4 weeks ago by
If you're doing the likelihood ratio test to distinguish between models with and without classes where sites can be assigned to groups with ω > 1 (e.g., following a chi-square distribution with the degrees of freedom equal to the difference in the number of free parameters between the models for most, or a mixture that results in a chi-square with one degree of freedom for the M8 vs. M8a comparison), then the p-value is simply providing evidence for selecting one model over the other.
If your best-fit model is one that supports some signature of positive selection (e.g., M8 as opposed to M7), you can then take a look at the results to see which codons have evidence for positive selection under as determined by the Naive Empirical Bayes and/or Bayes Empirical Bayes approaches.
Be aware that the one-rate models may not be the right tests for what you're actually trying to do in some cases. From the manual:
We suggest that The M0-M3 comparison should be used as a test of
variable ω among sites rather than a test of positive selection.
However, the model of a single ω for all sites is probably wrong in
every functional protein, so there is little point of testing.
If you're explicitly looking for other flavors of selection, perhaps across loci, you may find it helpful to explore other approaches. I recommend the Datamonkey Adaptive Evolution Server.