Department of Biology, Colorado State University, Fort Collins, CO 80523, USA.
Mol Phylogenet Evol. 2013 Oct;69(1):265-75. doi: 10.1016/j.ympev.2013.06.001. Epub 2013 Jun 10.
Empirical and simulated examples are used to demonstrate an artifact caused by undersampling optimal trees in data matrices that consist mostly or entirely of locally sampled (as opposed to globally, for most or all terminals) characters. The artifact is that unsupported clades consisting entirely of terminals scored for the same locally sampled partition may be resolved and assigned high resampling support-despite their being properly unsupported (i.e., not resolved in the strict consensus of all optimal trees). This artifact occurs despite application of random-addition sequences for stepwise terminal addition. The artifact is not necessarily obviated with thorough conventional branch swapping methods (even tree-bisection-reconnection) when just a single tree is held, as is sometimes implemented in parsimony bootstrap pseudoreplicates, and in every GARLI, PhyML, and RAxML pseudoreplicate and search for the most likely tree for the matrix as a whole. Hence GARLI, RAxML, and PhyML-based likelihood results require extra scrutiny, particularly when they provide high resolution and support for clades that are entirely unsupported by methods that perform more thorough searches, as in most parsimony analyses.
使用实证和模拟示例来说明一个伪影，该伪影是由数据矩阵中最优树的欠采样引起的，这些数据矩阵主要或完全由局部采样（相对于全局，对于大多数或所有末端）字符组成。该伪影是指完全由为同一局部采样分区打分的末端组成的不支持的分支可能会被解决，并被赋予高的重采样支持——尽管它们是正确不支持的（即在所有最优树的严格一致中未解决）。尽管应用了逐步末端添加的随机添加序列，但当仅持有单个树时，该伪影并不一定会被彻底的传统分支交换方法（甚至树二分连接）消除，因为在某些情况下，在简约 bootstrap 伪重复项中，以及在每个 GARLI、PhyML 和 RAxML 伪重复项和搜索整个矩阵最可能的树时，都会执行此操作。因此，基于 GARLI、RAxML 和 PhyML 的似然结果需要额外的审查，特别是当它们为那些完全不被执行更彻底搜索的方法支持的分支提供高分辨率和支持时，就像大多数简约分析一样。
