Why we should not abandon model selection in phylogeny reconstruction

A recent paper in Nature Communications (Abadi et al. 2019) investigated model selection in phylogeny reconstruction. Selecting an appropriate model of nucleotide substitution is considered best practice in phylogenetics, and indeed many studies have show that accurate modelling of substitution processes can substantially improve phylogenetic estimates. The authors quite surprisingly find that this practice may not be necessary after all. From multiple datasets of diverse simulated sequences, they find that the models chosen by commonly used criteria do not perform better than the most complex model. They conclude that cases model selection can be skipped altogether, and all phylogenetic inferences be performed with a complex model.

UPDATE (November 25, 2019): What is discussed here, together with some more in depth analyses, are now available as preprint here.

There are other interesting aspects of this paper, but I here want to focus on the claim (and that data that's used to support it) that model selection does not improve phylogenetic inference. Below, I will argue that

• the data shown in the paper are a bit more ambiguous than the authors acknowledge and do not really support this particular conclusion;
• upon closer inspection, the dataset used in the study does show that model choice makes a difference for inferring topologies;
• model choice is not as computationally expensive as claimed, and may even be cheaper than fitting overly complex models.

Read More