Substitution and site models of RNA virus

[Kritchakorn Sawakwongpra]

Hello, 

I have a question about modeling in RNA virus studies. I found that the HKY model is more suitable for generating Bayesian skyline plots compared to GTR+Gamma=4. Is it acceptable to use different models for different objectives? For instance, can I use GTR+Gamma=4 for root-to-tip divergence analysis and HKY for Bayesian skyline plots? If anyone has experience with this issue, I would greatly appreciate your insights, as I am new to this field. Additionally, how can I demonstrate which model is most suitable for my specific objective?

Thanks!

Kritchy

[Mamerto Jr Brina]

Hi Kritchakorn. I am unsure whether you can use different substitution models for different objectives. However, I can recommend using external applications for root-to-tip divergence analysis and Bayesian skyline plots. You can use IQTree with its Model Finder to know which substitution model is the most appropriate for your sequences. Additionally, IQTree will generate a TREEFILE, which you can use in TempEst for your root-to-tip divergence analysis. Hope this helps.

[Alexei Drummond]

Hi,

What do you mean by “more suitable”? What criteria did you judge this by?

Cheers

Alexei

--
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 view this discussion visit https://groups.google.com/d/msgid/beast-users/73dd981e-aaad-4af5-8d28-39f8bd05f745n%40googlegroups.com.

[Kritchakorn Sawakwongpra]

Hi Mamerto Jr. Brina,
Thank you for your suggestion.

Hi Alexi,
I have read some publications where they used different models for different genes of the same virus, such as HKY for RdRp and GTR for VP1. However, I am not sure how to evaluate which model is most suitable for the gene of interest. Could you provide some guidance on this?

Thanks

[Ziv Lieberman]

Hi Kritchy,

Do the publications say that they performed model-testing exercises, e.g., with ModelFinder as Mamerto mentioned? It sounds like these could be results pertaining to their particular dataset and modeling context. I expect relative model performance to be stated with specific reference to a specific optimality criterion, for example, with respect to AIC or BIC. It seems it would be very difficult to have enough independent information to know that a particular model is truly a more accurate representation of the process; one usually considers that, given no other information, a GTR model should be "more realistic" (more 'suitable'?) because of how it parameterizes parts of the process; on any given dataset, though, it may be overly complex.

All this to say, I agree with Alexei that such a statement about suitability needs to be made with regard to specific tests and criteria, and with Mamerto that if you want to evaluate models against each other, ModelFinder in IQ-TREE2 is a very good approach.

Cheers,

Ziv

To view this discussion visit https://groups.google.com/d/msgid/beast-users/67b3d5f1-0720-4095-b04b-b2857c8068e1n%40googlegroups.com.

[Alexei Drummond]

If you are uncertain about what model to use you can always run bModelTest, which will average over all time-reversible substitution models.

Cheers

Alexei

To view this discussion visit https://groups.google.com/d/msgid/beast-users/67b3d5f1-0720-4095-b04b-b2857c8068e1n%40googlegroups.com.

[Kritchakorn Sawakwongpra]

Thank you for all suggestion.