IQ-TREE: A Fast and Effec3ve Stochas3c Algorithm for Es3ma3ng Maximum-Likelihood Phylogenies Lam-Tung Nguyen, Heiko A. Schmidt, Arndt von Haeseler, Bui Quang Minh Mia Schoening November 27, 2018
Background • Phylogene3c inference by maximum likelihood – Es3ma3on of subs3tu3on model parameters, branch lengths, and tree topology • Finding op3mal tree topology is an NP-hard combinatorial op3miza3on problem
Exis3ng Methods • ML tree searches apply local tree rearrangements to improve current tree – Nearest neighbor interchange (NNI) – Subtree pruning and regraOing (SPR) – Tree bisec3on and reconnec3on (TBR) • Only “uphill” moves allowed – Prone to be stuck in local op3ma
Stochas3c Algorithms • Introduced to overcome the problem of local op3ma encountered by hill-climbing algorithms • Allow “downhill” moves or maintain a popula3on of candidate trees to avoid local op3ma • Found not to perform as well as SPR-based hill- climbing algorithms
IQ-TREE • Fast and effec3ve stochas3c algorithm to find ML trees • Perform an efficient sampling of local op3ma in the tree space • Best local op3mum found represents the reported ML tree
NNI Moves
Hill-Climbing NNI • For a given tree, compute the approximate likelihoods of each NNI-tree • Create a list of non-conflic3ng NNIs • Ini3alize the list with the best NNI • Add the next best NNI to the list if it does not conflict with any exis3ng NNI • Repeat un3l all NNIs have been processed
Hill-Climbing NNI • Apply all NNIs to current tree and compute likelihood of resul3ng tree • If worse than that of the best NNI tree, discard all topological modifica3ons except that of best NNI in the list • Replace current tree with new tree with higher likelihood • Conduct reduced NNI search on the new current tree – Only compute NNI trees on inner branches at most two branches away from tagged branches – If list is empty, a locally op3mal tree has been found and the hill-climbing search is finished – If list is not empty, combine the reduced NNI search with the beXer tree as described previously
Discussion • Success of IQ-TREE due to: – Tree search strategy helps to escape local op3ma – Phylogene3c likelihood library reduces the 3me for the likelihood computa3on
Reference Lam-Tung Nguyen, Heiko A. Schmidt, Arndt von Haeseler, Bui Quang Minh; IQ-TREE: A Fast and Effec3ve Stochas3c Algorithm for Es3ma3ng Maximum-Likelihood Phylogenies, Molecular Biology and Evolu2on , Volume 32, Issue 1, 1 January 2015, Pages 268-274, hXps://doi.org/10.1093/molbev/msu300
Recommend
More recommend