Systematic Errors In Maximum-Likelihood Tree Inference

Recent improvements of tools for inference of phylogenies from molecular data could not impede that tree topologies estimated from different data sets are often mutually incompatible (see e.g. recent discussions about placement of Porifera in the metazoan tree). There are obviously causes of error that remain undetected. Fundamental discrepancies between molecular phylogenies and morphological data (see e.g. Wagele and Kuck, this book) nourish major doubts concerning the reliability of established methods.A series of simulations with an 11-taxon setup designed to study conditions (branch length ratios, among-site variation, alignment lengths) that are prone to long-branch effects prove that three different classes of artifacts (Wagele and Mayer, 2007) can be discerned. Maximum likelihood inference is more successful if among-site variation is considered with a mixed-distribution model, however reconstruction fails when differences in branch lengths are very large, especially in topologies with short stem lineages. Long inner branches can lead to a systematic error with high support for the wrong tree caused by plesiomorphies even when long alignments and the correct model are used. Other errors are caused by signal erosion and by chance similarities. We present for a set of tree topologies the parameter ranges for which correct topologies can be obtained.