences between di erent phylogenies.
Our results should encourage ecologists to increasingly include phylogenetic analyses in community ecology studies given the growing accessibility of synthesis phylogenies and the robustness of phylogenetic diversity measures based on them. However, our results should not discourage the construction of purpose-built phylogenies, which are clearly valuable for many ecological and evolutionary questions. This is especially the case for purpose-built trees constructed from local DNA samples. The sequencing of species in a given community can yield data for species that have never been sequenced before. These new sequences can then be incorporated into synthesis trees, improving their resolution for future research. Direct sequencing of samples collected for a community is also important when the community contains

Data Accessibility
All phylogenies and R code used will be uploaded to gshare upon acceptance.