Partitioned coalescence support reveals biases in species-tree methods and detects gene trees that determine phylogenomic conflicts

Abstract

Genomic datasets sometimes support unconventional or conflicting phylogenetic relationships when different tree-building methods are applied. Coherent interpretations of such results are enabled by partitioning support for controversial relationships among the constituent genes of a phylogenomic dataset. For the supermatrix (= concatenation) approach, several simple methods that measure the distribution of support and conflict among loci were introduced over 15 years ago. More recently, partitioned coalescence support (PCS) was developed for phylogenetic coalescence methods that account for incomplete lineage sorting and use the summed fits of gene trees to estimate the species tree. Here, we automate computation of PCS to permit application of this index to genome-scale matrices that include hundreds of loci. Reanalyses of four phylogenomic datasets for amniotes, land plants, skinks, and angiosperms demonstrate how PCS scores can be used to: 1) compare conflicting results favored by alternative coalescence methods, 2) identify outlier gene trees that have a disproportionate influence on the resolution of contentious relationships, 3) assess the effects of missing data in species-trees analysis, and 4) clarify biases in commonly-implemented coalescence methods and support indices. We show that key phylogenomic conclusions from these analyses often hinge on just a few gene trees and that results can be driven by specific biases of a particular coalescence method and/or the extreme weight placed on gene trees with high taxon sampling. Attributing exceptionally high weight to some gene trees and very low weight to other gene trees counters the basic logic of phylogenomic coalescence analysis; even clades in species trees with high support according to commonly used indices (likelihood-ratio test, bootstrap, Bayesian local posterior probability) can be unstable to the removal of only one or two gene trees with high PCS. Computer simulations cannot adequately describe all of the contingencies and complexities of empirical genetic data. PCS scores complement simulation work by providing specific insights into a particular dataset given the assumptions of the phylogenetic coalescence method that is applied. In combination with standard measures of nodal support, PCS provides a more complete understanding of the overall genomic evidence for contested evolutionary relationships in species trees.

Plastome phylogenomic study of Gentianeae (Gentianaceae): widespread gene tree discordance and its association with evolutionary rate heterogeneity of plastid genes

“…We also computed the partitioned coalescence supports (PCSs) of the 76 PCGs. PCS can be positive, negative, or zero, indicating support, conflict, or ambiguity, respectively [ 43 ]. The results revealed 33 PCGs with positive PCS scores, 23 PCGs with negative PCS scores and 20 PCGs with zero PCS scores (Fig.…”

An empirical assessment of a single family‐wide hybrid capture locus set at multiple evolutionary timescales in Asteraceae

“…As well as biological processes, methodological artifacts create obstacles for phylogenetic reconstruction and can cause inaccurate gene tree estimations (Qu et al., ). Examples of such artifacts include alignment issues and homology errors, such as unrecognized paralogy (Gatesy et al., ) and long‐branch attraction, whereby long branches are erroneously grouped together in estimated trees (Felsenstein, ; Sanderson et al., ; Parks and Goldman, ; Qu et al., ; Mai and Mirarab, ). A number of approaches can help to improve the reliability of concatenation analyses, for example, the use of partitioning and best‐fit substitution models (Xi et al., ; Kainer and Lanfear, ; Lanfear et al., ), elimination of fast‐evolving sites, removal of long branches, or increasing taxon sampling; the latter approach is often challenging due to rare taxa or unknown extinction events (Pisani, ; Bergsten, ; Qu et al., ).…”

Divergence and support among slightly suboptimal likelihood gene trees

“…If so, then this difference is not a concern in the context of downstream coalescent analyses so long as both trees have the same topology. But if there are topological differences between these trees and the best‐reported tree is arbitrarily chosen or arbitrarily resolved then there is reason for concern given that individual outlier gene‐tree topologies can have disproportionate negative effects on two‐step coalescent analyses (Gatesy et al., , ).…”