Molecular phylogenetic studies are making increasing use of partitioned Bayesian analyses via software tools like MrBayes, version 3 (Ronquist and Huelsenbeck, 2003). Data partitioning is important because, as long as the same topology/history underlies all of the parti tions, it addresses some of the problems associated with the combination of data sets with heterogeneous rates (Bull et al., 1993) and eliminates the need to argue the validity of tests that have been used to judge data com binability (e.g., Huelsenbeck et al., 1994; Huelsenbeck and Bull, 1996; Yang, 1996; Cunningham, 1997; Barker and Lutzoni, 2002; Buckley et al., 2002; Dowton and Austin, 2002). In addition, new studies indicate that data partitioning and the use of mixed models often dramatically improve the fit of model to data without the cost of overparameterization (Yang, 1996; Nylan der et al., 2004; Brandley et al., 2005). While applying partitioned models to studies of protein-coding mito chondrial data, we have found that analyses using Mr Bayes may infer overly "long" trees if among-partition rate variation (APRV) is not explicitly accommodated and if data from different partitions evolve at differ ent average rates. We have subsequently confirmed this bias in analyses of simulated datasets, as explained be low. Because the erroneous tree lengths (TLs) can be nearly twice as high as those found when APRV is ex plicitly modeled, this finding should be of interest to those intending to apply previously calculated molec ular clock rates to branch lengths estimated under par titioned models, as well as to those inferring evolution ary rates from fossil-calibrated evolutionary trees. Here we discuss the importance of among-partition rate vari ation and potential pitfalls in implementation of mixed models that accommodate APRV in Bayesian analyses. Note that, for this study, "APRV" refers to rate varia tion across a priori defined partitions, and a "no-APRV" analysis is one that defines partitions but does not explic itly accommodate among-partition rate variation in the model.

中文翻译：

---

分区贝叶斯分析中的准确分支长度估计需要适应分区间速率变化并注意分支长度先验

分子系统发育研究越来越多地通过 MrBayes 等软件工具使用分区贝叶斯分析，版本 3（Ronquist 和 Huelsenbeck，2003）。数据分区很重要，因为只要相同的拓扑/历史构成所有分区的基础，它就可以解决与具有异构速率的数据集组合相关的一些问题（Bull 等人，1993 年），并且无需争论已用于判断数据组合性的测试的有效性（例如 Huelsenbeck 等人，1994 年；Huelsenbeck 和 Bull，1996 年；Yang，1996 年；Cunningham，1997 年；Barker 和 Lutzoni，2002 年；Buckley 等人，2002 年；道顿和奥斯汀，2002 年）。此外，新的研究表明，数据分区和混合模型的使用通常会显着提高模型对数据的拟合，而不会产生过度参数化的成本（Yang，1996年；尼兰德等人，2004 年；Brandley 等人，2005 年）。在将分区模型应用于蛋白质编码线粒体数据的研究时，我们发现如果分区间速率变化 (APRV) 未明确适应并且来自不同分区的数据演化为不同的平均利率。我们随后在模拟数据集的分析中证实了这种偏差，正如解释的那样低。由于错误的树长度 (TL) 几乎是显式建模 APRV 时发现的树长度的两倍，因此对于那些打算将先前计算的分子钟速率应用于分区模型下估计的分支长度的人来说，这一发现应该很有趣，以及那些从化石校准的进化树推断进化速率的人。在这里，我们讨论了分区间速率变化的重要性和在贝叶斯分析中适应 APRV 的混合模型的实施中的潜在缺陷。请注意，对于本研究，“APRV”是指先验定义分区之间的速率变化，而“无 APRV”分析是定义分区但未明确适应模型中分区间速率变化的分析。