Numeric issue in ancestral reconstruction ?

[benjami...@gmail.com]

Dear raxml team,

I have a dataset for which I do not managed to compute ancestral reconstruction.
It is a bacterial RNA tree, with around 13000 leaves, I often used similar alignments without issues in raxml-ng. You can find the dataset attached here.

1) If I evaluate the tree, the process is successful, I obtain the likelihood and models parameters.

$ raxml-ng --evaluate --model GTR+G4 --tree output_model.raxml.reduce.fasta.raxml.bestTree --msa output_model.raxml.reduce.fasta --threads 4 --seed 1 --redo
RAxML-NG v. 1.2.2 released on 11.04.2024 by The Exelixis Lab.
System: Intel(R) Core(TM) i7 CPU         920  @ 2.67GHz, 4 cores, 11 GB RAM
[...]
[00:35:44] Tree #1, final logLikelihood: -2621741.748007
$ cat output_model.raxml.reduce.fasta.raxml.bestModel
GTR{0.949240/2.202230/1.397029/0.841129/3.086234/1.000000}+FU{0.224835/0.227780/0.308182/0.239202}+G4m{0.466324}, noname = 1-5403

2) If I attempt ancestral reconstruction, it fails. Either with no error messages (older SSE3 CPU) or a message about infinite likelihood for the starting tree (recent AVX2 CPU).

$ raxml-ng --ancestral --model GTR{0.845653/2.573440/1.114568/0.750459/2.478924/1.000000}+FU{0.238629/0.216428/0.282286/0.262656}+G4m{0.334624} --tree output_model.raxml.reduce.fasta.raxml.bestTree --msa output_model.raxml.reduce.fasta --threads 4 --seed 1 --log DEBUG
[...]


   a) logged error in older SSE3 CPU (nothing)

Starting ML tree search with 1 distinct starting trees

Processus stopped


   b) logged error in recent AVX2 CPU
   
Starting ML tree search with 1 distinct starting trees
[00:00:01] Tree #1, initial LogLikelihood: -inf
[00:00:03 -inf] Initial branch length optimization
terminate called recursively
terminate called recursively
terminate called recursively
terminate called recursively
Abandon


The second case makes me think about numeric computation issues. However, the tree evaluation works, which means it is restricted to ancestral reconstruction.
There could be some too close to zero value, and division by zero during posterior probabilities computations. However, here the log shows that infinity seems related to the starting tree, which should not be an issue as its likelihood can be computed in evaluate command.

Do you have any idea of what is going on ?

Thank you for your help,

Benjamin Linard

[Oleksiy Kozlov]

Dear Benjamin,

thanks for reporting!

The problem is that this dataset requires per-rate scaling, which was not supported in the ancestral
reconstruction mode as of raxml-ng 1.2.2.

However, I recently implemented it in the dev branch, please feel free to try it out:

https://github.com/amkozlov/raxml-ng/wiki/Installation#building-development-branch

Best,
Oleksiy

> --
> 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>.
> To view this discussion on the web visit
> https://groups.google.com/d/msgid/raxml/2ace0bf2-6295-421a-a4bf-8b004fe40cbbn%40googlegroups.com
> <https://groups.google.com/d/msgid/raxml/2ace0bf2-6295-421a-a4bf-8b004fe40cbbn%40googlegroups.com?utm_medium=email&utm_source=footer>.

[benjami...@gmail.com]

Dear Oleksiy,

I confirm that the fix from the dev branch fixed the issue.

Thank you for your help,

Benjamin

[Oleksiy Kozlov]

Great, many thanks for testing!

> <https://groups.google.com/d/msgid/raxml/2ace0bf2-6295-421a-a4bf-8b004fe40cbbn%40googlegroups.com?utm_medium=email&utm_source=footer <https://groups.google.com/d/msgid/raxml/2ace0bf2-6295-421a-a4bf-8b004fe40cbbn%40googlegroups.com?utm_medium=email&utm_source=footer>>.

>
> --
> 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>.

> To view this discussion visit
> https://groups.google.com/d/msgid/raxml/0ce6b629-9d18-4ac5-b8ea-7721e7ed248dn%40googlegroups.com
> <https://groups.google.com/d/msgid/raxml/0ce6b629-9d18-4ac5-b8ea-7721e7ed248dn%40googlegroups.com?utm_medium=email&utm_source=footer>.