Bootstrap support in large phylogenies
47 views
Skip to first unread message
Dimitrios Parask
May 16, 2013, 2:02:11 PM5/16/13
to ra...@googlegroups.com
I would like to ask how the number of taxa affects the bootstrap values in the inferred trees.
From my experience in large alignments (>1,000 sequences) bootstrap values hardly get higher than > 75% mainly due to short distances between internal nodes. What is your experience how far about it?
Alexandros Stamatakis
May 17, 2013, 6:35:38 AM5/17/13
to ra...@googlegroups.com
Gia sou Dimitri,
> I would like to ask how the number of taxa affects the bootstrap
> values in the inferred trees.
In general the BS values decrease, in particular when you increase the
number of taxa while keeping the number of sites constant.
One reason for this is that as you add more taxa you will generally have
less signal and the BS procedure that just randomly re-samples some
sites will make the signal even weaker.
In other words, as you increase the #taxa and keep the #sites fixed the
bootstrap trees will be more and more different from the ML tree.
Please keep in mind that this is only an empirical observation, but
there are papers where people did simulations to show how reconstruction
accuracy decreases as a function of #taxa.
issue is that you are trying to infer the relationships among more taxa
for the same number of sites and the likelihood surface gets rougher ...
A good way for quantifying the roughness of the likelihood surface is to
do say 20 independent ML tree searches, then select those trees that are
statistically not significantly different from each other (using the
tests implemented in CONSEL) and then compute the RF distances between
those trees.
If they exceed 5% your likelihood surface will be pretty rough, for an
empirical single-gene dataset with 7000 taxa I observed an average RF
between equally plausible trees of around 25%.
Alexis
>
> --
> 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.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>
--
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
> I would like to ask how the number of taxa affects the bootstrap
> values in the inferred trees.
number of taxa while keeping the number of sites constant.
One reason for this is that as you add more taxa you will generally have
less signal and the BS procedure that just randomly re-samples some
sites will make the signal even weaker.
In other words, as you increase the #taxa and keep the #sites fixed the
bootstrap trees will be more and more different from the ML tree.
Please keep in mind that this is only an empirical observation, but
there are papers where people did simulations to show how reconstruction
accuracy decreases as a function of #taxa.
>
> From my experience in large alignments (>1,000 sequences) bootstrap
> values hardly get higher than > 75% mainly due to short distances
> between internal nodes. What is your experience how far about it?
It might also be because of short distances, but need not be. The basic
> From my experience in large alignments (>1,000 sequences) bootstrap
> values hardly get higher than > 75% mainly due to short distances
> between internal nodes. What is your experience how far about it?
issue is that you are trying to infer the relationships among more taxa
for the same number of sites and the likelihood surface gets rougher ...
A good way for quantifying the roughness of the likelihood surface is to
do say 20 independent ML tree searches, then select those trees that are
statistically not significantly different from each other (using the
tests implemented in CONSEL) and then compute the RF distances between
those trees.
If they exceed 5% your likelihood surface will be pretty rough, for an
empirical single-gene dataset with 7000 taxa I observed an average RF
between equally plausible trees of around 25%.
Alexis
>
> --
> 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.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>
--
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
Reply all
Reply to author
Forward
0 new messages