Bayesian Phylogenetics: Priors, Sampling and Judging Convergence

Classification groups organisms by shared ancestry, with each named rank nested inside a broader one. A phylogenetic tree makes those nested relationships explicit, showing which lineages split from which common ancestors. Bayesian phylogenetics treats the tree itself as an unknown to be estimated with uncertainty. Instead of returning one best tree, it returns a posterior distribution over trees. The posterior probability of a clade is the share of that distribution containing the clade. Beginners often confuse this with a bootstrap value, but the two answer different questions, one about belief given this data and model, the other about stability under resampling.

Every Bayesian analysis combines the likelihood of the data with prior beliefs. The tree prior encodes expectations about speciation and extinction, which matters most for node ages. The clock prior encodes how freely evolutionary rates vary across branches, from a strict shared rate to fully relaxed models. These choices genuinely move results, especially when sequences are short or calibrations are few. A defensible analysis therefore reports its priors openly and tests sensitivity by rerunning with reasonable alternatives. If dates or clades swing wildly, the data are not carrying the conclusion, the priors are.

Trees are hypotheses, so their presentation shapes their interpretation. Root placement decides which group looks ancestral, and branch lengths encode amounts of change or time. Bayesian output adds posterior probabilities as direct statements of support on each branch. Because the posterior is sampled by a Markov chain rather than computed exactly, the sampler must be checked. Trustworthy runs show high effective sample sizes, stable traces without trends, and independent runs that agree on clade posteriors. Chains that fail these checks have not converged, and their numbers should not be quoted.