Phylogeny estimation using mutation and selection

In 1998, Halpern and Bruno proposed a substitution model that combines a mutation process shared across sites with site-specific fitness parameters. The model has not yet been used for phylogeny estimation, mainly for two reasons: its many site-specific parameters are computationally expensive, and they are often poorly determined by the data, which makes inference sensitive to the choice of prior. The fast gradients provided by PhyloGrad now overcome the computational barrier. To address the second problem, we derive a prior from a principal component analysis (PCA) of large datasets. The resulting models enable phylogeny estimation that appears more robust to common sources of error than current models, for example long branch attraction.