RAxML LG4x: bootstraping works, ML search fails with error

[Tatiana Yurchenko]

Hello,

I am quite new to this, please don't bite :)

I am running RAxML on CIPRES (it says version 8.2.10)

it fails with an error:

raxmlHPC-HYBRID: optimizeModel.c:1571: optParamGeneric: Assertion `tr->likelihood >= initialLH' failed.

input params:

bootstop_	100
bootstrap_seed_	true
bootstrap_seed_val_	12345
convergence_criterion_	false
datatype_	protein
disable_seqcheck_	false
intermediate_treefiles_	false
mesquite_output_	false
mlsearch_	true
more_memory_	false
mulcustom_aa_matrices_	false
no_bfgs_	false
outsuffix_	eboC_LG4x
parsimony_seed_val_	12345
printbrlength_	true
prot_matrix_spec_	LG4X
prot_sub_model_	PROTGAMMA
provide_parsimony_seed_	true
rearrangement_yes_	false
runtime_	6
specify_bootstraps_	true
use_apobootstopping_	false
use_emp_freqs_	F

input contain protein sequences, matrix is 74 species x 333 aa positions.

in the output I can see that bootstrapping went fine, final lines are:

Overall Time for 108 Rapid Bootstraps 172.004880 seconds
Average Time per Rapid Bootstrap 1.592638 seconds

Starting ML Search ...

-40067.8138382497418206185102462768554687500000 -40051.0789550699555547907948493957519531250000
-40061.7894206660857889801263809204101562500000 -40050.0680180847775773145258426666259765625000

strange thing is this exact dataset works well with LG model, and very similar datasets (oanks ther 5 proteins in our operon) work well with both LG and LG4x.

Thanks for any help!

[Alexandros Stamatakis]

Hi Tatiana,

Could you try running this analysis with the new, totally re-designed
RAxML-NG (see here https://github.com/amkozlov/raxml-ng).

What you are encountering is probably a numerical issue that will
hopefully not occur in the new RAxML. The only problem is that RAxML-NG
is not available via CIPRES yet so you will have to install it yourself.

Alexis

PS: Mark, Wayne: any chance of installing RAxML-NG on CIPRES?

> --
> 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

www.exelixis-lab.org

[Mark Miller]

Yes it is true, Wayne is working on getting RAxML NG installed here, but we have had some issues.

I will post when we get it working in CIPRES.

Best,

Mark

[Tatiana Yurchenko]

Hi Alexis,

thanks for the reply!

as per your suggestion, I tried raxml-ng locally like this:

./raxml-ng --all --msa eboD_phycor2_test.fas --model LG4X+FO --tree pars{10} --bs-trees 200

including part of the dataset (44 taxa x 333 sites)

and during bootstrapping got following error:

[00:20:34] Starting bootstrapping analysis with 200 replicates.

[00:23:53] Bootstrap tree #1, logLikelihood: -10600.880278
[00:26:45] Bootstrap tree #2, logLikelihood: -12054.367126
[00:28:39] Bootstrap tree #3, logLikelihood: -11549.348246
[00:31:03] Bootstrap tree #4, logLikelihood: -11301.977838
[00:32:56] Bootstrap tree #5, logLikelihood: -11421.485884
[00:35:47] Bootstrap tree #6, logLikelihood: -10903.426757
raxml-ng: /home/alex/projects/hits/raxml-ng/src/Optimizer.cpp:33: double Optimizer::optimize_model(TreeInfo&, double): Assertion `cur_loglh - new_loglh < -new_loglh * RAXML_DOUBLE_TOLERANCE' failed.
raxml-ng: /home/alex/projects/hits/raxml-ng/src/Optimizer.cpp:33: double Optimizer::optimize_model(TreeInfo&, double): Assertion `cur_loglh - new_loglh < -new_loglh * RAXML_DOUBLE_TOLERANCE' failed.
raxml-ng: /home/alex/projects/hits/raxml-ng/src/Optimizer.cpp:33: double Optimizer::optimize_model(TreeInfo&, double): Assertion `cur_loglh - new_loglh < -new_loglh * RAXML_DOUBLE_TOLERANCE' failed.
zsh: abort      ./raxml-ng --all --msa eboD_phycor2_test.fas --model LG4X+FO --tree pars{10}

before the search for the tree with maximum likelihood went fine (but it was strange to me, as the previous version did it in opposite order: bootstraps first, ML second).
did I mess up something?

Thanks,
Tatiana

[Alexey Kozlov]

Hi Tatiana,

could you please send me you alignment file and full raxml-ng output / log file? I'll try reproduce and fix the problem.

Side note: you are using LG4X model with ML estimate of AA frequencies (LG4X+FO), but as far as I remember it is not
recommended (we have a corresponding warning in old RAxML). So perhaps you should try with just "LG4X" instead.

Thanks!

Alexey

> RAxML-NG (see here https://github.com/amkozlov/raxml-ng <https://github.com/amkozlov/raxml-ng>).

> > an email to raxml+un...@googlegroups.com <javascript:>
> > <mailto:raxml+un...@googlegroups.com <javascript:>>.
> > For more options, visit https://groups.google.com/d/optout <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
>

> www.exelixis-lab.org <http://www.exelixis-lab.org>

[Alexey Kozlov]

Hi Mark,

many thanks! Please let me know if I can help with those installation problems!

Best,
Alexey

[Tatiana Yurchenko]

Hi Alexey,

Please find shortened alignment and full alignment (i'm most interested to run) attached, as well as the log for the run.

Thanks for your help!

Tatiana

[Alexey Kozlov]

Hi Tatiana,

after further investigation, I can now confirm that the problem was caused by using LG4X model in combination with ML
estimate of AA frequencies (+FO). As I mentioned before, this setup is not recommended as it contradicts the idea behind
the LG4 models (which define 4 distinct sets of AA freqs, one per substitution matrix). I now added a corresponding
warning in RAxML-NG.

With the plain LG4X model, both your alignments (full and shortened) worked fine for me.

Thanks for testing, and please let me know if you encounter any further problems with RAxML-NG!

Best,
Alexey

> > www.exelixis-lab.org <http://www.exelixis-lab.org> <http://www.exelixis-lab.org>

> >
> > --
> > 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

> <javascript:>
> > <mailto:raxml+un...@googlegroups.com <javascript:>>.
> > For more options, visit https://groups.google.com/d/optout <https://groups.google.com/d/optout>.
>
> --