Title:Bayes estimators for phylogenetic reconstruction

Abstract: As observed in Holder et al 2008, the popular consensus tree has a natural interpretation as the Bayes estimator which minimizes the expected Robinson--Foulds distance to the unknown true tree. In general, for any loss function which measures dissimilarity between trees, we can define a Bayes estimator which minimizes the expected dissimilarity to the true tree. If tree uncertainty precludes finding the correct phylogeny, we want to choose a tree "close" to the true tree measured in terms of loss functions in Bayes estimators.
As part of an exploratory simulation study, we show that hill climbing can be used to find an empirical Bayes estimator in practice, given a sample of trees from the posterior distribution. In particular we use the square loss derived from the path difference metric described in Steel and Penny, 1993.
In our simulation study, we observed and show the empirical expected loss is a good proxy for the distance to the true tree, better at least than the likelihood score. Hill climbing optimization produced tree estimates which were closer to the true tree, outperforming Neighbor-Joining and ML. In the majority of cases, hill climbing improved expected loss even when the initial tree was obtained by hill climbing optimization of the likelihood. We consider this very encouraging for future work on hill climbing approaches for expected loss minimization.