A Simple Scheme for Mining Approximate Evolutionary Trees From Large Scale Data Sets

1 Determining the evolutionary relationship of a set of DNA or protein sequences is a notoriously important problem which becomes particularly difficult in the analysis of large sets of molecular sequence data. Presently used hierarchical clustering methods are not suited for such large scale data analysis due to their preventive algorithmic complexities. Here we discuss a simple approach, routed insertion, that overcomes this constraint. Sequences are added to a growing tree using the information provided by reconstructed ancestral sequences so that sequences move along internal nodes until a suitable branch is reached. Routed insertion enables the approximate reconstruction of a phylogenetic tree from character state data in quasi-linear time. Moreover, it is based on a specific model of evolution and employs local maximum likelihood and minimum evolution criteria. The method is applied to infer the evolutionary history of 1968 human and chimpanzee mtDNA HVRI/II sequences. 2 Introduction Statistical methods for reconstructing the evolutionary history of a set of DNA se-growing amount of available data it has become evident that the computational complexities of most of the currently used procedures render them inappropriate for large data sets. For example, the simple strategy of enumerating all possible tree topolo-gies in an exhaustive tree search takes exponential time in the number of sequences. Therefore this approach is impractical for larger numbers of sequences. Alternative strategies, e.g., divide-and-conquer procedures like quartet puzzling (Strimmer and von Haeseler, 1996) or DCM (Huson et al., 1998), have been proposed as heuristics to shortcut the time-consuming search for an optimal maximum-likelihood or par-simony tree. However, the computational complexity of these algorithms still scales in comparatively high polynomial orders (e.g., quartet puzzling requires time proportional to Ò where Ò is the number of sequences) which equally excludes very large data sets. Algorithms based on agglomerative hierarchical clustering (Arabie et al., 1996; Day and Edelsbrunner, 1984) that work on pairwise distances instead of character state data, among them the neighbor-joining method (Saitou and Nei, 1987), offer a faster though usually less accurate alternative. Neighbor-joining demands a computational effort proportional to the cube of the number of sequences analysed (Studier and Keppler, 1988). Though this scaling behaviour is sufficient for many applications it still does not meet the requirements for large scale data analysis. Hein has described a procedure (Hein, 1989a; Hein, 1989b) to select suitable parts from a distance matrix to allow the construction of large trees. For the more …