Search for a command to run...
Despite recent advances achieved by application of high-performance computing methods and novel algorithmic techniques to maximum likelihood (ML)-based inference programs, the major computational bottleneck still consists in the computation of bootstrap support values. Conducting a probably insufficient number of 100 bootstrap (BS) analyses with current ML programs on large datasets-either with respect to the number of taxa or base pairs-can easily require a month of run time. Therefore, we have developed, implemented, and thoroughly tested rapid bootstrap heuristics in RAxML (Randomized Axelerated Maximum Likelihood) that are more than an order of magnitude faster than current algorithms. These new heuristics can contribute to resolving the computational bottleneck and improve current methodology in phylogenetic analyses. Computational experiments to assess the performance and relative accuracy of these heuristics were conducted on 22 diverse DNA and AA (amino acid), single gene as well as multigene, real-world alignments containing 125 up to 7764 sequences. The standard BS (SBS) and rapid BS (RBS) values drawn on the best-scoring ML tree are highly correlated and show almost identical average support values. The weighted RF (Robinson-Foulds) distance between SBS- and RBS-based consensus trees was smaller than 6% in all cases (average 4%). More importantly, RBS inferences are between 8 and 20 times faster (average 14.73) than SBS analyses with RAxML and between 18 and 495 times faster than BS analyses with competing programs, such as PHYML or GARLI. Moreover, this performance improvement increases with alignment size. Finally, we have set up two freely accessible Web servers for this significantly improved version of RAxML that provide access to the 200-CPU cluster of the Vital-IT unit at the Swiss Institute of Bioinformatics and the 128-CPU cluster of the CIPRES project at the San Diego Supercomputer Center. These Web servers offer the possibility to conduct large-scale phylogenetic inferences to a large part of the community that does not have access to, or the expertise to use, high-performance computing resources.