Hybrid Tree Reconstruction Methods

Daniel Huson
Scott Nettles
Kenneth Rice
Tandy  Warnow
Shibu Yooseph

A major computational problem in Biology is the reconstruction
of evolutionary trees for species sets, and accuracy is measured by
comparing the {\em topology} of the reconstructed tree and the model
tree.  One of the major debates in the field  is whether
large evolutionary trees can be even approximately accurately
reconstructed from biomolecular sequences of realistically
bounded lengths (up to about 2000 nucleotides) using
standard techniques (polynomial time methods and heuristics
for NP-hard optimization problems).  Using both analytical and
experimental techniques, we show that on large trees, the two most
popular methods in systematic biology, neighbor-joining and
maximum parsimony heuristics, as well as two promising methods
introduced by theoretical computer scientists, are all likely to
have significant errors in the topology reconstruction of the
model tree. In particular, we observe that the maximum evolutionary
distance in the tree strongly negatively affects the
accuracy of polynomial time distance methods (including
neighbor-joining), an observation which has not yet appeared
in the systematic biology literature.  Our experimental results
also suggest that the major tree reconstruction methods produce
{\em different  types} of topological errors, suggesting that a
{\em combination} of these methods may have better performance than
the major techniques themselves. We use these observations to propose
a new general technique that combines the output of  existing
methods, thus producing  {\em hybrid methods}.  We study one such
hybrid experimentally, and show that in significant parts of the
parameter space, it achieves results that are equal to or better
than the best of its components.  Our experimental performance study
examined more than a thousand sets of sequences simulated
on more than a hundred different model trees.

Presented at: WAE'98. 1998.