EnsemblCompara GeneTrees: Complete, duplication-aware phylogenetic trees in vertebrates

The use of phylogenetic trees to describe the evolution of biological processes was established in the 1950s (Hennig 1952) and remains a fundamental approach to understanding the evolution of individual genes through to complete genomes; for example, in the mouse (Mouse Genome Sequencing Consortium 2002), rat (Gibbs et al. 2004), chicken (International Chicken Genome Sequencing Consortium 2004), and monodelphis (Mikkelsen et al. 2007) genome papers, and numerous papers on individual sequences. Now routine, the determination of vertebrate genome sequences provides a rich data source to understand evolution, and using phylogenetic trees of the genes is one of the best ways to organize these data. However, the increased set of genomes makes the compute and engineering tasks to form all the gene trees progressively more complex and harder for individual groups to use. The Ensembl project provides an accurate and consistent protein-coding gene set for all vertebrate genomes (International Human Genome Sequencing Consortium 2001; Dehal et al. 2002; Mouse Genome Sequencing Consortium 2002; Gibbs et al. 2004; Xie et al. 2005; Mikkelsen et al. 2007; Rhesus Macaque Genome Sequencing and Analysis Consortium 2007). Previously (until April 2006), Ensembl provided a basic method for tracing orthologs via the Best Reciprocal BLAST method, similar to approaches used in other genome analyses, such as Drosophila melanogaster (Adams et al. 2000) or human (International Human Genome Sequencing Consortium 2001). In June 2006 (Hubbard et al. 2007), we replaced this system with a phylogenetically sound, gene tree-based approach, providing a complete set of phylogenetic trees spanning 91% of genes across vertebrates. In addition to the vertebrates we have included a few important non-vertebrate species (fly, worm, and yeast) to act both as out groups and provide links to these model organisms. In this paper we provide the motivation, implementation, and benchmarking of this method and document the display and access methods for these trees. There have been a number of methods proposed for routine generation of genomewide orthology descriptions, including Inparanoid (Remm et al. 2001), MSOAR (Fu et al. 2007), OrthoMCL (Li et al. 2003), HomoloGene (Wheeler et al. 2008), TreeFam (Li et al. 2006), PhyOP (Goodstadt and Ponting 2006), and PhiGs (Dehal and Boore 2006). The first four, Inparanoid, MSOAR, OrthoMCL, and HomoloGene, focus on providing clusters (or linked clusters) of genes, without an explicit tree topology. PhyOP (Goodstadt and Ponting 2006) uses a tree-based method, but between pairs of closely related species, resolving paralogs accurately by using neutral substitution (as measured by d S, the synonymous substitution rate). TreeFam provides an explicit gene tree across multiple species, using both d S, d N (nonsynonymous substitution rate), nucleotide and protein distance measures, and the standard species tree to balance duplications vs. deletions to inform the tree construction, using the program TreeBeST (http://treesoft.sourceforge.net/treebest.shtml; L. Heng, A.J. Vilella, E. Birney, and R. Durbin, in prep.). The PhiGs method (Dehal and Boore 2006) is a leading phylogenetic-based method that produced a comprehensive phylogenetic resource for the genomes at the time it was run, and the basic outline of its analysis, which was clustering of protein sequences, followed by phylogenetic trees, is similar to the method presented here. However, the PhiGs resource covered a smaller number of species (23 vs. 45) and has been difficult to keep up to date with the advances in gene sets and genomes. Another major difference between PhiG-based phylogenetic trees and the phylogenetic trees presented here is that the former was calculated using a single maximum likelihood method based on protein evolution. In contrast, the Ensembl gene trees are calculated using a new method, TreeBeST, which integrates multiple tree topologies, in particular both DNA level and protein level models and combines this with a species-tree aware penalization of topologies, which are inconsistent with known species relationships. We show in this paper that this method produces trees that are more consistent with synteny relationships and less anomalous topologies than single protein-based phylogenetic methods. There are also many single phylogenetic tree-building approaches, many of them based on maximum likelihood methods; one leading method is PhyML (Guindon and Gascuel 2003). It is unclear what is the best method to use, in particular in the context of genome-wide tree building with constraints on computational costs and the need to robustly handle many complex scenarios usually involving large families with heterogeneous phylogenetic depths. In this paper, we benchmark in vertebrates the tree programs TreeBeST and PhyML, and the resulting trees to basic best reciprocal hit (BRH) methods, and cluster frameworks, in particular Inparanoid and HomoloGene. We also benchmark to a recent PhyOP data set. The PhyOP pipeline has recently switched to use the same tree-building program (TreeBeST) that we use, but differs in its input clusters. Although we adopted this same tree-building method, we describe here considerable novel engineering in the deployment of these methods across all vertebrates. Similar to the PhiGs resource, we have used the dense coverage of genomes to provide topologically based timings (i.e., the standard use of outgroups vs. subsequent lineages to bracket a duplication), in order to label duplication events.