Phylogenetics

COS551, Fall 2003

Mona Singh
 Phylogenetics
• Phylogenetic trees illustrate the
evolutionary relationships among groups of
organisms, or among a family of related
nucleic acid or protein sequences
• E.g., how might have this family been
derived during evolution
 Hypothetical Tree
Relating Organisms
 Phylogenetic Relationships
Among Organisms
• Entrez: www.ncbi.nlm.nih.gov/Taxonomy
• Ribosomal database project:
rdp.cme.msu.edu/html/
• Tree of Life:
phylogeny.arizona.edu/tree/phylogeny.html
Globin Sequences
 Phylogeny Applications
• Tree of life: Analyzing changes that have
occurred in evolution of different organisms
• Phylogenetic relationships among genes can
help predict which ones might have similar
functions (e.g., ortholog detection)
• Follow changes occuring in rapidly
changing species (e.g., HIV virus)
 Phylogeny Packages
• PHYLIP, Phylogenetic inference package
– evolution.genetics.washington.edu/phylip.html
– Felsenstein
– Free!
• PAUP, phylogenetic analysis using
parsimony
– paup.csit.fsu.edu
– Swofford
 What data is used to
build trees?
• Traditionally: morphological features (e.g.,
number of legs, beak shape, etc.)
• Today: Mostly molecular data (e.g., DNA
and protein sequences)
 Data for Phylogeny
• Can be classified into two categories:
– Numerical data
• Distance between objects
e.g., distance(man, mouse)=500,
distance(man, chimp)=100
Usually derived from sequence data
– Discrete characters
• Each character has finite number of states
e.g., number of legs = 1, 2, 4
DNA = {A, C, T, G}
 Rooted vs Unrooted
Trees
Internal node

Root
External node

Rooted tree Unrooted tree

Note: Here, each node has three neighboring nodes

 Terminology
• External nodes: things under comparison;
operational taxonomic units (OTUs)
• Internal nodes: ancestral units; hypothetical; goal is
to group current day units
• Root: common ancestor of all OTUs under study.
Path from root to node defines evolutionary path
• Unrooted: specify relationship but not evolutionary
path
– If have an outgroup (external reason to believe certain
OTU branched off first), then can root
• Topology: branching pattern of a tree
• Branch length: amount of difference that occurred
along a branch
 How to reconstruct trees
• Distance methods: evolutionary distances are
computed for all OTUs and build tree where
distance between OTUs “matches” these distances
• Maximum parsimony (MP): choose tree that
minimizes number of changes required to explain
data
• Maximum likelihood (ML): under a model of
sequence evolution, find the tree which gives the
highest likelihood of the observed data
 Number of possible trees
Given n OTUs, there are unrooted trees

OTUs unrooted trees

3 1
4 3
5 15
10 2,027,025
 Number of possible trees
Given n OTUs, there are rooted trees

OTUs Rooted trees

Bottom Line: an 3 3
enumeration strategy
4 15
over all possible trees to
find the best one under 5 105
some criteria is not
feasible! 10 34,459,425
 Parsimony
Find tree which minimizes number of changes needed to
explain data

Ex:
123456
A GTCGTA
B GTCACT
C GCGGTA
D ACGACA
E ACGGAA
 Parsimony
• For given example tree and alignment, can do this
for all sites, and get away with as few as 9 changes
• Changing the tree (either the topology or labeling
of leaves) changes the minimum number of
changes need
• Two computational problems
– (Easy) Given a particular tree, how do you find
minimum number of changes need to explain data?
(Fitch)
– (Hard) How do you search through all trees?
 Parsimony: Fitch’s algorithm

Idea: construct set of possible nucleotides for internal nodes,

based on possible assignments of children
 Parsimony: Fitch’s algorithm
• For each site:
– Each leaf is labeled with set containing observed
nucleotide at that position
– For each internal node i with children j and k with
labels Sj and Sk

• Total # changes necessary for a site is # of union

operations
 Parsimony
• How do you search through all trees?
– Enumerate all trees (too many…)
– Can use techniques to try to limit the search space (e.g., branch and bound)
– or use heuristics (many possibilities)
• E.g., nearest neighbor interchange. Start with a tree and consider neighboring
trees. If any neighboring tree has fewer changes, take it as current tree. Stop when
no improvements

a b a b a c

c d d c b d
 Parsimony weaknesses
Parsimony analysis implicitly assumes that rate of change
along branches are similar

G G
G A

G A
A A
Real tree: two long branches Inferred tree
where G has turned to A independently
 Distance Methods
• Input: given an n x n matrix M where
Mij>=0 and Mij is the distance between
objects i and j
• Goal: Build an edge-weighted tree where
each leaf (external node) corresponds to one
object of M and so that distances measured
on the tree between leaves i and j
correspond to Mij
 Distance Methods
A B C D E
A 0
B 12 0
C 14 12 0
D 14 12 6 0
E 15 13 7 3 0

A tree exactly fitting the matrix does not always exist.

Distance Method Criteria
• Try to find the tree with distances dij which
“best fits” the distance data Mij
• Different possibilities for “best”
– Cavalli-Sforza criterion: minimize

– Fitch-Margoliash criterion: minimize

• Unfortunately, both lead to computationally

intractable problems (e.g., enumerating)
 Distance Method
Heuristic: UPGMA
• UPGMA (Unweighted group method with
arithmetic mean)
– Sequential clustering algorithm
– Start with things most similar
• Build a composite OTU
– Distances to this OTU are computed as
arithmetic means
– From new group of OTUs, pick pair with
highest similarity etc.
• Average-linkage clustering
 UPGMA: Visually

4
3 1 2 3 5 4
5
 UPGMA Example
A B C D
A 0
B 8 0
C 7 9 0
D 12 14 11 0

M B(AC) = (MBA + MBC)/2 = (8+9)/2=8.5

M D(AC) = (MDA + MDC)/2= (12+11)/2=11.5
 UPGMA Example
AC B D
AC 0
B 8.5 0
D 11.5 14 0

M (ABC)D = (MAD + MBD + MCD)/3 = (12+14+11)/3

 UPGMA: Example

ABC D
ABC 0
D 12.33 0
 UPGMA weaknesses
A B C D
A 0
B 8 0
C 7 9 0
D 12 14 11 0

In fact, exact fitting tree exists !

 UPGMA weaknesses
• UPGMA assumes that the rates of evolution
are the same among different lineages
• In general, should not use this method for
phylogenetic tree reconstruction (unless
believe assumption)
• Produces a rooted tree
• As a general clustering method (as we
discussed in an earlier lecture), it is better…
 Distance Method:
Neighbor Joining
• Most widely-used distance based method
for phylogenetic reconstruction
• UPGMA illustrated that it is not enough to
just pick closest neighbors
• Idea here: take into account averaged
distances to other leaves as well
• Produces an unrooted tree
 Neighbor Joining (NJ)

Start off with star tree; pull out pairs at a time

 NJ Algorithm
Step 1: Let
– (Almost) “average” distance to other nodes
Step 2: Choose i and j for which Mij – ui –uj is
smallest
– Look for nodes that are close to each other,
and far from everything else
– Turns out minimizing this is minimizing sum
of branch lengths
 NJ algorithm
Step 3: Define a new cluster (i, j), with a
corresponding node in the tree
i
(i,j)
j

Distance from i and j to node (i,j):

di, (i,j) = 0.5(Mij + ui-uj) Default: split distance but
if on average one is further
dj, (i,j) = 0.5(Mij +uj-ui) away, make it longer
 NJ Algorithm
Step 4: Compute distance between new cluster
and all other clusters:
M(ij)k = Mik+Mjk – Mij
2 i

k
(i,j)
j
Step 5: Delete i and j from matrix
and replace by (i, j)

Step 6: Continue until only 2 leaves remain

 NJ Performance
• Works well in practice
• If there is a tree that fits the matrix, it will
find it
• Can sometimes get trees with negative
length edges (!)
Computing Distances
Between Sequences

Could compute fraction of mismatches between

two sequences; however, this is an underestimate
of actual distance
 Computing Distances
Between Sequences

E.g., many
underlying
substitutions
possible

Use models of
substitution to
correct these values
Computing Distances
Between Sequences
Jukes & Cantor model
-Each position in DNA
sequence is independent
-Each position can mutates
with same probability to
any another base

Correction to observed
substitution rate (see notes):
Ex: Computing Distances
Between Sequences
• Alignment of two DNA sequences
– Length of alignment (non gapped positions): 100
– Number of differences: 25
• Naïve distance calculation = 25/100 = ¼
• Correction

• Other models for DNA, also protein (e.g.,

PAM)
 Maximum Likelihood
• Given a probabilistic model for nucleotide
(or protein) substitution (e.g., Jukes &
Cantor), pick the tree that has highest
probability of generating observed data
– I.e., Given data D and model M, find tree T
such that Pr(D|T, M) is maximized
• Models gives values pij(t), the probability of
going from nucleotide i to j in time t
 Maximum Likelihood
• Makes 2 independence assumptions
– Different sites evolve independently
– Diverged sequences (or species) evolve
independently after diverging
• If Di is data for ith site
 Maximum Likelihood
How to calculate Pr(Di|T,M) ?

pxy(t) ~ prob
of going from x
to y in time t
 Maximum Likelihood
• Given tree topology and branch lengths, can
efficiently calculate Pr(D|T, M) using dynamic
programming
– I.e., don’t have to enumerate over all internal states
• Finding best maximum likelihood tree is expensive
– Must consider all topologies
– Find best edge lengths for each topology
• Idea: use some search procedure, e.g., EM, to optimize these
lengths
 Assessing Reliability:
Bootstrap
Say we’ve inferred the following tree

Would like to get confidence

levels that 1 & 2 belong together,
and 3&4 belong together

1 2 3 4
 Assessing Reliability:
Bootstrap
Say we’re given following alignment:
12345678
1 GCAGTACT We’ll create a pseudosample
2 GTAGTACT by choosing sites randomly
until N sites are chosen
3 ACAATACC (N is length of alignment)
4 ACAACACT
 Assessing Reliability:
Bootstrap
Say chose 6th, 1st, 6th, 8th, …
12345678 6168 …
1 GCAGTACT AGAT …
2 GTAGTACT AGAT …
3 ACAATACC AAAC …
4 ACAACACT AAAT …
 Assessing Reliability:
Bootstrap
• Use pseudosample to construct tree
• Repeat many times
• Confidence of (1) and (2) together is
fraction of times they appear together in
trees generated from pseudosamples
95
90

1 2 3 4
 Phylogeny Flowchart
Family of Build Strong Y MP
sequences alignment similarity Methods

Recognizable Y
Distance
similarity Methods

N
ML
Methods

(Mount, Bioinformatics)
 Difference in Methods
• Maximum-likelihood and parsimony methods
have models of evolution
• Distance methods do not necessarily
– Useful aspect in some circumstances
• E.g., trees built based on whole genomes, presence or absence
of genes
• Religious wars over which methods to use
– Most people now believe ML based methods are best:
most sensitive at large evolutionary distances – but also
most time-consuming & depend on specific model of
evolution used
• Most commonly used packages contain software
for all three methods: may want to use more than 1
to have confidence in built tree
 Phylip
• Parsimony
– DNApenny or Protpars
• Distance
– Compute distance measure using DNAdist or
Protdist
– Neighbor (can use NJ or UPGMA)
• ML
– DNAml