Automated removal of noisy data in phylogenomic analyses

Noisy data, especially in combination with misalignment and model misspecification can have an adverse effect on phylogeny reconstruction; however, effective methods to identify such data are few. One particularly important class of noisy data is saturated positions. To avoid potential errors related to saturation in phylogenomic analyses, we present an automated procedure involving the step-wise removal of the most variable positions in a given data set coupled with a stopping criterion derived from correlation analyses of pairwise ML distances calculated from the deleted (saturated) and the remaining (conserved) subsets of the alignment. Through a comparison with existing methods, we demonstrate both the effectiveness of our proposed procedure for identifying noisy data and the effect of the removal of such data using a well-publicized case study involving placental mammals. At the least, our procedure will identify data sets requiring greater data exploration, and we recommend its use to investigate the effect on phylogenetic analyses of removing subsets of variable positions exhibiting weak or no correlation to the rest of the alignment. However, we would argue that this procedure, by identifying and removing noisy data, facilitates the construction of more accurate phylogenies by, for example, ameliorating potential long-branch attraction artefacts.