TreeFormer: A transformer-based tree rearrangement operation for phylogenetic reconstruction

Phylogenetic inference is a fundamental problem in biology, which studies the origins and evolutionary relationships among species. Popular phylogenetic inference methods, such as IQ-TREE, RAxML, and PHYML, typically utilize heuristic tree search algorithms to seek a phylogenetic tree that maximizes the likelihood of the observed genetic data. However, tree search is time-consuming and often prone to local optima. To address these issues, we introduce TreeFormer, a new Transformer-based tree rearrangement operation for tree search. Experimental results show that TreeFormer achieves higher accuracy than FastTree 2 when reconstructing trees from real alignments with fewer than 1000 sites.