TriMouNet: An Algorithm for Inferring Level-1 Phylogenetic Networks from Multi-Locus Gene Tree Distributions.

Abstract

With the availability of full genomes of an ever-increasing number of species, many recent phylogenetic analyses have focused on datasets with thousands of loci. In the presence of incomplete lineage sorting (ILS), many gene trees will be discordant with the species tree, which can then be estimated with a supertree method. In a separate stream of research, the reconstruction of phylogenetic networks, which aim to detect and visualize reticulations in addition to the dominant phylogenetic tree, has become common practice. Level-1 networks have the feature that every node can be interpreted as an ancestor of a subset of the taxa of interest and no two distinct cycles share a common node. TriLoNet is a method that constructs a level-1 network on all taxa from 3-taxa networks, so-called trinets. This approach is similar to modern supertree methods, but the trinets are assigned based on a single sequence alignment. Here we present TriMouNet (Trinet Multilocus Network), which uses the tree topology and branch length distribution in gene trees of a multilocus dataset to infer best-fitting trinets, together with scores quantifying their statistical support. These trinets are then puzzled together into a network on all taxa in a TriLoNet fashion. Experiments on simulated and real datasets show that TriMouNet can identify reticulations with low false positive rate, if the gene trees are accurate. On the other hand, TriLoNet applied to the concatenation of all loci, tends to predict wrong reticulations as the consequence of violations of model assumptions.