rjmcmc [ reversible jump in MrBayes ]

nst=mixed. This is one of my favorite commands in MrBayes. It's used instead of designating a nucleotide substitution model a priori. Instead, reversible jump is a form of model-averaging (across different dimensions of parameter space) where all possible time-reversible substitution models are explored, incorporating uncertainty in model selection.

Background:
​Of six the possible nucleotide substitution rates in the GTR family, MrBayes typically makes use of only half: designated using nst = 1 (F81/JC), nst = 2 (HKY/K2P), & nst = 6 (GTR/SYM).
So, nst = 3, 4, & 5 had been unavailable... until rjMCMC.

Reversible jump Markov chain Monte Carlo is a method to accommodate uncertainty in model selection. Using rjMCMC, MrBayes will incorporate all substitution models (6 rates, 203 models) and the Markov chain will sample a nucleotide substitution model in proportion to its marginal likelihood (spending the most time in best likelihood).  figure, Huelsenbeck et al. 2004

prior density for the rates when using rjMCMC; the number of possible rates (k) = 6

posterior density of rates (nst=1:6) for gene region 28S D2-D3; the highest density is in rates 3, 4, & 5

After a rjMCMC analysis, you could load MrBayes parameter files (.p) into Tracer to visualize which substitution rates were sampled the most often. [Stats are also output in the .pstat file.] This example shows the rate categories estimated for one subset in a partitioned analysis of a clade of eucharitid parasitoid wasps. All sampled rates contribute to the model. 

Likelihood-based analyses results are dependent on the model-fit to the data. The credible set of substitution models typically contains several different models and the benefit of ​rjMCMC is that it allows integration of all. Though tree topology isn't extremely sensitive to model misspecification, other parameters may be (Alfaro & Huelsenbeck 2006).

---

example of mrbayes block using rjMCMC:

[after the nexus alignment of data, use these commands for rjMCMC in MrBayes; this shows data grouped in two subsets]
BEGIN MRBAYES;
log start filename=murray-rj_log;  
    charset codon_pos1 = 1-1041\3;
    charset codon_pos2 = 2-1041\3;
    charset codon_pos3 = 3-1041\3;
partition two_subsets = 2: codon_pos1 codon_pos2, codon_pos3;
set partition = two_subsets;
lset applyto=(all) nst=mixed rates=gamma;   [each of the two subsets will be treated separately]
[use 'nst= mixed' for rjMCMC and 'rates=gamma' to incorporate rate heterogeneity; I don't use parameter for invariant sites]
    unlink shape=(all) pinvar=(all) statefreq=(all) revmat=(all);
    prset applyto=(all) ratepr=variable;
    mcmc ngen=1000000 samplefreq=100 filename=murray_rj;
sumt;  
sump;
[gives a default 25% burnin for tree and parameter files]
log stop;
end;

Also a great tool -- rjMCMC in BEAST 2! Just install the RBS plugin while in BEAUti and you are set to go. The drawback -- as of now (Dec. 2016) you cannot run the XML file in CIPRES if you are using reversible jump models from the RBS plugin.

note:

• Implementing different models of nst=1,2, & 6 in MrBayes is dependent on fixing base frequencies to equal or unequal. For instance, nst=2 codes the HKY model, because the default here is unequal base frequencies, however, entering nst=2 but with equal state frequencies gives the model K2P. Here's a nice site on MrBayes substitution model commands: https://gist.github.com/brantfaircloth/895282 ​.
• Assuming rate variation across sites (using model + I + G) is not a different substitution model than with no I + G.

references:
Huelsenbeck, J.P., Larget, B. & Alfaro, M.E. (2004) Bayesian phylogenetic model selection using reversible jump Markov chain Monte Carlo. Mol Biol Evol, 21, 1123-33.
Alfaro, M.E., & Huelsenbeck, J.P. (2006). Comparative performance of Bayesian and AIC-based measures of phylogenetic model uncertainty. Syst Biol, 55, 89-96.