Friday, June 27, 2008

Highlights from Evolution 2008: Part II

There were a few notable advances in comparative methods. Emma Goldberg and our own Boris Igic illustrated how several recent rejections of Dollo's law may have resulted from a failure to consider differential rates of species diversification and several other violations of the Mk2 model (see our related previous post on Maddison et al.'s work and the BiSSE method). 

If lineages with the character state A are more likely to speciate, or less likely to go extinct, than lineages with character state B, there will be a tendency to infer an ancestor with state A. Emma and Boris illustrate the practical implications of this sort of differential diversification by showing how it may have resulted in the incorrect inference of re-evolution of wings in stick insects (Whiting et al. 2003; in Nature).  If there is any truth to the widespread belief that vagility is inversely associated with speciation rate, it seems logical to suggest that the non-winged lineages have undergone more species diversification and biased the conclusions of standard methods toward the reconstruction of a non-winged ancestor and repeated re-evolution of wings. 

For those of you who missed the talk, Boris was sporting a mustache to symbolize his allegiance with Dollo.

1 comment:

Brian R. Moore said...

I agree that this is a wonderful and important insight. I believe that this work—in addition to that of Maddison (2006) and Maddison et al. (2007)—highlights the need to develop heterogeneous stochastic process models of morphological evolution. In other words, the assumption that traits can be appropriately modeled under a homogeneous process through time and across lineages is apt to be invalid for many inference scenarios, particularly when the traits under consideration are not selectively neutral.

Imagine, for example, that a group of species exhibiting state A occurs in an environment in which A is advantageous, such that transitions from B->A are apt to occur more frequently than A->B. Further imagine that the environment (and associated selection regime) of lineages outside this subclade do not experience selection for A (or B).

Accordingly, the attempt to model the evolution of this trait (i.e., to infer ancestral states and associated parameter estimates) over the entire tree under a uniform stochastic process will be biased because any such model will necessarily provide a poor fit to part of the data (e.g., where the MK1 assumes symmetric transition probabilities, and the MK2/AsymmK assumes asymmetric rate coefficients).

I believe that there may be a fairly straightforward solution to the problem motivated by this insight. Very cool stuff!