A large-scale evaluation of tree shape indices reveals potential pitfalls

Abstract

While there exists a plethora of prior research on phylogenetic tree shape indices, a large-scale analysis of the behavior of these indices on empirical trees has not yet been conducted.Here, we address this by computing 54 indices on more than 45,000,000 empirical trees retrieved from the EvoNAPS and RAxML Grove databases. To calculate the indices, we use our novel, comprehensive open-source Python library called treeshapy. The results of our large-scale evaluation indicate that there exist several potential pitfalls when conducting tree shape studies. For $14$ indices we find clear indications, that they are highly sensitive to the position of the tree's root. Only $5$ indices appear stable in that sense, while we observe a medium degree of rooting instability for the remaining $35$ indices under study. Therefore, uncertainties pertaining to the root placement directly affect tree shape values. Furthermore, the values of all except $5$ indices are inherently correlated with tree size, even so, when applying adequate normalization techniques. As a consequence, tree shape values for trees of different sizes should generally not be compared. We further observe that several groups of indices are strongly correlated with one another. Hence, to conduct a representative tree shape study, an appropriate subset of uncorrelated indices should be selected to capture as many aspects of the tree shape as possible while avoiding essentially redundant results at the same time. Our study serves as a guide for conducting tree shape studies in a more cautious and comprehensive manner on empirical data. Apart from our open-source treehshapy Python package, we devise appropriate guidelines for selecting suitable indices and cautiously interpreting respective results.