vol. 176, no. 5 the american naturalist november 2010 ൴ Phylogenetic Targeting of Research Effort in Evolutionary Biology Christian Arnold1,2,* and Charles L. Nunn1 1. Department of Human Evolutionary Biology, Harvard University, Cambridge, Massachusetts 02138; 2. Bioinformatics Group, Department of Computer Science, and Interdisciplinary Center for Bioinformatics, University of Leipzig, Ha¨rtelstraße 16–18, D-04107 Leipzig, Germany Submitted April 18, 2010; Accepted July 13, 2010; Electronically published September 21, 2010 Online enhancements: appendixes, data file. males on the basis of their ability to resist parasites (Ham- abstract: Many questions in comparative biology require that new ilton and Zuk 1982), and it has been used to probe the data be collected, either to build a comparative database for the first time or to augment existing data. Given resource limitations in col- origins of both parasitic and symbiotic associations (e.g., lecting data, the question arises as to which species should be studied Hugot 1999; Lutzoni et al. 2001). More recently, com- to increase the size of comparative data sets. By taking hypotheses, parative methods have been applied to study phylogenetic existing data relevant to the hypotheses, and a phylogeny, we show community ecology (Webb et al. 2002), for example, in that a method of “phylogenetic targeting” can systematically guide the context of the phylogenetic overdispersion of mam- data collection while taking into account potentially confounding malian communities (Cooper et al. 2008). The compar- variables and competing hypotheses. Phylogenetic targeting selects ative method can also be used to address conservation potential candidates for future data collection, using a flexible scoring system based on differences in pairwise comparisons. We used sim- issues (Fisher and Owens 2004), such as questions in- ulations to assess the performance of phylogenetic targeting, as com- volving the factors that influence rates of extinction pared with the less systematic approach of randomly selecting species (Purvis et al. 2000b) and how the phylogenetic clumping (as might occur when data have been collected without regard to of conservation threat status can lead to greater loss of phylogeny and variation in the traits of interest). The simulations phylogenetic diversity when species become extinct (Purvis revealed that phylogenetic targeting increased the statistical power et al. 2000a). to detect correlations and that power increased with the number of A comparative analysis requires data on a set of species species in the tree, even when the number of species studied was held constant. We also developed a Web-based computer program relevant to a hypothesis of interest. Usually, however, data called PhyloTargeting to implement the approach (http:// are available for only a fraction of the species in a clade, phylotargeting.fas.harvard.edu). and data collection in both the field and the laboratory is expensive and time consuming. A proper selection of spe- Keywords: comparative method, phylogeny, correlated evolution, cies to study is a nontrivial and multifaceted problem (Gar- taxon sampling, pairwise comparison. land 2001; Westoby 2002) that has rarely been addressed in a systematic way. Instead, species are often chosen either randomly or subjectively (Westoby 1999; Faustino et al. Introduction 2010) because they are of “particular (and perhaps irra- tional) interest” (Garland 2001, p. 119). Two problems are The comparative method has played a major role in un- introduced when species are chosen in an unsystematic covering adaptive trait evolution in biological systems way. First, the full range of variation is not used to test (Ridley 1983; Harvey and Pagel 1991; Pagel 1999; Martins the hypotheses. Second, taxonomic gap bias may occur, 2000). The comparative method has, for example, revealed meaning that data collection has been focused on a few links between mating systems and sperm competition in “popular” lineages. These different kinds of biases—in- primates (Harcourt et al. 1981) and other animals (Møller complete variation and gap biases—can make a momen- 1991; Hosken 1997). The comparative method also sup- tous difference to the conclusions one draws. In studies ports a model of sexual selection in which females choose of primates, for example, results of comparative research * Corresponding author; e-mail: [email protected]. are likely to change when the sample is tilted toward ter- Am. Nat. 2010. Vol. 176, pp. 601–612. ᭧ 2010 by The University of Chicago. restrial species rather than those that live in trees, because 0003-0147/2010/17605-52091$15.00. All rights reserved. terrestrial species possess larger body masses, exhibit dif- DOI: 10.1086/656490 ferent locomotor patterns, and live in larger social groups 602 The American Naturalist (Clutton-Brock and Harvey 1977; Martin 1990; Nunn and netically independent contrasts (PIC; Felsenstein 1985; van Schaik 2002). Harvey and Pagel 1991; Garland et al. 1992), pairwise To address these issues, methods are required that quan- comparison does not require a specific model of evolution tify potential biases in comparative data sets and identify or the estimation of states at interior nodes. In addition, the species that should be studied in the future. Indeed, some sets of species within a larger clade might not be it is common to read in articles of comparative research directly comparable in standard implementations of com- that further sampling is required to validate the findings, parative methods, such as PIC. In regard to mammalian because either the sample sizes were small or the sample sleep, for example, some cetaceans sleep with only one- was biased toward particular species within a clade (e.g., half of their brains (Lyamin et al. 2008), making it difficult in the study of sleep patterns; Roth et al. 2006; Capellini to compare the measurements of sleep in cetaceans with et al. 2009; Nunn et al. 2009). Unfortunately often, how- those in other mammals. The method of selecting specific ever, only general guidelines for this selection process have pairwise comparisons provides a way to limit comparisons been given, and these guidelines are often specific to the so that cetaceans are compared only with other cetaceans question of interest (Westoby 2002). To our knowledge, and noncetaceans are compared only with noncetaceans. no method yet exists that is flexible and specific enough Similarly, some behavioral experiments might require sim- to address the crucial task of prioritizing future research ilar sensory modalities or cognitive ability among species in light of specific hypotheses about the apportionment in the data set. Pairwise comparisons of some close rela- of variation in relation to one or more ecological factors. tives may be more appropriate for selecting species for Only a handful of studies have investigated ways of focused comparative experiments that take these factors systematically identifying species to study. For example, into account. Ackerly (2000) compared the performance of different When using the method of pairwise comparisons, it is taxon sampling strategies and found their statistical per- important that all pairs are phylogenetically independent, formances to differ substantially. One of the algorithms that is, that no branches are shared among the compari- he examined is based on the pairwise comparison ap- sons (Felsenstein 1985; Maddison 2000). In figure 2, for proach (Felsenstein 1985, p. 13; Møller and Birkhead 1992; example, different sets of phylogenetically independent Oakes 1992; Read and Nee 1995; Purvis and Bromham pairs (which we call a “pairing”; see Maddison 2000) are 1997; Maddison 2000) and identifies meaningful com- shown for each tree. Thus, when selecting phylogenetically parisons by selecting species pairs that differ by a certain independent pairs, the selection of a particular pair con- amount in the independent variable, following the sug- strains which other pairs can be selected. gestion of Westoby (1999). Although it overestimates the Here we present a new approach, which we call “phy- magnitude of the correlation, Ackerly (2000) showed that logenetic targeting,” to systematically identify the species this design increases the statistical power to detect cor- to be studied. Phylogenetic targeting is a taxon sampling related evolution (see also Garland 2001). Major weak- approach that aims to prioritize future research by iden- nesses of the method are that the threshold for when dif- tifying species that should be studied in a target-oriented ferences are large is arbitrary, that it is dependent on the way under consideration of the specific hypotheses and data set, and that it must be set manually, which limits its data. It is not a new way to analyze comparative data or applicability considerably. Mitani et al. (1996) considered a substitute for existing analysis methods, but rather it sampling strategies in relation to testing competing hy- draws on existing methods in comparative biology. This potheses, while Read and Nee (1995) discussed the need method uses the pairwise comparison approach and is to identify pairs that contribute for or against hypotheses. based on a scoring system that incorporates phylogeny and Similarly, Maddison (2000) presented a methodology for data on variables relevant to testing hypotheses, specifically choosing species pairs in which each pair is “a comparison involving the predictor and response variables in a com- relevant for the question