Critical bugfix in RAxML-NG v0.5.0b: potentially incorrect bootstrap support values

[Alexey Kozlov]

Dear RAxML-NG users,

please read two important announcements regarding the recently released RAxML-NG v0.5.0b:

1. We discovered a serious bug in RAxML-NG v0.4.1b and below (many thanks to Fabricia Nascimento!): due to an overflow error, bootstrap support values were severely underestimated if you used more than 250 bootrstrap replicates (e.g. '--bs-trees 1000`).

This bug is now fixed, so please upgrade to RAxML-NG v0.5.0b as soon as possible!

PLEASE NOTE:

- regular tree searches ('--search') are NOT affected
- combined analyses (`--all`) with default settings (= 100 BS trees) are NOT affected
- all bootstrap trees are correct (regardless of the settings); only the support values plotted on the best ML tree are broken

 If your analysis was affected, please use the newly introduced '--support' command to re-compute support values from the existing bootstrap trees, e.g.:

 ./raxml-ng --support --tree oldRun.raxml.bestTree --bs-trees oldRun.raxml.bootstraps --prefix newBS

I really apologize for any inconvenience caused by this nasty bug!

2. We introduced the "site repeats" optimization technique described in this paper: https://doi.org/10.1093/sysbio/syw075.  According to our preliminary tests, this optimization can accelerate tree search by 10%-60%,
However, we also saw a few cases where it led to likelihood discrepancies (most probably for numerical reasons) and even to different resulting topologies. 

For now, this feature is considered experimental and it is disabled by default. However, you are encouraged to try it out on your datasets (please use --site-repeats on switch). We would appreciate your feedback on site repeats performance in terms of both speedups and result discrepancies compared to the default version (--site-repeats off). It would be particularly interesting to collect some datasets on which we get different results with and without site repeats, such that we can investigate this issue further.

Best wishes and have a good week,
Alexey

[Péter Poczai]

Hi,

How can I cite RAxML-NG?

Best,

Péter

[Alexey Kozlov]

Hi Peter,

RAxML-NG paper is still in preparation, so for the time being please cite github repository via provided DOI:

https://github.com/amkozlov/raxml-ng#license-and-citation

Thanks,
Alexey

On 24.01.2018 16:46, Péter Poczai wrote:
> Hi,
>
> How can I cite RAxML-NG?
>
> Best,
> Péter
>
> 2017. szeptember 11., hétfő 17:02:28 UTC+3 időpontban Alexey Kozlov a következőt írta:
>
> Dear RAxML-NG users,
>
> please read two important announcements regarding the recently released RAxML-NG v0.5.0b:
>
> 1. We discovered a serious bug in RAxML-NG v0.4.1b and below (many thanks to Fabricia Nascimento!): due to an

> overflow error, bootstrap support values were severely underestimated *if you used more than 250 bootrstrap
> replicates* (e.g. '--bs-trees 1000`).

>
> This bug is now fixed, so please upgrade to RAxML-NG v0.5.0b as soon as possible!
>
> PLEASE NOTE:
>

> - regular tree searches ('--search') are *NOT* affected
> - combined analyses (`--all`) with default settings (= 100 BS trees) are *NOT* affected
> - all /bootstrap trees/ are correct (regardless of the settings); only the /support values/ plotted on the best ML

> tree are broken
>
>  If your analysis was affected, please use the newly introduced '--support' command to re-compute support values
> from the existing bootstrap trees, e.g.:
>

> * ./raxml-ng --support --tree oldRun.raxml.bestTree --bs-trees oldRun.raxml.bootstraps --prefix newBS
>
> *I really apologize for any inconvenience caused by this**nasty bug*!
>
> *2. We introduced the "site repeats" optimization technique described in this paper:
> https://doi.org/10.1093/sysbio/syw075 <https://doi.org/10.1093/sysbio/syw075>.  According to our preliminary tests,

> this optimization can accelerate tree search by 10%-60%,
> However, we also saw a few cases where it led to likelihood discrepancies (most probably for numerical reasons) and
> even to different resulting topologies.
>
> For now, this feature is considered experimental and it is disabled by default. However, you are encouraged to try

> it out on your datasets (please use *|--site-repeats on|* switch). We would appreciate your feedback on site repeats
> performance in terms of both speedups and result discrepancies compared to the default version (*|--site-repeats
> off|*). It would be particularly interesting to collect some datasets on which we get different results with and

> without site repeats, such that we can investigate this issue further.
>
> Best wishes and have a good week,
> Alexey
>

> --
> 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.