A 567-taxon data set for angiosperms: The challenges posed by bayesian analyses of large data sets

Bayesian analyses of a three-gene, 567-taxon (560 angiosperms, seven outgroups) data set revealed the analytical challenges posed by such large data sets. Determining stationarity in Markov chains for such large data sets is difficult. In more than 30 analyses of up to 10 million generations each (with an average run time of 45 d), log-likelihood plots showed that runs can stabilize for several million generations before making jumps in likelihood scores. Simultaneous independent runs reached apparent stationarity as early as 2 million generations and as late as 9.7 million generations, suggesting that (a) 10 million generations are insufficient for data sets of this size and (b) periods of stationarity even as long as 6 million generations should not be taken as an indication that the tree is fully optimized. Our Bayesian analyses recovered a topology highly similar to that found previously with parsimony. However, a few topological differences were found between the Bayesian and shortest parsimony trees obtained for the same data set, the most noteworthy of which is that a clade (posterior probability [pp] = 0.99) of Amborellaceae + Nymphaeaceae is sister to all other extant angiosperms (pp 1: 0) in the Bayesian tree, whereas Amborellaceae alone are sister to all other extant angiosperms with parsimony. Additionally, the Bayesian analysis indicates that the magnoliids and Chloranthaceae are sister to Ceratophyllum and eudicots rather than to monocots, as indicated by the parsimony analyses. Many clades receiving moderate to low jackknife support in parsimony analyses received pp values of 1.0.