Completed on 1 May 2017 by Robert Lanfear . Sourced from http://biorxiv.org/content/early/2017/04/29/132324.
Login to endorse this review.
This is pretty amazing stuff, very much enjoyed the ms. I would love to see two things here:
1. A discussion in the main ms (perhaps it is buried in the supps somewhere, which I didn't read in detail) of the power of the new method. E.g. roughly how many mutations would one have to measure to have a hope of using the models you propose (e.g. 192 rates in the trinucleotide model presumably requires a decent amount of variation as input; this is compounded by trying to estimate mutation rates for each gene, even using a binomial regression).
2. Implementations of the methods, with READMEs etc to get people started. I notice that you mention this on Twitter, but I think the impact of the ms would be much higher with available implementations.
A question related to point 1 on power. It seems like a standard model selection framework (e.g. hLRT or AICc) should work well here to decide how many parameters one can/should be estimating for a given dataset, and thus whether fitting a full 192 rate model is justified. Does this seem right to you?