Setting parameters when using bModelTest

[essa...@gmail.com]

Hi,

I'm a BEAST newbie trying to get divergence times for a small group of invertebrates in the same Family (5 spp in one genus, 2 in another). My dataset consists of 3 mitochondrial protein coding genes, 3 nuclear housekeeping genes, and nuclear 28s rRNA (~7500 bp in all with 9 partitions). I also have 1 Order-level outgroup for calibration (my ingroup has been found in Baltic amber c. 44 mya). I wish to use the "bModelTest" option under substitution models, but I can't find any information on how this influences other settings. It seems logical that I would not link site models, even though I've chosen the same option for each partition. A strict clock was rejected for these data, so I assume I do not link clocks? Should the trees be linked or unlinked? Selecting "bModelTest" also presents a lot of choices under the Priors tab and I have no idea what to select here. Any guidance is greatly appreciated!

Regards,

SAL 

[Remco Bouckaert]

Hi SAL,

I assume you are doing a *BEAST analysis (using the StarBeast template in BEAUti). Trees should be linked if genes share the same history — so mitochondrial genes all have the same history and should be linked, but nuclear genes typically don’t share a history so in general should have one tree each.

The parameters for bModelTest are all estimated, and the priors on rates, the models and other parameters are fairly broad, meaning they cover most cases. If you have any extra information about rates, rate heterogeneity or proportion invariant, you can use that information to change the priors on these parameters in the priors panel. More details in the paper http://biorxiv.org/content/early/2015/06/11/020792. Having the site models unlinked seems sensible (if there is sufficient data) since each gene will have its own rate and mechanism of evolution,  but if you find that estimates for all parameters of two or more site models are close, you could link them and save estimating a few parameters, resulting in faster convergence and better fit of the marginal likelihood.

With clock models it is probably best to start with the simplest case; a shared strict clock with estimated substitution rates. If that converges, try a relaxed clock shared by all genes, and if that runs fine you can unlink the clock (but stop estimating substitution rates, since independent clock rates make it impossible to estimate both these parameters, more here: http://beast2.org/2015/06/23/help-beast-acts-weird-or-how-to-set-up-rates/) and see how that fares.

Cheers,

Remco

--
You received this message because you are subscribed to the Google Groups "beast-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at https://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.