Attempting to estimate divergence dates on molecular phylogenies is a messy affair. Bayesian methods (for example, BEAST and MrBayes) have the advantage of being able to accommodate a lot of unknowns and still often are able to generate distributions that seem to contain the true likely divergence date of interest. Three classes of information can be included in a Bayesian analysis of divergence dating to calibrate the phylogeny: molecular rates, biogeographic information, and fossils. Fossils tend to be thought of as the gold standard among these three – researchers take the date for the oldest known fossil from their group of interest and apply it as a calibration on a node in their phylogeny. However, this presents a few issues. First, fossils are often imperfectly preserved, making them difficult to place taxonomically. Second, it can be difficult to come up with appropriate prior distributions to apply in the dating machinery. In addition, in current dating methods, researchers have to assume that they know for certain on which group’s node to place a fossil calibration and that the group is monophyletic – both of these can be strong assumptions.
In a new paper by Ronquist et al (2012), researchers present a new approach for dealing with these issues, using MrBayes 3.2. Instead of placing a calibration on a node in the phylogeny, the fossil is included as a species in the data matrix. How it works is that extant species in the data set have molecular and morphological characters included, while the extinct species are coded for all the morphological characters that can be assigned (usually a subset of all of the characters because it is often not possible to see all of the characters on a fossil). Rather than calibrating nodes of groups the user defines, the user can enter the dates of the strata the fossils came from as a calibration on the taxa itself. The Mr.Bayes machinery then takes this information and estimates a phylogenetic tree, placing the fossils alongside the extant taxa based on their morphological characters and using their dates to estimate divergence times for the rest of the tree. This approach is neat because it removes error associated with improperly classifying fossils and having to restrict the node it is placed on to be monophyletic.
Ronquist et al (2012) test the method on a data set of the early Hymenoptera radiation that includes 45 fossils, many of which are poorly preserved, and also run a comparative analysis using the more traditional node-dating technique. One issue they had to deal with in the development of this approach was that they did not feel that existing tree priors (e.g. birth-death, Yule) could be reasonably applied to their data set. So, they developed and described a new, uninformative tree prior that allows the tree to have terminals of different ages, which allows the branch length information to come from the data.
Because this was a proof-of-concept paper, they did extensive exploration of the method, sensitivity analyses, and took a lot of care with selecting their prior distributions. They first ran an uncalibrated analysis of the data, from which they could detect significant rate variation among the lineages. Because of this, they did not want to use a strict clock for the analysis (which assumes rates are the same along all of the branches in the tree). Instead, they wanted to allow for the rates to vary among the branches, so used a relaxed clock approach. There are several relaxed clock models available that model how rates vary in very different ways. As it was not clear which model fit their data best, they ran the analysis with three different models (two autocorrelated models and one uncorrelated model) and compared them using Bayes Factors. In addition to helping them identify the best model for their data, this allowed them to showcase one of the many new features in MrBayes 3.2, calculation of the marginal likelihoods using stepping-stone sampling. This is a major advance for MrBayes users, allowing them to do model comparisons and hypothesis testing without having to use the harmonic mean estimator of the likelihood, which is known to produce biased estimates. Plus, it was a really nice description for users of a method for going about accounting for rate variation for their own data sets.
In their final comparison of the Total-Evidence approach versus the node dating approach, Ronquist et al found several interesting things. First, they found that their method produced a tree that compared well topologically and in its estimates of divergence times with previous studies. They also found that their method is both more precise (smaller error bars) and less sensitive to prior choice than the node dating approach. Finally, the Total-Evidence analysis produced posterior probability distributions of less than 50% for over half of the fossils used as node calibrations in the node dating approach (even though the authors had sub-selected only the best-understood fossils in the group), indicating that their placement in the tree is highly uncertain, and thus indicating that specifying which node they belong to is inappropriate.
Ronquist et al’s findings suggest that this is a method worth continuing to explore. Carefully selecting models and priors and running sensitivity analyses will be important for users as they begin testing this on new data sets. The authors provide their nexus file with all of the blocks of commands as supplementary information, so you can see for yourself how these elegant analyses are set up.
Kari Goodman
Ronquist, F., Kloppstein, S., Vilhelmsen, L, Schulmeister, S., Murray, D.L. and A.P. Rasnitsyn 2012. A Total-Evidence Approach to Dating with Fossils, Applied to the Early Radiation of the Hymenoptera. Systematic Biology 61 (6): 973-999.
In a new paper by Ronquist et al (2012), researchers present a new approach for dealing with these issues, using MrBayes 3.2. Instead of placing a calibration on a node in the phylogeny, the fossil is included as a species in the data matrix. How it works is that extant species in the data set have molecular and morphological characters included, while the extinct species are coded for all the morphological characters that can be assigned (usually a subset of all of the characters because it is often not possible to see all of the characters on a fossil). Rather than calibrating nodes of groups the user defines, the user can enter the dates of the strata the fossils came from as a calibration on the taxa itself. The Mr.Bayes machinery then takes this information and estimates a phylogenetic tree, placing the fossils alongside the extant taxa based on their morphological characters and using their dates to estimate divergence times for the rest of the tree. This approach is neat because it removes error associated with improperly classifying fossils and having to restrict the node it is placed on to be monophyletic.
Ronquist et al (2012) test the method on a data set of the early Hymenoptera radiation that includes 45 fossils, many of which are poorly preserved, and also run a comparative analysis using the more traditional node-dating technique. One issue they had to deal with in the development of this approach was that they did not feel that existing tree priors (e.g. birth-death, Yule) could be reasonably applied to their data set. So, they developed and described a new, uninformative tree prior that allows the tree to have terminals of different ages, which allows the branch length information to come from the data.
Because this was a proof-of-concept paper, they did extensive exploration of the method, sensitivity analyses, and took a lot of care with selecting their prior distributions. They first ran an uncalibrated analysis of the data, from which they could detect significant rate variation among the lineages. Because of this, they did not want to use a strict clock for the analysis (which assumes rates are the same along all of the branches in the tree). Instead, they wanted to allow for the rates to vary among the branches, so used a relaxed clock approach. There are several relaxed clock models available that model how rates vary in very different ways. As it was not clear which model fit their data best, they ran the analysis with three different models (two autocorrelated models and one uncorrelated model) and compared them using Bayes Factors. In addition to helping them identify the best model for their data, this allowed them to showcase one of the many new features in MrBayes 3.2, calculation of the marginal likelihoods using stepping-stone sampling. This is a major advance for MrBayes users, allowing them to do model comparisons and hypothesis testing without having to use the harmonic mean estimator of the likelihood, which is known to produce biased estimates. Plus, it was a really nice description for users of a method for going about accounting for rate variation for their own data sets.
In their final comparison of the Total-Evidence approach versus the node dating approach, Ronquist et al found several interesting things. First, they found that their method produced a tree that compared well topologically and in its estimates of divergence times with previous studies. They also found that their method is both more precise (smaller error bars) and less sensitive to prior choice than the node dating approach. Finally, the Total-Evidence analysis produced posterior probability distributions of less than 50% for over half of the fossils used as node calibrations in the node dating approach (even though the authors had sub-selected only the best-understood fossils in the group), indicating that their placement in the tree is highly uncertain, and thus indicating that specifying which node they belong to is inappropriate.
Ronquist et al’s findings suggest that this is a method worth continuing to explore. Carefully selecting models and priors and running sensitivity analyses will be important for users as they begin testing this on new data sets. The authors provide their nexus file with all of the blocks of commands as supplementary information, so you can see for yourself how these elegant analyses are set up.
Kari Goodman
Ronquist, F., Kloppstein, S., Vilhelmsen, L, Schulmeister, S., Murray, D.L. and A.P. Rasnitsyn 2012. A Total-Evidence Approach to Dating with Fossils, Applied to the Early Radiation of the Hymenoptera. Systematic Biology 61 (6): 973-999.
RSS Feed
