Cédric Notredame (03/01/22) Recent Progress in Multiple Sequence Alignments: A Survey Cédric Notredame
Jan 11, 2016
Cédric Notredame (21/04/23)
Recent Progress in Multiple Sequence
Alignments:A Survey
Cédric Notredame
Cédric Notredame (21/04/23)
Our Scope
What are The existing Methods?
How Do They Work: -Assemby Algorithms-Weighting Schemes.
When Do They Work ?
Which Future?
Cédric Notredame (21/04/23)
Outline
-Introduction
-A taxonomy of the existing Packages
-A few algorithms…
-Performance Comparison using BaliBase
Cédric Notredame (21/04/23)
Introduction
Cédric Notredame (21/04/23)
What Is A Multiple Sequence Alignment?
A MSA is a MODEL
It Indicates the RELATIONSHIP between residues of different sequences.
It REVEALS-Similarities-Inconsistencies
LIKE ANYMODEL
Cédric Notredame (21/04/23)
chite ---ADKPKRPLSAYMLWLNSARESIKRENPDFK-VTEVAKKGGELWRGLKDwheat --DPNKPKRAPSAFFVFMGEFREEFKQKNPKNKSVAAVGKAAGERWKSLSEtrybr KKDSNAPKRAMTSFMFFSSDFRS----KHSDLS-IVEMSKAAGAAWKELGPmouse -----KPKRPRSAYNIYVSESFQ----EAKDDS-AQGKLKLVNEAWKNLSP ***. ::: .: .. . : . . * . *: *
chite AATAKQNYIRALQEYERNGG-wheat ANKLKGEYNKAIAAYNKGESAtrybr AEKDKERYKREM---------mouse AKDDRIRYDNEMKSWEEQMAE * : .* . :
How Can I Use A Multiple Sequence Alignment?
Extrapolation
Motifs/Patterns
Phylogeny
Profiles
Struc. Prediction
Multiple Alignments Are CENTRAL to MOST Bioinformatics Techniques.
Cédric Notredame (21/04/23)
How Can I Use A Multiple Sequence Alignment?
Multiple Alignments Is the most INTEGRATIVE Method Available Today.
We Need MSA to INCORPORATE existing DATA
Cédric Notredame (21/04/23)
Why Is It Difficult To Compute A multiple Sequence Alignment?
A CROSSROAD PROBLEM
BIOLOGY:What is A Good Alignment
COMPUTATIONWhat is THE Good Alignment
chite ---ADKPKRPLSAYMLWLNSARESIKRENPDFK-VTEVAKKGGELWRGLKDwheat --DPNKPKRAPSAFFVFMGEFREEFKQKNPKNKSVAAVGKAAGERWKSLSEtrybr KKDSNAPKRAMTSFMFFSSDFRS----KHSDLS-IVEMSKAAGAAWKELGPmouse -----KPKRPRSAYNIYVSESFQ----EAKDDS-AQGKLKLVNEAWKNLSP ***. ::: .: .. . : . . * . *: *
Cédric Notredame (21/04/23)
Why Is It Difficult To Compute A multiple Sequence Alignment ?
BIOLOGY
CIRCULAR PROBLEM....
GoodSequences
GoodAlignment
COMPUTATION
Cédric Notredame (21/04/23)
A Taxonomy of Multiple Sequence Alignment Methods
Cédric Notredame (21/04/23)
Grouping According to the assembly Algorithm
Cédric Notredame (21/04/23)
SimultaneousAs opposed to Progressive
Exact As opposed to Heursistic
Stochastic As opposed to Determinist
Iterative As opposed to Non Iterative
[Simultaneous: they simultaneously use all the information]
[Heuristics: cut corners like Blast Vs SW]
[Heuristics: do not guarranty an optimal solution]
[Stochastic: contain an element of randomness]
[Stochastic: Example of a Monte Carlo Surface estimation ]
[Iterative: Most stochastic methods are iterative]
[Iterative: run the same algorithm many times]
Cédric Notredame (21/04/23)Iterative
Iteralign
Prrp
SAM HMMer
SAGAGA
Clustal
Dialign
T-Coffee
ProgressiveSimultaneous
MSA
POA OMA
PralineMAFFT
DCA
Combalign
Non tree based
GAs
HMMs
Cédric Notredame (21/04/23)Iterative
Iteralign
Prrp
SAM HMMer
GA
Clustal
Dialign
T-Coffee
ProgressiveSimultaneous
MSA
POA OMA
PralineMAFFT
DCA
Combalign
StochasticSAGA
Cédric Notredame (21/04/23)
NEARLY EVERY OPTIMISATIONALGORITHM
HAS BEEN APPLIED TO THEMSA PROBLEM!!!
Cédric Notredame (21/04/23)
Grouping According to the Objective Function
Cédric Notredame (21/04/23)
Scoring an Alignment: Evolutionary based
methods
BIOLOGYHow many events separate my sequences?
Such an evaluation relies on a biological model.
COMPUTATIONEvery position musd be independant
Cédric Notredame (21/04/23)
REAL Tree
Model: ALL the sequences evolved from the same ancestor
A
A
A C
Tree: Cost=1C
AAACC
A CA
PROBLEM: We do not know the true tree
Cédric Notredame (21/04/23)
STAR Tree
Model: ALL the sequences have the same ancestor
A
A
A CStar Tree: Cost=2
C
AAACC
A
PROBLEM: the tree star is phylogenetically wrong
Cédric Notredame (21/04/23)
Sums of Pairs
Model=Every sequence is the ancestor of every sequence
A
A
A CSums of Pairs: Cost=6
CAAACC
PROBLEM: -over-estimation of the mutation costs-Requires a weighting scheme
lk
li
kii mmsmS ,
[s(a,b): matrix]
[i: column i]
[k, l: seq index]
Cédric Notredame (21/04/23)
Sums of Pairs: Some of itslimitations (Durbin,
p140)
LLLLL
GCost=5*N*(N-1)/2-(5)*(N-1) - (-4)*(N-1)
[glycine effect]
Cost=5*N*(N-1)/2-(9)*(N-1)
Cost= 5*N*(N-1)/2[5: Leucine Vs Leucine with Blosum50]
Cédric Notredame (21/04/23)
Sums of Pairs: Some of its limitations (Durbin,
p140)
LLLLL
G
Delta=2*(9)*(N-1)
5*N*(N-1)=
(9)
5*N
N
Delta
Conclusion: The more Leucine, the less expensive it gets to add a Glycin to the column...
Cédric Notredame (21/04/23)
Enthropy based Functions
Model: Minimize the enthropy (variety) in each Column
AAACC
PROBLEM: -requires a simultaneous alignment-assumes independant sequences
j
jiia amc [number of Alanine (a) in column i]
a
iaiai PcmS log* [Score of column i][a: alphabet]
[P can incorporate pseudocounts]
S=0 if the column is conserved
Cédric Notredame (21/04/23)
Consistency based Functions
Model: Maximise the consistency (agreement) with a list of constraints (alignments)
PROBLEM: -requires a list of constraints
AAACC
lk
li
kii mmS , [kand l are sequences, i is a column]
Existsmmmm li
ki
li
ki ,1,
[the two residues are found aligned in the list of constraints]
Cédric Notredame (21/04/23)
Concistency Based
Iteralign
Dialign
T-Coffee
Praline
Combalign
Prrp
ClustalPOA
MSA
MAFFTOMA
DCA
SAGA
WeightedSums
of Pairs
EnthropySAM HMMer
GIBBS
Cédric Notredame (21/04/23)
A few Multiple Sequence Alignment Algorithms
Cédric Notredame (21/04/23)
A Few Algorithms
MSA and DCA
ClustalW
Dialign IIPrrp
SAGA
GIBBS Sampler
MAFFT
POA
Cédric Notredame (21/04/23)
Simultaneous: MSA and DCA
Cédric Notredame (21/04/23)
Simultaneous Alignments : MSA
1) Set Bounds on each pair of sequences (Carillo and Lipman)
2) Compute the Maln within the Hyperspace
-Few Small Closely Related Sequence.
-Do Well When They Can Run.
-Memory and CPU hungry
Cédric Notredame (21/04/23)
MSA: the carillo and Lipman bounds
chite ---ADKPKRPLSAYMLWLNSARESIKRENPDFK-VTEVAKKGGELWRGLKDwheat --DPNKPKRAPSAFFVFMGEFREEFKQKNPKNKSVAAVGKAAGERWKSLSEtrybr KKDSNAPKRAMTSFMFFSSDFRS----KHSDLS-IVEMSKAAGAAWKELGPmouse -----KPKRPRSAYNIYVSESFQ----EAKDDS-AQGKLKLVNEAWKNLSP
chite ---ADKPKRPLSAYMLWLNSARESIKRENPDFK-VTEVAKKGGELWRGLKDwheat --DPNKPKRAPSAFFVFMGEFREEFKQKNPKNKSVAAVGKAAGERWKSLSE
chite ---ADKPKRPLSAYMLWLNSARESIKRENPDFK-VTEVAKKGGELWRGLKDtrybr KKDSNAPKRAMTSFMFFSSDFRS----KHSDLS-IVEMSKAAGAAWKELGP
S( )=
S(S(
)
)+
…[Pairwise projection of sequences k and l]
Cédric Notredame (21/04/23)
MSA: the carillo and Lipman bounds
a(k,l)=score of the projection k l in the optimal MSA
â(k,l)=score of the optimal alignment of k l
(a(x,y))=score of the complete multiple alignment
a(k,l) â(k,l) a(k,m) â(k,m)
?
Upper
Lower
Cédric Notredame (21/04/23)
MSA: the carillo and Lipman bounds
LM: a lower bound for the complete MSA
a(k,l)>=LM +â(k,l)-(â(x,y))
LM<=(â(x,y)) - (â(k,l)-a(k,l))
a(k,l) â(k,l)
â(k,l)
LM+ â(k,l)-(â(x,y))
?
Cédric Notredame (21/04/23)
MSA: the carillo and Lipman bounds
LM: can be measured on ANY heuristic alignment
a(k,l) â(k,l)
â(k,l)
LM+ â(k,l)-(â(x,y)) ä(k,l)
LM = (ä(x,y))
The better LM, the tighter the bounds…
Cédric Notredame (21/04/23)
MSA: the carillo and Lipman bounds
backward Forward
Best( M-i, N-j) Best( 0-i, 0-j)
0
M
N 0
M
N
+
Cédric Notredame (21/04/23)
Simultaneous Alignments : MSA
1) Set Bounds on each pair of sequences (Carillo and Lipman)
2) Compute the Maln within the Hyperspace
-Few Small Closely Related Sequence.
-Do Well When They Can Run.
-Memory and CPU hungry
Cédric Notredame (21/04/23)
Simultaneous Alignments : DCA
-Few Small Closely Related Sequence, but less limited than MSA
-Do Well When Can Run.
-Memory and CPU hungry, but less than MSA
Cédric Notredame (21/04/23)
Simultaneous With a New Sequence Representaion:
POA-Partial Ordered Graph
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
POA
POA makes it possible to represent complex relationships:
-domain deletion-domain inversions
Cédric Notredame (21/04/23)
Progressive: ClustalW
Cédric Notredame (21/04/23)
Progressive Alignment: ClustalW
Feng and Dolittle, 1988; Taylor 198ç
Clustering
Cédric Notredame (21/04/23)
Dynamic Programming Using A Substitution Matrix
Progressive Alignment: ClustalW
Cédric Notredame (21/04/23)
Tree based Alignment : Recursive Algorithm
Align ( Node N){
if ( N->left_child is a Node)A1=Align ( N->left_child)
else if ( N->left_child is a Sequence)A1=N->left_child
if (N->right_child is a node)A2=Align (N->right_child)
else if ( N->right_child is a Sequence)A2=N->right_child
Return dp_alignment (A1, A2)}
A D E F GCB
Cédric Notredame (21/04/23)
Progressive Alignment : ClustalW
-Depends on the ORDER of the sequences (Tree).
-Depends on the CHOICE of the sequences.
-Depends on the PARAMETERS:
•Substitution Matrix.
•Penalties (Gop, Gep).
•Sequence Weight.
•Tree making Algorithm.
Cédric Notredame (21/04/23)
Weighting Within ClustalWProgressive Alignment : ClustalW Weighting
Cédric Notredame (21/04/23)
Position Specific GOPProgressive Alignment : ClustalW GOP
Cédric Notredame (21/04/23)
ClustalW is the most Popular Method
-Fast
-Greedy Heuristic (No Guarranty).
Progressive Alignment : ClustalW
-Scales Well: N, N L3 2 2
Cédric Notredame (21/04/23)
Progressive Alignment With a Heuristic DP:
MAFFT
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
ProgressiveAnd
Concistency BasedDialign II
Cédric Notredame (21/04/23)
Dialign II
1) Identify best chain of segments on each pair of sequence. Assign a Pvalue to each Segment Pair.
3) Assemble the alignment according to the segment pairs.
2) Ré-évaluate each segment pair according to its consistency with the others
Cédric Notredame (21/04/23)
Dialign II
-May Align Too Few Residues
-No Gap Penalty-Does well with ESTs
Cédric Notredame (21/04/23)
ProgressiveAnd
Concistency BasedT-COFFEE
Cédric Notredame (21/04/23)
Mixing Local and Global Alignments
Local Alignment Global Alignment
Extension
Multiple Sequence Alignment
Cédric Notredame (21/04/23)
What is a library?
Extension+T-Coffee
2Seq1 MySeqSeq2 MyotherSeq#1 21 1 253 8 70….
3Seq1 anotherseqSeq2 atsecondoneSeq3 athirdone#1 21 1 25#1 33 8 70….
Cédric Notredame (21/04/23)
Iterative
Cédric Notredame (21/04/23)
7.16.1 ProgressiveIterative Methods
-HMMs, HMMER, SAM.
-Slow, Sometimes Inaccurate-Good Profile Generators
Cédric Notredame (21/04/23)
7.16.2 PrrpInitial Alignment
Tree and weights computation
Weights converged End
Realign two sub-groups
Alignment converged
YES
NO
YES NO
Inner Iteration
Outer Iteration
Iterative Methods : Prrp
Cédric Notredame (21/04/23)
Iterative Sochastic:SAGA, The Genetic
Algorithm
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
Automatic scheduling of the operators
Cédric Notredame (21/04/23)
Cédric Notredame (21/04/23)
Weighting Schemes
Cédric Notredame (21/04/23)
The Problem
The sequences Contain Correlated Information
Most scoring Schemes Ignore this Correlation
Cédric Notredame (21/04/23)
Weighting Sequence Pairs with a Tree:
Carillo and LipmanRationale I
Cédric Notredame (21/04/23)
A D E F GCB
E=EDGE
P=Evolutive Path from A to X
E must contribute the same weight to every path P that goes throught it.
QUESTION: Which Weight for a Pair of Sequences
All the weights using E must sum to 1: (WP,E)=1.
Wp=Nk-1)
1
Nk: Number of Edges meeting on Node k.
Cédric Notredame (21/04/23)
USAGE
]][[*),( yB
xAAB
yB
xA RRMatWRRScore
Cédric Notredame (21/04/23)
PROBLEM: Weight Depends only on the Tree topology
B
A C
AB: 0.5AC: 0.5BC: 0.5.
B
A C
AB: 0.5AC: 0.5BC: 0.5.
Cédric Notredame (21/04/23)
Weighting Sequences with a Tree
Clustal WWeights
Cédric Notredame (21/04/23)
GA D E FCB
QUESTION: Which Weight for Sequences ?
W=Length *1/4
W=Length *1/2
W=Length *1
GG W=W)
Number Sequences Sharing Edge
Edge LengthWseq =
Cédric Notredame (21/04/23)
USAGE
]][[**),( yB
xABA
yB
xA RRMatWWRRScore
Cédric Notredame (21/04/23)
PROBLEM: Overweight of distant sequences
D E F G
C-C Will dominate the Alignment
-C Will be very Difficult to align
Cédric Notredame (21/04/23)
Performance Comparison Using
Collections of Reference
Alignments: BaliBase and
Ribosomal RNA
Cédric Notredame (21/04/23)
What Is BaliBaseBaliBase
BaliBase is a collection of reference Multiple Alignments
The Structure of the Sequences are known and were used to assemble the MALN.
Evaluation is carried out by Comparing the Structure Based Reference Alignment With its Sequence Based Counterpart
Cédric Notredame (21/04/23)
What Is BaliBaseBaliBase
DALI, Sap …
Method X
Comparison
Cédric Notredame (21/04/23)
What Is BaliBaseBaliBase
DescriptionPROBLEM
Source: BaliBase, Thompson et al, NAR, 1999,
Even Phylogenic Spread.
One Outlayer Sequence
Two Distantly related Groups
Long Internal Indel
Long Terminal Indel
Cédric Notredame (21/04/23)
Choosing The Right Method
Cédric Notredame (21/04/23)
Choosing The Right Method (POA Evaluation)
Cédric Notredame (21/04/23)
Choosing The Right Method (POA Evaluation)
Cédric Notredame (21/04/23)
Choosing The Right Method (MAFFT evaluation)
Cédric Notredame (21/04/23)
Choosing The Right Method (MAFFT evaluation)
Cédric Notredame (21/04/23)
Choosing The Right Method (MAFFT evaluation)
Cédric Notredame (21/04/23)
Conclusion
Cédric Notredame (21/04/23)
What Is BaliBaseWhich Method ?
PROBLEM
Source: BaliBase, Thompson et al, NAR, 1999,
Strategy
Strategy
ClustalW, T-coffee,MSA, DCA
PrrP,T-Coffee
Dialign
T-Coffee
T-Coffee
Dialign
T-Coffee
Cédric Notredame (21/04/23)
Methods /Situtations
1-Carillo and Lipman:-MSA, DCA.
-Few Small Closely Related Sequence.
2-Segment Based:-DIALIGN, MACAW.
-May Align Too Few Residues-Good For Long Indels
-Do Well When They Can Run.
3-Iterative:-HMMs, HMMER, SAM.
-Slow, Sometimes Inaccurate-Good Profile Generators
4-Progressive: -ClustalW, Pileup, Multalign…-Fast and Sensitive
Cédric Notredame (21/04/23)
Addresses
MAFFT Progressive www.biophys.kyoto-u.jp/katoh POA Progressive/Simulataneous www.bioinformatics.ucla.edu/poa MUSCLE Progressive/Iterative www.drive5.com/muscle/