Dear Albert,

In Ziheng's PAML notation:

MG94CUSTOM uses F3x4 estimator (see this paper for complete details [url]http://www.hyphy.org/sergei/2gamma.pdf[/url]), as do MG94W9 and GY94W9.

MG94 and GY94 use F1x4.

Fcodon and 1/61 each are not implemented (mostly for reasons that they Fcodon will almost surely bias the estimation for small data sets, where some codons will have very small counts and for large datasets Fcodon and F3x4 tend to yield similar estimates, and 1/61 is a very unrealistic model). If you are interested in adding them, I could easily do that, at least for the GY94 models (rate matrix parameterization makes it trickier for MG94). Let me know if you want these options for your comparisons, and I'll post a file which implements them.

That F3x4MG is similar MG94custom (and not MG94w9) is right, because PAML always corrects for transition/transversion biases (as does MG94customx010010), whereas MG94, as originally published, did not.

HyPhy and PAML also differ in how they treat gaps and ambiguous characters, which could contribute to small differences in log likelihoods.

HTH,
Sergei