The biasing effect of compositional heterogeneity on phylogenetic estimates may be underestimated.

The effect of compositional heterogeneity in sequence data on phylogenetic inference was first identified as a potential problem in the late 1980s and early 1990s (Chang and Campbell, 2000; Conant and Lewis, 2001; Foster and Hickey, 1999; Hasegawa et al., 1993; Klenk et al., 1994; Lockhart et al., 1992a, 1992b; Loomis and Smith, 1990; Olsen and Woese, 1993; Penny et al., 1990; Sogin et al., 1993; Tarrio et al., 2001; Van Den Bussche et al., 1998; Weisburg et al., 1989), and by 1993 the first methods had been developed to measure the extent of the problem (Lockhart et al., 1993, 1994; Steel et al., 1993, 1995) or to overcome it (Foster, 2004; Galtier and Gouy, 1995, 1998; Galtier et al., 1999; Gu and Li, 1996, 1998; Lake, 1994; Steel, 1994; Steel et al., 1993, 1995; Tamura and Kumar, 2002; Yang and Roberts, 1995). It is now widely accepted that compositional heterogeneity in aligned sequence data can mislead methods commonly used to infer phylogenetic trees, but it is still unclear (i) why phylogenetic studies based on the LogDet (or paralinear) distance (Lockhart et al., 1994; Steel, 1994) sometimes fail to recover the expected tree topology from compositionally heterogeneous alignments (e.g., Foster and Hickey, 1999; Tarrio et al., 2001), and (ii) how much compositional convergence is necessary before the phylogenetic methods fail to recover the correct topology. Using Monte Carlo simulations to address the second point, Conant and Lewis (2001) concluded that “rather extreme amounts of convergence are necessary before parsimony begins to prefer the incorrect tree.” Other simulation studies have reached similar conclusions (e.g., Galtier and Gouy, 1995; Rosenberg and Kumar, 2003; Van Den Bussche et al., 1998). Based on the study by Galtier and Gouy (1995), it would appear that it is safe to use DNA for phylogenetic inference as long as the difference in GC content is less than 8% to 10%. This im-