Positive Selection And Evolutionary Model Selection

Question

[Pappu]

I am trying to identify the amino acids which are under positive selection based on the codon alignment and phylogenetic tree using PAML and SLR. More complex models like M8 or higher seem to have better (closer to 0) log-likelihood compared to M1/M2/M3. Therefore I selected the M8 model.

I also calculated the amino acid residue conservation in each column of the MSA. I expected to obtain dN/dS <1 for conserved sites and dN/dS>1 for the sites which has no conservation. But it turns out that in M8 model the dN/dS is >=1 in rare cases in the completely non-conserved sites. However in M2 all these sites are having dN/dS=1. I am a bit confused as a result.

[Asaf]

I don't know about the models but maybe this paper: http://www.plosgenetics.org/article/info%3Adoi%2F10.1371%2Fjournal.pgen.1004239 will help you in understanding what to expect from dN/dS values.

Answer 1

[Brice Sarver]

More complex (i.e., more parameterized) models will fit the data better and will have better likelihoods by definition. Appropriately choosing from among models requires model-selection approaches. In codeml, compare nested models using the likelihood ratio test. For non-nested models, compare them using information-theoretic approaches, like the AIC or BIC, which penalize likelihoods based on the number of parameters.

To test for selection, codeml assigns codons to site classes. The key is to compare among models that allow for a class with omega > 1 and one that doesn't allow for that. If a model with omega > 1 can be selected as the best-fit model, you may have evidence that some codons are under positive selection.