Ritchie Andrew M, Lo Nathan, Ho Simon Y W
School of Life and Environmental Sciences, University of Sydney, Sydney, NSW 2006, Australia.
Syst Biol. 2017 May 1;66(3):413-425. doi: 10.1093/sysbio/syw095.
In Bayesian phylogenetic analyses of genetic data, prior probability distributions need to be specified for the model parameters, including the tree. When Bayesian methods are used for molecular dating, available tree priors include those designed for species-level data, such as the pure-birth and birth-death priors, and coalescent-based priors designed for population-level data. However, molecular dating methods are frequently applied to data sets that include multiple individuals across multiple species. Such data sets violate the assumptions of both the speciation and coalescent-based tree priors, making it unclear which should be chosen and whether this choice can affect the estimation of node times. To investigate this problem, we used a simulation approach to produce data sets with different proportions of within- and between-species sampling under the multispecies coalescent model. These data sets were then analyzed under pure-birth, birth-death, constant-size coalescent, and skyline coalescent tree priors. We also explored the ability of Bayesian model testing to select the best-performing priors. We confirmed the applicability of our results to empirical data sets from cetaceans, phocids, and coregonid whitefish. Estimates of node times were generally robust to the choice of tree prior, but some combinations of tree priors and sampling schemes led to large differences in the age estimates. In particular, the pure-birth tree prior frequently led to inaccurate estimates for data sets containing a mixture of inter- and intraspecific sampling, whereas the birth-death and skyline coalescent priors produced stable results across all scenarios. Model testing provided an adequate means of rejecting inappropriate tree priors. Our results suggest that tree priors do not strongly affect Bayesian molecular dating results in most cases, even when severely misspecified. However, the choice of tree prior can be significant for the accuracy of dating results in the case of data sets with mixed inter- and intraspecies sampling. [Bayesian phylogenetic methods; model testing; molecular dating; node time; tree prior.].
在对遗传数据进行贝叶斯系统发育分析时，需要为模型参数指定先验概率分布，包括树。当使用贝叶斯方法进行分子定年时，可用的树先验包括为物种水平数据设计的那些，如纯生和生死先验，以及为种群水平数据设计的基于溯祖的先验。然而，分子定年方法经常应用于包含多个物种的多个个体的数据集。这样的数据集违反了基于物种形成和基于溯祖的树先验的假设，使得不清楚应该选择哪一个，以及这种选择是否会影响节点时间的估计。为了研究这个问题，我们使用模拟方法在多物种溯祖模型下生成具有不同种内和种间抽样比例的数据集。然后在纯生、生死、恒定大小溯祖和天际线溯祖树先验下对这些数据集进行分析。我们还探讨了贝叶斯模型检验选择表现最佳的先验的能力。我们证实了我们的结果适用于来自鲸类、海豹科动物和白鲑科白鱼的经验数据集。节点时间的估计通常对树先验的选择具有鲁棒性，但树先验和抽样方案的某些组合导致年龄估计存在很大差异。特别是，纯生树先验经常导致对包含种间和种内抽样混合的数据集的估计不准确，而生死和天际线溯祖先验在所有情况下都产生稳定的结果。模型检验提供了一种拒绝不适当树先验的适当方法。我们的结果表明，即使在严重错误指定的情况下，树先验在大多数情况下也不会强烈影响贝叶斯分子定年结果。然而，对于包含种间和种内抽样混合的数据集，树先验的选择对于定年结果的准确性可能很重要。[贝叶斯系统发育方法；模型检验；分子定年；节点时间；树先验。]
