Our previously published results showed how well weighbor performs on small, simple trees. Here we show how well weighbor 1.2 performs on randomly generated 50-taxon trees.

Methods

Trees were generated using evolver from Ziheng Yang's PAML 3.0c package. Parameters were chosen as 6.7 for species birth, 2.5 for species death, and 0.3333 for sampling rate, corresponding to Yang and Rannala's estimates for primate speciation [Mol. Biol. Evol. 14:717 (1997)]. The trees were then made to violate the molecular clock by multiplying each branch length by a random number chosen from an exponential distribution. Branches were then rescaled by a constant, so that the longest branch length equaled each value to be plotted. A Jukes-Cantor simulation was done for each such tree using 1000 nucleotide positions. Distances were calculated and input order of the taxa was randomized. Infinite values were replaced by the largest obtainable finite value plus .75 ln(4).

Results

Relative performance was measured by number of correct branches (i.e. bipartitions) found in by one method and not another. In the following plot the small black circles are weighbor 1.2's mean number of missing branches. Other symbols show the median relative performance of other methods. The verticle bars represent the quartile boundaries for performance relative to weighbor 1.2. These are not error bars but rather show the range of performance for different random trees; for 25% of the trees the pefromance was above this range, for 25% it was below.

The plot shows that none of the other methods has median peformance than weighbor 1.2 on sets with longest edges of more than 0.3 replacements per site. Weighbor 1.2 becomes dramatically better than all of these other methods for longest edges of 2.0 to 5.0. In fact, not only is weighbor's mean performance significantly better (by several sigma) in that range; weighbor's score is outside the quartile boundaries, meaning that for more than 75% of random trees, weighbor found more correct branches.

When the longest edge is shorter than about 0.07, parsimony becomes significantly better than the other methods. Weighbor appears to be better than NJ down to about 0.03, below which all the distance methods are about equal. Weighbor 1.0.1-alpha is much worse than weighbor 1.2 for distances greater than 1.

This test was inspired by the work of Miklos Csuros presented at RECOMB 2001. Thanks to Miklos for providing some of the software used in these tests.