Alpha parameter exceeds 10

[sah...@iisertvm.ac.in]

When I run the dataset using GTR+G model on RAxML, the error comes as:

WARNING the alpha parameter with a value of 10.107210 estimated by RAxML for partition number 5 with the name "Partition_6"
is larger than 10.000000.
You should do a model test and confirm that you actually need to incorporate a model of rate heterogeneity!
You can run inferences with a plain substitution model (without rate heterogeneity) by specifyng the CAT model and the "-V" option!

The same error comes even I use GTR+CAT or GTR+CAT with -V option.
But the RAxML runs fine for the same dataset when I run it on my computer.
I think the problem is that the RAxML on the CIPRES server is not working whenever the alpha parameter exceeds 10.
Is there any way to change this, i.e. to let the programmme function work even when the alpha parameter exceeds 10. Such option is available on RAxML using the command -u. But such facility to use -u on RAxML is not available on CIPRES. Is there any alternate way to solve this problem.

Thanks in advance.

[Heiko]

Dear sahoork,

please read the message you got from RAxML carefully. It is NOT an error but a warning.

What it means is the following - if the value for the Gamma-shape parameter alpha is high, the model behaves as if the data is not heterogeneous but actually homogeneous. This can well be the case if you do not have many mutations and/or if they are distributed evenly (=homogeneously) along your sequence. In such cases enforcing rate heterogeneity does not make sense because the additional parameter will not improve the fit to your data significantly.
One way typically used to check this is applying a modeltest / mode selection approach (possible tools: jmodeltest for DNA, prottest for AA, IQ-Tree for both and much faster, RAxML for AA) to test what might be the best evolutionary model to use.
Using more parameters may give a better fit, but on the other hand takes more time and increase the variance of your results.

Modeltesting/modelselection tries to figure out the point at which using a more parameter-rich model does not significantly improve the likelihood of your estimated tree. Thus, RAxML automatically points out here that although you made it use a Gamma model that you should check with a model-test whether it is actually necessary to use the +G because with its high alpha (>10.0) it already behaves almost like a model without +G.

So this is neither an error nor a problem in CIPRES or RAxML, but in any case you should do a modeltest to determine the best suited model for your data.

I hope I could help,
Heiko

[Alexandros Stamatakis]

Thanks Heiko,

Just to add: the -u option does something different, than you think,
it uses the median instead of the mean for discretizing the Gamma
distribution, but still uses Gamma.

As to why the warning still appears when using -m GTRCAT -V please see:

https://groups.google.com/forum/?hl=de#!topic/raxml/CkgUkEcBjWs

Alexis

On 06.07.2015 15:15, Heiko wrote:
> Dear sahoork,
>

> please read the message you got from RAxML carefully. It is NOT an error e

> but a warning.
>
> What it means is the following - if the value for the Gamma-shape
> parameter alpha is high, the model behaves as if the data is not
> heterogeneous but actually homogeneous. This can well be the case if you
> do not have many mutations and/or if they are distributed evenly
> (=homogeneously) along your sequence. In such cases enforcing rate
> heterogeneity does not make sense because the additional parameter will
> not improve the fit to your data significantly.
> One way typically used to check this is applying a modeltest / mode
> selection approach (possible tools: jmodeltest for DNA, prottest for AA,

> IQ-Test for both and much faster, RAxML for AA) to test what might be

> the best evolutionary model to use.
> Using more parameters may give a better fit, but on the other hand takes
> more time and increase the variance of your results.
>
> Modeltesting/modelselection tries to figure out the point at which using
> a more parameter-rich model does not significantly improve the
> likelihood of your estimated tree. Thus, RAxML automatically points out
> here that although you made it use a Gamma model that you should check
> with a model-test whether it is actually necessary to use the +G because
> with its high alpha (>10.0) it already behaves almost like a model
> without +G.
>
> So this is neither an error nor a problem in CIPRES or RAxML, but in any
> case you should do a modeltest to determine the best suited model for
> your data.
>
> I hope I could help,
> Heiko
>

> --
> 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

[Heiko Schmidt]

Dear all,

one comment: our software which I listed among those you can use to do the DNA model test/selection is named IQ-Tree (not IQ-Test ;) You can find that freely available at http://www.cibiv.at/software/iqtree/ .

I had corrected that in my post, but Alexis has quoted the one with the typo.

Best regards,
Heiko