CrossWA: A new approach of combining pairwise and three-sequence alignments to improve the accuracy for highly divergent sequence alignment

1
CrossWA A new approach of combining pairwise and
three-sequence alignments to improve the accuracy
for highly divergent sequence alignment

• Che-Lun Hung, Chun-Yuan Lin, Yeh-Ching Chung, and
  Chuan Yi Tang
• National Tsing Hua University, Taiwan

Sixth International Conference on
Bioinformatics InCoB2007 HKUST, Hong Kong
2
Outline

• Introduction
• Motivation
• Algorithm
• Experiments
• Conclusions

3
Introduction

• Multiple sequence alignment (MSA)
• NP-hard problem
• The heuristic methods for MSA
• Progressive method
• ClustalW, T-Coffee, POA, and etc.
• Iterative method
• Muscle, DIALIGN, and etc.
• Probabilistic method
• Probcons, Hmmt, Muscle, and etc.
• Anchor-based method
• MAFFT, Align-m , and etc.

4
Introduction (cont)

• Pairwise alignment
• Use Dynamic programming to find the optimal
  alignment. Needleman, J. Mol. Biol 1970 Smith,
  J. Mol. Biol 1981
• Three-sequence alignment
• More accurate than pairwise alignment. Murata,
  PNAS 1985
• Introduce linear gap penalty. Gotoh, J. Theor.
  Biol 1986
• Space has been reduced from O(N3) to O(N2) with
  affine gap penalty. Huang, ACM 1994
• Useful for MSA. Makoto, Bioinformatics 1993 CY
  Lin, CMCT 2006, ICPP 2007

5
Introduction (cont)

• Progressive multiple sequence alignment
  (Progressive pairwise MSA)
• To align pair sequences following the branching
  order of the guide tree until all sequences are
  aligned.
• The resulting alignment is affected by Initial
  branching order.
• Problems of Gap
• Gap will not be removed.
• Insertion gap may be calculated multiple times.
  Loytynoja, PNAS 2005

6
Introduction (cont)

• Progressive triple MSA - aln3nn
• Published on Matthias, BMC Bioinformatics July,
  2007.
• Any alignment step is three-sequence alignment.
• The three-sequence alignment uses the affine gap
  penalty same as Huang, ACM 1994.
• Use Huangs three-sequence alignment algorithm.

7
Motivation

• CrossWA - combine three-sequence and pairwise
  alignments
• Minimize the problem of Progressive pairwise MSA
• Use three-sequence alignment to reduce the
  affection of initial branching order.
• Increase the accuracy of alignment
• Three-sequence alignment may obtain more accurate
  alignments.
• Keep pairwise alignment because three-sequence
  alignment is not always better than pairwise
  alignment.
• For pairwise, using position-specific gap penalty
  is more accurate than affine gap penalty.
  Thompson, Bioinformatics 1995
• Introduce position-specific gap penalty into
  three-sequence alignment which is different to
  the algorithm aln3nn.
• Avoid increasing the computing time

8
Motivation (cont)

• Comparison of three protein sequences among
  different methods

9
Motivation (cont)

• Three-sequence alignment VS Progressive pairwise
  MSA with three sequences (430 test sets, random
  selected from BAliBase 2.0 Ref1 -5)
• Three-sequence alignment with position-specific
  gap penalty and sequence weighting

10
Motivation (cont)

• Progressive pairwise MAS (ClustalW) VS
  Progressive Triple MSA (aln3nn) reference set
  1, BAliBase 2.0 Matthias, BMC Bioinformatics
  2007, 7

11
General Process of Progressive Multiple sequence
alignment




. .
. . .
Step 2. Constructing guide tree
Unaligned sequences
Step 1. Calculating distance matrix
Aligning pair sequence or group along the
branching order
. .
Aligned sequences
Step 3. Alignment
12
Algorithm

• Process of CrossWA
• Step 1. construct distance matrix.
• Step 2. build guide tree Neighbour-Joining.
• Sequence weights will be calculated.
• Step 3. build a new guide tree modified from the
  guide tree.
• Branches will be changed for three-sequence and
  pairwise alignments.
• Sequence weights will be recalculated.
• Step 4. Alignment.
• Pairwise alignment
• Three-sequence alignment
• Compare with the alignment produced by
  progressive pairwise alignment with same three
  sequences and select better one.

13
Algorithm (cont)




. .
. . .
Unaligned sequences
Step 1. Calculating distance matrix
Step 2. Constructing guide tree
Aligning pair or three sequences (or groups)
along the branching order of new tree
. .
. . .
Aligned sequences
VS
Step 3. Constructing new tree modified from the
guide tree in step 2
Progressive Pairwise MSA
Three-sequence alignment
Step 4. Alignment
14
Algorithm (cont)

• The branch changing rule

Type I
Type II
Type III
15
Algorithm (cont)

• The evaluation of three-sequence alignment

• If SP(S) gt SP(T) then keep S
• IF SP(T) gt SP(S) then keep T

A
B
C
A
B
C
S Align(B, C)
S Align(A, S)
T Align(A, B, C)
16
Algorithm (cont)

• Modification of sequence weights
• The calculation of sequence weight is same as
  ClustalW.

D
D
B
A
C
A
C
Weight of Hba_Human 0.055 0.219/2 0.061/4
0.015/5 0.062/6 0.194
Length between node A and node C 0.219 0.061
0.280 Weight of Hba_Human 0.055 0.280/2
0.077/5 0.210

• The strategy of Gap penalty
• Introduce position-specific gap penalty into
  three-sequence alignment (modified from ClustalW).

17
Experiments

• System environment
• Linux (AMD opteron 250 2.4G with 512MB of memory)
• Data source
• BAliBASE 2.0
• Reference sets (1 5). T-Coffee, Muscle,
  Probcons, aln3nn, and etc
• Reference sets (6 8) contain repeats,
  inversions and transmembrane helices, for which
  none of the tested algorithms is designed.
  Muscle

18
Experiments (cont)

• Scoring functions
• Sum-of-pair (SP)
• Total Column Score (TC)
• Proportion probability ()
• No. of best alignment of the method/No. of total
  test sets
• Comparing algorithms
• CrossWAfast, CrossWAfull, ClustalW 1.83, T-Coffee
  5.05, Muscle 3.6.
• CrossWAfast only use the type I in the branch
  changing rule.
• CrossWAfull use all types in the branch
  changing rule.

19
Experiments (cont)

• The comparison of SP scores among different
  alignment methods

Ref1 (81) Ref2 (19) Ref3 (12) Ref4 (7) Ref5 (12)
CrossWAfast 0.774 22 0.872 5 0.669 8 0.657 0 0.741 0
CrossWAfull 0.777 30 0.877 15 0.685 32 0.658 0 0.762 17
ClustalW 0.773 11 0.876 16 0.656 0 0.674 0 0.762 0
T-Coffee 0.787 34 0.884 37 0.692 8 0.718 57 0.825 50
Muscle 0.776 28 0.891 37 0.713 67 0.728 43 0.822 33
20
Experiment (cont)

• The comparison of TC scores among different
  alignment methods

Ref1 (81) Ref2 (19) Ref3 (12) Ref4 (7) Ref5 (11)
CrossWAfast 0.665 25 0.498 16 0.368 8 0.333 29 0.529 18
CrossWAfull 0.671 33 0.515 32 0.390 42 0.301 29 0.546 18
ClustalW 0.673 33 0.489 42 0.358 17 0.320 14 0.543 0
T-Coffee 0.676 30 0.434 21 0.323 17 0.409 29 0.625 45
Muscle 0.679 22 0.475 32 0.408 42 0.396 29 0.658 36
21
Experiments (cont)

• The SP scores for each method of variant average
  identities in Reference 1 data set

lt 25 20 - 40 gt 35
CrossWAfast 0.491 35 0.825 19 0.941 18
CrossWAfull 0.500 40 0.827 31 0.942 25
ClustalW 0.493 10 0.817 12 0.938 11
T-Coffee 0.477 32 0.840 35 0.954 57
Muscle 0.487 25 0.824 27 0.948 18
22
Experiments (cont)

• The TC scores for each method of variant average
  identities in Reference 1 data set

lt 25 20 - 40 gt 35
CrossWAfast 0.208 28 0.714 20 0.891 21
CrossWAfull 0.303 33 0.712 24 0.893 46
ClustalW 0.324 38 0.713 12 0.887 29
T-Coffee 0.274 33 0.744 36 0.905 33
Muscle 0.299 19 0.725 36 0.907 46
23
Experiments (cont)

• The performance of CrossWA with 20 sequences

24
Experiments (cont)

• The Performance of CrossWA with 40 sequences

25
Experiments (cont)

• Comparison of performance among different methods
  with 20 sequences

26
Experiments (cont)

• Comparison of performance among different methods
  with 40 sequences

27
Conclusions

• Three-sequence alignment can obtain better
  resulting alignment than pairwise alignment, but
  not for all data sets.
• Combining three-sequence alignment and pairwise
  alignment can keep better alignment at any
  alignment step in progressive MSA.
• From the experimental results, CrossWA can be
  another useful tool to align multiple sequence.
• CrossWA can be used to align DNA sequences.
• For aligning Genome data, computing time is a
  problem. It can be solved by parallel
  programming. CY Lin, ICPP 2007

28
Web service
Http//140.114.91.10/Genome
29
Reference

• Needleman SB, Wunsch CD A general method
  applicable to the search for similarities in the
  amino acid sequence of two proteins. J Mol Biol
  1970, 48443-453.27. Needleman, J Mol Biol 1970
• Smith TF, Waterman MS Identification of common
  molecular subsequences. J. Mol. Biol. 1981,
  147195-197. Smith, J Mol Biol 1981
• Murata M, Richardson JS, Sussman JL Simultaneous
  comparison of three protein sequences. Proc Natl
  Acad Sci U S A. 1985, 823073-3077. Murata,
  PNAS 1985
• Gotoh O Alignment of three biological sequences
  with an efficient traceback procedure, J Theor
  Biol 1986, 327-337. Gotoh, J Theor Biol 1986
• Huang X Alignment of three sequences in
  quadratic space. Applied Computing Review 1993,
  17-11. Huang, ACM 1993
• Makoto H, Maski H, Masato I, Tomoyuki T MASCOT
  multiple alignment system for protein sequences
  based on three-way dynamic programming, J Mol
  Biol 1993, 2161-167. Makoto, Bioinformatics
  1993

30
Reference (cont)

• CY Lin, CT Huang, YC Chung, Chuan YT Parallel
  Three-sequence Alignment with Space-efficient,
  Proceedings of the 23th Workshop on
  Combinatorial Mathematics and Computation Theory,
  Chang-Hua, Taiwan, April 2006, 160-165. CY Lin,
  CMCT 2006
• CY Lin, CT Huang, YC Chung, Chuan YT Efficient
  Parallel Algorithm for Optimal Three-Sequences
  Alignment. International Conference on Parallel
  Processing 2007. CY Lin, ICPP 2007
• Loytynoja A, Goldman N An algorithm for
  progressive multiple alignment of sequences with
  insertions. Proc Natl Acad Sci U S A. 2005,
  102(30)10557-10562. Loytynoja, PNAS 2005
• Matthias K, Peter FS Progressive multiple
  sequence alignments from triplets. BMC
  Bioinformatics 2007. matthias, BMC
  Bioinformatics July, 2007
• Thompson JD Introducing variable gap penalties
  to sequence alignment in linear space.
  Bioinformatics 1995, 11181-186. Thompson,
  Bioinformatics 1995

31

• Thank you for your attention

Che-Lun Hung allen_at_sslab.cs.nthu.edu.tw Chun-Yuan Lin cylin_at_sslab.cs.nthu.edu.tw
Yeh-Ching Chung ychung_at_cs.nthu.edu.tw Chuan Yi Tang cytang_at_cs.nthu.edu.tw
32

• Position-specific gap penalty V.S. Affine gap
  penalty in three-sequence alignment (300 test
  sets)
• Three Align use Position-specific gap penalty
• Three Align use Affine gap penalty

Ave Ave
SP TC
Three Align 0.81 0.75
Three Align 0.75 0.70
33

• Algorithm of three-sequence alignment with
  position-specific gap penalty