Bayesian inference of phylogeny: a non-technical primer

commented upon the impact that phylogeny reconstruction has had on systematic botany. We further indicated that we feel that plant systematics is currently in a period of reevaluation of the data we have used as well as the methodology employed to estimate phylogeny. This is not to say that issues such as the use of morphological data versus molecular data or whether or not to combine data for phylogenetic analyses have not been debated in the past. Clearly, these and many other issues have received a considerable amount of debate in the phylogenetic literature. Possibly one of the most rigorously debated topics is the choice of an optimality criterion. In general, there are three basic methods that have been used to estimate phylogeny, including distance, maximum parsimony (MP), and maximum likelihood (ML). The relative merits and shortcomings of these methods have been debated for a number of years (e.g., Faith, 1985; Swofford & Olsen, 1990; Kunhner & Felsenstein, 1994; Huelsenbeck, 1995; Farris & al., 1996; Lewis, 1998; Steel & Penny, 2000), and it is not within the scope of this column to reiterate these discussions. However, it is noteworthy that numerous comparative studies employing both known phylogenies and simulated data have been very useful in determining under what set of conditions each of the methods “out performs” the others. For example, it is now generally accepted that when rates of change along branches vary greatly, employing a parsimony optimality criterion may be misleading due to “long branch attraction” (Felsenstein, 1978; but see Siddall, 1998); whereas additional studies have shown that ML may be inconsistent in other situations, such as when the chosen model of evolution is inappropriate (e.g., Farris, 1999). Simulation studies indicate that distance methods (especially UPGMA) are highly susceptible to variations in evolutionary rates and typically perform more poorly than either MP or ML (e.g., Huelsenbeck & Hillis, 1993). Studies such as these have been important in laying a theoretical foundation for making decisions on how best to estimate phylogeny given the data in hand. However, under most sets of realistic conditions, comparison of ML and MP indicates that these methods perform similarly and often result in highly concordant topologies (e.g., Reed & al., 2002; Kimball & al., 2003). Recently, another round of comparative studies has begun to address a new approach for phylogeny reconstruction (e.g., Suzuki & al., 2002; Wilcox & al., 2002; Alfaro & al., 2003; Douady & al., 2003). This new approach, Bayesian analyses, was proposed in 1996 (Rannala & Yang, 1996; Mau, 1996; Li, 1996) and is now receiving much attention in the literature [e.g., see Systematic Biology 51 (5)]. Several excellent technical reviews have recently been provided by Huelsenbeck & al. (2001, 2002) and Lewis (2001). Although this approach is now a “hot” topic in systematics, Bayesian statistics actually dates back to the 18th century and its utility for reconstructing phylogeny was suggested initially in 1968 by Felsenstein (see Huelsenbeck & al., 2002). It is only recently, however, that these methods have become more widely known and that relevant computer programs have become available. Internet links for downloadable programs for Bayesian analyses (and other methods) are available from the websites of J. Felsenstein (http://evolution.genetics.washington.edu/ phylip/software.html), and P. Lewis (http://lewis. eeb.uconn.edu/lewishome/), or directly from the MrBayes (http://morphbank.ebc.uu.se/mrbayes/; Huelsenbeck & Ronquist, 2001) or BAMBE (http://www. mathcs.duq.edu/larget/bambe.html; Simon & Larget, 1998) websites. A helpful introduction on how to use MrBayes is provided by Hall (2001). Here we attempt to provide a basic introduction to Bayesian approaches to phylogeny reconstruction. In doing so, we point out what we feel are some of the most significant attributes of this new approach. It is important to note that we do not consider ourselves to be experts on this topic, but merely are interested in how this approach differs from other methods and how to implement this methodology into our own research, should we feel it appropriate. Thus, we hope this column will serve as a primer for those of you curious about Bayesian methods.