Large-scale molecular phylogenies have been thought to be non-tractable due to computer limitation (extensive searches), resulting in the need for years of CPU time or a parallel computing environment. Using simulated and real 3-gene data sets for ca. 150 angiosperms, we evaluated various tree-building algorithms. We used minimum evolution (neighbor-joining with corrected distances), no swapping or NNI and TBR swapping with maximum parsimony, and maximum likelihood. We compared tree lengths versus topological distances of individual searches to the best trees (viz. combined tree with real data and model tree in simulations). We found that the simplest tree-building algorithms (e.g. fast swapping) performed equally well or better compared to other methods. These results imply that large-scale phylogenies can be built easily and accurately without extensive computer power. It also explains why incorrect trees are built when not enough data are used, no matter how complex and time-consuming the tree-building methods employed.

Key words: large-scale phylogeny, tree-building methods