[R-sig-phylo] Ancestral state reconstruction

[Szymek Drobniak]

Hello all,

I'm trying to reconstruct ancestral states in a 3-state trait evolving on a
big (>2000 tips) phylogeny. I've tried both BayesTraits and tools from Liam
Revell's phytools (continuous time Markov model and stochastic mapping)
package. Surprisingly (or maybe not?) both give different answers,
especially with respect to the ancestral states of deep nodes (including
root node) and transition rates. Is it something I should expect - do
anyone have some idea why these are so different?

Cheers
szymon

[Liam J. Revell]

Hi Szymon.

Can you give us some more details on the analyses that you ran? Marginal
ancestral states (from rerootingMethod) and posterior frequencies (from
make.simmap, using describe.simmap to summarize the results across
stochastic maps) tend to be very highly correlated (although they should
not expected to be exactly the same as give marginal ancestral states
and the other samples sets of states from their joint probability
distribution). There are several possible reasons why ancestral states
might be different. For instance, the default substitution model in
rerootingMethod is "ER" (equal-rates) whereas the default model in
make.simmap is "SYM" (symmetrical rates). In addition, the methods can
use different prior probabilities for the root. By default, both
functions use a flat prior; however make.simmap can also use the
stationary distribution given the transition matrix Q, or an arbitrary
prior. Finally, make.simmap can use the MLE of Q, an arbitrary value of
Q, or sample Q from its posterior probability distribution conditioned
on a model. All of these different options could potentially result in
different ancestral reconstructions.

All the best, Liam

Liam J. Revell, Assistant Professor of Biology
University of Massachusetts Boston
web: http://faculty.umb.edu/liam.revell/
email: ***@umb.edu
blog: http://blog.phytools.org

Post by Szymek Drobniak
Hello all,
I'm trying to reconstruct ancestral states in a 3-state trait evolving on a
big (>2000 tips) phylogeny. I've tried both BayesTraits and tools from Liam
Revell's phytools (continuous time Markov model and stochastic mapping)
package. Surprisingly (or maybe not?) both give different answers,
especially with respect to the ancestral states of deep nodes (including
root node) and transition rates. Is it something I should expect - do
anyone have some idea why these are so different?
Cheers
szymon
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-phylo

[Szymek Drobniak]

Hello, thank you Liam for your quick answer, I might have phrased my
question not very clearly. The point is that both simmap and
rerootingMethod (using bort EM and SYM) give almost identical results
(there are some slight differences which is obvious but the overall picture
- and the root ancestral state - are the same). I was wondering why
BayesTraits (which, at least in theory, uses similar approach) gives
different result than phytools - in BayesTraits the transition rates are
entirely different, and the ancestral state of the root is different. I was
wondering that maybe somebody had more experience with BayesTraits and came
across this difference between BayesTraits and some other packages. As for
the BT analysis - I used both ML and MCMC methods, in the latter I used a
flat exponential hyperprior. For the time being I used a consensus tree
rather than a sample from the distribution of trees.

Cheers
sz.

Post by Liam J. Revell
Hi Szymon.
Can you give us some more details on the analyses that you ran? Marginal
ancestral states (from rerootingMethod) and posterior frequencies (from
make.simmap, using describe.simmap to summarize the results across
stochastic maps) tend to be very highly correlated (although they should
not expected to be exactly the same as give marginal ancestral states and
the other samples sets of states from their joint probability
distribution). There are several possible reasons why ancestral states
might be different. For instance, the default substitution model in
rerootingMethod is "ER" (equal-rates) whereas the default model in
make.simmap is "SYM" (symmetrical rates). In addition, the methods can use
different prior probabilities for the root. By default, both functions use
a flat prior; however make.simmap can also use the stationary distribution
given the transition matrix Q, or an arbitrary prior. Finally, make.simmap
can use the MLE of Q, an arbitrary value of Q, or sample Q from its
posterior probability distribution conditioned on a model. All of these
different options could potentially result in different ancestral
reconstructions.
All the best, Liam
Liam J. Revell, Assistant Professor of Biology
University of Massachusetts Boston
web: http://faculty.umb.edu/liam.revell/
blog: http://blog.phytools.org

Post by Szymek Drobniak
Hello all,
I'm trying to reconstruct ancestral states in a 3-state trait evolving on a
big (>2000 tips) phylogeny. I've tried both BayesTraits and tools from Liam
Revell's phytools (continuous time Markov model and stochastic mapping)
package. Surprisingly (or maybe not?) both give different answers,
especially with respect to the ancestral states of deep nodes (including
root node) and transition rates. Is it something I should expect - do
anyone have some idea why these are so different?
Cheers
szymon
_______________________________________________
https://stat.ethz.ch/mailman/listinfo/r-sig-phylo
Searchable archive at http://www.mail-archive.com/r-

[Liam J. Revell]

Hi Szymon.

I don't know about BayesTraits; however it is possible to do marginal
ancestral state reconstruction using multiple functions in R (e.g., ee
my blog:
http://blog.phytools.org/2013/03/marginal-ancestral-state-reconstruction.html).
If the same model is used, these different approaches should give the
same probabilities at internal nodes. However, a word of caution. With
such a very large tree, it's possible that different implementations of
the same method may give different results simply for having failed to
converge to the MLEs.

All the best, Liam

Liam J. Revell, Assistant Professor of Biology
University of Massachusetts Boston
web: http://faculty.umb.edu/liam.revell/
email: ***@umb.edu
blog: http://blog.phytools.org

Post by Szymek Drobniak
Hello, thank you Liam for your quick answer, I might have phrased my
question not very clearly. The point is that both simmap and
rerootingMethod (using bort EM and SYM) give almost identical results
(there are some slight differences which is obvious but the overall
picture - and the root ancestral state - are the same). I was wondering
why BayesTraits (which, at least in theory, uses similar approach) gives
different result than phytools - in BayesTraits the transition rates are
entirely different, and the ancestral state of the root is different. I
was wondering that maybe somebody had more experience with BayesTraits
and came across this difference between BayesTraits and some other
packages. As for the BT analysis - I used both ML and MCMC methods, in
the latter I used a flat exponential hyperprior. For the time being I
used a consensus tree rather than a sample from the distribution of trees.
Cheers
sz.
Hi Szymon.
Can you give us some more details on the analyses that you ran?
Marginal ancestral states (from rerootingMethod) and posterior
frequencies (from make.simmap, using describe.simmap to summarize
the results across stochastic maps) tend to be very highly
correlated (although they should not expected to be exactly the same
as give marginal ancestral states and the other samples sets of
states from their joint probability distribution). There are several
possible reasons why ancestral states might be different. For
instance, the default substitution model in rerootingMethod is "ER"
(equal-rates) whereas the default model in make.simmap is "SYM"
(symmetrical rates). In addition, the methods can use different
prior probabilities for the root. By default, both functions use a
flat prior; however make.simmap can also use the stationary
distribution given the transition matrix Q, or an arbitrary prior.
Finally, make.simmap can use the MLE of Q, an arbitrary value of Q,
or sample Q from its posterior probability distribution conditioned
on a model. All of these different options could potentially result
in different ancestral reconstructions.
All the best, Liam
Liam J. Revell, Assistant Professor of Biology
University of Massachusetts Boston
web: http://faculty.umb.edu/liam.__revell/
<http://faculty.umb.edu/liam.revell/>
blog: http://blog.phytools.org
Hello all,
I'm trying to reconstruct ancestral states in a 3-state trait evolving on a
big (>2000 tips) phylogeny. I've tried both BayesTraits and tools from Liam
Revell's phytools (continuous time Markov model and stochastic mapping)
package. Surprisingly (or maybe not?) both give different answers,
especially with respect to the ancestral states of deep nodes (including
root node) and transition rates. Is it something I should expect - do
anyone have some idea why these are so different?
Cheers
szymon
_________________________________________________
https://stat.ethz.ch/mailman/__listinfo/r-sig-phylo
<https://stat.ethz.ch/mailman/listinfo/r-sig-phylo>
Searchable archive at
--
*
*
*Dr Szymon Drobniak*
Anthropological Institute and Museum
Office Y42-K-66
University of Zürich - Irchel
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland