How and Why Overcome the Impediments to Resolution: Lessons from rhinolophid and hipposiderid Bats

Foley, Nicole M.; Thong, Vu Dinh; Soisook, Pipat; Goodman, Steven M.; Armstrong, Kyle N.; Jacobs, David S.; Puechmaille, Sébastien J.; Teeling, Emma C.

Phylogenetic analysis was carried out using RAxML (Stamatakis 2006) for ML and BEAST v1.7.2. , MrBayes 3.2.1 (Ronquist et al. 2012) , and PhyloBayes 3.3 incorporating the CAT model (Lartillot et al. 2009 ) for Bayesian analysis. The RAxML analysis was carried out via the RAxMLgui v.0.95 (Silvestro and Michalak 2012) using the thorough bootstrap method with 500 replicates. The analysis used a partitioned model in which each gene was allowed its own model of sequence evolution. The analysis carried out in BEAST was as described above for single gene trees. BEAST employs a single Markov chain in its analysis, whereas MrBayes uses two independent chains. To highlight any differences or avoid shortcomings in either program arising from these different tree-searching methods, we used both programs as representatives of Bayesian Analysis. The MrBayes analysis was carried out under default settings, and the resulting run was analyzed as above for single-gene trees generated using BEAST. Bayesian analysis was also carried out in PhyloBayes to implement the CAT model of site rate heterogeneity, which is not yet implemented in BEAST or MrBayes. The mixture model options in PhyloBayes for nucleotide data are restricted to variants of either a general time reversible (GTR) or Poisson process of sequence evolution. For this analysis, the more complicated GTR was chosen in combination with the CAT model and a Dirichlet process to describe rate variation across sites. Two simultaneous MCMC chains were run, sampling and checking for convergence every 100 generations. This was done by automatically discarding 20% of the trees from each independent run and comparing, every 100 generations, the resulting 50% majority rule consensus trees until the observed maximum difference in split frequencies between the two chains converged on a value of less than 0.1. To assess further the convergence of the run, the analysis was repeated three times, resulting in the same topology.