NJ, Treefinder and PhyMl gave similar results, in many cases showing clear differences when compared with BI and RaxML trees. As a general trend, the first (NJ, Treefinder and PhyMl) result in some groupings contradicted by widely accepted clades (see examples and a discussion in figure 3 in the electronic supplementary material), mostly showing a tendency to group long branches together. A distance-based method such as NJ was expected to be prone to artefactual groupings, but to our surprise, similar results were found for many PhyMl and Treefinder topologies. Both programs start the heuristic search with a NJ inferred tree that could trap the search in local minima close to the topology of the NJ algorithm, pointing to inefficiency in the heuristic search. Given the worries about artefacts affecting these methods, we relied on RaxML as well as BI results (from all datasets), together with comparison of topologies, to define which clades were robustly recovered in our analyses.
BI and RaxML
phylogenies agreed for all the datasets analysed, recovering the same basic topology across all datasets. No differences were seen in the topologies recovered by BI when covarion was used or when BI estimates were unlinked for the SSU and LSU partitions of the matrix. shows the BI and RaxML
topology obtained from the All-set; the problematic groups are boxed (see §2a(iii)
; further analysed in figure 4 in the electronic supplementary material). Figure 4 in the electronic supplementary material (a
) to (e
) shows the trees obtained when the subsets including only one of the problematic groups were analysed. Finally, another dataset excluding all the problematic taxa (Basic-set, figure 4f
in the electronic supplementary material) was used to test the effect on the support values when all the problematic groups were excluded.
Figure 1 Bayesian and RaxML topology (GTR+Γ+I) for All-set (104 metazoan representatives, Cnidaria as the outgroup). Posterior probabilities (PP) and ML bootstrap values (BV) are indicated with a bullet (PP=1.0 and BV>90%) or a square (PP=1.0 and (more ...)
The overall topology of the tree was consistent between the All-set (), the subsets and Basic-set (figure 4 in the electronic supplementary material). However, most of the nodal supports increased in the subsets and even more so in the Basic-set (except for the Deuterostomia), as shown in . The fact that the supports did not decrease when long branches were removed clearly indicates that high supports in the All-set are not a consequence of a LBA artefact misleading the method. The position of the problematic groups in the All-set tree is consistent with their position in the subset phylogenies (compare with figure 4 in the electronic supplementary material), although again the subsets showed notably higher support for these groups ().
(a) Comparison of topologies
For each dataset, the best tree was statistically compared against alternative trees (). Concerning the subsets, all the alternative topologies tested were significantly worse than the original tree for all the sets with two exceptions: (i) the test based on the Gast-set (hypotheses 9–11) rejects the original polyphyletic Gastrotricha in favour of their monophyly, despite the fact that the former is found in BI and ML trees, and (ii) the hypothesis placing chaetognaths as a sister group to ecdysozoans (hypothesis 13) can not be rejected. The All-set allowed the same alternative hypotheses tested in the previous datasets to be studied, as well as new ones. In general, the All-set allowed the rejection of fewer hypotheses than the previous analysis ().
Topology test results. AU (approximately unbiased test) p-values; in bold, the original tree obtained by BI against which alternative hypotheses were tested.