Phylogenomics: the beginning of incongruence?

Until recently, molecular phylogenies based on a single or few orthologous genes often yielded contradictory results. Using multiple genes in a large concatenation was proposed to end these incongruences. Here we show that single-gene phylogenies often produce incongruences, albeit ones lacking statistically significant support. By contrast, the use of different tree reconstruction methods on different partitions of the concatenated supergene leads to well-resolved, but real (i.e. statistically significant) incongruences. Gathering a large amount of data is not sufficient to produce reliable trees, given the current limitation of tree reconstruction methods, especially when the quality of data is poor. We propose that selecting only data that contain minimal nonphylogenetic signals takes full advantage of phylogenomics and markedly reduces incongruence.

Introduction

Molecular characters, primarily DNA and derived protein sequences, provide a wealth of new information that sheds light on many parts of the ‘Tree of Life’. However, molecular phylogenies based on single genes often lead to apparently conflicting results. To overcome this limitation, it is tempting to apply a genome-scale approach to phylogenetic inference (phylogenomics) by combining many genes. The number of published yeast genomes offers the opportunity to test this proposition. Indeed, using 106 genes from yeast genomes, a fully supported phylogeny has been obtained by the analysis of their concatenation 1, 2. Following from this, it has been anticipated that using large amounts of genomic data will mark the end of incongruence in phylogenetics [3].

The incongruence between two phylogenies can be the result of: (i) violations of the orthology assumption generated by mechanisms such as gene duplication, horizontal gene transfer or lineage sorting [4]; (ii) stochastic error related to the length of the genes; and (iii) systematic error leading to tree reconstruction artifacts generated by the presence of a nonphylogenetic signal in the data. Adopting a genome-scale approach theoretically overcomes incongruence because of the first two reasons: nonorthologous comparisons are gene-specific and will probably be buffered in a multigene analysis; and stochastic error naturally vanishes when more and more genes are considered. By contrast, systematic error is not expected to disappear with the addition of data [5]. Systematic error results from nonphylogenetic signals being present in the data, such as heterogeneity of nucleotide compositions among species (compositional signal), rate variation across lineages (rate signal) and also within-site rate variation (heterotachous signal) [6]. The bias causing systematic error creates a signal because, contrary to stochastic noise, it does not average out over several sites. If a bias is strong enough, it can dominate the true phylogenetic signal causing the tree reconstruction method to be inconsistent and lead to an incorrect, but highly supported tree 5, 7. Therefore, phylogenomics, instead of ending incongruence, might open an era of real, statistically significant incongruence resulting from the use of different methods, different taxon samplings, or different character partitions of the same data set.

To illustrate this paradox, we used the large data set of 106 genes (120 762 nucleotides) from 14 yeast species assembled by Rokas and Carroll [1]. Phylogenetic trees were inferred by maximum parsimony (MP) from nucleotide sequences, as in Ref. [1], and alternatively by probabilistic methods – Bayesian inference (BI) [8] or maximum likelihood (ML) 9, 10, 11 – because these methods are generally considered the most accurate 12, 13. In addition, because the diversification of these yeasts is ancient (>250 Mya [14]) and amino acid sequences evolve more slowly than nucleotide sequences, the translated protein sequences were also used to construct trees. Phylogenies were inferred from each of the 106 genes and from their concatenation, using two different methods (MP and BI) and two types of characters (nucleotides and amino acids), yielding a total of 428 trees. We estimated the level of incongruence as the number of bipartitions (or splits, i.e. groups of species defined by a branch of a phylogenetic tree), supported by more than a given bootstrap value, that are different between two trees. Our aim was to compare the level of among-gene incongruence for a given tree reconstruction method with the level of among-method incongruence for a given data set.