Introduction to

Dynamic Programming
The sequence alignment problem

Wilson Leung 08/2015

 Outline
Overview of the sequence alignment problem
Calculate the optimal global alignment
Characteristics of dynamic programming algorithms
Calculate the optimal local alignment
 Learning objectives
Understand the theory behind sequence alignment
Become a better informed user of NCBI BLAST

This presentation will not cover:

The BLAST algorithm
Parameter optimizations
Statistics for similarity searches (Karlin-Altschul theory)

Korf, I., Yandell, M., and Bedell, J. (2003). BLAST. O’Reilly Media, Inc.
 Design goals

Generate an alignment between two sequences

Identify the “best” (most parsimonious) alignment
Generate the best alignment “quickly”
 Strategy #1: Visual inspection

Query: ATTACCAG
|| |||||
Subject: ATCACCAG
Sequences must have high percent identity
Applications:
PAM scoring matrix (align sequences with >= 85% identity)
Align mononucleotide runs during sequence improvement
Strategy #2: Enumerate all alignments

Guaranteed to find the best alignment

Does not scale
Combinatorial explosion
Two 300 bp sequences have ~10179 possible alignments (Eddy
2004)

Brute-force algorithm
Establish baseline performance and test cases
Identify patterns in the problem space
Apply the brute force algorithm to a single column of
the alignment

Homologous Not homologous

Query: A A- -A
Subject: A -A A-
Three possible alignments for two 1 bp sequences
Query length (M) = 1; Subject length (N) =1

Only two biological interpretations:

A in the query is homologous to A in the subject
A in the query is not homologous to A in the subject
 Six possible relationships between the
query and subject for M=2, N=2
2 aligned bases 1 aligned base 0 aligned bases

Query: AT A-T -AT AT-- --AT

Subject: AT -AT A-T --AT AT--
AT- A-T A-T- -A-T
Each color denotes a
different evolutionary
A-T AT- -A-T A-T-
relationship
AT- -AT A--T -AT-
-AT AT- -AT- A--T
Observations from the brute force
alignment strategy
Many of the possible alignments are redundant
Imply the same evolutionary relationship

Large number of possible alignments

13 possible alignments for sequences of length 2

Can ignore many possible alignments

Many are suboptimal compared to the best alignment
 Strategy #3: Dot plot
Deletion in subject Cell position (i,j):
i = Query position (x-axis)
j = Subject position (y-axis)

Draw a dot at (i,j) if the

Subject (y)

Align
two bases are identical

Connect the dots to make a

line (alignment)
Level of noise depends on
Insertion in subject repeat density
Use longer words and higher
cutoff scores to reduce noise

Query (x)
 Assessment of the three sequence
alignment strategies
Infeasible to examine all possible alignments
Need to reduce the search space

Only a small subset of alignments are “interesting”

Many alignments are redundant

Connect the dots in the dot plot to create an alignment

Consider the cumulative levels of similarity
 The optimal alignment is composed of
smaller optimal alignments
Query: AT Subject: AT

Query: A T A - T A T -
Subject: A T - A T A - T

Only the best alignment at each position A - T -

could be part of the final optimal alignment
- A - T

Align Deletion in subject Insertion in subject

 Partition the alignment problem into
smaller subproblems
1 100
1
Subject (y)

Subject
100 Query
Query (x)
Assume the query and subject sequences are the same
 Three different ways to reach cell (i,j)
in the alignment matrix
A
Subject (y)

(i-1,j-1) (i,j-1) Align with subject A

(i-1, j-1) A
Gap in subject A
(i-1, j) -
A Gap in query -
(i-1,j) (i,j) (i, j-1) A
Query (x) Arrow = alignment
 Construct a scoring system to measure
similarity between two sequences
Scoring system for the aligned state: 𝛔
𝛔(a, b) = Score for aligning a in query with b in subject
𝛔(A, A) = Bonus for aligning A in query with A in subject
𝛔(A, T) = Penalty for aligning A in query with T in subject

Penalty for adding a gap: 𝛾

More sophisticated scoring systems take transitions,

transversions, affine gap penalty into account
Pearson WR. Selecting the Right Similarity-Scoring Matrix. Curr Protoc
Bioinformatics. 2013;43:3.5.1-3.5.9.
 Recursive definition for the optimal cumulative
alignment score S(i,j)
a
Subject (y)

(i-1,j-1) (i,j-1)
S(i,j) = max {
𝛾 S(i-1,j-1) + 𝛔(a,b)
𝛔(a,b)
S(i-1,j ) + 𝛾
𝛾
S(i ,j-1) + 𝛾
b }
(i-1,j) (i,j)
Query (x)
Align Gap in subject Gap in query
Determine the best way to reach cell (i,j) if
it were part of the optimal alignment
(i,j)
Query
?
Subject
Optimal alignment
S(i,j) = max { a
Align
b
a
Gap in subject

Gap in query
b
}
Use the maximum score at each cell to eliminate entire
branch of suboptimal alignments

(i,j)

Gap in query Align Gap in subject

Cumulative score S(i,j) encapsulates the
alignment decisions up to position (i,j)
All potential optimal alignments that go through cell
(i,j) have the same ancestry
Re-use the cumulative alignment score (memoization)

Gaps are described by the cumulative score

Do not affect the coordinates of the alignment matrix

Do not know the optimal alignment until we complete

the entire alignment matrix
Optimal alignment has the highest cumulative score
Needleman-Wunsch algorithm (global alignment)
(Query length: M; Subject length: N)
Construct a (M+1) x (N+1) matrix
Extra column and row = gaps at the beginning of the alignment

Fill in the cells in the first row and first column with the
cumulative gap costs
Calculate the maximum score for subsequent cells (i,j)
Keep track of the decision that leads to the maximum score (S)

S(i-1,j-1) + 𝛔(a,b)
S(i,j) = max S(i-1,j ) + 𝛾
S(i ,j-1) + 𝛾
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 Mar;48(3):443-53.
 Initialize the alignment matrix
(Match = +5; Mismatch = -2; Gap = -6)

0 1 2 3 4 5 6 7 8
T G C T C G T A
0 0 -6 -12 -18 -24 -30 -36 -42 -48
1 T -6
2 T -12

Subject
3 C -18
4 A -24
5 T -30
6 A -36
Query (Eddy, 2004)
 Calculate the possible scores for the cell
at position (1,1)
T
Subject (y)

(0,0) (1,0)

0 -6 S(1,1) = max {
𝛔(T,T) 𝛾 S(0,0) + 𝛔(T,T)
S(0,1) + 𝛾
-6 S(1,0) + 𝛾
T 𝛾 }
(0,1) (1,1)
Query (x)
Align Gap in subject Gap in query
 Calculate the optimal score for the cell
at position (1,1)
T
Subject (y)

S(1,1) = max {
0 -6 0 + (+5) = 5
-6
+5 -6 + (-6) = -12
-6 + (-6) = -12
5 -12
-6 }
T -6
-12 5 S(1,1) = 5
Query (x)
(Match = +5; Mismatch = -2; Gap = -6)
 Calculate the possible scores for the cell
at position (2,1)
T G
Subject (y)

(1,0) (2,0)

-6 -12 S(2,1) = max {

𝛔(T,G) 𝛾 S(1,0) + 𝛔(T,G)
S(1,1) + 𝛾
5 S(2,0) + 𝛾
T 𝛾 }
(1,1) (2,1)
Query (x)
Align Gap in subject Gap in query
 Calculate the optimal score for the cell
at position (2,1)
T G
Subject (y)

S(2,1) = max {
-6 -12 -6 + (-2) = -8
-6
-2 5 + (-6) = -1
-12 + (-6) = -18
-8 -18
5 }
T -6
-1 -1 S(2,1) = -1
Query (x)
(Match = +5; Mismatch = -2; Gap = -6)
Align Alignment matrix after two iterations
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subject
Gap in 0 1 2 3 4 5 6 7 8
query T G C T C G T A
0 0 -6 -12 -18 -24 -30 -36 -42 -48
1 T -6 5 -1
2 T -12

Subject
3 C -18
4 A -24
5 T -30
6 A -36
Query
 Calculate the optimal score for the cell
at position (3,1)
G C
Subject (y)

S(3,1) = max {
-12 -18 -12 + (-2) = -14
-6
-2 -1 + (-6) = -7
-18 + (-6) = -24
-14 -24
-1 }
T -6
-7 -7 S(3,1) = -7
Query (x)
(Match = +5; Mismatch = -2; Gap = -6)
Align Matrix after three iterations
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subject
Gap in 0 1 2 3 4 5 6 7 8
query T G C T C G T A
0 0 -6 -12 -18 -24 -30 -36 -42 -48
1 T -6 5 -1 -7
2 T -12

Subject
3 C -18
4 A -24
5 T -30
6 A -36
Query
 Calculate the optimal score for the cell
at position (1,2)
T
S(1,2) = max {
T -6 5 -6 + (+5) = -1
Subject (y)

-6
+5 -12 + (-6) = -18
5 + (-6) = -1
-1 -1 }
T
-12
-6
-18 -1 S(1,2) = -1
Query (x)
(Match = +5; Mismatch = -2; Gap = -6)
Align Complete alignment matrix
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subject
Gap in 0 1 2 3 4 5 6 7 8
query T G C T C G T A
0 0 -6 -12 -18 -24 -30 -36 -42 -48
1 T -6 5 -1 -7 -13 -19 -25 -31 -37
2 T -12 -1 3 -3 -2 -8 -14 -20 -26

Subject
3 C -18 -7 -3 8 2 3 -3 -9 -15
4 A -24 -13 -9 2 6 0 1 -5 -4
5 T -30 -19 -15 -4 7 4 -2 6 0
6 A -36 -25 -21 -10 1 5 2 0 11
Query
 Use traceback to recover the
optimal alignment
Start from the cell within the last row and last column that
has the highest score
Recall the step (color) that leads to this optimal score
Report this step in the alignment output
All the alignment decisions have already been made

Repeat until we reached the beginning of the sequence

Two options if multiple paths produce the same score

Report only one of the paths (pick arbitrarily)
Report all paths with the optimal score
 Query: T C G T A
Subject: T C A T A
Traceback:
Query T G C T C G T A
0 -6 -12 -18 -24 -30 -36 -42 -48
T -6 5 -1 -7 -13 -19 -25 -31 -37
T -12 -1 3 -3 -2 -8 -14 -20 -26
Subject

C -18 -7 -3 8 2 3 -3 -9 -15
A -24 -13 -9 2 6 0 1 -5 -4
T -30 -19 -15 -4 7 4 -2 6 0
A -36 -25 -21 -10 1 5 2 0 11
 Calculate the optimal score for the cell at
position (5,3)
T C
Subject (y)

S(5,3) = max {
-2 -8 -2 + (+5) = 3
-6
+5 2 + (-6) = -4
-8 + (-6) = -14
3 -14
2 }
C -6
-4 3 S(5,3) = 3
Query (x)
(Match = +5; Mismatch = -2; Gap = -6)
Traceback must follow the steps that produce the
optimal cumulative global alignment score

T C

T -2 -8

Subject (y)
C 2 3

Query (x)
 Query: T G C T C G T A
Subject: T - - T C A T A
Traceback:
Query T G C T C G T A
0 -6 -12 -18 -24 -30 -36 -42 -48
T -6 5 -1 -7 -13 -19 -25 -31 -37
T -12 -1 3 -3 -2 -8 -14 -20 -26
Subject

C -18 -7 -3 8 2 3 -3 -9 -15
A -24 -13 -9 2 6 0 1 -5 -4
T -30 -19 -15 -4 7 4 -2 6 0
A -36 -25 -21 -10 1 5 2 0 11
 The Needleman-Wunsch algorithm is an example
of a dynamic programming algorithm

Problem must satisfy two criteria:

Optimal substructure
Optimal solution to the complete problem is composed of optimal
solutions to the subproblems
Overlapping problems
Re-use the results for the subproblems (e.g., lookup table)

Many bioinformatics problems satisfy these criteria

Sequence alignment, gene prediction, RNA-folding

Bellman B. The theory of dynamic programming. Bulletin of the American Mathematical

Society. 1954; 60(6):503–516
 Smith-Waterman algorithm (local alignment)
(Query length: M; Subject length: N)

Three changes to the Needleman-Wunsch algorithm:

The minimum score for a cell is zero
Initiate a new alignment when the cumulative score is negative
Begin traceback from the cell within the entire matrix that has
the highest score
Terminate traceback when the score is zero

S(i-1,j-1) + 𝛔(a,b)
S(i-1,j ) + 𝛾
S(i,j) = max
S(i ,j-1) + 𝛾
0
Smith TF, Waterman MS. Identification of common molecular subsequences.
J Mol Biol. 1981 Mar 25;147(1):195-7.
 Global versus local alignments
Global alignment
Optimal alignment along the entire length of two sequences
Compare protein sequences to identify orthologs

Local alignment
Optimal alignment between parts of two sequences
Identify conserved domains within protein sequences

Glocal (semi-global) alignment

Optimal global alignment for one sequence; optimal local alignment
for the other sequence
Map a coding exon against a genomic sequence
 Initialize the local alignment matrix
(Match = +5; Mismatch = -2; Gap = -6)

0 1 2 3 4 5 6 7 8
T G C T C G T A
0 0 0 0 0 0 0 0 0 0
1 T 0
2 T 0

Subject
3 C 0
4 A 0
5 T 0
6 A 0
Query
 Calculate the possible local alignment
scores for the cell at position (1,1)
T
Subject (y)

(0,0) (1,0)
S(1,1) = max {
0 0 S(0,0) + 𝛔(T,T)
𝛔(T,T) 𝛾
S(0,1) + 𝛾
S(1,0) + 𝛾
0 0
T 𝛾 0
(0,1) (1,1) }
Query (x)
Align Gap in subject Gap in query
 Calculate the optimal local alignment
score for the cell at position (1,1)
T
Subject (y)

S(1,1) = max {
0 0 0 + (+5) = 5
-6 0 + (-6) = -6
+5
0 + (-6) = -6
5 -6 0
0 }
T -6
0
-6 5 S(1,1) = 5
Query (x)
(Match = +5; Mismatch = -2; Gap = -6)
Align Local alignment matrix
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subject
Gap in 0 1 2 3 4 5 6 7 8
query T G C T C G T A
0 0 0 0 0 0 0 0 0 0
1 T 0 5 0 0 5 0 0 5 0
2 T 0 5 3 0 5 3 0 5 3

Subject
3 C 0 0 3 8 2 10 4 0 3
4 A 0 0 0 2 6 4 8 2 5
5 T 0 5 0 0 7 4 2 13 7
6 A 0 0 3 0 1 5 2 7 18
Query
 Query: T C G T A
Subject: 0 T C A T A
Traceback:
Query T G C T C G T A
0 0 0 0 0 0 0 0 0
T 0 5 0 0 5 0 0 5 0
T 0 5 3 0 5 3 0 5 3
Subject

C 0 0 3 8 2 10 4 0 3
A 0 0 0 2 6 4 8 2 5
T 0 5 0 0 7 4 2 13 7
A 0 0 3 0 1 5 2 7 18
Techniques to improve the performance of
sequence alignment
Time and space complexity: O(MN)
Double the size of the two sequences leads to a four-fold
increase in the amount of time and space required

Reduce memory requirement

Myers EW, Miller W. Optimal alignments in linear space. Comput Appl Biosci.
1988 Mar;4(1):11-7.

Fill the matrix in parallel (SIMD, CUDA)

Farrar M. Striped Smith-Waterman speeds database searches six times over other
SIMD implementations. Bioinformatics. 2007 Jan 15;23(2):156-61.

Find high-scoring instead of the best alignment

Altschul SF, Gish W, Miller W, Myers EW, Lipman DJ. Basic local alignment
search tool. J Mol Biol. 1990 Oct 5;215(3):403-10.
 Questions?

Eddy SR. What is dynamic programming? Nat Biotechnol. 2004 Jul;22(7):909-10.

Rationale for calculating the scores for
the entire alignment matrix
Cannot determine the best global alignment without
aligning the entire query and subject sequences
Cannot evaluate all possible alignments

If the alignment before we reached cell (i,j) is part of

the optimal alignment:
Identify the next step (i.e. align, gap in query, gap in subject)
that will be part of the optimal alignment

Use traceback to determine the final alignment

Different alignments could produce the same score
 Overview of the BLAST algorithm
Heuristic algorithm to find local regions of similarity
between the query and subject sequences

Consists of four main stages:

Find common subsequences (words)
Extend the word matches into longer alignments
Evaluate the significance of the high-scoring segment pairs
(HSPs)
Combine multiple HSPs into a longer alignment

Korf, I., Yandell, M. and Bedell, J. (2003). The BLAST Algorithm. In BLAST (76-87).
Sebastopol, CA: O’Reilly Media, Inc.
Number of alignments for two sequences with length N

Stirling’s approximation
Number of alignments for two sequences with length N
Number of alignments for two sequences with length N
 Brute force alignment approach is
computationally intractable
Sequence # possible
length (N) alignments
10 1.87E+05
50 1.01E+29
100 9.07E+58
200 1.03E+119
300 1.35E+179
400 1.88E+239
500 2.70E+299