Choosing the tree which actually best explains the data: another look at the bootstrap in phylogenetic reconstruction

We consider the problem of phylogenetic reconstruction, which consists in estimating the evolutionary history of a set of species. This unknown history is modelled as a tree and estimated from nucleotide sequences taken from the species’ genome. The rst goal of the estimation is to produce a tree which is structurally as close as possible to the true tree. However, most phylogenetic tree-building methods rely on optimization criteria which lead to infering fully resolved trees, i.e. models of maximal complexity. Thus, such trees usually contain some wrong edges, too specic to the data, i.e., resulting from an overtting eect. We rst introduce a structural goodness-of-t criterion based on quartets of species. Then we describe a tree-building method inferring a fully resolved tree by optimizing this criterion. We present two descending approaches to remove unreliable edges from this tree. The rst one relies on the bootstrap process (Efron, 1979) as introduced in the phylogenetic eld by Felsenstein (1985). The second one is original in this context but analogous to usual methods in model calibration. Simulations show the eciency of both approaches, in that the structural distance between the true tree and the estimated tree is signicantly reduced. c 2000 Elsevier Science B.V. All rights reserved.