Parsimony and model-based phylogenomic analyses: a comparison based on empirical datasets

AMBROSIO TORRES; SANTIAGO A. CATALANO; GOLOBOFF, PABLO A.

Buenos Aires

Encuentro; XXXV Meeting of the Willi Hennig Society; 2016

Willi Hennig Society

The advent of the genomic era in recent years has brought the debate about the suitability of differentmethods of tree reconstruction into a new stage. Most of the phylogenomic studies published have usedmodel-based methods, possibly driven by the supposed superiority of those approaches over parsimonymethods in terms of realism of the models considered. Many empirical and theoretical studies havecompared the behavior of parsimony and model-based methods in the past. However, it is not clearhow the results of the different approaches differ in the case of phylogenomic analyses where theamount of data is impressively higher. For example, the assumption of all characters having acorrelated branch lengths (i.e. probability of change along a branch) is more and more likely to beviolated when a data set includes many different genes, which may lead model-based methods furtheraway from statistical consistency than parsimony (which assumes no such correlation). As part of anongoing project for the comparison of parsimony and model-based methods in phylogenomic studies,we collected 71 phylogenomic matrices (from 2006 to 2016) that included more than 20 000 characters(amino acids or nucleotides). From those we selected the datasets that were analyzed by both Bayesianinference (BI) and Maximum Likelihood (ML). In most of the original analyses MP trees were notcalculated, hence we performed the MP searches in TNT. Topological comparisons (SPR-distance andFarris's distortion coefficient) indicated that in 75% of the datasets the topologies found by the differentmethods -MP, ML and BI- were identical or differ in a single SPR move. On average, ML tree differfrom MP in 0.7 SPR-moves more than they differ from BI trees. In those datasets where the resultswere more incongruent, the differences were generally associated to nodes that presented low support.In addition the main conclusions drawn in each of the analyses were not affected by those differences.If these preliminary results are confirmed, and considering the computational efficiency of MP methodsrelative to ML and BI, Parsimony may have an important role in phylogenomics in the future.