Puzzling Rocks and Complicated Clocks: How to Optimize Molecular Dating Approaches in Historical Phytogeography.
Qian Wang,Kang-Shan Mao
DOI: https://doi.org/10.1111/nph.13676
IF: 9.4
2016-01-01
New Phytologist
Abstract:In recent decades, molecular dating has become an important component of phytogeographic studies because the establishment of an evolutionary timescale enables the testing of hypotheses about the distribution and evolution of plants across time and space (Renner, 2005; Crisp et al., 2011; Ho, 2014). Based on molecular sequences and fossil calibration points, divergence times between living species can be estimated using a molecular clock. This approach was first proposed by Zuckerkandl & Pauling (1962) and assumes that the genetic distance between molecular sequences increases linearly from time of divergence between any pair of species. Great efforts have been made in recent years to improve molecular dating approaches for timing divergence events (Box 1). Despite these advances, there remain various problems with the acquisition (and interpretation) of age estimates of lineage diversification by means of molecular dating approaches (Pulquério & Nichols, 2007; Sauquet et al., 2012). In this regard, Wilf & Escapa (2015) recently pointed out that a number of newly discovered fossils provide evidence that previously published molecular dating estimates have underestimated lineage divergence times, which, in turn, has greatly affected ideas about how the breakup of the Gondwana supercontinent (from c. 180 to 30 Ma) shaped biogeographic patterns in the Southern Hemisphere (Sanmartín & Ronquist, 2004). They argued that the use of ‘megabiased clocks’ has resulted in the underestimation of divergence dates (younger bias) and wrongly led to the formulation of a ‘Green Web’ hypothesis that dispersal (over oceanic barriers), rather than vicariance due to the breakup of Gondwana, played a major role in the diversification of plant genera and families. Consequently, they concluded that ‘fossils and geochronology provide the only rigorous, enduring temporal framework for evolutionary radiations’. Although Wilf & Escapa (2015) showed that ‘… the (fossil) record from Patagonia convincingly demonstrates that Gondwanan history remains fundamental to the evolutionary radiations, distributions, survival, and conservation of Southern Hemisphere plants and plant associations …’, their views about the value of molecular dating are, in our opinion, partial and misleading. … Wilf & Escapa ( 2015 ) were right to draw attention to the potential younger biases in traditional ‘node dating’ approaches that were introduced via unjustified but routine fossil calibrating protocols (a list of ways to reduce these potential biases is provided in Table 1) … Second, Wilf & Escapa (2015) based their molecular dates on mean values. Although this is a common approach, it results in inappropriate interpretation and should be avoided (Sauquet, 2013). Instead, interpretation must take account of the uncertainty of dates (e.g. credible intervals of Bayesian estimates) as this facilitates statistically sound testing of different hypotheses for comparing age estimates from different studies, between molecular dates and the fossil record, and especially between geological events and molecular dates (Crisp et al., 2011). Third, much of the evidence presented by Wilf & Escapa (2015) in support of their argument of a ‘megabiased clock’ is flawed. In fact, for 19 plant lineages, they compared fossil age with mean molecular age (rather than credible interval). The credible interval of estimated molecular age was compatible with the oldest fossils of the respective groups (Todea, Acompyle, Athrotaxis, Papuacedrus, Bailiancarpus, and Ripogonum), whereas a younger bias existed for estimated molecular dates relative to fossil age in nine cases (Agathis, Dacrycarpus, Tripylocarpa, Monimiophyllum, Raiguenrayun, Gymnostoma, Paracacioxylon, Eucalyptus, Orites). In the nine latter cases, however, older fossils of these groups were reported after those molecular dating studies had been published. These examples of ‘megabiased clocks’ will be invalid if these fossils are simply integrated into a Bayesian molecular dating analysis as minimum age calibrations. Concerning this point, we consider that, because the fossil record is constantly updated, it makes sense that molecular dates whose estimation relies on fossil data should be updated when new fossils become available. It does not seem justified to use up-to-date fossil data to criticize molecular dates, when the latter were based on the less complete fossil record that was available at the time of node age estimation. A further point is that informative traits of fossils are frequently missing or ambiguous, and convergent evolution among extinct and living lineages is common. Therefore, inappropriate age constraints related to the phylogenetic placement and evolutionary meaning of controversial fossils must also be considered as a source of any discrepancies between molecular dates and fossil ages. In this respect, exploring the effect of ‘outlier’ (wrongly-placed) fossil calibrations is a crucial step in generating reliable estimates of the diversification timescale of organisms (Warnock et al., 2015). We argue that the four cases (Kurtziana spp. and three Araucaria spp.) cited in Wilf & Escapa (2015) where presumed fossil calibrations are significantly older than molecular dates may represent a set of ‘outlier’ fossil calibrations. In Araucariaceae, empirical studies have suggested that adoption of ‘outlier’ fossil calibrations (e.g. assigning stem fossils to crown nodes) will lead to unrealistic age estimates for other conifer lineages (Kranitz et al., 2014) or a significant change in the evolutionary rate between stem and crown Araucariaceae lineages (Biffin et al., 2010). Meanwhile, there are also abundant examples where molecular dates are older than the unequivocal fossil record, such as the crown age of angiosperms (Smith et al., 2010; Zeng et al., 2014), land plants (Smith et al., 2010; Magallón et al., 2013), and the crown age of mammals (Meredith et al., 2011; Dos Reis et al., 2014a), among others. Therefore, caution should be employed when integrating ambiguous fossils as minimum age constraints, or integrating unambiguous fossils as maximum age constraints into molecular dating (see also later). Generally speaking, potential errors (or biases) may be introduced into molecular dating in nearly all steps of the process, and it would be helpful to identify these errors and to propose strategies to reduce them. What then are the key factors likely to introduce age estimation errors into mainstream molecular dating approaches (‘node dating’, see later), such as Bayesian relaxed clocks (as implemented in the programs BEAST, MCMCTree and Multidivtime)? These factors can be divided into three main categories: the estimation of branch length and phylogenetic topology, clock models and priors, and most importantly, the uncertainty of fossil calibration (Inoue et al., 2010; Dos Reis & Yang, 2013). Until now, most molecular dating studies did not follow a systematic protocol to justify their fossil calibration scheme. As pointed out by Wilf & Escapa (2015) and others (Sauquet et al., 2012), molecular dating is strongly dependent on the selection and placement of fossils in phylogenetic trees in order to calibrate the timescale of diversification. A problem arises with the selection of relatively ancient fossils in that traits that might link ancient fossils to living taxa are often likely to be absent or ambiguous. By contrast, it is easier to select younger fossils linked to extant lineages because better identification of synapomorphies or higher resolution of phylogenetic placement in the ‘total evidence’ tree are more likely for younger fossils. Sauquet et al. (2012) emphasized this point by describing the use of ancient and more recent fossils as ‘early but risky’ and ‘safe but late’, respectively, and further demonstrated that selecting younger fossils for molecular dating often leads to underestimation of the dates of lineage divergence. Wilf & Escapa (2015) further emphasized that the conservative placement of fossils in a phylogenetic tree, such as at the stem rather than crown node, is likely to introduce a younger bias into molecular dating. In this respect, however, we argue that this is a two-way logic, if fossilized taxa cannot be assigned to crown nodes confidently, then they most likely represent extinct relatives of the stem lineage, in which case radical placement of these fossils on crown nodes may potentially introduce overestimation biases. Moreover, since each calibration point tends to have a greater influence on the age estimates of the closest nodes in a phylogeny, a deficiency in number and coverage of calibration points usually leads to younger estimates of node ages in other parts of the phylogeny (Duchêne et al., 2014). To reduce these errors in ‘node dating’, we strongly recommend that when choosing and placing fossils for calibration purposes, a strict and systematic protocol is followed, as outlined by Parham et al. (2012). This widely accepted protocol aims to justify why a certain set of fossils is employed for the calibration points, and to facilitate replication and assessment of age estimation results in the future. Turning to other fossil calibration issues that can cause errors when commonly used models are employed for molecular dating (e.g. uncorrelated lognormal and exponential relaxed-clock models in BEAST), inappropriate parameter settings may be common, especially with regard to the uncertainty of ages of fossils that were adopted for calibration. Usually, this parameter is integrated into the analysis as a form of prior probability distribution (henceforth simply referred to as a ‘prior’) for a particular ancestral node age (see Ho & Phillips, 2009). Inappropriate setting of this parameter can potentially lead to biased estimates of node ages (Ho & Phillips, 2009; Heath, 2012). For example, use of an exponential prior or a lognormal prior assumes a rapidly declining probability of older ages, which may introduce a ‘younger bias’ to dating lineage divergence if the true node age is significantly older than the major part of the probability distribution for this prior (Clarke et al., 2011). It has been proposed that many phylogenetic studies that reject the link between the Gondwana breakup and lineage divergence in the Southern Hemisphere possibly underestimate dates of lineage divergence only because particular calibrating node age priors have been inappropriately used in the Bayesian dating process (Heads, 2012). In addition to setting calibration priors, interactions between calibration priors at different nodes, and between the calibration prior and other priors in the clock model (such as tree priors) may also affect age estimations in Bayesian relaxed-clock dating (Inoue et al., 2010; Heled & Drummond, 2012; Warnock et al., 2015). To overcome the problem described earlier, we recommend a strategy for assigning the appropriate calibration prior (for a summary of calibration priors or densities, see Ho & Phillips, 2009) to specific nodes (Warnock et al., 2012). First, a minimum constraint (uniform prior) should be the first choice for all relatively shallow calibration nodes; if other priors are applied, justification should be provided based on a thorough survey of related fossils, and Bayesian priors should be evaluated with appropriate approaches (Nowak et al., 2013; Norris et al., 2015). For deeper nodes, choosing a conservative (hard) maximum constraint prior may lead to a ‘younger bias’, and therefore in most cases the use of a gamma hyperprior (Heath, 2012), a lognormal prior or a soft maximum constraint is preferable (Warnock et al., 2012). In addition, it would always be helpful to compare and discuss age estimates from different calibration scenarios (Mao et al., 2012; Sauquet et al., 2012). Molecular dating results may differ between different clock models (Duchêne et al., 2014) and between different prior settings in the same model with the same set of fossil calibration points (Condamine et al., 2015). Various molecular clock models relating to the heterogeneity of rate among lineages are now available, such as the strict clock, local multirate clock, discrete multirate clock, autocorrelated relaxed clock and uncorrelated relaxed clock, and empirical studies suggest that different clock models suit different situations (for a review, see Ho & Duchene, 2014). Empirical studies suggest that uncorrelated rate-relaxed clocks (as implemented in BEAST) usually produce older age estimates than penalized likelihood relaxed clocks (as implemented in R8S) (Mao et al., 2010, 2012). In the same clock model, for example, the uncorrelated relaxed clock, a recent empirical study demonstrated that adoption of two branching process priors (Yule vs birth–death priors) resulted in strikingly different diversification timescales, with mean age estimates differing by a factor of three (Condamine et al., 2015). We therefore recommend that sufficient justification should be provided when certain clock models and priors are selected, for example, according to likelihood-based or Bayesian-based criteria (Drummond & Suchard, 2010; Paradis, 2013; Ho & Duchene, 2014). The estimation of branch length and phylogenetic topology may introduce errors into molecular dating analyses, but in recent years the increased feasibility of obtaining genome-scale data has progressively improved our ability to establish phylogenetic topology and genetic distance (Yang & Rannala, 2012). Nevertheless, challenges remain about how to integrate genomic data efficiently into molecular dating approaches (Ho, 2014). In one empirical example, when many loci were analyzed in molecular dating based on genome-scale data, the traditional oversimplified substitution rate prior may have dominated posterior age estimation, much more so than the Dirichlet prior (Dos Reis et al., 2014b). Molecular dating approaches are constantly being updated (Ho, 2014; Ho & Duchene, 2014) and recent developments in Bayesian dating methods, for example, ‘total evidence’ (Ronquist et al., 2012a) and ‘fossilized birth–death’ (Heath et al., 2014) methods, may lead to further improvements in the accuracy of dating lineage diversification. In comparison to the traditional Bayesian methods of ‘node dating’ within a molecular phylogeny (e.g. Bayesian relaxed clocks that were implemented in the programs BEAST, MCMCTree, Multidivtime), ‘total evidence’ methods incorporate fossilized taxa into a molecular-and-morphological phylogeny as an extinct side branch (Pyron, 2011; Ronquist et al., 2012a,b; Grimm et al., 2015), while the ‘fossilized birth–death’ method considers living and fossil taxa together as part of the same macro-evolutionary process covering speciation, extinction and fossilization rates (Heath et al., 2014). It is crucial to note that the ‘fossilized birth–death’ approach integrates all available fossilized taxa, and therefore avoids potential biased selection and placement of fossil calibrations (Heath et al., 2014; Grimm et al., 2015). There is no need for prior age densities to be applied to fossils as is the case for ‘node dating’ methods, nor is a morphological data matrix required as for ‘total evidence’ dating methods (Heath et al., 2014; Grimm et al., 2015). Hence, ‘fossilized birth–death’ is completely different from ‘node dating’ and ‘total evidence’ dating methods, both of which employ only a selected portion of fossils for calibration. These new developments might lead to improved accuracy of molecular dating, although their merits and shortcomings still require further evaluation (Arcila et al., 2015; Grimm et al., 2015). Time will tell whether their use will help to overcome some of the problems inherent in traditional Bayesian methods of ‘node dating’. Although suffering from a number of possible error-introducing factors (see Table 1), molecular dating approaches provide invaluable insights into the diversification and biogeographic histories of plants. Molecular-dated timescales of lineage divergence are a reliable complement to timescales based solely on fossils, if molecular dating is conducted carefully and age estimates are interpreted cautiously (Sauquet et al., 2012; Sauquet, 2013; Hipsley & Müller, 2014). The alleged ‘megabiased clock’ issue identified by Wilf & Escapa (2015) referred to 19 cases where molecular dates are younger than corresponding fossil ages; these were mainly drawn from comparisons between outdated molecular dates and up-to-date fossil records. Except for a small number of controversial cases, molecular dates of most plant lineages will be compatible with fossil records by integrating up-to-date fossil calibrations. Nevertheless, Wilf & Escapa (2015) were right to draw attention to the potential younger biases in traditional ‘node dating’ approaches that were introduced via unjustified but routine fossil calibrating protocols (a list of ways to reduce these potential biases is provided in Table 1). Moreover, their systematic identification of key fossil taxa supporting the Gondwana vicariance hypothesis rather than the Green Web hypothesis for plant lineage diversification in the Southern Hemisphere is an important contribution to research on this topic. However, it is important not to ‘throw the baby out with the bathwater’ and dismiss the value of the molecular dating approach. It is our belief that closer cooperation between paleontologists and molecular phylogeneticists will steadily refine our understanding of the temporal framework of diversification and biogeographic history in both the Southern and Northern Hemispheres of the Earth. Most likely S1 > S2 The authors thank Drs Richard J. Abbott, Richard I. Milne and Jian-Quan Liu for their encouragement and insightful comments on earlier versions of the manuscript. The authors appreciate the constructive suggestions from Drs Susanne S. Renner, H. Peter Linder and two anonymous reviewers. The authors are supported by funds from the National Basic Research Program of China (grant 2014CB954100), the National Natural Science Foundation of China (grants 31370261, 31100488), Sichuan Provincial Department of Science and Technology (grant 2015JQ0018) and Sichuan University. K-S.M. planned and designed the research. K-S.M. and Q.W. collected, compared and summarized data in previous publications. Q.W. and K-S.M. wrote the manuscript.