Re: [raxml] Protein models in RaxML.

[Alexandros Stamatakis]

Gia sou Panagioti,

> I was wondering whether it is among your plans (and whether it is
> possible to implement) to add protein mixture models to RaxML,

We are just doing a complete-redesign of RAxML, hence it will take some
time until we can start thinking about new models, again, but I will
take a note of this.

> particularly C10-C60
> (http://bioinformatics.oxfordjournals.org/content/24/20/2317.short), and
> UL,EX,EHO
> (http://rstb.royalsocietypublishing.org/content/363/1512/3965.short).
> The aim is, if feasible, to be able to apply the mixture model in
> conjunction with protein GTR (separately estimated for each profile, or
> universally)

Are you sure you want to use a protein GTR? In general all 189 free rate
parameters are difficult to estimate and the optimization does take a
lot of time, mostly this leads to an over-parametrization of the model.

> or different substitution matrices. Also, any plans on
> adding freerate heterogeneity as an option?

Not sure, thus far I have not seen any convincing evidence that it makes
a substantial difference compared to GTR+Gamma. The fact that it simply
yields better likelihoods does not mean anything by itself.

All the best,

Alexis



> Thanks a lot.
>
> --
> You received this message because you are subscribed to the Google
> Groups "raxml" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to raxml+un...@googlegroups.com
> <mailto:raxml+un...@googlegroups.com>.
> For more options, visit https://groups.google.com/d/optout.

--
Alexandros (Alexis) Stamatakis

Research Group Leader, Heidelberg Institute for Theoretical Studies
Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology
Adjunct Professor, Dept. of Ecology and Evolutionary Biology, University
of Arizona at Tucson

www.exelixis-lab.org

[Bui Quang Minh]

Hi Alexis, regarding this:

> or different substitution matrices. Also, any plans on
> adding freerate heterogeneity as an option?

Not sure, thus far I have not seen any convincing evidence that it makes
a substantial difference compared to GTR+Gamma. The fact that it simply
yields better likelihoods does not mean anything by itself.

We have evidence for the impact of the freerate model on the topology. Can share you the data later.

Moreover, the number of rate categories for the Gamma model alone also have a strong impact. Thus, there should be a possibility for users to change this, instead of always using the default 4 categories. I opt for more accuracy instead of speed.

Cheers, Minh

 

[Alexey Kozlov]

Dear Panos and all,

as Alexis pointed out, we're currently working hard towards the new raxml release.
Here are some details regarding the features you asked for:

- both free rates and variable no. of GAMMA rates are supported by the library, so I see no reasons for not offering
them to the users

- UL/EX/EHO models are not there yet, but should be straightforward to include (just adding new matrices/freqs, I guess?)

- protein GTR is already available in raxml, or do you want a mixture of protein GTRs?

- regarding C10-C60 I'm not so sure, it wasn't our priority thus far

So please stay tuned :)

Alexey

[Alexandros Stamatakis]

Dear All,

> as Alexis pointed out, we're currently working hard towards the new
> raxml release.
> Here are some details regarding the features you asked for:
>
> - both free rates and variable no. of GAMMA rates are supported by the
> library, so I see no reasons for not offering them to the users

Yes :-)

> - UL/EX/EHO models are not there yet, but should be straightforward to
> include (just adding new matrices/freqs, I guess?)

Not really, I think it's a bit more involved, I just cross-read the paper.

Alexis

[Alexandros Stamatakis]

Dear Minh,

> > or different substitution matrices. Also, any plans on
> > adding freerate heterogeneity as an option?
>
> Not sure, thus far I have not seen any convincing evidence that it
> makes
> a substantial difference compared to GTR+Gamma. The fact that it simply
> yields better likelihoods does not mean anything by itself.
>
>
> We have evidence for the impact of the freerate model on the topology.
> Can share you the data later.

Yes that would be nice, but maybe not via the RAxML google group, but
that's up to you.

> Moreover, the number of rate categories for the Gamma model alone also
> have a strong impact. Thus, there should be a possibility for users to
> change this, instead of always using the default 4 categories. I opt for
> more accuracy instead of speed.

It really depends, I guess this is a long discussion, every ML
implementation involves several trade-offs, one could also argue that 4
discrete rate cats are okay, and that more time should be invested in
the tree search doing TBRs and other heavy tree moves.

It's just really difficult to get the trade-offs right.

Alexis

>
> Cheers, Minh