Samuel Blanquart (CNRS, LIRMM, Montpellier)

We combined the "CAT" mixture model presented by Lartillot and Philippe
(2004) and the non-stationary "BP" model presented by Blanquart and
Lartillot (2006) into a model accounting for variations in the
evolutionary process both along the sequence and across lineages. As in
CAT, the model implements a mixture of distinct Markovian processes of
substitution to be distributed among sites, thus accomodating
site-specific biochemical properties. Furthermore, and as in BP, these
processes are non-stationary, and their equilibrium frequencies are
allowed to change along lineages in a correlated way, through discrete
shifts in global amino-acid composition distributed across the
phylogenetic tree.

Using posterior predictive tests, based on two different test statistics,
we investigate the four models' behaviour concerning their ability to
recover (1) site-specific and (2) taxon specific amino-acid frequencies.
We show that, in contrast to GTR, CAT and BP, CAT-BP is the only model not
to be rejected by either of these two posterior predictive tests on
several data sets, suggesting that both lineage and site-specific
variations have to be accounted for, and that this is done successfully by
the CAT-BP combination.

We further show that the CAT-BP model is able to suppress a phylogenetic
reconstruction artefact, positioning the bees Apis and Melipona among
chelicerates, in an arthropod phylogeny obtained with a concatenation of
mitochondrial protein sequences. Importantly, all other models tested,
GTR, but also CAT and BP taken separately, all display the artefact, thus
indicating that the combination between CAT and BP may result in
synergistic improvements of phylogenetic inference.