approximate equivalences between gamma distribution and R/free-rate models

[conxor...@gmail.com]

Hi everyone,

I've been using IQ-Tree for phylogenomic analysis of a multigene dataset (~20000 aa positions). The TESTNEW analysis yielded LG+R7 as the best fitting model. Now I want to replicate the results using a bayesian approach in PhyloBayes, but free-rate R models don't exist there.

So, I'm wondering which is the closest approximation to R7 in the gamma distributions supported by Phylobayes and other software. My guess is that a gamma distribution with 7 categories is the best approach (but maybe gamma and free-rate distributions are too fundamentally different for this to work).

Any ideas?

Thanks a lot in advance!

Xavi

[Bui Quang Minh]

Dear Xavi (CC Lars Jermiin, in case you want to add something),

You are right that LG+G7 is the closest match to LG+R7, though they are still quite different.

Moreover, since you have such long sequences (20K aa), I suggest that you look at the composition test result (in the log file) and if many sequences failed the test, you would try the C10-C60 models as well (they are the ML version for PhyloBayes’ CAT model). You can use LG+C10+R7 model, for example, and look if such model provides better fit to the data. This is because -m TESTNEW by default does not include complex models due to computational reason. In fact you can instruct IQ-TREE to consider these models by e.g.:

iqtree … -m TESTNEW -madd LG+C10+R7,LG+C20+R7

And finally if the C10 model series takes too much time and memory during subsequent analysis, you can use the recently proposed site frequency model (see http://www.iqtree.org/doc/Complex-Models/#site-specific-frequency-models)

Cheers, Minh

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

--
Bui Quang Minh
Center for Integrative Bioinformatics Vienna (CIBIV)
Campus Vienna Biocenter 5, VBC5, Ebene 1
A-1030 Vienna, Austria
Phone: ++43 1 4277 74326
Email: minh.bui (AT) univie.ac.at

[conxor...@gmail.com]

Hi Minh,

Thanks a lot for your prompt answer!

Regarding the composition bias test, you are totally right: many sequences appear to fail to pass it. Thanks for the suggestion of the C10-60 and PMSF models, I've been playing around with those too.

Is there a way to tell which is the best C-XX value for the C10-C60 and PMSF models? Would the following command work for that purpose?

iqtree … -m TESTNEW -mset LG+R7 -madd LG+C10+R7,LG+C20+R7,LG+R7+PMSF

Cheers,

Xavi

[Bui Quang Minh]

Hi Xavi,

> On Nov 23, 2016, at 3:35 PM, conxor...@gmail.com wrote:
>
> Hi Minh,
>
> Thanks a lot for your prompt answer!

you are welcome ;-)

>
> Regarding the composition bias test, you are totally right: many sequences appear to fail to pass it. Thanks for the suggestion of the C10-60 and PMSF models, I've been playing around with those too.
>
> Is there a way to tell which is the best C-XX value for the C10-C60 and PMSF models? Would the following command work for that purpose?
>
> iqtree … -m TESTNEW -mset LG+R7 -madd LG+C10+R7,LG+C20+R7,LG+R7+PMSF

do this:

iqtree … -m TESTNEW -mset LG+R7 -madd LG+C10+R7,LG+C20+R7,LG+C30+R7,LG+C40+R7,LG+C50+R7,LG+C60+R7

Remove LG+R7+PMSF. This is not valid syntax and, btw the PMSF model was not designed for model testing. Later on if LG+C20+R7 provides the best fit, then you can apply PMSF by:

iqtree … -m LG+C20+R7 -ft <guide_tree>

Minh

[Xavi Grau]

Hi Minh,

Perfect, thanks!

B,

Xavi

>>> To unsubscribe from this group and stop receiving emails from it, send an email to iqtree+unsubscribe@googlegroups.com.