Rapid bootstrapping and standard bootstrapping
688 views
Skip to first unread message
Ding
Oct 13, 2011, 5:05:30 AM10/13/11
to raxml
Hi Alexis,
Could you help me to clarify a question regarding the novel rapid
bootstrapping (RBS) method?
Please could you correct me if I misunderstood the differences from
RBS and Standard BootStrapping (SBS).
According to your paper in 2008 (Stamatakis A, Hoover P, Rougemont J.
A rapid bootstrap algorithm for the RAxML Web servers. Syst Biol 2008
Oct.;57(5):758–771.):
1. The novel RBS search generally explore less tree space (randomly
assign radius, 5 <= r <= 15) comparing to SBS (determined
automatically by program). Could it potentially compromises the LSR
capability on individual bootstrap tree search?
2. In every 10 RBS searches, program uses previous final tree as the
starting tree of next bootstrap search. Could it also diminish the
power of bootstrapping since the trees are gradually improved hence
somehow correlate with each other in every 10 RBS (i.e. 100 RBS search
actually in a way only represent 10 SBS searches?) ?
Ultimately what I'd like to do is an ML tree with 100 BS throughly
rather than rapidly, should I always use SBS instead of RBS?
best,
Ding
Could you help me to clarify a question regarding the novel rapid
bootstrapping (RBS) method?
Please could you correct me if I misunderstood the differences from
RBS and Standard BootStrapping (SBS).
According to your paper in 2008 (Stamatakis A, Hoover P, Rougemont J.
A rapid bootstrap algorithm for the RAxML Web servers. Syst Biol 2008
Oct.;57(5):758–771.):
1. The novel RBS search generally explore less tree space (randomly
assign radius, 5 <= r <= 15) comparing to SBS (determined
automatically by program). Could it potentially compromises the LSR
capability on individual bootstrap tree search?
2. In every 10 RBS searches, program uses previous final tree as the
starting tree of next bootstrap search. Could it also diminish the
power of bootstrapping since the trees are gradually improved hence
somehow correlate with each other in every 10 RBS (i.e. 100 RBS search
actually in a way only represent 10 SBS searches?) ?
Ultimately what I'd like to do is an ML tree with 100 BS throughly
rather than rapidly, should I always use SBS instead of RBS?
best,
Ding
Alexis
Oct 13, 2011, 5:25:51 AM10/13/11
to raxml
Dear Ding,
> Could you help me to clarify a question regarding the novel rapid
> bootstrapping (RBS) method?
>
> Please could you correct me if I misunderstood the differences from
> RBS and Standard BootStrapping (SBS).
>
> According to your paper in 2008 (Stamatakis A, Hoover P, Rougemont J.
> A rapid bootstrap algorithm for the RAxML Web servers. Syst Biol 2008
> Oct.;57(5):758–771.):
>
> 1. The novel RBS search generally explore less tree space (randomly
> assign radius, 5 <= r <= 15) comparing to SBS (determined
> automatically by program). Could it potentially compromises the LSR
> capability on individual bootstrap tree search?
I don't know what LSR stands for, but since this is a more
approximate
and fast search algorithm, it is more prone to get stuck in local
optima of tree spaces.
Thus, if in doubt and if enough computational resources are available
you should
use the standard slow bootstrap if you can.
> 2. In every 10 RBS searches, program uses previous final tree as the
> starting tree of next bootstrap search.
Exactly.
> Could it also diminish the
> power of bootstrapping since the trees are gradually improved hence
> somehow correlate with each other in every 10 RBS (i.e. 100 RBS search
> actually in a way only represent 10 SBS searches?) ?
I wouldn't expect that, since the trees are not gradually improved
because there
is the intermediate step of bootstrap re-sampling, i.e., a tree with a
good likelihood on
bootstrap replicate i does not need to be a tree with a good
likelihood score for replicate i+1.
This is especially the case if you don't have many sites, say a single
gene with 500 or 1000 taxa.
However, on very broad phylogenomic datasets with say 100 genes or
more and about 100-200 taxa,
what you state above may well apply.
> Ultimately what I'd like to do is an ML tree with 100 BS throughly
> rather than rapidly, should I always use SBS instead of RBS?
If you can (in terms of available computational resources) I'd
definitely recommend that. The RBS has mainly been designed
to be able to offer a web-service that can do bootstrapping and for
bootstrapping very large datasets.
I wouldn't however just compute 100 SBS replicates but deploy the
bootstopping option, see:
http://www.springerlink.com/content/d026w742647500h4/
To automatically determine how many SBS replicates are necessary to
obtain stable support values.
Alexis
> Could you help me to clarify a question regarding the novel rapid
> bootstrapping (RBS) method?
>
> Please could you correct me if I misunderstood the differences from
> RBS and Standard BootStrapping (SBS).
>
> According to your paper in 2008 (Stamatakis A, Hoover P, Rougemont J.
> A rapid bootstrap algorithm for the RAxML Web servers. Syst Biol 2008
> Oct.;57(5):758–771.):
>
> 1. The novel RBS search generally explore less tree space (randomly
> assign radius, 5 <= r <= 15) comparing to SBS (determined
> automatically by program). Could it potentially compromises the LSR
> capability on individual bootstrap tree search?
approximate
and fast search algorithm, it is more prone to get stuck in local
optima of tree spaces.
Thus, if in doubt and if enough computational resources are available
you should
use the standard slow bootstrap if you can.
> 2. In every 10 RBS searches, program uses previous final tree as the
> starting tree of next bootstrap search.
> Could it also diminish the
> power of bootstrapping since the trees are gradually improved hence
> somehow correlate with each other in every 10 RBS (i.e. 100 RBS search
> actually in a way only represent 10 SBS searches?) ?
because there
is the intermediate step of bootstrap re-sampling, i.e., a tree with a
good likelihood on
bootstrap replicate i does not need to be a tree with a good
likelihood score for replicate i+1.
This is especially the case if you don't have many sites, say a single
gene with 500 or 1000 taxa.
However, on very broad phylogenomic datasets with say 100 genes or
more and about 100-200 taxa,
what you state above may well apply.
> Ultimately what I'd like to do is an ML tree with 100 BS throughly
> rather than rapidly, should I always use SBS instead of RBS?
definitely recommend that. The RBS has mainly been designed
to be able to offer a web-service that can do bootstrapping and for
bootstrapping very large datasets.
I wouldn't however just compute 100 SBS replicates but deploy the
bootstopping option, see:
http://www.springerlink.com/content/d026w742647500h4/
To automatically determine how many SBS replicates are necessary to
obtain stable support values.
Alexis
Martin Dohrmann
Oct 13, 2011, 12:28:32 PM10/13/11
to ra...@googlegroups.com
Hi Alexis,
how can I use the standard bootstrap in 7.2.8? I couldn't figure it out since the -f i option changed to something different compared to what's in the old manual...
how can I use the standard bootstrap in 7.2.8? I couldn't figure it out since the -f i option changed to something different compared to what's in the old manual...
Cheers,
Martin
Alexis
Oct 14, 2011, 3:10:43 AM10/14/11
to raxml
if you want to do 100 bootstraps in v728 you'd type:
./raxmlHPC-SSE3 -b 12345 -p 12345 -# 100 -m GTRGAMMA -s dna.phy -n T1
This hasn't really changed since v704
Also, the most up to date RAxML versions are now available via
github:
https://github.com/stamatak
Alexis
On 13 Okt., 18:28, Martin Dohrmann <m.dohrm...@lrz.uni-muenchen.de>
wrote:
./raxmlHPC-SSE3 -b 12345 -p 12345 -# 100 -m GTRGAMMA -s dna.phy -n T1
This hasn't really changed since v704
Also, the most up to date RAxML versions are now available via
github:
https://github.com/stamatak
Alexis
On 13 Okt., 18:28, Martin Dohrmann <m.dohrm...@lrz.uni-muenchen.de>
wrote:
Reply all
Reply to author
Forward
0 new messages