What is the best strategy for building gene trees and species trees?

Originally presented at the ISMB 2026 conference in Washington D.C., USA.



Many methods have been proposed for phylogenetic reconstruction, yet it remains unclear which approach is most appropriate for empirical data. Recent studies quantified gene tree accuracy using gene-to-species tree discordance and have unexpectedly shown that computationally intensive approaches, such as maximum likelihood and Bayesian inference, are less accurate than faster distance-based methods. Here, we introduce Treeline, a unified optimization framework for phylogenetic inference that facilitates comparison of trees optimized under alternative criteria. Treeline employs a novel strategy for exploring tree space based on structured perturbation of patristic distances, allowing optimization of balanced minimum evolution, maximum likelihood, and maximum parsimony objectives within a single framework. Benchmarking on large empirical datasets shows that Treeline often achieves accuracy, runtime, or memory efficiency comparable to programs designed for individual optimization objectives. Using Treeline, we confirm that the balanced minimum evolution criterion produces gene trees with greater concordance to species trees than maximum likelihood. Moreover, we show that maximum parsimony is competitive with maximum likelihood when biologically realistic, non-uniform cost matrices are applied. These advantages extend to species tree inference, where minimum evolution gene trees yield increased species tree consistency across diverse empirical datasets. Together, our results support the growing evidence that distance-based methods offer superior performance for gene and species tree inference and challenge the prevailing assumption that greater computational complexity necessarily leads to more accurate phylogenetic trees.