Title: http://creativecommons.org/licenses/by-sa/2.0/
1http//creativecommons.org/licenses/by-sa/2.0/
2CIS786, Lecture 3
3Maximum Parsimony
- Character based method
- NP-hard (reduction to the Steiner tree problem)
- Widely-used in phylogenetics
- Slower than NJ but more accurate
- Faster than ML
- Assumes i.i.d.
4Maximum Parsimony
- Input Set S of n aligned sequences of length k
- Output A phylogenetic tree T
- leaf-labeled by sequences in S
- additional sequences of length k labeling the
internal nodes of T - such that is minimized.
5Maximum parsimony (example)
- Input Four sequences
- ACT
- ACA
- GTT
- GTA
- Question which of the three trees has the best
MP scores?
6Maximum Parsimony
ACT
ACT
ACA
GTA
GTT
GTT
ACA
GTA
GTA
ACA
ACT
GTT
7Maximum Parsimony
ACT
ACT
ACA
GTA
GTT
GTA
ACA
ACT
2
1
1
3
3
2
GTT
GTT
ACA
GTA
MP score 7
MP score 5
GTA
ACA
ACA
GTA
2
1
1
ACT
GTT
MP score 4
Optimal MP tree
8Maximum Parsimony computational complexity
9Local search strategies
10Local search for MP
- Determine a candidate solution s
- While s is not a local minimum
- Find a neighbor s of s such that MP(s)ltMP(s)
- If found set ss
- Else return s and exit
- Time complexity unknown---could take forever or
end quickly depending on starting tree and local
move - Need to specify how to construct starting tree
and local move
11Starting tree for MP
- Random phylogeny---O(n) time
- Greedy-MP
12Greedy-MP
Greedy-MP takes O(n3k) time
13Faster Greedy MP3-way labeling
- If we can assign optimal labels to each internal
node rooted in each possible way, we can speed up
computation by order of n - Optimal 3-way labeling
- Sort all 3n subtrees using bucket sort in O(n)
- Starting from small subtrees compute optimal
labelings - For each subtree rooted at v, the optimal
labelings of children nodes is already computed - Total time O(nk)
14Faster Greedy MP3-way labeling
- If we can assign optimal labels to each internal
node rooted in each possible way, we can speed up
computation by order of n - Optimal 3-way labeling
- Sort all 3n subtrees using bucket sort in O(n)
- Starting from small subtrees compute optimal
labelings - For each subtree rooted at v, the optimal
labelings of children nodes is already computed - Total time O(nk)
With optimal labeling it takes constant Time to
compute MP score for each Edge and so total
Greedy-MP time Is O(n2k)
15Local moves for MP NNI
- For each edge we get two different topologies
- Neighborhood size is 2n-6
16Local moves for MP SPR
- Neighborhood size is quadratic in number of taxa
- Computing the minimum number of SPR moves between
two rooted phylogenies is NP-hard
17Local moves for MP TBR
- Neighborhood size is cubic in number of taxa
- Computing the minimum number of TBR moves between
two rooted phylogenies is NP-hard
18- Tree Bisection and Reconnection (TBR)
19- Tree Bisection and Reconnection (TBR)
Delete an edge
20- Tree Bisection and Reconnection (TBR)
21- Tree Bisection and Reconnection (TBR)
Reconnect the trees with a new edge that
bifurcates an edge in each tree
22Local optima is a problem
23Iterated local search escape local optima by
perturbation
Local optimum
Local search
24Iterated local search escape local optima by
perturbation
Local optimum
Local search
Perturbation
Output of perturbation
25Iterated local search escape local optima by
perturbation
Local optimum
Local search
Perturbation
Local search
Output of perturbation
26ILS for MP
- Ratchet
- Iterative-DCM3
- TNT
27Iterated local search escape local optima by
perturbation
Local optimum
Local search
Perturbation
Local search
Output of perturbation
28Ratchet
- Perturbation input alignment and phylogeny
- Sample with replacement p of sites and reweigh
them to w - Perform local search on modified dataset starting
from the input phylogeny - Reset the alignment to original after completion
and output the local minimum
29Ratchet escaping local minimaby data
perturbation
Local optimum
Local search
Ratchet search
Local search
Output of ratchet
30Ratchet escaping local minimaby data
perturbation
Local optimum
Local search
Ratchet search
Local search
Output of ratchet
But how well does this perform? We have to
examine this experimentally on real data
31Experimental methodology for MP on real data
- Collect alignments of real datasets
- Usually constructed using ClustalW
- Followed by manual (eye) adjustments
- Must be reliable to get sensible tree!
- Run methods for a fixed time period
- Compare MP scores as a function of time
- Examine how scores improve over time
- Rate of convergence of different methods (not
sequence length but as a function of time)
32Experimental methodology for MP on real data
- We use rRNA and DNA alignments
- Obtained from researchers and public databases
- We run iterative improvement and ratchet each for
24 hours beginning from a randomized greedy MP
tree - Each method was run five times and average scores
were plotted - We use PAUP---very widely used software package
for various types of phylogenetic analysis
33500 aligned rbcL sequences (Zilla dataset)
34854 aligned rbcL sequences
352000 aligned Eukaryotes
367180 aligned 3domain
3713921 aligned Proteobacteria
38Comparison of MP heuristics
- What about other techniques for escaping local
minima? - TNT a combination of divide-and-conquer,
simulated annealing, and genetic algorithms - Sectorial search (random) construct ancestral
sequence states using parsimony randomly select
a subset of nodes compute iterative-improvement
trees and if better tree found then replace - Genetic algorithm (fuse) Exchange subtrees
between two trees to see if better ones are found - Default search (1) Do sectorial search starting
from five randomized greedy MP trees (2) apply
genetic algorithm to find better ones (3) output
best tree
39Comparison of MP heuristics
- What about other techniques for escaping local
minima? - TNT a combination of divide-and-conquer,
simulated annealing, and genetic algorithms - Sectorial search (random) construct ancestral
sequence states using parsimony randomly select
a subset of nodes compute iterative-improvement
trees and if better tree found then replace - Genetic algorithm (fuse) Exchange subtrees
between two trees to see if better ones are found - Default search (1) Do sectorial search starting
from five randomized greedy MP trees (2) apply
genetic algorithm to find better ones (3) output
best tree
How does this compare to PAUP-ratchet?
40Experimental methodology for MP on real data
- We use rRNA and DNA alignments
- Obtained from researchers and public databases
- We run PAUP-ratchet, TNT-default, and
TNT-ratchet each for 24 hours beginning from
randomized greedy MP trees - Each method was run five times on each dataset
and average scores were plotted
41500 aligned rbcL sequences (Zilla dataset)
42854 aligned rbcL sequences
432000 aligned Eukaryotes
447180 aligned 3domain
4513921 aligned Proteobacteria
46Can we do even better?
- Yes! But first lets look at
- Disk-Covering Methods
47Disk Covering Methods (DCMs)
- DCMs are divide-and-conquer booster methods. They
divide the dataset into small subproblems,
compute subtrees using a given base method, merge
the subtrees, and refine the supertree. - DCMs to date
- DCM1 for improving statistical performance of
distance-based methods. - DCM2 for improving heuristic search for MP and
ML - DCM3 latest, fastest, and best (in accuracy and
optimality) DCM
48DCM2 technique for speeding up MP searches
49DCM2 decomposition
- DCM2
- Input distance matrix d, threshold
, sequences S - Algorithm
- 1a. Compute a threshold graph G using q and d
- 1b. Perform a minimum weight triangulation of G
- Find separator X in G which minimizes max
where are the connected components of G
X - Output subproblems as .
50Threshold graph
- Add edges until graph is connected
- Perform minimum weight triangulation
- NP-hard
- Triangulated graphperfect elimination ordering
(PEO) - Max cliques can be determined in linear time
- Use greedy triangulation heuristic compute PEO
by adding vertices which minimize largest edge
added - Worst case is O(n3) but fast in practice
51Finding DCM2 separator
- Find separator X in G which minimizes max
where are the connected components of G
X - Output subproblems as
- This takes O(n3) worst case time perform depth
first search on each component (O(n2)) for each
of O(n) separators
52DCM2 subsets
53DCM3 decomposition - example
54DCM1 vs DCM2
DCM1 decomposition NJ gets better accuracy on
small diameter subproblems (which we shall return
to later)
DCM2 decomposition Getting a smaller number of
smaller subproblems speeds up solution
55We saw how decomposition takes place, now on to
supertree methods
56Supertree Methods
57Optimization problems
- Subtree Compatibility Given set of trees
- ,does there exist tree
,such that, (we
say contains ). - NP-hard (Steel 1992)
- Special cases are poly-time (rooted trees, DCM)
- MRP also NP-hard
58Direct supertree methods
- Strict consensus supertrees, MinCutSupertrees
59Indirect supertree methods
60MRP---Matrix Representation using Parsimony (very
popular)
61Strict Consensus Merger---faster and used in DCMs
62Strict Consensus Merger compatible subtrees
63Strict Consensus Merger compatible but collision
64Strict Consensus Merger incompatible subtrees
65Strict Consensus Merger incompatible and
collision
66Strict Consensus Merger difference from Gordons
SC method
67We saw how decomposition takes place, now on to
supertree methods
68Tree Refinement
- Challenge given unresolved tree, find optimal
refinement that has an optimal parsimony score - NP-hard
69Tree Refinement
70We saw how decomposition takes place, now on to
supertree methods
71Comparing DCM decompositions
72Study of DCM decompositions
Comparison of MP scores
Comparison of running times
DCM2 is faster and better than DCM1
73Best DCM (DCM2) vs Random
Comparison of MP scores
Comparison of running times
DCM2 is better than RANDOM w.r.t MP scores and
running times
74DCM2 (comparing two different thresholds)
Comparison of MP scores
Comparison of running times
75Threshold selection techniques
Biological dataset of 503 rRNA sequences.
Threshold value at which we get two subproblems
has best MP score.
76Comparing supertree methods
77MRP vs. SCM
Comparison of MP scores
Comparison of running times
- SCM is better than MRP
78Comparing tree refinement techniques
79Study of tree refinement techniques
Comparison of MP scores
Comparison of running times
Constrained tree search had best MP scores but is
slower than other methods
80Next time
- DCM1 for improving NJ
- Recursive-Iterative-DCM3 state of the art in
solving MP and ML