Phylogenetic trees evaluating

The number of unique phylogenetic trees increases exponentially with the number of taxa, becoming astronomical even for, say, 50 sequences (Swofford et al., 1996 Li, 1997). In most cases, computational limitations permit exploration of only a small fraction of possible trees. The exact number will depend mainly on the nmnber of taxa, the optimality criterion (e.g., MP is much faster than ML), the parameters (e.g., unweighted MP is much faster than weighted ML with fewer preset parameters is much faster than with more and/or simultaneously optimized parameters), computer hardware, and computer software (some algorithms are faster than others some software allows multiprocessing some software limits the number and kind of trees that can be stored in memory). The search procediue is also affected by data structure poorly resolvable data produce more nearly optimal trees that must be evaluated to find the most optimal. 

The methods described above produce imrooted trees (i.e., trees having no evolutionary polarity). To evaluate evolutionary hypotheses, it is often necessary to locate the root of the tree. Rooting phylogenetic trees is not a trivial problem (Nixon and Carpenter, 1993). 

Bootstrapping is a resampling tree evaluation method that works with distance, parsimony, likelihood, and just about any other tree derivation method. It was invented in 1979 (Efron, 1979) and introduced as a tree evaluation method in phylogenetic analysis by Felsenstein (1985). The result of bootstrap analysis is typically a number associated with a particular branch in the phylogenetic tree that gives the proportion of bootstrap replicates that supports the monophyly of the clade. 

Application of the test to alternative phylogenetic trees is problematic, especially because of the irregularity of [the] parameter space (Yang et al., 1995), but its use has been advocated for evaluating optimality of the substitution model when the number of parameters between models is known. 

Sometimes attempts have been made to rank methods according to the probability of them giving the right result from simulated data. This is an old canard in statistical inference, and in connection with phylogenetic trees I tried to shoot it down in a letter to Science (Edwards, 1995). The approach confuses before-trial evaluation methods that are usually right and after-trial evaluation methods that are credible on each occasion of use. I can do no better than quote the last paragraph of my letter ... 

Relationship between geological time and the events of phylogenetic trees constructed from protein sequences have been evaluated, assuming that divergence between fish and man represents about 400 million years. Each... 

Several procedures are available that evaluate the phylogenetic signal in the data and the robustness of trees (Swofford et al., 1996 Li, 1997). The most popular of the former class are tests of data signal versus randomized data (skewness and permutation tests). The latter class includes tests of tree support from resampling of observed data (nonparametric bootstrap). The likelihood ratio test provides a means of evaluating both the substitution model and the tree. 

Since our phylogenetic analysis system permits the predetermination of all or part of the tree structure, we are able to evaluate alternate, sub-optimal topologies. In such an experiment with the tRNA data, we have placed the Bacillus branch (with Mycoplasma) on the principal bacterial root and put the mitochondria under the constraint that they are monophyletic. With these two constraints, the minimized topoloav of Fia. 4 was... 

The central tenets of the falsificationist philosophy of Karl R. Popper are reviewed in detail, and the way they do or do not apply to systematics and phylogeny reconstruction is analyzed. Cladistic analysis, cast in either maximum parsimony or in maximum likelihood approaches, is not compatible with Popperian falsificationism. The main reasons are the absence of a deductive link between a hypothesis of phylogenetic relationships and character distribution on a tree, which translates into the absence of the basic asymmetry of falsification versus verification. This sets Popper s philosophy of science apart from inductive systems. In cladistic analysis, falsification (disconfirmation) is symmetrical to verification (confirmation), which reveals an inductive and hence probabilistic background. The basic problem of systematics as an empirical science resides in character conceptualization and its critical evaluation. 

---