Claremont Colleges

Scholarship @ Claremont
Scripps Senior Theses Scripps Student Scholarship

4-9-2012

Constructing Phylogenetic Trees Using Maximum

Likelihood
Anna Cho
Scripps College

Recommended Citation
Cho, Anna, "Constructing Phylogenetic Trees Using Maximum Likelihood" (2012). Scripps Senior Theses. Paper 46.
http://scholarship.claremont.edu/scripps_theses/46

This Open Access Senior Thesis is brought to you for free and open access by the Scripps Student Scholarship at Scholarship @ Claremont. It has been
accepted for inclusion in Scripps Senior Theses by an authorized administrator of Scholarship @ Claremont. For more information, please contact
[email protected].
Constructing Phylogenetic Trees using
Maximum Likelihood

Anna Cho

Christopher Towse, Advisor

Ami Radunskaya, Reader

Submitted to Scripps College in Partial Fulfillment

of the Degree of Bachelor of Arts

March 9, 2012

Department of Mathematics
Abstract

Maximum likelihood methods are used to estimate the phylogenetic trees

for a set of species. The probabilities of DNA base substitutions are mod-
eled by continuous-time Markov chains. We use these probabilities to es-
timate which DNA bases would produce the data that we observe. The
topology of the tree is also determined using base substitution probabili-
ties and conditional likelihoods. Felsenstein [2] introduced this method of
finding an estimate for the maximum likelihood phylogenetic tree. We will
explore this method in detail in this paper.
Contents

Abstract iii

Acknowledgments vii

1 The Stochastic Process of DNA Base Substitutions 1

1.1 Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Evolution and DNA mutations . . . . . . . . . . . . . . . . . 4
1.3 Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 The Memoryless Property . . . . . . . . . . . . . . . . . . . . 8
1.5 Stationary Distributions . . . . . . . . . . . . . . . . . . . . . 10

2 The Likelihood of a Phylogenetic Tree 13

2.1 Computing the Likelihood of a Phylogenetic Tree . . . . . . 13
2.2 Time Reversibility . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 The Pulley Principle . . . . . . . . . . . . . . . . . . . . . . . 19

3 Finding a Maximum Likelihood Tree 23

3.1 How Many Possible n-species Trees are There? . . . . . . . . 24
3.2 Which Tree Topology? . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Maximum Likelihood Branch Lengths . . . . . . . . . . . . . 31
3.4 Putting Topologies and Branch Lengths Together to find the
Maximum Likelihood Tree . . . . . . . . . . . . . . . . . . . . 34

4 Finding the Maximum Likelihood Tree - An Example 37

5 The Limitations of Our Model 45

Bibliography 49
Acknowledgments

I’d like to thank my parents for their love and support. I wouldn’t have
gotten to the point I am at today without them. I’d also like to thank my
advisor, Professor Towse, for all of the advice and emotional support that
he offered me on my thesis. Thank you for for your energy and helping me
see the beauty of math. I’d also like to thank my reader, Professor Radun-
skaya, for all of her brilliant advice. I would have been lost without your
help.
Chapter 1

The Stochastic Process of DNA

Base Substitutions

More and more DNA sequences are being analyzed today than ever be-
fore. With this increase in the accumulation of DNA sequences comes a
demand for the study of the ancestry of organisms and their phylogenetic
trees. Scientists are interested in how closely related one species is to an-
other. Studying phylogenetic trees and the evolutionary processes that they
model allow scientists to gain a better understanding of how organisms
have arrived at the state they are in today. Phylogenetic trees give us the
ability to see how species evolve and adapt throughout different time peri-
ods with different conditions and needs. Studying these evolutionary pro-
cesses is clearly important to the advancement of biology, but finding the
correct phylogenetic tree for a set of related species is very difficult consid-
ering that we are only given the data that we can observe today, namely the
DNA sequences of those species. We have overcome much of this difficulty
using statistical inference. Statistical models and Markov models allow us
to estimate how similar a phylogenetic tree is to the actual, unknown phy-
logenetic tree for a given set of DNA sequences. We use the maximum like-
lihood method to infer what the true phylogenetic tree of our set of data
looks like. Maximum likelihood uses an explicit evolutionary model. We
assume that the data we observe is identically distributed from this model.
Before defining maximum likelihood, we review some of the terminol-
ogy used in this statistical approach. We have been using the term phylo-
genetic tree to indicate a branching diagram describing a set of species and
their common ancestors. The terms phylogenetic tree and evolutionary tree
are often used interchangeably in the field of computational biology. In
2 The Stochastic Process of DNA Base Substitutions

this paper, we will be using the term phylogenetic tree exclusively. The set
of data, namely the DNA sequences of the species we are observing, will
be at the tips of the phylogenetic tree. The internal nodes of the tree repre-
sent the DNA sequences of the ancestors of the species we are examining.
The segments of the diagram that connect one DNA sequence to another
are called the branches of the tree. Finally, the root of the tree represents the
DNA sequence of the sole common ancestor of all of the species we observe
in our set of data. Each species at the tip of the tree can be traced back to
this common ancestor at the root of the tree. Figure 1.1 shows an example
of a phylogenetic tree.

Figure 1.1: An example of a phylogenetic tree. Note: This tree represents

the phylogenetic tree for one site in the DNA sequences (i.e., the DNA base
of each species’ DNA sequence that is located at the same place).

We refer to the tree shape or the way in which the tips, nodes, and root
are connected by branches as topology.It is important to note that this use
of the word topology is different from the branch of mathematics known
as topology. To topologists, each phylogenetic tree would have the same,
trivial topology and would be indistinguishable. In evolutionary biology,
two topologies are considered different from one another if one topology
cannot be cannot be recreated from the second topology without disassem-
bling a connection between two nodes or between a node and a tip. Fig-
ure 1.2 shows both an example of equivalent topologies and an example of
topologies that are different from one another.
 Likelihood 3

Figure 1.2: Tree A and Tree B have equivalent topologies. Tree C has the
same number of tips as Tree A and Tree B, but the branches connecting the
two middle tips cannot be changed to look like Tree A and Tree B without
detatching them. Thus, Tree C has a different topology from Tree A and
Tree B.

1.1 Likelihood
The process of finding a phylogenetic tree using maximum likelihood in-
volves finding the topology and branch lengths of the tree that will give us
the greatest probability of observing the DNA sequences in our data. After
each step, we take the likelihood of each tree that we examine. The tree
that gives us the largest likelihood is then chosen to be examined in the
next step. We will describe this process in more detail in Chapters 2 and 3.

Definition 1. The likelihood of a set of data, D, is the probability of the data,

given a hypothesis, θ. The hypothesis will usually come in the form of
different parameters. We denote the likelihood, L, of a set of data, D, as
L = P (D | θ).

This definition seems quite simple, but we also need to be careful not to use
the term, likelihood as we otherwise would in English. When we say, ”the
likelihood of a phylogenetic tree,” we are not referring to the probability of
seeing that particular tree. Rather, we are referring to the probability of see-
ing the DNA sequences that we have in front of us, given that phylogenetic
tree.
In order to see how likelihood works, we will first consider a very sim-
ple example.

Example 1. Suppose there is a gameshow that uses a robot to decide whether

or not a contestant will win a prize of $5,000. If the robot raises its left arm,
then the contestant wins the prize, and if the robot raises its right arm the
contestant goes away with nothing. The robot is controlled by a computer
4 The Stochastic Process of DNA Base Substitutions

that randomly chooses the probability with which the robot will raise its
left arm for the entire season of the game show. Out of ten episodes of
the gameshow, six contestants left with a prize while 4 contestants went
away with nothing. One ambitious frequent viewer of the show is inter-
ested in finding the likelihood of the robot’s decision to raise its left arm
for this season. Let p be the probability that the robot raises its left arm,
and let X be the proportion of times the robot raises its left arm. One hy-
pothesis we could consider is that the robot is fair. Then, the likelihood that
p = 0.5 is P (X = 3/5 | p = 0.5). Clearly, this probability is less than one.
Another hypothesis we could consider is that the robot has a 3/5 proba-
bility of raising its left arm and a 2/5 probability of raising its right arm.
The likelihood of this probability is P (X = 3/5 | p = 3/5) = 1. Then,
P (X = 3/5 | p = 0.5) < P (X = 3/5 | p = 3/5). The viewer’s best guess is
that the probability of the robot raising its left arm is 53 .

Similarly, when inferring phylogenetic trees using maximum likelihood,

we are searching for the tree that gives the highest probability of producing
the observed DNA sequences. For another example see [3], page 249.

1.2 Evolution and DNA mutations

Before we go any further into detail about phylogenetic trees, we describe
the evolutionary process of DNA mutations that phylogenetic trees rep-
resent. DNA is short for deoxyribonucleic acid, and this genetic material
makes up the chromosomes that give organisms the characteristics they
have. DNA is made up of nucelotides. The four different kinds of nu-
cleotides are distinguished by their different nitrogenous bases: adenine,
guanine, cytosine, and thymine. We will denote these four different bases
with A, G, C, and T, respectively. When an organism has offspring, its DNA
is replicated and passed on to its offspring.
During DNA replication, changes in the DNA, which we call mutations,
can occur. The changes in the genes that the offspring inherits gives them
a different phenotype from their parents. The changes in the DNA occur at
the level of the bases; different bases are substituted for the bases that the
parents’ DNA originally had. These alterations change the characteristics
of the offspring and eventually, after a few generations have passed, may
lead to the production of a new species [8].
In this paper, we look at the DNA sequences of a set of species and
use maximum likelihood methods to determine the how closely related the
 Stochastic Processes 5

species are to each other. We model the probability of a DNA base substi-
tution as a continuous time Markov process (see [4]). We will describe this
in more detail in the following section.

1.3 Stochastic Processes

Recall that for a specific site in a DNA sequence, the bases (A, C, T, or G)
observed in each sequence of DNA in our set of data are placed at the tips of
the phylogenetic tree. In order to compute the likelihood of a phylogenetic
tree producing the set of bases seen in our data, we must have a probability
model describing the event of a mutation from one base to another. In other
words, for each site in a DNA sequence, we must find a model describing
the evolution of a site from a previous DNA sequence into a base we see at
the tip of our tree.
The next issue we encounter is the difficulty of putting all of the DNA
sites together to form the DNA sequences that we are actually concerned
with in our data. In order to make the computation of a tree’s likelihood
more feasible, we assume that the probability of mutation in one site of a se-
quence of DNA is independent of the probability of mutation in another site
of that sequence of DNA. This means that given any information about site
n, the probability of site m having state s is not effected by the information
we have received about site n. The converse is also true.

Definition 2. We say that event A is independent of event B when the occur-

rence of B gives no information and does not change the probability that A
will occur. In other words,

P (A | B) = P (A).

By the definition of conditional likelihood, P (A | B) = P (A ∩ B)/P (B).

When A is independent of B, this is equivalent to

P (A ∩ B) = P (A)P (B).
This allows us to compute the likelihood of a given set of DNA se-
quences one site at a time. Once we have computed the likelihoods of each
site of the sequences, the product of the likelihoods of each individual site
gives us the likelihood of the set of DNA sequences as a whole. Thus, the
bulk of the maximum likelihood method lays in finding the optimal phylo-
genetic tree for each single site of DNA.
6 The Stochastic Process of DNA Base Substitutions

For a single site of DNA, we denote the probability of a site in state i

evolving to state j in t units of time as Pij (t). States i and j can be either 1,
2, 3, or 4, which correspond to the bases A, C, G, or T, respectively. We say
that the state space, S = {1, 2, 3, 4}. When a site undergoes a mutation, its
state is called a random variable in a stochastic process [5].

Definition 3. A random variable is a real-valued function of the outcome of

an experiment, or a process described by a probablistic model.

Definition 4. A probability mass function (PMF), denoted pX , of a random

variable, X, is the function pX (x) = P (X = x).

Example 2. Suppose that at site n of the DNA sequence of a unicorn there

are equal probabilities of finding base A, C, T, or G. Just as we did earlier,
we will represent A, C, T, and G with 1, 2, 3, and 4, repectively. Then, we are
applying a real-valued function to the possible outcomes for the base found
at site n. The base we find at site n is a random variable which we will call
X. The PMF for X is pX (x). We have that pX (1) = 0.25, pX (2) = 0.25,
pX (3) = 0.25, and pX (4) = 0.25.

Definition 5. A stochastic process is an indexed family of random variables.

In other words, it is a set of random variables whose values are assigned
a t from the index, T . For each t ∈ T , the state of the stochastic process at
step t of the process is a random variable denoted by X(t). We denote a
stochastic process by {X(t) : t ∈ T }.

When the indexing parameter, t, can take on a continuous range of values,

we call the process a continuous-time process. The random variable X(t) can
take on a continuous or a discrete range of values.

Example 3. In the case of base substitutions, at any time, t, during the evo-
lutionary process of a site of DNA, the state of the site can take on the
random variable, 1, 2, 3, or 4. The process of substituting bases is in-
dexed by the time, T , which has a continuous range. Thus, the process
is a continuous-time process. The stochastic process, {X(t) : t ∈ T } can
take on the values 1, 2, 3, and 4 at each time, t, in our index, T .

The probability of a base substitution occuring at time t, which we de-

note as Pij (t), is the probability that a base, i, will undergo a mutation and
be substituted by base, j, at time, t. Again, the states, i and j, are random
variables, and the index of these random variables is time, T . This reflects
a process with the Markov property: given the present state, the future states
 Stochastic Processes 7

are not dependent on past states. This means that a future state only de-
pends on the present state, X(s). It is not affected by any past state, X(u),
where 0 ≤ u < s. Since there are not any discrete indicators of when a base
will mutate and the index, T , takes on a continuous range, the process of
base substitution is a continuous-time Markov process, as described below.
Definition 6. A stochastic process, {X(t) : t ≥ 0}, is called a continuous-time
Markov processes when it posesses these properties:
1. Each event of the process is independent of previous events. (In other
words, the process has the Markov property).
2. When entering a state, the process will stay in that state for a random
amount of time before transitioning to another state.
We denote the probability that a continuous Markov process currently in
state i will be in state j after t time units as
Pij = P (X(t + s) = j | X(s) = i).
Matrices are used to represent Markov processes with rows representing all
possible current states and columns representing all possible future states.
Each term, Pij , of a matrix representing a Markov process represents the
probability of the process moving from state i to state j in a span of time, t.
Example 4. Suppose the robot in Example 1 works without any problems
for a period which is exponentially distributed with parameter, λ. Then, it
breaks down and a replacement robot will have to substitute for the bro-
ken robot on the gameshow for a period which is exponentially distributed
with parameter, µ. Since the time that each robot spends on the gameshow
is exponentially distributed with different parameters, the times that each
robot spends on the show are independent of each other. Let X(t) = 1 if the
robot is working at time t. Let X(t) = 2 if a replacement robot is being used
at time t. Then, {X(t) : t ≥ 0} is a continuous-time Markov process with
each event being independently distributed and the following transition
probabilities: P11 = P22 = 0, P12 = P21 = 1.
As we stated earlier, sometimes, we represent base substitution proba-
bilities in a matrix called a transition matrix. The rows of the matrix repre-
sent the initial state, and the columns represent the state after a base sub-
stitution has occurred. Our transition matrix for this problem would be:
   
P11 P12 0 1
Pij = = .
P21 P22 1 0
8 The Stochastic Process of DNA Base Substitutions

1.4 The Memoryless Property

Suppose that a continuous-time Markov chain enters state i at time 0. Next,
suppose that the process does not leave state i in the following ten minutes.
Then, what would the probability of the process remaining in state i in the
following 20 minutes (30 minutes total) be? Since we have a continuous-
time Markov chain, the process has the Markov property. Thus, the prob-
ability of the process remaining in state i from the time interval [10, 30] is
simply the probability that it stays in state i for 20 minutes. In other words,
if we let Ti represent the amount of time that the process remains in state i
before transitioning into a different state, then

P (Ti > 30 | Ti > 10) = P (Ti > 20)

and in general,

P (Ti > s + t | Ti > s) = P (Ti > t)

for all s and t ≥ 0. We say that a random variable X with this property is
memoryless. The process’s determination of what state it is in at time t + s
is not affected by the state it was in at time t for all 0 ≤ t < s; it does not
remember its prior states. If we restate this property using the definition of
conditional properties, we have

P (X > s + t ∩ X > t)
P (X > s + t | X > t) = = P (X > s + t)
P (X > t)

which is equivalent to

P (X > s + t ∩ X > t) = P (X > s + t)P (X > t). (1.1)

We will use a stochastic process called the Poisson process in conjunction
with Markov processes to model the base substitution probability Pij (t).

Definition 7. A stochastic process is a Poisson process with rate, λ for some

λ ≥ 0, if it posesses these properties:

1. At time t = 0, the number of events that have occurred is 0.

2. The time increments t ∈ T are independent of each other.

3. The number of events in any interval of length t is a random variable

with a Poisson distribution with mean λt.
 The Memoryless Property 9

If the Poisson process has N (t) events at time t, then we represent the process
by:
(λt)n
P (N (t + s) − N (s) = n) = e−λt , for n = 0, 1, . . . .
n!
We will assume that in a small interval of time of length dt, there is a
probability µdt that the current base at a site will transform, where µ is the
rate of base substitution per unit of time. It is clear that at time t = 0, there
will have been 0 base substitutions. Thus, the first property of our defini-
tion of a Poisson process is fulfilled. The probability of transitioning bases,
µdt, is the same for all intervals of time with length dt. This is the second
property of a Poisson process. Further, the number of base substitutions
during a particular time interval is independent of the history of changes
outside of this interval, so each time increment is independent from other.
time increments Finally, the probability of a change in base in a time in-
terval is very small and the number of base changes can be modeled by
a Poisson distribution, satisfying the third property of a Poisson process.
Thus, the process of changing from the current base to another base in a
time interval dt is a Poisson Process. If we let N (t) be the number of tran-
k
sitions from base i to base j at time t, then P (N (t) = k) = e−µt (µk)
k! . Then,
P (N (t) = 0) = e−µt and P (N (t) > 0) = 1 − e−µt . If we let X be the time
of the first event, then the probability of the first event occuring before a
time t > 0 is P (X ≤ t) = P (N (t) > 0) = 1 − e−µt . The complement of this
probability, the probability that no mutation occurs before a time t > 0, is
1 − (1 − e−µt ) = e−µt . Then,

Pij (t) = e−µt δij + (1 − e−µt )πj (1.2)

where πj is the probability that a mutation will result in the current base
being replaced with base j [2]. The δ is taken from the Kronecker delta
formula in which (
1 if i = j
δij =
0 if i 6= j
In other words,
(
e−µt + (1 − e−µt )πj if i = j
Pij (t) = .
(1 − e−µt )πj if i 6= j
10 The Stochastic Process of DNA Base Substitutions

1.5 Stationary Distributions

When determining the likelihood of the root of a phylogenetic tree, it is
helpful to know the proportion of base substitutions that results in a spe-
cific base. In this section, we will be examining general ideas about Markov
models. We will see that for some continuous Markov chains, a limiting
probability, limt→∞ Pij (t), which we denote, πj , exists. This means that as
time goes on to infinity, the probability of switching from state i to state j
in time t approaches a constant.
The stationary distribution of an ergodic Markov process, πi , is also
interpreted as the proportion of time the process is in state i.

Definition 8. Ergodicity is the property that the limiting probability limt→∞ Pij (t)
exists and is independent of the initial state, i. We call a Markov process
with this property ergodic.

Definition 9. Let {X(t) : t ≥ 0} be a continuous Markov process with state

space S and transition probability matrix P . A state j is said to be accessible
from state i if there is a positive probability that starting from i, the Markov
process will be in state j after a finite time. In other words, Pij (t = n) > 0,
for some n < ∞.

Example 5. At the Pacific Science Center, there is an exciting exhibit that

allows visitors to learn about Markov chains. When the line for the exhibit
is empty, there is a positive probability that after n time units, the line will
have one visitor. Then, if we represent the state of the line when it is empty
with 0 and when it has one visitor with 1, we have that P01 (t = n) > 0.
Thus, 1 is accessible from 0.

Definition 10. Let {X(t) : t ≥ 0} be a continuous Markov process with

state space S and transition probability matrix P . If i, j ∈ S, and i and j are
accessible from each other, then we say that i and j communicate.

Example 6. Suppose that in Example 5, there is also a positive probability

that when the line has one visitor, there will be no visitors in line after m
time units. Then, we have that P10 (t = m) > 0 along with P01 (t = n) > 0.
Thus, 0 and 1 communicate.

Definition 11. If all of the states of a Markov process communicate with

each other, then we call the Markov process irrreducible.

Example 6 is an example of an irreducible Markov process.

 Stationary Distributions 11

Definition 12. Let {X(t) : t ≥ 0} be a continuous Markov process with

state space S and transition probability matrix P . Let Pii (n) be the proba-
bility, that starting from state i, the process will return to state i, for the first
time, after finite time n. Let pi be the probability that, starting from state i,
the process
P∞ will return to state i after a finite number of transitions. Then,
Pi = n=1 Pii (n). If Pi = 1, then the state i is called recurrent.

Example 7. In the Pacific Science Center problem presented in Example 5

and Example 6, It is clear that P01 < 1 and P10 < 1. This implies that
there is a positive probability, that starting with an empty line, the line will
remain empty for a finite amount of time, n. In other words, P00 (n) > 0.
There is also a positive probability that after starting with an empty line,
and having one visitor in the line, the line will be empty again after a finite
amount of time, m. Then, P11 (m) > 0. Thus, at each time that the line is in
state 0, after a finite amountPof time, there is a chance that it will return to
being in state 0. Then, Pi = ∞ n=1 Pii = 1 and state 0 is recurrent.

Definition 13. Let i be a recurrent state of a Markov process. The state i is

called positive recurrent if the expected amount of time between two consec-
utive returns to i is finite.

The Pacific Science Center examples are all positive recurrent since the
line will not stay in state 0 or state 1 for an infinite amount of time. For
a nonexample, consider the robot gameshow example once more. Sup-
pose that after a robot malfunctions, the gameshow completely replaces the
robot forever. Then, if we let 1 denote the state that robot 1 is being used
on the gameshow, the expected amount of time between two consecutive
returns to state 1 is infinite. We never see robot 1 again since it is replaced.
Therefore, state 1 is not positive recurrent in this case.
Consider a continuous Markov process, {X(t) : t ≥ 0}, with state space,
S. Suppose that Pij > 0 for each i, j ∈ S. Restated, this means that for
all i, j ∈ S, the states, i and j, are accessible to each other. Thus, all of
the states in S communicate with each other, making the Markov process
irreducible. Also, suppose that, starting from state i, the process will return
to state i with probability 1, and the expected number of transitions before
a first return to i is finite. This means that the Markov process is recurrent.
An irreducible continuous Markov process in which each state is positive recurrent
is ergodic [5].
To understand the reasoning behind this argument we will consider
a continuous Markov process. For a continuous Markov process, we call
the limiting probabilities for transitions into state j, denoted πj , stationary
12 The Stochastic Process of DNA Base Substitutions

distributions. We want to show that πj = limt→∞ Pij (t). Suppose that the
Markov process is reducible, and that there are two irreducible partitions of
S, say s1 and s2 . Let j be an element of s2 . Also, suppose that limt→∞ Pij (t)
converges to a limiting probability for an i in s1 . But since none of the
elemetns of s1 are accessible from s2 and vice versa, Pij (t) = 0 for all t ≥ 0.
This means that depending on whether the initial state i is in s1 or s2 , the
limiting probabilities could differ. The limiting probabilities would not be
independent of state i, a necessary property for an ergodic Markov process.
Therefore, the Markov process will not be ergodic. The Markov process
must also be recurrent. Otherwise, if i is the initial state, Pii (t) = 0 for all
t ≥ 0 and limt→∞ Pii (t) = 0. Again, the limiting probability depends on
whether or not the initial state is i. Hence, the Markov process needs to be
irreducible, and each of its states must be positive recurrent.

1.5.1 Base Frequencies

Our model for base substitution probabilities represent a continuous Markov
process. We will assume that this Markov process is ergodic, so as t ap-
proaches ∞, the probability that the DNA site is in some state, j, is non-
zero and independent of the starting state, i. In other words, there are pos-
itive values, π1 , π2 , π3 , and π4 , such that, for all i and j in our state space
S = {1, 2, 3, 4},

limt→∞ Pij (t) = πj .

Remember that we are representing DNA bases A, C, T, and G here with 1,
2, 3, and 4, respectively. Furthermore, for all t ≥ 0, these values satisfy
X
πj = πi Pij (t). (1.3)
i∈S

We call πj the base frequency of base j because it represents the proportion

of base substitutions that result in base j.
Chapter 2

The Likelihood of a
Phylogenetic Tree

When calculating likelihood in Example 1, we did not know the probability

with which the robot would raise its left or right arm. Similarly, we do not
know the probability with which a base will mutate. Thus, calculating the
likelihood of a tree is equivalent to finding the probability of the data we
see at the tips of our tree given a hypothesis for the shape and bases of
the nodes in the tree. In this chapter, we will focus on describing how to
calculate the likelihood of a phylogenetic tree for a set of species.

2.1 Computing the Likelihood of a Phylogenetic Tree

Figure 2.1: 2-Species Phylogenetic Tree.

Let us first consider a tree with tips, 1 and 2, root, 0, and branch lengths,
14 The Likelihood of a Phylogenetic Tree

v1 and v2 , as in Figure 2.1. If the state, S0 , of node 0 was known, the like-
lihood of the tree would simply be the product of the probabilities of base
substitution in each tree branch and the base frequency, πS0 , of state, S0 :

L = πS0 PS0 S1 (v1 )PS0 S2 (v2 )

Here, we are simply multiplying the independent events that would give
us the tree we see in Figure 2.1, given the hypothesis that node 0 has state
S0 . More specifically, these independent events are the event that the root
of the tree is in state S0 , with probability πS0 , and the event that the root
bifurcates to give us species 1 and species 2 at the tips of our tree. The event
that the root transitions from S0 to the state of tip 1, S1 , in a branch length,
v1 , has probability, PS0 S1 (v1 ). Similarly the event that the root transitions
to the state of tip 2, S2 , in a branch length, v2 , has probability, PS0 S2 (v2 ).
Since, in reality, we will not know the state of the interior node, 0, we
must sum the likelihoods for each possible state, S0 . The state, S0 , can take
on the values 1, 2, 3, or 4, which correspond to A, C, T, and G, respectively.
So, our likelihood calculation for the tree in Figure 2.1 will actually look
like this:

X 4
X
L= πS0 PS0 S1 (v1 )PS0 S2 (v2 ) = πS0 PS0 S1 (v1 )PS0 S2 (v2 ).
S0 1

The states S1 , S2 , v1 , and v2 , are given (from the tree), and πi , for i =
{1, 2, 3, 4}, is given as a fixed constant, independent of the tree. Notice that
when we are calculating the likelihood of a phylogenetic tree, we are only
able to determine the likelihood of the shape of the tree (i.e, the likelihood
of the locations of the branches and nodes of the tree). Thus, we cannot
determine the states of interior nodes simply by calculating the likelihood
of the tree.

Example 8. If we know that tip 1 is an A and that tip 2 is a G in the tree in

Figure 2.1, S1 = 1 and S2 = 4. Then, the likelihood of the tree is:
X
L= πS0 PS0 1 (v1 )PS0 4 (v2 ).
S0

Now, let us consider a slightly more complicated example in Figure 2.2.

For this tree, the likelihood is the product of the sums of the transition
probabilities for all possible states for the interior nodes and πS0 (recall that
 Computing the Likelihood of a Phylogenetic Tree 15

Figure 2.2: 4-Species Phylogenetic Tree.

states 1, 2, 3, and 4 correspond to DNA bases A, C, T, and G, respectively).

In this tree, we are given observable data at the four tips. Thus, our likeli-
hood caculation will look like this:

4 X
X 4 X
4
L= πS0 PS0 S5 (v5 )PS5 1 (v1 )PS5 2 (v2 )PS0 S6 (v6 )PS6 3 (v3 )PS6 4 (v4 )
S0 =1 S5 =1 S6 =1
(2.1)
This calculation sums over all four possible states for each node with an
unknown state.
Unfortunately, this calculation is extremely long. The likelihood calcu-
lation for Figure 2.1 has 4 terms, while the calculation for Figure 2.2 has 64
terms. Phylogenetic trees with more species have even more terms. It is
helpful to move each summation as far to the right in the likelihood cal-
culation as possible, allowing us to find the likelihoods of each individual
segment of the tree. If we do this with (2.1), our calculation would look like
this:

X X
L= πS0 [PS0 S5 (v5 )(PS5 1 (v1 ))(PS5 2 (v2 ))]
S0 S5
X
× [PS0 S6 (v6 )(PS6 3 (v3 ))(PS6 4 (v4 ))] (2.2)
S6

We have placed the parentheses and brackets in our equation [( )( )][( )(

)] so that they model the topology of our tree. Each segment of the tree, rep-
resented in our equation by the terms in the parentheses ( ), contains a base
16 The Likelihood of a Phylogenetic Tree

substitution probability Pij (v). Within each set of parentheses, we have the
likelihood for each branch connected to a tip. Within each bracket, we have
the likelihood for each of the two branches that stem from the root, node
0. It should be clear that this relationship between the equation and the
topology of the tree is a result of our construction of the likelihood calcula-
tion for the tree, but this construction also allows us to compute likelihood
using the conditional likelihoods of each segment of the tree.
(k)
We will use the notation, Ls , for the likelihood of the tree at and above
(k)
node k on the tree, given that node k has state s. Then, each Ls corre-
sponds to segments of the tree beginning with the tips of the tree. For
example, in Figure 2.2,
4
(5) (1) (2)
X
LS = (PS5 1 (v1 )L1 )(PS5 2 (v2 )L2 ). (2.3)
s5 =1

Since the tips of our tree contain our set of data, we know the states of
(k)
the tips of the tree. Thus, if k is a tip of our phylogenetic tree, Ls will be
0 for all states except for its observed state. If s∗ is the observed state at tip
(k)
k, then Ls∗ = 1. Remember that this simply means that the probability of
tip k having its observed state s∗ is 1 given the hypothesis that tip k has
probability s∗ and 0 given any other hypothesis. For example, in Figure
(1) (2) (3) (4)
2.2, L1 = 1, L2 = 1, L3 = 1, and L4 = 1. Now that we have an
easy calculation for the tips of our tree we are able to begin our likelihood
calculation for the entire tree in Figure 2.2 at its tips.
Since our likelihood calculation has been reduced to conditional likeli-
hood calculations and we can easily find the likelihoods of the tips of the
tree, we begin the computation from the tips of the tree and work our way
down to the root. We can compute the conditional likelihoods of the nodes
in the tree with tips as their immediate descendents. In Figure 2.2, nodes 5
and 6 satisfy this property.
Example 9. For node 6 in Figure 2.2, the likelihood that node 6 has state S
is:
4 h X 4 4
 X i
(6) (1) (4)
X
LS = PS6 S3 (v3 )LS1 PS6 S4 (v4 )LS4
S6 =1 S1 =1 S6 =1
(1) (2)
But we know that L1 = 1 and L2= 1, so the calculation reduces to:
(2)
X
LS = [(PSk 1 (v1 )(1))(PSk 2 (v2 )(1))].
S
 Computing the Likelihood of a Phylogenetic Tree 17

Now, we can state the process of finding conditional likelihoods as a

general formula. For any node x, whose immediate descendants are tips y
(x)
and z, we can compute Ls :
4 h X
4 4
 X i
(x) (y) (z)
X
LS = PSx Sy (vy )LSy PSx Sz (vz )LSz
S=1 Sy =1 Sz =1

Once we have found the conditional likelihoods for all nodes with tips
as immediate descendants, we can think of those nodes as our new ”tips”
by using their conditional likelihoods to compute the likelihoods of their
ancestor nodes. Thus, for any node x with immediate descendants y and z,
the conditional likelihood at node k is
(x)
X h X (y)
 X
(z)
i
LS = PSx Sy (vy )LSy PSx Sz (vz )LSz . (2.4)
s Sy Sz

We continue calculating conditional likelihoods, replacing our new ”tips”,

and finding conditional likelihoods of ancestor nodes until we reach the
root of our tree. At the root, node 0, of the tree, we will also compute the
(0)
conditional likelihood LS0 given each possible state for root 0. Then, the
overall likelihood of the tree for the DNA site we are considering is
(0)
X
L= πS0 LS0 .
S0

The base frequency for the root of our tree, πS0 , gives us the probability that
the root is in state S0 .
Example 10.  
0.8 0.1 0.1 0.1
0.1 0.8 0.1 0.1
Pij = 
0.1

0.1 0.8 0.1
0.1 0.1 0.1 0.8
Suppose the terms in the matrix above represent the base substitution prob-
abilities of a phylogenetic tree with two species that are both one branch
length unit away from a common ancestor. As we have seen before, i will
represent a prior state, and j will represent a later state for the DNA site un-
der consideration. Both i and j can take on values 1, 2, 3, and 4 corespond-
ing to DNA bases, A, C, T, and G, respectively. Suppose we are studying
orcs and trolls in our set of data. For the DNA site that we are examin-
ing, orcs have an A, and trolls have a T. Then, the likelihood of seeing our
18 The Likelihood of a Phylogenetic Tree

observed data, or in other words, the probability of seeing our observed

data given a hypothesis for the base seen at the DNA site in the common
ancestor, will be as follows:
(0)
LS = PS1 PS3

where 0 is the node in our tree representing the common ancestor and S
is the state of node 0. We take the base substitution probabilities from our
(0) (0) (0) (0)
matrix and see that L1 = 0.08, L2 = 0.01, L3 = 0.08, and L4 = 0.01.
Now, say that the common ancestor is the root of our tree. Then, we sim-
ply need to find the sum of the products of these conditional likelihoods
and the base frequencies for each hypothesized ancestor base. If the base
frequencies are 0.25 for each base, then
4
(0) (0) (0) (0)
X
L= πs L(0)
s = π1 L1 + π2 L2 + π3 L3 + π4 L4 = 0.045
S=1

In order to get a tree with a greater likelihood, we could find a branch

length that would provide us with greater base substitution probabilities,
increasing the likelihood of the subsequent tree. We will see how to do this
in Chapter 3.

2.2 Time Reversibility

While computing the likelihoods of the examples in the previous section,
we actually made quite a few assumptions. For instance, in Example 5, we
assumed that the root of the tree was at node 0 in order to compute the like-
lihood of the entire tree. In reality, when given a set of DNA sequences, we
oftentimes will not know where the root of the tree lays. Another assump-
tion we made was that the DNA bases at the tips of our tree were exactly
one branch length away from their common ancestor. Again, in actual fam-
ilies of species, we will have to search for this information. One property
that helps us find the actual branch lengths and location of the root of our
tree is time reversibility.
Definition 14. A Markov chain is time reversible if the rate at which it goes
from state i directly to state j is the same as the rate at which it goes from
state j directly to state i.
Recall that the base frequency of base i is πi , and this term represents
the proportion of base substitutions that will result in base i. We can restate
 The Pulley Principle 19

our definition of time reversibility, using our model of base substitution, as

πi Pij (t) = Pji (t)πj for all i, j, and t. In order to determine whether or
not our Markov process is time reversible, we have to show that these two
products are equivalent. This is made clear by utilizing (1.2). By (1.2),

πi Pij (t) = πi e−µt δij + (1 − e−µt )πi πj

We also see that,

πj Pji (t) = πj e−µt δji + (1 − e−µt )πi πj

The only difference between these two equations is the δij and δji . If
we recall the origin of these terms in (1.2), the Kronecker delta function,
we know that the values of δij and δji are either both 1 or 0 depending on
whether i = j or i 6= j, respectively. Then, δij = δji for all i and j. Hence,
we have that the Markov process modeling base substitution is in fact time
reversible.
But where exactly does (1.2) depend on t? The probability of a base
substitution from base i to base j, is only dependent on the product µt.
Suppose we divided µ in half and doubled t. Then, we would have µ2 2t =
µt. Next, suppose we divided µ by 3 and tripled t. We would be left with
µ
3 3t = µt. We see that if we divide the rate µ by an amount and multiply
the time t by the same amount, we get the same product µt. This makes it
difficult for us to determine the values of µ and t separately in (1.2). All we
are able to take away from (1.2) is the product, µt.
If we assume that the rate of base substitution, µ, is the same for all
branches of the tree and all times, then µt should be proportional to the
time that has passed so far during the Markov process, where time is rep-
resented in the number of mutations that have occurred. Since we do not
have any indication of what µ or t could be, we will assume that µ = 1.
This allows us to think of t as units of expected numbers of substitutions.

2.3 The Pulley Principle

The time-reversibility and ambiguity of branch lengths turn out to be quite
useful in finding the best phylogenetic tree for a set of DNA sequences. We
can actually place the root of our tree anywhere in our tree. We will look
at our calculation of the likelihood of the tree seen in Figure 2.2 to see why.
The last two steps of our algorithm for computing the likelihood of the tree
involved nodes 0, 5, and 6. In the last step of the algorithm, we had
20 The Likelihood of a Phylogenetic Tree

4
X
L= πs0 ((Ps0 s5 (v5 )L(5) (6)
s5 )(Ps0 s6 (v6 )Ls6 )). (2.5)
s0 =1

What would happen to the likelihood of the tree, L, if we added a length

x to v5 and subtracted x from v6 ? We will try rewriting (2.5) using what
we know about time reversibility to find out. With the time reversibility
property, we have

πs0 Ps0 s5 (v5 ) = πs5 Ps5 s0 (v5 )

which allows us to rewrite (2.5) as
4
X
L = πs5 Ls5 Ps5 s0 (v5 )Ps0 s6 (v6 ). (2.6)
s0 =1

This form of our equation allows us to experiment with different branch

lengths. In order to do this, we must first introduce the Chapman-Kolmogorov
equation. We will consider a continuous Markov process {X(t) : t ≥ 0} with
state space S and transition probability matrix P . If the process begins in
state i and moves to state j in t units of time, then we have

Pij (t) = P (X(s + t) = j | X(s) = i), for i, j ∈ S; s, t ≥ 0.

Since X(0) = i, it should be clear that after s units of time, the Markov
process will enter another state, we will call it k, before entering state j in
s + t units of time. This gives us the Chapman Kolmogorov equation:
∞
X
Pij (s + t) = Pik (s)Pkj (t). (2.7)
k=0

We will prove (2.7) using the Law of Total Probability. Since {X(s) =
k | k ∈ S} is a set of mutually exclusive events, we can apply the Law of
Total Probability to get:
 The Pulley Principle 21

Pij (s + t) = P (X(s + t) = j | X(0) = i)

X∞
= P (X(s + t) = j | X(0) = i, X(s) = k)P (X(s) = k | X(0) = i)
k=0
∞
X
= P (X(s + t) = j | X(s) = k)P (X(s) = k | X(0) = i)
k=0
∞
X ∞
X
= Pkj (t)Pik (s) = Pik (s)Pkj (t)
k=0 k=0

Now that weP have the Chapman-Kolmogorov Equation, we can rewrite

the summation 4s0 =1 Ps5 s0 (v5 )Ps0 s6 (v6 ) from (2.6) as Ps5 s6 (v5 + v6 ). This
means that the likelihood of our tree is only dependent on the total lenth of
the tree branches; it only depends on the sum, v5 +v6 . This makes it possible
for us to increase and decrease the lengths of v5 and v6 as long as the sum,
v5 + v6 , remains constant. Thus, the root of the tree can be placed anywhere
between node 5 and node 6. Since these nodes were chosen arbitrarily, we
are able to think of each branch of our tree in a similar manner and place the
root of the tree anywhere in the tree. Since Section 2.1 only described how
to find the likelihood of a rooted tree, how do we find the likelihood of an
unrooted tree? By the Pulley Principle, we can root the tree anywhere and
get the same likelihood. Figure 2.3 shows three trees that are equivalent
because of the pulley principle.

Figure 2.3: Tree A, Tree B, and Tree C are all equivalent due to the pulley
principle. Tree A shows the tree rooted at node 0. Tree B demonstrates how
the tree root can be moved. Tree C shows the unrooted tree.
Chapter 3

Finding a Maximum
Likelihood Tree

The Pulley Principle forces us to think of each segment of the phylogenetic

tree as a possible location for the root of the tree. This actually becomes
a very useful property of the tree because it allows us to manipulate or
change branch lengths and the location of the root of the tree. This is help-
ful because it simplifies the process of finding the tree with the maximum
likelihood out of the set of trees that we test for a set of DNA sequences.
In Chapter 2, we described a long but still practical process for finding the
likelihood of a phylogenetic tree. However, in order to find the tree that
produces the largest likelihood out of all of the trees for which we calcu-
late a likelihood, we still need to experiment with possible topologies and
branch lengths and then, evaluate the likelihoods of the trees they produce.
After we have found all possible topologies for a particular site in a set
of DNA sequences, we can evaluate the likelihood of each topology. As
described in Chapter 2, in each computation for the likelihood of a phy-
logenetic tree, we get the base substitution probability, Pij (vj ), of a site in
state i transitioning to state j in vj time units by finding the transition prob-
abilities specified for a time of length vj . But so far we have only found
possible tree topolgies; we do not know what vj is. Furthermore, direct
search would not be a feasible method of finding that vj because it would
require us to evaluate the likelihood of the tree under consideration for each
distinct vj . The Pulley Principle gives us an algorithm which allows us to
change one branch length, vj , at a time (see section 3.2). Each branch length
is altered to the value that provides the highest likelihood out of all of the
other branch lengths tested. Once we find each branch length, vj , we can
24 Finding a Maximum Likelihood Tree

complete the likelihood calculations for each possible topology. Finally, the
topology producing the greatest likelihood out of all of the other topologies
we examine is the tree that we choose as the best phylogenetic tree for our
set of data.

3.1 How Many Possible n-species Trees are There?

Now that we have a general idea of how to find a tree with a good likeli-
hood, we still need to find a feasible way to find a good tree topology with-
out examining every possible tree topology. For the purposes of this paper,
we will be focusing on finding the topology of bifurcating trees. Although
it is possible to find every tree topology available for a set of DNA se-
quences, there would be too many possible tree topologies to test since we
would have to compute the likelihood for each topology. This direct search
strategy would require us to look at each possible unrooted tree topology,
iterate each branch length to its optimal value for each branch, and then
choose the topology that has given us the greatest likelihood. However, by
looking at even a simple tree, we see that this method is not ideal. We will
later see that the number of unrooted bifurcating trees with n labelled tips
is (2n − 5)!/[(n − 3)!(2n−3 )] (see section 3.1.1). For a tree with 10 labelled
tips, for example, there are more than 2 million such topologies to evaluate.
This is much too big a number of trees to evaluate the likelihood for.
Therefore, we need to use a more focused strategy to search among pos-
sible tree topologies. Felsenstein [2] suggests building the tree starting with
a 2-species tree and successively adding one species at a time to the tree un-
til all of the species are on the tree. One difficulty with this method is deter-
mining where to place each species as it is added to the tree. I will describe
the process for deciding where to place a species on both a rooted bifurcat-
ing phylogenetic tree and an unrooted bifurcating phylogenetic tree. For
both types of trees, we will add one species at a time in some predeter-
mined order (e.g. alphabetical order by species name).

3.1.1 Rooted Bifurcating Trees

In Chapter 1, both Figure 1.1 and Figure 2.1 show examples of rooted bifur-
cating trees. Each of the phylogenetic trees has a root at the bottom of the
tree which can be traced back to from any node and tip, and when a base
undergoes a mutation, there are two possible bases that are substituted for
the original base. Figure 3.1 also shows some rooted bifurcating trees. How
 How Many Possible n-species Trees are There? 25

can we determine whether or not Figure 3.1 contains all possible labeled,
rooted bifurcating trees for a tree with three tips?

Figure 3.1: All possible labeled, rooted bifurcating trees for three species.

We use the following straightforward reasoning. Consider constructing

one possible phylogenetic tree by beginning with a tree with two of the n
species and then adding one species at a time to that tree in some prede-
termined order (e.g. species name in alphabetical order). It is clear that if
we add one more species to a tree that already has n species added to it,
we only need to place the (n + 1)-st species in each possible location in or-
der to get the total number of trees for an n + 1-species tree. Since the tree
is bifurcating before and after each node, and hence before and after the
location of the addition, we are unable to add the new species to a node.
Instead, the new species must be added to an existing branch, resulting in
the creation of a new node and two new branches. This implies that each
existing branch in the n-species tree is a possible location for the addition
of the new species. Our reasoning for this is depicted in Figure 3.2.
How can we be sure that this process will give us all possible labeled
rooted, bifurcating trees? With the information we have so far, we also can-
not be completely sure whether or not adding a new species to two differ-
ent branches will produce two different trees. In order to understand this
further, consider the process of adding species k to a tree with k − 1 species
and also the process of removing species k from a k-species tree. Suppose
we have an n-species tree. If we remove n − k species in the reverse order
of which they were added to the tree, we will be left with a k-species tree.
This k-species tree will actually be the tree that we should have obtained
before when we were adding species k to the tree. This indicates that there
exists a particular sequence of places to add species k + 1, k + 2, k + 3, . . .
onto the k-species tree in order to return to the n-species tree we had before
26 Finding a Maximum Likelihood Tree

Figure 3.2: Tree A shows a 3-species tree. When we want to add species
4 to the tree, we must add it to a new node on an existing branch in the
3-species tree as shown in trees B-F . If we decide to add species 4 to an
existing node, as we have done in Tree G, then our tree will no longer be a
rooted bifurcating tree.

we began removing species. In addition, no other k-species tree can turn

into that same n-species tree after adding the n − k species that are miss-
ing. Suppose there was another k-species tree that could produce the same
n-species tree. Then, this k-species tree should also be produced through
the removal of the n − k species that were added after it. However, this
presents a contradiction, since the same sequence of removals cannot yield
two different trees. Therefore, any n-species tree can be produced from one
and only one k-species tree [3].
This result implies that each possible addition sequence leads to a dif-
ferent n-species tree. Since everytime we add a new species to a tree, it can
only be added to one of the existing branches, the number of ways in which
we can add that species to the tree is equal to the number of branches in the
existing tree. This includes the branch connected to the root of the tree.

Example 11. In Figure 3.2, tree A has 5 branches, so species 4 can be added
in 5 ways. Notice that when we add species 4 to tree A, as shown in tree
B, there are 2 new branches and 1 new interior node. Then, when we add
 How Many Possible n-species Trees are There? 27

species 5 to tree B, there will be 7 ways that we can add it.

Using this reasoning, we see that there are 3×5×7×9×11×· · ·×(2n−3)
different ways to add species to a rooted bifurcating phylogenetic tree in
order to produce an n-species tree. The product, 3 × 5 × 7 × 9 × 11 × · · · ×
(2n − 3), looks similar, but not equal, to (2n − 3)!. The even divisors have
simply been removed from (2n − 3)!. In order to correct for this, we take
out the even divisors by dividing by 2 × 22 × · · · × 2n−1 = 2n−1 (n − 1)!. This
(2n−3)!
shows that 3 × 5 × 7 × 9 × 11 × · · · × (2n − 3) = 2n−1 (n−1)!
Now, we are able to check that at each addition of a new species to a
phylogenetic tree, we have added the new species to each possible location.
For example, if we had 20 species (n = 20), the total possible number of
trees would be:

(2n − 3)!
= 8, 200, 794, 532, 637, 891, 559, 375.
2n−1 (n− 1)!
This means that if a computer was able to evaluate the likelihood of a 20-
species tree in one hour, it would take us about 1.64 × 1023 hours, or about
1.87 × 1019 years. Remember that this is only for the calculation of the
likelihood of each tree, without testing various branch lengths.
Obviously, there are way too many total possible trees to examine all
of them when we have a tree with more than ten species. We must also
keep in mind that we have to examine all possible trees for each site of the
DNA sequences, making this process even less feasible. Thus, we must
use a different algorithm for examining as many possible tree topologies as
possible.

3.1.2 Unrooted Bifurcating Trees

As we saw in section 2.2, we do not actually know where the root of our tree
is. Therefore, instead of considering rooted, bifurcating trees as we did in
the previous section, we must consider unrooted, bifurcating trees. Then,
how many possible unrooted, bifurcating trees would we have to consider?
Well, if we root the tree at one of its species, we will have a rooted,
bifurcating tree. Suppose we had n tips on an unrooted tree. Then, after
rooting our tree at one of those n species, we would end up with n − 1 tips.
This resulting tree would be a rooted, bifurcating tree with n − 1 tips. Thus,
the number of possible trees would be:

3 × 5 × 7 × · · · × (2(n − 1) − 3) = 1 × 3 × 5 × 7 × · · · × (2n − 5)
28 Finding a Maximum Likelihood Tree

This equation gives us the same number of possible rooted, bifurcating

trees with n − 1 tips. We see that every rooted tree with n − 1 labeled tips
is analogous to an unrooted, bifurcating tree with n tips. Likewise, every
unrooted tree with n tips is analagous to one rooted tree with n − 1 tips.
Figure 3.3 shows an example of rooting an unrooted, bifurcating tree at one
of its species.

Figure 3.3: Tree A shows an labeled unrooted, bifurcating tree. We can root
Tree A at species 1, resulting in Tree B, a rooted, bifurcating tree.

3.2 Which Tree Topology?

Even though there are fewer possible unrooted, bifurcating trees for n species
than there would be for rooted, bifurcating trees with n species, it is not by
much. Clearly, we must introduce a method that is not as computationally
expensive. Felsenstein [2] has come up with a method that is much more
simple. When the k-th species is being added to the tree, there will be 2k −5
different branches to which it could be added. Each of these different lo-
cations is tried, and the likelihood of each resulting topology is evaluated.
The topology that produces the largest likelihood out of all of the other
likelihoods computed is accepted, and the rest are thrown out.
In order to ensure that the order in which we add each species is not
limiting the number of topologies we produce using this strategy, we use
local rearrangements of the tree (described in more detail in the following
section). These local rearrangements will not lead us to every topology, for
that would give us the same problem presented in 3.1.1 and 3.1.2: there
would be way too many trees to examine. Instead, local rearrangements
 Which Tree Topology? 29

give us a greedy algorithm. In a greedy algorithm, we decide when the

algorithm has produced desirable enough results to stop performing the
algorithm. We determine how ”greedy” we want to be in assuming that
when the results are sufficient. With local rearrangements, we are simply
checking to see if those topologies that are similar to the topology we begin
with can result in a higher likelihood. As the likelihoods of the successive
trees increase, we can determine when to stop performing local rearrange-
ments.
If the tree has more than four species, before we add the next species,
local rearrangements are done to see if any of the resulting topologies in-
creases the likelihood of the tree. If a topology resulting from these local
rearrangements does in fact increase the tree’s likelihood, it is accepted.
These local rearrangements continue until a tree is found for which no lo-
cal rearrangement can increase the likelihood significantly.

3.2.1 Local Rearrangements

Since the order in which we add a new species to a tree will affect the max-
imum likelihood tree we find, we use local rearrangements in an attempt
to correct for this. This means that if we use a different sequence of adding
species to our tree, we may end up with a different maximum likelihood
tree.
Remark 1. The tree with the likelihood that we ultimately decide on after
performing local rearrangements will only provide a local maximum like-
lihood. This means that there may be a possible topology with an even
greater likelihood that we have not considered because its topology is so
different from the original topology that we began with. We cannot be
guaranteed that we have found the tree with the greatest likelihood of all.
There are several different kinds of local rearrangements that we can
use. Since our goal is simply to understand how these local rearrange-
ments can guide us to a tree with a local maximum likelihood, we will use
a simple local rearrangement process called Nearest Neighbor Interchange
(NNI). This process basically involves switching adjacent branches with
each other. More specifically, we are switching adjacent subtrees connected
to an interior branch. We erase the interior branch and the branches con-
nected to it at each end. For example, in an unrooted, bifurcating tree, we
erase a total of five branches. The four adjacent trees are then disconnected
from each other. They can then be reconnected into a tree in three possible
ways. These are the only three possible ways to reconnect the tree without
30 Finding a Maximum Likelihood Tree

repeating a topology. These rearrangements are depicted in Figure 3.4.

Figure 3.4: Nearest Neighbor Interchange. Here we see our original tree with
subtrees A, B, C, and D. For the particular interior branch that we are con-
cerned with, we erase that particular branch along with all of the branches
directly connected to it. Then, we reassemble the subbranches in the two
different ways shown above. This gives us a total of three possible ways to
construct our tree, including the tree before local rearrangements.

Once we have completed this local rearrangement at one of our interior

branches, we evaluate the likelihood of each of the three resulting trees. We
accept the tree that gives us the highest likelihood and throw out the other
two possiblities. We then continue using NNI on all of the other interior
branches and accept those rearrangements that give our tree a higher like-
lihood. We continue this process until no more local rearrangements can
increase the likelihood of the tree. Once we have reached this point, we
add the next species on to our tree.
Before we go any further, we can’t forget that we need branch lengths in
order to calculate the likelihood of a phylogenetic tree. The term, PSk Si (vi )
in the general equation for the likelihood calculation (2.4) is a transition
probability specific to the branch length, vm . But we do not yet know the
branch lengths that make up our phylogenetic tree. This means that before
we can actually evaluate any of the likelihoods we have discussed above,
we need to figure out each branch length.
 Maximum Likelihood Branch Lengths 31

3.3 Maximum Likelihood Branch Lengths

For each of the topologies we will consider when looking for the maximum
likelihood tree topology, we have to find the branch lengths that will maxi-
mize the likelihoods of those topologies. Felsenstein [2] created an iteration
technique that simplifies this at first sight, daunting task.
We will consider Tree C in Figure 2.3 in order to understand this strat-
egy better. This is the unrooted phylogenetic tree for these three species,
and we do not know where the root of the tree is located. Consider branch
v3 . Suppose the root is located somewhere along v3 . Then, we can use
the method presented in Chapter 2 to compute the likelihood of the tree.
We will assume that the root is directly to the right of node 4. Then, the
likelihood of the tree for one site of DNA is:

(4) (3)
XXX
L= πS0 (PS0 S3 (0)LS4 )(PS0 S3 (v3 )LS3 ) (3.1)
S0 S4 S3
(4) (3)
X X
= πS0 LS0 ( PS0 S3 (v3 )LS3 ) (3.2)
S0 S3

The equality in (3.2) comes from the fact that PS0 S4 (0) = 1. Now, if we
substitute (1.2) into (3.2), letting µ = 1, we get:
X  X 
(4) (3) (4) (3)
X
−v3 −v3
L=e πS LS LS + (1 − e ) πS4 LS4 πS3 LS3 (3.3)
S S4 S3
(4) (3)
Now, we must figure out how to find LS4 and LS3 without knowing
the branch lengths in the tree. Recall that in Section 2.1, for a tip, k, of our
(k) (k)
tree with base, i, Li = 1, while Lj6=i = 0. Since species 1, 2, and 3 are all
located at the tips of our tree, they have a known conditional likelihood,
(k)
LSk . In the 2-species tree that we added species 3 to in order to get Tree C,
we should have evaluated branch lengts, v1 and v2 . We can then use these
branch lengths and the known conditional likelihoods of the tips of our tree
(4)
to compute LS4 .
Recall that this calculation for likelihood only gives us the likelihood
for the phylogenetic tree representing one DNA site. For all K DNA sites
in the sequences of our data set, the likelihood is:
K
Y
(Ai q + Bi p) (3.4)
i=1
32 Finding a Maximum Likelihood Tree

where q = e−v3 and p = 1 − q = 1 − e−v3 . Ai and Bi represent the following:

(4) (3)
X
Ai = πS LS LS
S

and
X  X 
(4) (3)
Bi = πS4 LS4 πS3 LS3
S4 S3

for the i-th DNA site in the set of sequences. We are interested in finding
the value of v3 that will maximize the likelihood. We can find this value of
v3 by finding the value of p that will maximize (3.3). Then, we can solve for
v3 since v3 = − ln(1 − p).
Now, we will explore some properties that follow from these equations
that will help us understand Felsenstein’s iteration formula for finding seg-
ment lengths that will maximize the likelihood. If we take the logarithm of
(3.3), we get
K
X
ln(L) = ln(Ai q + Bi p).
i=1

If we take the derivative of this equation and set it equal to zero, we get
K
d ln(L) X Bi − Ai
= = 0. (3.5)
dp (Ai q + Bi p)
i=1

Now, notice that if we have K sites total,

K K K
X X Ai q + B i p X Bi − (Bi − A)q
K= 1= = . (3.6)
Ai q + B i p Ai q + B i p
i=1 i=1 i=1

We can use (3.5) to take out the terms in the numerator of (3.6) containing
q. We end up with
K K K
X Bi − (Bi − A)q X Bi X Bi − Ai
= −q .
Ai q + Bi p Ai q + B i p (Ai q + Bi p)
i=1 i=1 i=1

The last term is 0 because of (3.5). Then,

K
X Bi
K= . (3.7)
Ai q + Bi p
i=1
 Maximum Likelihood Branch Lengths 33

The equation presented in (3.7) must be satisfied at a maximum in the like-

lihood function for the tree, or equivalently, when d ln(L)/dp = 0. Now, by
multiplying both sides of (3.7) by p, we get the iteration-formula:
K
1 X Bi p(k)
p(k+1) = (3.8)
K
i=1
Ai q (k) + Bi p(k)

where q (k) = 1 − p(k) . This iteration expression is a specific case of the

general EM algorithm presented by Dempster, et al [1]. The first step of the
algorithm is to make an estimated for the value of p, calling it p(1) . This is
simply an educated guess. This value of p(1) generates another estimate for
p, namely p(2) . This value of p(2) is then used to find p(3) , and this process
continues until the estimates for p converge to some p(k) . It can be shown
that the successive segment lengths produced never decrease the likelihood
of the tree [1]. Thus, we can iterate this process until the p(k) s converges.
This iteration is another greedy algorithm because it is only ensured to lead
to a branch length that produces a local maximum likelihood. Once the p(k)
converges, we may continue the iteration to see whether or not it will lead
us to an even larger likelihood, or we may choose the p(k) we already have.
It is important to observe that summing over many (K) DNA sites is
significant here. Let us consider what happens when K = 1. In this case,
the iteration equation will be:

Bp(k)
p(k+1) = .
Aq (k) + Bp(k)
Now, we will substitute an x for p(k) . Then, we have a function in terms of
x:

Bx Bx
f (x) = = .
A(1 − x) + Bx (B − A)x + A
Notice that this equation is in the form of a fractional linear transformation
such that it has only two fixed points. These two fixed points are x = p =
0 and x = p = 1. Then, the iteration will converge to either p = 1 or
p = 0. Hence, if we use this iteration algorithm for the single site, we will
get that the branch length we are estimating is either 0 or undefined since
p = 1 − e−v , for branch length, v. If p = 0 = 1 − e−v , then v = 0. If
p = 1 = 1 − e−v , then v = ∞. Therefore, in practice, we only use the EM
algorithm with DNA sequences with two or more sites.
We must do this iteration technique for each of the branches on each
topology. Once we iterate and optimize for vx , we fix this value of vx and
34 Finding a Maximum Likelihood Tree

move on to optimize for another branch length, vy . We continue optimizing

branch lengths in this fashion until the branch lengths each converge to
some value.

3.4 Putting Topologies and Branch Lengths Together

to find the Maximum Likelihood Tree
Now that we have a method for finding all of the branch lengths that
will maximize each topology’s likelihood, we can use the resulting branch
lengths to actually compute the likelihoods of each topology we are in-
terested in. Once we compute each topology’s likelihood, we follow the
procedure described in 3.2. We will accept those topologies that increase
the likelihood of the tree until the topology cannot be altered to increase
the likelihood of the tree. This will give us a sequence of trees with in-
creasingly better likelihoods. Remember that these topologies will all be
unrooted bifurcating trees. The Pulley Principle allows us to compute the
likelihoods of unrooted trees. The tree that results in the highest likelihood
is accepted. Once we find a phylogenetic tree whose topology and branch
lengths cannot be altered to significantly increase the likelihood of the tree,
we choose that topology as our maximum likelihood tree. Figure 3.5 illus-
trates this entire process for us.
Putting Topologies and Branch Lengths Together to find the Maximum Likelihood Tree
35

Figure 3.5: Finding an estimate for the maximum likelihood tree.

Chapter 4

Finding the Maximum

Likelihood Tree - An Example

In this chapter, we will demonstrate the process of finding a maximum like-

lihood phylogenetic tree using a relatively simple example. We will only
consider two DNA sites for a family of five species, namely the wizards,
elves, hobbits, dwarves, and humans from Middle Earth. At site 1, we set
the DNA data to be A, C, T, G, and T, respectively in our five species. At
site 2, let us set our DNA data to be A, A, C, G, and G. In other words, the
first species has base A at site 1 and base A at site 2. The following table
illustrates which base each species has at each DNA site.
DNA Site Wizard Elf Hobbit Dwarf Human

1 A C T G T
2 A A C G G

We will assume that we can model base substitution probabilities using the
Jukes-Cantor model (see [7]). Under this model, base frequencies are all
equal. In other words, π1 = π2 = π3 = π4 = 1/4.
First, we will focus on site 1 of the DNA sequences. We will add each
base to the tree in alphabetical order (A, C, G, T, T). Our 2-species tree will
look like this:

Now, since there is only one branch in our tree, there is only one possi-
ble location for us to add species 3. Then, our 3-species tree will look like
38 Finding the Maximum Likelihood Tree - An Example

this:

In order to begin the process of finding branch lengths, suppose the

root of the tree lays along v1 . Suppose it lays directly to the left of the node
labelled with *. Then, as in Section 3.3, the likelihood of the tree is:
(∗)
XX
L= πS0 (PS0 S∗ (0)LS∗ )(PS0 1 (v1 )).
S0 S∗

Since PS0 S∗ (0) = 1, this is equivalent to:

(∗)
X
L= πS0 LS0 (PS0 1 (v1 )).
S0

Now, we can substitute (1.2) for PS0 1 (v1 ) to get:

(∗) (∗)
X
L = π1 L1 (e−v1 + (1 − e−v1 )π1 ) + πS0 LS0 (1 − e−v1 )π1 (4.1)
S0 6=1
(∗) (∗)
X
−v1 −v1
=e π 1 L1 + (1 − e ) πS0 LS0 π1 . (4.2)
S0

Our equation is now in the same form as (3.3), so we can label our terms
(∗)
as we did in (3.7) for site 1. In this example, term A1 = π1 L1 , term B1 =
P (∗) −v1 , and term p = 1 − q = 1 − e−v1 .
S0 πS0 LS0 π1 , term q = e
Next, we consider DNA site 2, so that we can find A2 and B2 . If we
order the bases in site 2 in the same order as we ordered the bases in site
1, we have A, A, G, C, and G. The phylogenetic tree for site 2 will look like
this:

Then, the likelihood for the phylogenetic tree for site 2 is:
 39

(∗) (∗)
X
L = e−v1 π1 L1 + (1 − e−v1 ) πS0 LS0 π1 .
S0
(∗) P (∗)
In this equation, A2 = π1 L1 and B2 = S0 πS0 LS0 π1 .
Since the number of sites in the DNA sequences of our data is K = 2,
our iteration formula will simply be:
2
1 X Bi p(k)
p(k+1) = (4.3)
2
i=1
Ai q (k) + Bi p(k)
In order to use this iteration formula, we must first determine what Bi
and Ai are for i = 1 and for i = 2. For each Ai , we know that π1 = 1/4 under
(∗)
the Jukes-Cantor model. Then, we are left with finding L1 . Similarly, for
each Bi , we know that πS0 = π1 = 1/4. Thus, we are left with the task of
(∗)
finding LS0 . By (2.4), we have that
(∗)
L1 = (P13 (v3 ))(P12 (v2 ))
and
(∗)
X
LS0 = (PS0 4 (v3 ))(PS0 2 (v2 ))
S0

at site 1 in our DNA sequence.

At site 2 in our DNA sequence,
(∗)
L1 = (P14 (v3 ))(P11 (v2 ))
and
(∗)
X
LS0 = (PS0 4 (v3 ))(PS0 1 (v2 )).
S0

Here we find that we cannot calculate these likelihoods without know-

ing v1 and v2 . Recall that the algorithm described in Section 3.3 is a special
case of the EM algorithm that never produces branch lengths that will de-
crease the likelihood of the tree. Thus, we are able to set v2 and v3 at an
initial estimate of the branch lengths while trying to find v1 . We will es-
timate that v2 = v3 = 1. Now, we are able to use these values in (1.2) to
get
(∗)
L1 = ((1 − e−1 )(1/4))2 ≈ 0.025
40 Finding the Maximum Likelihood Tree - An Example

in A1 ,
(∗)
L1 = ((1 − e−1 )(1/4))(e−1 + (1 − e−1 )(1/4)) ≈ 0.083
in A2 , and

(∗)
LS0 = 2((1 − e−1 )(1/4))2 + 2((e−1 + (1 − e−1 )(1/4))((1 − e−1 )(1/4))) ≈ 0.216

in both B1 and B2 .
Now, we must take an initial estimate of p(1) . We will also estimate that
p = 1 − e−1 . Once we run the EM algorithm a few times, it converges to
(1)

p ≈ 0.999653874. Since p = 1 − e−v1 ≈ 0.999653874, v1 ≈ 7.969. We still

need to find v2 and v3 , and since the calculation of v1 depended on these
two branch lengths, it may change again.
Next, we must do the same computation for v2 . We will root the tree
directly to the right of node * so that the node lays in v2 . Then, for site 1,
the likelihood of the tree is:
(∗) (∗)
X
L = e−v2 π2 L2 + (1 − e−v2 ) πS0 LS0 π2 .
S0
(∗) P (∗)
This means that A1 = π2 L2 and B1 = S0 πS0 LS0 π2 .
For site 2, the likelihood of the tree is:
(∗) (∗)
X
L = e−v2 π1 L1 + (1 − e−v2 ) πS0 LS0 π1 .
S0
(∗) P (∗)
This means that A2 = π1 L1 and B2 = S0 πS0 LS0 π1 .
(∗) (∗)
We are again left with the difficulty of calculating L2 and LS0 for site
(∗) (∗)
1 and L1 and LS0 for site 2. Fortunately, we now have a better estimate
for v1 . We will estimate that v1 = 7.969 and that v3 = 1. Then, for site 1,
(∗)
L2 = [(1 − e−7.969 )(1/4)][(1 − e−1 )(1/4)] ≈ 0.0395
and

(∗)
LS0 = 2[(1 − e−7.969 )(1/4)][(1 − e−1 )(1/4)]
+ [e−7.969 + (1 − e−7.969 )(1/4)][(1 − e−1 )(1/4)]
+ [(1 − e−7.969 )(1/4)][e−1 + (1 − e−1 )(1/4)] ≈ 0.250. (4.4)
 41

For site 2,

(∗)
L1 = [e−7.969 + (1 − e−7.969 )(1/4)][(1 − e−1 )(1/4)] ≈ 0.395
and

(∗)
LS0 = 2[(1 − e−7.969 )(1/4)][(1 − e−1 )(1/4)] + [e−7.969
+ (1 − e−7.969 )(1/4)][(1 − e−1 )(1/4)]
+ [(1 − e−7.969 )(1/4)][e−1 + (1 − e−1 )(1/4)] ≈ 0.250. (4.5)
Now, we have all of the components to find v2 using our iteration for-
mula (4.3). Again, we will estimate p(1) = 1 − e−1 . Once we run the EM
algorithm a few times, it converges will converge to a number as it did for
v1 . We will then be able to use this number to find v2 .
We follow the same procedure to find an estimate for v3 . Using our new
estimates for v1 and v2 , the EM algorithm will converge to some value. We
will use this value to find an estimation for v3 . We use these approxima-
tions of v1 , v2 , and v3 in the EM algorithm again until each of the values
converges.
We continue to use the resulting branch length estimates from the EM
algorithm to produce better estimates of the branch lengths until the succes-
sive estimates begin to converge. Then, the branch lengths that the iteration
converges to are accepted as the best estimates for the actual branch lengths
of the phylogenetic tree in this step of the maximum likelihood process.
We can now proceed to add our fourth species to the tree. Since there
are three branches in the 3-species tree, there are three possible locations
for species 4. For site 1 of the DNA sequences, the 4-species trees that can
result from the 3-species tree are:

Notice that a new node and two new branches are created after the addition
of species 4 to the tree.
42 Finding the Maximum Likelihood Tree - An Example

Now, we proceed as we did with the three-species tree. For each new
tree produced, we evaluate the branch lengths using the EM algorithm.
After the iteration converges and gives an estimate of each branch length,
we evaluate the likelihoods of each topology. We accept the topology that
gives us the greatest likelihood and dismiss the other two topologies. After
we have found our optimal 4-species tree, we can add species 5 to each
possible location in the accepted 4-species tree. After evaluating the branch
lengths and likelihoods of each resulting 5-species tree topology, we have a
5-species tree.

Figure 4.1: The 5-species tree resulting in the greatest likelihood before local
rearrangements.

Before we can accept this tree as our maximum likelihood tree, we must
perform local rearrangements to help test some possible phylogenetic trees
that we may have missed due to the order in which we added each species
to the tree. Suppose the tree in Figure 4.1 is the tree we obtain. For each
interior branch of the tree, we can perform Nearest Neighbor Interchange
(NNI). For interior branch, v7 . the resulting topologies are:
Next, we evaluate the branch lengths and likelihoods of these two topolo-
gies, accepting the tree that results in the largest likelihood. If this tree’s
likelihood is greater than the likelihood of the tree in Figure 4.1, we accept
the new tree. In the tree that we accept, whether it be the tree in Figure
4.1 or the new tree, we continue to perform NNI on any remaining interior
branches. Each time we do NNI, we evaluate the lengths of the branches
and find the likelihood of the new trees. Then, we accept each tree that
increases the likelihood. Remember that this is a greedy algorithm, so we
continue this process of local rearrangements until we feel that a tree gives
 43

us a sufficiently large likelihood for us to choose it as our maximum like-

lihood tree. The tree that we choose is our best guess for the phylogenetic
trees of the species from Middle Earth. This tree allows us to see how
closely related each of the species is to one another. Figure 4.2 illustrates
this.

Figure 4.2: Suppose this is our maximum likelihood tree for wizards, elves,
hobbits, dwarves, and humans. The branch lengths, v1 , v2 , . . . , v7 , are
known from the EM algorithm calculations. They help us see how closely
related one species is to another.
Chapter 5

The Limitations of Our Model

We have examined how probability and statistics can describe the process
of estimating the most likely phylogenetic tree for a set of DNA sequences.
As with any mathematical modeling problem, there are several assump-
tions that we made in order to ease the computational burden of finding
the maximum likelihood tree. It is important to consider the assumptions
that we have made and investigate ways in which to make our approach
better.
In Chapter 4, we took quite a large leap in modeling base substitution
with the Jukes-Cantor model. There are several models of DNA base sub-
stitution, including Kimura, Felsenstein, and Tamura (see [3] for descrip-
tions of these models). We did not collect any actual information about the
DNA sequences in our data besides the bases that we could observe. Af-
ter running a few experiments, one is able to determine which model of
base substitution best fits the set of species one is examining. For exam-
ple, in some DNA sequences, the rate of base substitution of a transition
can be different from the rate of base substitution of a transversion. Tran-
sitions and transversions categorize base substitutions. Both states before
and after the substitution can be purines (A or G) or pyramidines (C or
T), in which case we call the mutation a transversion. If the base was a
purine and was substituted with a pyramidine, or vice versa, we call the
base substitution a transition (note: the name, transition, does not relate
to the transition probabilties we discussed earlier). There are also cases in
which we are unable to observe which specific base we have at the tips of
our trees (i.e., in our set of DNA sequences). Sometimes, we are only able
to observe whether or not we have a purine or a pyrimadine at the tips of
the tree. In order to determine which model of base substitution best de-
46 The Limitations of Our Model

scribes the set of DNA sequences we have, we must perform experiments

and observations on the DNA.
Our model of DNA mutations also neglects the fact that deletions and
insertions can occur. This means that if we begin with K DNA sites, muta-
tions can cause an addition or removal of a DNA site. Then, we may end
up with K + 1, K − 1, K + 2, etc. DNA sites.
Another limitation of our model is that it does not consider whether or
not rates of substitution vary. We assumed that each DNA site mutated
independently from other DNA sites. Gillespie and Langley [6] observed
that the number of substitutions can vary in the two branches descending
from a common node depending on the location of the DNA site being
inspected. This means that depending on how close together two DNA
sites are, they may not be independent of each other.
We also made the assumption that the phylogenetic tree is bifurcating.
This is not always the case. We may have a multifurcating tree, in which
case, we would have to think of a different method for finding the possible
locations to add a new species when constructing a tree.
A few of these ambiguities can be solved using our current model of
base substitution. For example, when we have a set of DNA sequences in
which we are only able to observe whether we have purines or pyrimadines
at the tips of our branches, we can adjust the likelihoods of the tips of our
(i) (i) (i) (k)
tree. If we know that tip i is a purine, then L1 = L4 = 1 and L2 = L3 =
0. Then, to accomodate for some of the unknown parameters presented
here, we can either tweak the Jukes-Cantor model or use a totally different
model of base substitution.
The computational burden for the simple example we presented in Chap-
ter 4 was large, and factoring in these limitations can make the computation
even more difficult. Fortunately, this method of maximum likelihood esti-
mation has been programmed in PASCAL by Mark Moehring. It estimates
the terms of the p(k) during the branch length iteration process. Unfortu-
nately, the program is known to be quite slow. However, since the process
of iterating p(k) s can become extremely long and tedious, this program is
very useful (for more information about the program, see [2].
In this paper, we have described a method to find an estimate of the
topology of a set of DNA sequences for a group of species. In our exam-
ple in Chapter 4, we were able to find a maximum likelihood topology for
the family of species (see Figure 4.2). This information allows us to an-
alyze how closely related one species is to another in relation to genetic
makeup. We have not, however, determined the DNA of the species’ com-
 47

mon ancestors. The likelihoods of each topology must be analyzed further

to determine the states of each internal node of the tree.
Bibliography

[1] Dempster, A. P., Laird, N. M., and Rubin, D. (1977). Maximum like-
lihood from incomplete data via the em algorithm. Journal of the Royal
Statistical Society, 39:1–38.

[2] Felsenstein, J. (1981). Evolutionary trees from dna sequences: A maxi-

mum likelihood approach. Journal of Molecular Evolution, (17):368–376.

[3] Felsenstein, J. (2004). Inferring Phylogenies. Sinauer Associates, Inc.,

Sunderland, MA.

[4] Gascuel, O., editor (2007). Mathematics of evolution and phylogeny. Ox-
ford University Press, Oxford.

[5] Ghahramani, S. (2000). Fundamentals of Probability. Prentice Hall, Upper

Saddle River, NJ, second edition.

[6] Gillespie, J. H. and Langley, C. H. (1979). Are evolutionary rates really

variable? Journal of Molecular Evolution, pages 27–34.

[7] Jukes, T. H. and Cantor, C. (1969). Evolution of protein molecules. Mam-

malian protein metabolism, pages 21–123),.

[8] Sadava, D., Heller, C., Orians, G., Purves, B., and Hills, D. (2008). Life
the Science of Biology. The Courier Companies, Inc., Sunderland, MA,
eighth edition.