Abstract Detail
Recent Topics Posters Charboneau, Joseph [1], Sanderson, Michael [1]. Terraces in phylogenetic tree space caused by missing data: an example from the grasses (Poaceae). Large plant phylogenies utilizing supermatrices with thousands or tens of thousands of taxa often have incomplete taxon coverage for each locus used. This missing data can lead to the problem of terraces in phylogenetic tree space. Trees exist on a terrace in multidimensional tree space if they have the same likelihood (or Bayesian posterior probability) thereby lying on a flat likelihood surface together. In some circumstances hundreds of thousands or millions of trees can exist on a single terrace simply because data is missing from some taxa, and tree search methods must traverse these terraces while attempting to find an optimal tree. We explore methods to reduce terrace size by removing a small number of taxa from a supermatrix using a previously constructed data set consisting of 298 grass (Poaceae) taxa with three loci (rbcL, matK, and trnL-F) and 33% missing data. The maximum likelihood tree for this data set is on a terrace with 61 million other trees, and a previously developed heuristic method reduces the terrace size to one tree by removing 12 taxa from the supermatrix. We attempted to account for uncertainty in the original tree estimated by applying the same heuristic to a set of bootstrap trees. We identified eight problematic taxa that when removed produce a supermatrix with an ML tree on a terrace with 54,000 others. This bootstrapping method to identify and remove problematic taxa may allow tree search methods to better evaluate uncertainty caused by actual sequence data itself rather than missing data. Log in to add this item to your schedule
1 - University of Arizona, Ecology and Evolutionary Biology, Tucson, AZ, 85721, USA
Keywords: phylogenetic methods supermatrix.
Presentation Type: Recent Topics Poster Session: P Location: / Date: Monday, July 28th, 2014 Time: 5:30 PM Number: PRT031 Abstract ID:1269 Candidate for Awards:None |