mutation rate priors

[chris blair]

Hi all, 

I am in the process of conducting a divergence time analysis in BEAST 2.1.3. I am looking at the traces in Tracer all all traces for mutationRate parameters seem odd. I did not change any of the default priors in Beauti so they are still set to the 1/x. Should these be modified to improve mixing? Perhaps give the same prior as ucldMean?

I am also experimenting with reversible jump model selection and unfortunately many of these traces don't look so hot either. I have looked on the BEAST website for some literature and guidance regarding RJ in BEAST, but could not find much. Perhaps the chain simply needs to be run for a long time?

Thanks a bunch!

Chris

[Remco Bouckaert]

Hi Chris,

Is it possible you have the estimate box checked for the substitution rate (in the Site Model panel in BEAUti)? If so, you should not estimate the clock rate (ucldMean) as well, unless you fix the mean substitution rate to 1 (using the Fix mean substitution rate checkbox in the Site Model pane, which only shows when there is more than 1 partition). If both substitution and clock rate are estimated both become unidentifiable, since the total rate is the product of these two and there is no information other than the prior to inform their individual values. You can find more details in the book on http://beast2.org/book.html.

Regarding your reversible jump question, I assume you mean the reversible jump substitution model (RB in the substitution model tab). If your data does not support a model with many parameters, not all parameters will be sampled very often. Say, your data supports HKY, corresponding to RBCount=1, then parameters RBrates 2, 3 and up will not be sampled very often resulting in very low ESSs. This is not a problem, since these parameters are not used in the model (most of the time).

Hope this helps,

Remco

[chris blair]

Hi Remco, 

Thanks for the reply and congrats to you, Alexei and everyone who contributed to the book. I am sure it will make a great reference to many of us. 

Yes, I did have the 'estimate' box checked in the Site Model tab for each partition. 

Regarding RJ model selection, is there a key for what each of the parameters represent? You state that the RBCount=1 parameter represents HKY. Perhaps after the analysis is complete there is a file that states the partitions, models, etc.? 

Thanks again.

Chris

[Remco Bouckaert]

Hi Chris,

The model jumps between F81, HKY85, TAN93, TIM, EVS and GTR and RBCount should be interpreted as
0 = F81,
1 = HKY85,
2 = TAN93,
3 = TIM,
4 = EVS and
5 = GTR

If RBCount = X, then all rates numbered up to X are used, so for HKY, one rate is used, for GTR all 5 rates are used, etc.
As I mentioned before, rates that are not used will be sampled with low frequency, so even though all rates are reported in the log these will probably have show up with a low ESS in Tracer. It depends on RBCount whether that is a problematic or not -- if the rate is not used which you can tell by looking at RBCount a low ESS can be ignored.

Cheers,

Remco

[Jacob Berv]

Does the RB model search also search +I and +G model space? If I select these extra parameters for each of my partitions, will the model selection be performed on each partition automatically?

Does the RB model option in the site model tab also auto-partition? So this would negate the need to split my data into a priori partitions? 

Jake

[Remco Bouckaert]

On Saturday, 16 August 2014 08:14:31 UTC+12, Jacob Berv wrote:

Does the RB model search also search +I and +G model space?

No it does not.
 

If I select these extra parameters for each of my partitions, will the model selection be performed on each partition automatically?

If you select the RB model for each of your partitions, the model parameters will be independently estimated for each of the partitions, so you can end up with a HKY model dominating partition 1, and GTR dominating partition 2. You still have to set +I and +G where appropriate.

 

Does the RB model option in the site model tab also auto-partition?

Not by itself. You can auto-partition an alignment by using the AutoPartition template in BEAUti.

So this would negate the need to split my data into a priori partitions? 

If there is sufficient signal in the data, the auto-partition model will find split points in the alignment. However, if you have prior information on where these split-points are, it is more desirable to add that information to the analysis and split the partitions yourself to reduce uncertainty in any of the parameters of interest.

Cheers,

Remco

 

[Jacob Berv]

Can you clarify how this would be implemented? If I select the RB model and then also tell beauti to estimate gamma, what will happen?
Jake

Sent from my iPhone

[Remco Bouckaert]

The substitution model will still be selected using reversible jump. If you specify say 4 gamma categories, each of the substitution models will use these 4 categories. Likewise, if you specify that the proportion of invariant sites needs to be estimated this will be used by each of the substitution models that the RB model jumps between.

Remco

[Jacob Berv]

So would you say a valid approach might be to first do model testing with some other method, like partition finder or model test, to determine the necessity of the gamma shape or invariant sites parameters? And then to use the RB method to select a model? 

Jake

Sent from my iPhone

--
You received this message because you are subscribed to a topic in the Google Groups "beast-users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/beast-users/HuXpOeYfsH0/unsubscribe.
To unsubscribe from this group and all its topics, send an email to beast-users...@googlegroups.com.
To post to this group, send email to beast...@googlegroups.com.
Visit this group at http://groups.google.com/group/beast-users.
For more options, visit https://groups.google.com/d/optout.