1. Estimating phylogenetic trees from
discrete morphological data
Thesis Defense
April M. Wright
April 14, 2015
Supervising Professor: David M. Hillis
1
2. 2
Bayesian Analysis
Using a Simple
Likelihood Model
Outperforms
Parsimony for
Estimation of
Phylogeny from
Discrete
Morphological Data
Modeling character
change heterogeneity
through the use of
priors
Use of an
Automated
Method for
Partitioning
Morphological
Data
"Bayes' Theorem MMB 01" by mattbuck, PartitionFinder logo by Ainsley Seago
5. Why Morphology?
● Estimates put > 99% of biota that has ever existed as extinct
● Probably, we won’t wring DNA from a stone
5
6. Why Morphology?
● Estimates put > 99% of biota that has ever existed as extinct
● Probably, we won’t wring DNA from a stone
● Inclusion of fossils acknowledged to improve phylogenetic trees
(Huelsenbeck 1991, Wiens 2001 & 2004), divergence dating (Heath, Stadler and Huelsenbeck
2012) and comparative method estimates (Slater and Harmon 2012)
6
9. Image:WikimediaCommons
● Smaller
● Selection bias
● Morphology has to be
interpreted
● Sequence data sets large
● Take the whole sequence
● Characters have more
clearly-defined properties
11. Image:WikimediaCommons
● Smaller
● Morphology has to be
interpreted
● ...not so much
● DNA data sets large
● Clearly-defined properties
● Explicit, well-developed
models for sequence
evolution
14. Chapter One: Bayesian Analysis Using a Simple
Likelihood Model Outperforms Parsimony for Estimation of
Phylogeny from Discrete Morphological Data
14
15. Likelihood models
● There is one practical published model for
phylogenetic estimation from discrete
morphological data, Mk
15
16. Likelihood models
● There is one practical published model for
phylogenetic estimation from discrete
morphological data, Mk
○ Like all methods, Mk has assumptions
16
17. Likelihood models
● There is one practical published model for
phylogenetic estimation from discrete
morphological data, Mk
○ Like all methods, Mk has assumptions
■ Change can occur at any instant along a branch
■ Change is symmetrical between states
17
18. Likelihood models
● There is one practical published model for
phylogenetic estimation from discrete
morphological data, Mk
○ Like all methods, Mk has assumptions
■ Change can occur at any instant along a branch
■ Change is symmetrical between states
● This model is statistically consistent
18
19. Likelihood models
● There is one practical published model for
phylogenetic estimation from discrete
morphological data, Mk
○ Like all methods, Mk has assumptions
■ Change can occur at any instant along a branch
■ Change is symmetrical between states
● This model is statistically consistent
○ Caveat: As long as the assumptions hold
○ Also, we don't have infinite data 19
20. Does a parametric approach (Mk)
outperform non-parametric approach
(parsimony) when model assumptions are
violated?
20
27. ● Some types of characters are more likely to be missing
from fossil taxa
● Missing characters may be more likely to fall into certain
rate classes
Char. 1 Char. 2 Char. 3 Char. 4 Char. 5 Char. 6 Char. 7 Char. 8 Char. 9
Species 1
(Fossil)
0 1 1 0 ? ? ? ? ?
Species 2
(Fossil)
1 0 1 0 ? ? ? ? ?
Species 3 1 0 1 0 0 0 1 1 0
Species 4 0 0 0 1 0 1 0 0 0
Species 5 0 0 0 1 0 1 1 0 0
28. Missing data
● In these conditions, some model
assumptions have been violated
● Given these model violations, is it preferable
to use parsimony?
28
32. A simulation framework
● Simulate characters
○ 350 & 1000 character data sets
● Estimate topology using the Mk model and
parsimony
32
33. Rate heterogeneity
● Different rates of character evolution
○ Low rates of change mean each character is likely to
have changed rarely, if at all
○ High rates mean there are likely reversals and
parallel evolution in the data
33
36. Rate heterogeneity
● Diversity of rate classes can be helpful for
resolving different regions of phylogenetic
trees
○ Likelihood models account for superimposed
changes
36
37. Rate heterogeneity
● Diversity of rate classes can be helpful for
resolving different regions of phylogenetic
trees
○ Likelihood models account for superimposed
changes
○ In analysis, rate heterogeneity is often modeled as
gamma-distributed
37
43. Summary - Chapter One
Does a parametric approach (Mk)
outperform non-parametric approach
(parsimony) when model assumptions are
violated?
43
44. Summary - Chapter One
Does a parametric approach (Mk) outperform
non-parametric approach (parsimony) when
model assumptions are violated?
Yes
44
45. Summary - Chapter One
Caveats: We’ve really only looked at one type
of model violation here
There are other reasons you might use
parsimony, or might think the contrast of
likelihood and parsimony methods is telling you
something interesting
45
50. Model Assumptions
● Change is symmetrical between states
○ We know this is not always true
What if we could relax this assumption?
50
51. Relaxing this assumption
● In Bayesian estimation, we can put priors on
the parameters in our analyses
51
52. Relaxing this assumption
● In Bayesian estimation, we can put priors on
the parameters in our analyses
● Transition probabilities are the product of
exchangeabilities and frequencies
52
68. Empirical Datasets
● 206 datasets
○ 5 to 279 taxa
○ 11 to 364 characters
○ Biased towards vertebrates
68
69. Empirical Datasets
● 206 datasets
○ 5 to 279 taxa
○ 11 to 364 characters
○ Biased towards vertebrates
● Modeled character change asymmetry
according to the 6 distributions
69
74. Simulations
● Simulated data according to 4 distributions
○ Modeled the data according to each of the four
distributions
○ One generating model, 3 misspecified models
74
77. Simulations
● Simulated data according to 4 distributions
○ Modeled the data according to each of the four
distributions
○ One generating model, 3 misspecified models
● Also simulated missing data
77
78. Simulations
● Estimated trees according to each of the 4
values of alpha
● Used Bayes Factor model selection to
choose the best-fit value
● Used Robinson-Foulds distance to assess
topological correctness
78
79. Simulations
Can we detect the generating value of alpha
among misspecified values of alpha?
Does using the correct alpha result in a more
correct tree?
79
84. Empirical Datasets
Which priors best match empirical data?
About half: best fit is the α = ∞ prior
Strength of support for different values of α
varies
84
98. Simulations
Does using the best-fit α result in a more
correct tree?
Yes, and the importance of doing so is greater
when the problem is harder
98
99. Chapter Two: Conclusions
● Appropriate fit of α parameter improves
phylogenetic estimation
● Bayes Factor model selection performs well
at choosing the best-fit value of α among a
set of α values
99
100. Chapter Three: Use of an Automated Method for
Partitioning Morphological Data
100
101. Partitioning
● Refers to breaking a dataset into smaller
subsets that can be analyzed under different
phylogenetic models
101
102. Partitioning
● Refers to breaking a dataset into smaller
subsets that can be analyzed under different
phylogenetic models
○ Well-explored in a molecular context
102
103. Partitioning
● Refers to breaking a dataset into smaller
subsets that can be analyzed under different
phylogenetic models
○ Well-explored in a molecular context (Brown and Lemmon
2007)
○ Often, partition schemes are tested as a stage in
model-fitting
103
105. Partitioning
● Less well-explored in morphology
○ Clarke and Middleton (2008) is one of the few
explorations of partitioning in a likelihood context for
morphology
105
106. Partitioning
● Less well-explored in morphology
○ Clarke and Middleton (2008) is one of the few
explorations of partitioning in a likelihood context for
morphology
○ Used anatomical subregion partitioning
106
107. Partitioning
● Less well-explored in morphology
○ Clarke and Middleton (2008) is one of the few
explorations of partitioning in a likelihood context for
morphology
○ Used anatomical subregion partitioning
○ Found improved model fit and different topology with
partitioned data
107
108. Partitioning
● Less well-explored in morphology
○ Clarke and Middleton (2008) is one of the few
explorations of partitioning in a likelihood context for
morphology
○ Used anatomical subregion partitioning
○ Found improved model fit and different topology with
partitioned data
108
113. PartitionFinder Morphology
● Estimate a phylogenetic tree from the
unpartitioned data matrix
● Fit parameters of the evolutionary model to
the whole dataset as a single set of sites
113
114. PartitionFinder Morphology
● Estimate a phylogenetic tree from the
unpartitioned data matrix
● Fit parameters of the evolutionary model to
the whole dataset as a single set of sites
● Calculate the score of the data given this
model according to an information theoretic
criterion (AIC, BIC or AICc)
114
116. PartitionFinder Morphology
● Generate rates of evolution for each site in
the dataset
● Use k-means clustering to split the subset in
two based on these rates
116
117. PartitionFinder Morphology
● Generate rates of evolution for each site in
the dataset
● Use k-means clustering to split the subset in
two based on these rates
● Fit parameters of the model for these new
subsets
117
118. PartitionFinder Morphology
● Calculate the score of this new partitioned
data matrix according to the same
information theoretic criterion used in step 3.
118
119. PartitionFinder Morphology
● Calculate the score of this new partitioned
data matrix according to the same
information theoretic criterion used in step 3.
● If smaller subsets are supported by this
criterion, continue to divide them, repeating
steps 5-7. If not, terminate the search.
119
121. PartitionFinder Morphology
Are partitioned models often the best-fit model
for empirical datasets?
When they are, does this make a difference to
the tree estimated?
121
122. Modeling
● We used PartitionFinder Morphology to
partition 209 datasets
○ 3 criteria: AIC, BIC and AICc
2k- 2lnL -2 (lnL + 2k * n-k-1)
-2lnL + k * lnn
n
122
123. Modeling
● We used PartitionFinder Morphology to
partition 209 datasets
○ 3 criteria: AIC, BIC and AICc
Least conservative
Most conservative
123
124. Estimation
● Estimate trees under likelihood and
Bayesian implementations of the Mk model
using the partitioned data and unpartitioned
data
124
132. When a partitioned model is the best fit, does
this make a difference to the tree estimated?
Yes, in empirical datasets we often estimate
different trees.
132
138. Conclusions
● Chapter One: Likelihood-based methods are
effective for estimating phylogeny for
morphological data, even in the presence of
biased missing data
138
139. Conclusions
● Chapter Two: Use of a prior on equilibrium
state frequencies can improve the
performance of Bayesian estimation using
morphological data
139
140. Conclusions
● Chapter Three: PartitionFinder Morphology
is a promising lead for evaluation of
partitioning schemes
140
141. Thank you!
Committee
David Hillis
Martha Smith
David Cannatella
Randy Linder
Bob Jansen
Labmates
Ben, Emily Jane, Thomas, Patricia,
Becca, Mariana, Shannon, Patrick, Chris,
J9, Matt, Sandi, Carlos, Taylor, Katie,
Devon, Anne, Jim, Jeremy, Tracy
Other
UT vert paleo, especially Julia, Robert,
Zach
And the people who make
this worth doing:
You know who you are.
And Jason and unnamed baby girl Wright Stinnett.
Collaborators:
Graeme Lloyd, Paul Fransden, David Bapst, Nick
Matzke, Matt Brandley, Rob Lanfear
141