Computational resources. To make thorough maximum-likelihood and boot- strap searches (involving, for example, hundreds of maximum-likelihood search replicates per data matrix) possible under a GTR model, we performed analyses in parallel using grid computing38,42, through The Lattice Project43. Additionally, a dedicated cluster of Linux servers with 40 cores at 2.5 GHz each and 80 GB RAM was set up to allow the computationally very demanding codon-model analyses. Additional analyses were carried out using the CIPRES cluster at the San Diego Supercomputing Center, and on the Duke Shared Resource Cluster.