Phylogenetic trees III Maximum Parsimony . Gerhard Jger ESSLLI - - PowerPoint PPT Presentation

. .

Phylogenetic trees III Maximum Parsimony

Gerhard Jäger ESSLLI 2016

Gerhard Jäger Maximum Parsimony ESSLLI 2016 1 / 30

Background

. .

Background

Gerhard Jäger Maximum Parsimony ESSLLI 2016 2 / 30

Background

Character-based tree estimation

distance-based tree estimation has several drawbacks:

very strong theoretical assumptions - e.g., all characters evolve at the same rate Neighbor Joining and UPGMA produce good but sub-optimal trees no solid statistical justification for those algorithms distances are black boxes — we get a tree, but we learn nothing about the history of individual characters

character-based tree estimation

estimates complete scenario (or distribution over scenarios) for each character finds the tree that best explains the observed variation in the data (at least in theory, that is...)

Gerhard Jäger Maximum Parsimony ESSLLI 2016 3 / 30

Parsimony

. .

Parsimony

Gerhard Jäger Maximum Parsimony ESSLLI 2016 4 / 30

Parsimony

Parsimony of a tree

background reading: Ewens and Grant (2005), 15.6 suppose a character matrix and a tree are given parsimony score: minimal number of mutations that has to be assumed to explain the character values at the tips, given the tree

Gerhard Jäger Maximum Parsimony ESSLLI 2016 5 / 30

Parsimony

Parsimony of a tree

Kopf "head" kop "head" head "head" tête "head" testa "head" cap "head"

Gerhard Jäger Maximum Parsimony ESSLLI 2016 6 / 30

Parsimony

Parsimony of a tree

Kopf "head" kop "head" head "head" tête "head" testa "head" cap "head"

Gerhard Jäger Maximum Parsimony ESSLLI 2016 6 / 30

Parsimony

Parsimony of a tree

Kopf "head" kop "head" head "head" tête "head" testa "head" cap "head" "head"

? ? ? ? ?

Gerhard Jäger Maximum Parsimony ESSLLI 2016 6 / 30

Parsimony

Parsimony of a tree

Kopf "head" kop "head" head "head" tête "head" testa "head" cap "head" *kop "head" testa "head"

Gerhard Jäger Maximum Parsimony ESSLLI 2016 6 / 30

Parsimony

Parsimony of a tree

Kopf "head" kop "head" head "head" tête "head" testa "head" cap "head" *kop "head" *haubud- "head" testa "head" caput "head" *kaput- "head"

Gerhard Jäger Maximum Parsimony ESSLLI 2016 6 / 30

Parsimony

Parsimony reconstruction

A C C B A B B A B B C

Parsimony = 2

Gerhard Jäger Maximum Parsimony ESSLLI 2016 7 / 30

Parsimony

Parsimony reconstruction

A C C A B B A B C

Parsimony = 3

A A

Gerhard Jäger Maximum Parsimony ESSLLI 2016 7 / 30

Parsimony

Parsimony reconstruction

A C C A B B A C

Parsimony = 3

A C C

Gerhard Jäger Maximum Parsimony ESSLLI 2016 7 / 30

Parsimony

Weighted parsimony reconstruction

A C C B A B B A B B C Weighted Parsimony = 3

Weight matrix A B C A 1 2 B 1 2 C 2 2

Gerhard Jäger Maximum Parsimony ESSLLI 2016 8 / 30

Parsimony

Weighted parsimony reconstruction

A C C A B B A B C A A Weighted Parsimony = 4

Weight matrix A B C A 1 2 B 1 2 C 2 2

Gerhard Jäger Maximum Parsimony ESSLLI 2016 8 / 30

Parsimony

Weighted parsimony reconstruction

A C C A B B A C Weighted Parsimony = 5 A C C

Weight matrix A B C A 1 2 B 1 2 C 2 2

Gerhard Jäger Maximum Parsimony ESSLLI 2016 8 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞ Gerhard Jäger Maximum Parsimony ESSLLI 2016 9 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

0 2 4 4 4 0

Gerhard Jäger Maximum Parsimony ESSLLI 2016 9 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 2 4 1 4 4 4 0 2 2

Gerhard Jäger Maximum Parsimony ESSLLI 2016 9 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 4 2 4 1 4 4 4 0 2 2 3 5

Gerhard Jäger Maximum Parsimony ESSLLI 2016 9 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 4 2 4 1 4 4 4 0 2 2 3 5

Gerhard Jäger Maximum Parsimony ESSLLI 2016 10 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 4 2 4 1 4 4 4 0 2 2 3 5

Gerhard Jäger Maximum Parsimony ESSLLI 2016 10 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 4 2 4 1 4 4 4 0 2 2 3 5

Gerhard Jäger Maximum Parsimony ESSLLI 2016 10 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 4 2 4 1 4 4 4 0 2 2 3 5

Gerhard Jäger Maximum Parsimony ESSLLI 2016 10 / 30

Parsimony

Dynamic Programming (Sankoff Algorithm)

wp(mother, s) = ∑

d∈daughters

min

s′∈states(w(s, s′) + wp(d, s′))

A C C A B B

0 ∞ ∞ 0 ∞ ∞

∞ ∞ ∞ ∞ ∞ ∞ ∞ ∞

1 3 4 2 4 1 4 4 4 0 2 2 3 5

Gerhard Jäger Maximum Parsimony ESSLLI 2016 10 / 30

Parsimony

Searching for the best tree

total parsimony score of tree: sum over all characters note: if weight matrix is symmetric, location of the root doesn’t matter Sankoff algorithm efficiently computes parsimony score of a given tree goal: tree which minimizes parsimony score no efficient way to find the optimal tree → heuristic tree search

Gerhard Jäger Maximum Parsimony ESSLLI 2016 11 / 30

Searching the tree space

. .

Searching the tree space

Gerhard Jäger Maximum Parsimony ESSLLI 2016 12 / 30

Searching the tree space

How many rooted tree topologies are there?

2 1

n=2

Gerhard Jäger Maximum Parsimony ESSLLI 2016 13 / 30

Searching the tree space

How many rooted tree topologies are there?

2 3 1 3 2 1 2 3 1 2 1

n=2 n=3

2 3 1

Gerhard Jäger Maximum Parsimony ESSLLI 2016 13 / 30

Searching the tree space

How many rooted tree topologies are there?

2 3 1 2 4 3 1 2 4 3 1 2 4 3 1 2 3 4 1 3 2 1 2 3 1 2 1

n=2 n=3 n=4

2 3 1 2 4 3 1

Gerhard Jäger Maximum Parsimony ESSLLI 2016 13 / 30

Searching the tree space

How many rooted tree topologies are there?

f(2) = 1 f(n + 1) = (2n − 3)f(n) f(n) = (2n − 3)! 2n−2(n − 2)!

2 1 3 3 4 15 5 105 6 945 7 10395 8 135135 9 2027025 10 34459425 11 654729075 12 13749310575 13 316234143225 14 7.9e + 12 15 2.1e + 14 16 6.1e + 15 17 1.9e + 17 18 6.3e + 18 19 2.2e + 20 20 8.2e + 21 21 3.1e + 23 22 1.3e + 25 23 5.6e + 26 24 2.5e + 28 25 1.1e + 30 26 5.8e + 31 27 2.9e + 33 28 1.5e + 35 29 8.6e + 36 30 4.9e + 38 31 2.9e + 40 32 1.7e + 42 33 1.1e + 44 34 7.2e + 45 35 4.8e + 47 36 3.3e + 49 37 2.3e + 51 38 1.7e + 53 39 1.3e + 55 40 1.0e + 57

Gerhard Jäger Maximum Parsimony ESSLLI 2016 14 / 30

Searching the tree space

How many unrooted tree topologies are there?

2 3 1

n=3 Gerhard Jäger Maximum Parsimony ESSLLI 2016 15 / 30

Searching the tree space

How many unrooted tree topologies are there?

2 3 4 1 3 2 4 1 4 2 3 1 2 3 1

n=3 n=4 Gerhard Jäger Maximum Parsimony ESSLLI 2016 15 / 30

Searching the tree space

How many unrooted tree topologies are there?

2 5 4 3 1 2 3 4 5 1 2 4 5 3 1 2 5 3 4 1 5 2 3 4 1 3 5 2 4 1 2 4 3 5 1 3 2 5 4 1 3 4 2 5 1 5 3 4 2 1 4 5 2 3 1 4 2 3 5 1 4 3 5 2 1 4 5 2 3 1 5 4 2 3 1 2 3 4 1 3 2 4 1 4 2 3 1 2 3 1

n=3 n=4 n=5 Gerhard Jäger Maximum Parsimony ESSLLI 2016 15 / 30

Searching the tree space

How many unrooted tree topologies are there?

f(3) = 1 f(n + 1) = (2n − 3)f(n) f(n) = (2n − 5)! 2n−3(n − 3)!

3 1 4 3 5 15 6 105 7 945 8 10395 9 135135 10 2027025 11 34459425 12 654729075 13 13749310575 14 316234143225 15 7.90e + 12 16 2.13e + 14 17 6.19e + 15 18 1.91e + 17 19 6.33e + 18 20 2.21e + 20 21 8.20e + 21 22 3.19e + 23 23 1.31e + 25 24 5.63e + 26 25 2.53e + 28 26 1.19e + 30 27 5.84e + 31 28 2.98e + 33 29 1.57e + 35 30 8.68e + 36 31 4.95e + 38 32 2.92e + 40 33 1.78e + 42 34 1.12e + 44 35 7.29e + 45 36 4.88e + 47 37 3.37e + 49 38 2.39e + 51 39 1.74e + 53 40 1.31e + 55

Gerhard Jäger Maximum Parsimony ESSLLI 2016 16 / 30

Searching the tree space

Heuristic tree search

tree space is too large to do an exhaustive search if n (number of taxa) is larger than 12 or so heuristic search:

start with some tree topology (e.g., Neighbor-Joining tree) apply a bunch of local modifications to the current tree if one of the modified tree has lower or equal parsimony, move to that tree stop if no further improvement is possible

⇒ standard approach for optimization problems in computer science

Gerhard Jäger Maximum Parsimony ESSLLI 2016 17 / 30

Searching the tree space

Tree modifications

three tree modifications commonly in use:

. .

1

Nearest Neighbor Interchange (NNI) . .

2

Tree Bisection and Reconnection (TBR) . .

3

Subtree Pruning and Regrafting (SPR)

local modifications are better than arbitrary moves in tree space because partial parsimony computations can be re-used in modified tree

Gerhard Jäger Maximum Parsimony ESSLLI 2016 18 / 30

Searching the tree space

Nearest Neighbor Interchange

Gerhard Jäger Maximum Parsimony ESSLLI 2016 19 / 30

Searching the tree space

Tree Bisection and Reconection

1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 10 9 1 2 3 4 5 6 7 8 10 9

Gerhard Jäger Maximum Parsimony ESSLLI 2016 20 / 30

Searching the tree space

Subtree Pruning and Regrafting

1 2 3 4 5 6 7 8 10 9 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 10 9

Gerhard Jäger Maximum Parsimony ESSLLI 2016 21 / 30

Searching the tree space

Heuristic tree search

NNI is very local → only O(n) possible moves SPR and TBR are more aggressive → O(n2)/O(n3) possible moves NNI search is comparatively fast, but prone to get stuck in local

• ptima

Gerhard Jäger Maximum Parsimony ESSLLI 2016 22 / 30

Searching the tree space

Running example: SPR search with cognate data

parsimony=1984

Spanish Portuguese H i n d i Nepali Bengali Greek Breton Welsh I r i s h Dutch G e r m a n English Swedish Danish Icelandic Polish Czech Russian Ukrainian Bulgarian Lithuanian French Italian Romanian C a t a l a n

starting with Neighbor Joining tree . . .

Gerhard Jäger Maximum Parsimony ESSLLI 2016 23 / 30

Searching the tree space

Running example: SPR search with cognate data

parsimony=1979

Spanish Portuguese H i n d i Nepali Bengali Greek Breton Welsh I r i s h Dutch G e r m a n English Swedish Danish Icelandic Polish Czech R u s s i a n Ukrainian Bulgarian Lithuanian French Italian Romanian Catalan

Gerhard Jäger Maximum Parsimony ESSLLI 2016 23 / 30

Searching the tree space

Running example: SPR search with cognate data

parsimony=1975

Spanish Portuguese Hindi N e p a l i Bengali Greek Breton Welsh I r i s h Dutch G e r m a n English Swedish Danish Icelandic Polish Czech Russian U k r a i n i a n Bulgarian Lithuanian French Italian Romanian Catalan

Gerhard Jäger Maximum Parsimony ESSLLI 2016 23 / 30

Searching the tree space

Running example: SPR search with cognate data

parsimony=1973

Spanish Portuguese Hindi N e p a l i Bengali Greek Breton Welsh I r i s h D u t c h German English Swedish Danish Icelandic Polish Czech R u s s i a n Ukrainian Bulgarian Lithuanian French Italian Romanian Catalan

Gerhard Jäger Maximum Parsimony ESSLLI 2016 23 / 30

Searching the tree space

Running example: SPR search with cognate data

parsimony=1969

Spanish Portuguese Hindi N e p a l i Bengali Greek Breton Welsh Irish D u t c h German English Swedish Danish Icelandic Polish Czech R u s s i a n Ukrainian Bulgarian Lithuanian French Italian R

• m

a n i a n Catalan

. . . Maximum Parsimony tree

Gerhard Jäger Maximum Parsimony ESSLLI 2016 23 / 30

Searching the tree space

Running example: SPR search with cognate data

there are actually 16 different trees with minimal parsimony score

Greek Irish Welsh Breton French Italian Catalan Portuguese Spanish Romanian Icelandic Swedish Danish German Dutch English Hindi Nepali Bengali Ukrainian Russian Polish Czech Bulgarian Lithuanian

Gerhard Jäger Maximum Parsimony ESSLLI 2016 24 / 30

Searching the tree space

MP tree for WALS characters

Bengali Breton Irish Welsh Bulgarian Greek Czech Lithuanian Polish Russian Ukrainian Catalan Romanian Italian Spanish Portuguese French Danish Swedish Icelandic Dutch German English Hindi Nepali

Gerhard Jäger Maximum Parsimony ESSLLI 2016 25 / 30

Searching the tree space

MP tree for sound-concept characters

Greek Bulgarian Russian Ukrainian Polish C z e c h Lithuanian I c e l a n d i c Swedish Danish Dutch G e r m a n English Catalan P

• r

t u g u e s e S p a n i s h I t a l i a n Romanian French B r e t

• n

Welsh Irish Hindi Bengali Nepali

Gerhard Jäger Maximum Parsimony ESSLLI 2016 26 / 30

Searching the tree space

Dollo parsimony

previous trees were estimated with a symmetric weight matrix if weights are asymmetric, location of the root matters extreme case: Dollo Parsimony w(0 → 1) = ∞

Spanish Portuguese Hindi Nepali Bengali Greek Breton Welsh Irish Dutch German English Swedish Danish Icelandic Polish Czech Russian Ukrainian Bulgarian Lithuanian French Italian Romanian Catalan

Gerhard Jäger Maximum Parsimony ESSLLI 2016 27 / 30

Searching the tree space

Maximum Parsimony: Discussion

Once we have found the best tree (or, in any event, which is very close to the best tree), we can reconstruct ancestral states via the Sankoff algorithm this allows to compute statistics about stability of characters, frequency and location of parallel changes etc.

⇒ much more informative than distance-based inference

Gerhard Jäger Maximum Parsimony ESSLLI 2016 28 / 30

Searching the tree space

Maximum Parsimony: Discussion

disadvantages of MP:

simulation studies: capacity to recover the true tree is decent but not

• verwhelming

possibility of multiple mutations on a single branch is not taken into consideration all characters are treated equal; no discrimination between stable and volatile characters ties are common, especially if you have few data values for weight matrix are ad hoc no real theoretical justification

Why should the true tree minimize the total number of mutations? Rests on a valid intuition: Mutations are unlikely, so assuming fewer mutations increases the likelihood of the data. Likelihood is not formally derived from a probabilistic modell though.

Next step: Maximum Likelihood tree estimation

Gerhard Jäger Maximum Parsimony ESSLLI 2016 29 / 30

Searching the tree space

Hands on

Install the software Paup*. Go to the directory where you have the put the nexus files and type > paup4 ielex.bin.nex At Paup’s command prompt, type paup> hsearch. Display tree with paup> describetree /plot=phylo Save result with paup> savetree format=newick file = ielex.mp.tre \ brlen=yes Leave Paup* with paup> q Install Dendroscope or FigTree and load ielex.mp.tre.

Gerhard Jäger Maximum Parsimony ESSLLI 2016 30 / 30

Searching the tree space

Ewens, W. and G. Grant (2005). Statistical Methods in Bioinformatics: An

• Introduction. Springer, New York.

Gerhard Jäger Maximum Parsimony ESSLLI 2016 30 / 30