How should species phylogenies be inferred from sequence data?

levels as follows: (1) Separate gene trees are inferred from each linkage partition, and (2) the species phylogeny is then inferred from the set of gene trees. A method (Maddison, 1997; Page and Charleston, 1997a, 1997b; Slowinski et al., 1997) termed gene tree parsimony by Slowinski et al. (1997) is the appropriate method for implementing the second step. Gene tree parsimony operates by finding the species tree or trees that minimizes the number of hypothesized gene tree/ species tree conflict-producing events required to fit each gene tree to the species tree(s). Central to gene tree parsimony is the concept of fitting trees to other trees (Page, 1994a). Gene tree parsimony implements Doyle’s (1992) insightful concept that nucleotides are characters of gene trees, whereas gene trees are characters of species trees. An important caveat relates to the serious question raised by Maddison (1997) of just what a species phylogeny is meant to represent. The search for a species phylogeny assumes that such a diagram has some meaning. This is a difficult issue that we leave to other workers. Below, we briefly discuss the sources of conflict between gene and species trees, identify problems with previous methods for inferring species trees from molecular sequence data, define gene tree parsimony, and then illustrate the application of gene tree parsimony by using the computer program GeneTree (Page, 1998), which is free and available at http://taxonomy.zoology.gla.ac.uk/rod/genetree/genetree.html; it requires Mac OS 7.5 or later running on a PowerMac, or an Intel-based PC running Windows 95/NT 4.0 or later. The issues explored in this article bear directly on the controversial question of whether or not There are two levels of potential error in the inference of species phylogenies from molecular sequence data: (1) A gene tree (herein, we use the term gene for any contiguous block of nucleotides, regardless of whether it codes for a protein or not) for a set of molecular sequences will be incorrectly inferred if there is sufficient random or systematic error (Swofford et al., 1996), and (2) even if a gene tree is correctly inferred, the phenomena of deep coalescence, gene duplication, and horizontal gene transfer can produce a gene tree different from the true species tree (Goodman et al., 1979; Avise et al., 1983; Pamilo and Nei, 1988; Doyle, 1992; Maddison, 1996, 1997). The second level of error would be quite worrisome if all nucleotides of genomes were historically linked (as is generally true with organellar genomes). In this situation, there would only be one gene tree, which might not be the same as the true species tree. But happily, because of intraand interchromosomal recombination, the nuclear genome is composed of many historically linked sets of nucleotides with different histories. We call these linkage partitions (the cgenes of Doyle, 1992, 1997). Sequences sampled from several species and forming a single linkage partition are related by a unique, binary gene tree. Contrary to claims that natural data partitions do not exist (e.g., Kluge and Wolf, 1993; Siddall, 1997), linkage partitions are natural partitions of molecular sequence data and can be considered as independent estimators of the overlying species phylogeny. This strongly suggests that the molecular phylogenetic analysis of species by using 814 SYSTEMATIC BIOLOGY VOL. 48