Introduction to

Dynamic
Programming
The sequence alignment problem

November 06, 2023

 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: ATTACCA
||
G
Subject ATCACCA
|||||
:
G
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

Homologou Not homologous

s
Query: 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- - AT-- --AT

Subject AT T
- AT
A- -- AT-
:
AT T AT -
AT A- A-T- -A-T
Each color
denotes a
-
A- T
AT -A- A-
different T - T T-
evolutionary AT - A--T -AT-
relationship
- AT - A--
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 Cell position (i,j):
i = Query position (x-axis)
subject
j = Subject position (y-
axis)
Subject (y)

Align Draw a dot at (i,j)

if the two bases are
identical

Connect the dots to

make a line
Insertion in (alignment)
subject
Level of noise depends
on repeat density
Query (x) Use longer words and
 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 Insertion in

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

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

𝛔(i-1,j
S(i-1,j-1) +

𝛾 ,j-1)
S(i,j) = max S (a,b) )

+ 𝛾
+
S(i
Needleman SB, Wunsch CD. A general method applicable to the search
for similarities in the amino acid sequence of two proteins. J Mol Biol.
 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,
 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
𝛔
S(0,0) +
(T,T) + 𝛾
S(0,1)
S(1,0) + 𝛾
𝛾
T -6
}
(0,1) (1,1)
Query (x)
Align Gap in Gap in
 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) = -
5 -12
-6 } 12
T -6
- 5 S(1,1) = 5
Query 12
(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 - S(2,1) = max {
𝛔(T,G) 12𝛾
𝛔
S(1,0) +
(T,G) + 𝛾
S(1,1)
S(2,0) + 𝛾
𝛾
T 5
}
(1,1) (2,1)
Query (x)
Align Gap in Gap in
 Calculate the optimal score for
the cell at position (2,1)
T G
Subject (y)

S(2,1) = max {
-6 - -6 + (-2) = -8
12 -6
-2 5 + (-6) = -1
-12 + (-6) = -
-8 -18
5 }18
T -6
-1 -1 S(2,1) = -
Query (x) 1
(Match = +5; Mismatch = -2; Gap = -6)
Align Alignment matrix after two iterations
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subjec
tGap 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 + (-2) = -
12-2 18 -6 14
-1 + (-6) = -7
-18 + (-6) = -
- -24
-1 }24
T -6 14
-7 -7 S(3,1) = -
Query (x) 7
(Match = +5; Mismatch = -2; Gap = -6)
Align Matrix after three iterations
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subjec
tGap 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 {
-6 5 -6 + (+5) = -
Subject (y)

T
+5
-6 1
-12 + (-6) = -
185 + (-6) = -
-1 -1 }1
T
-
-6
12 - -1 S(1,2) = -
Query 18
(x) 1
(Match = +5; Mismatch = -2; Gap = -6)
Align Complete alignment matrix
Gap in (Match = +5; Mismatch = -2; Gap = -6)
subjec
tGap 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) = -
3 -14
2 }14
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

𝛔(i-1,j
S(i-1,j-1) +

𝛾 ,j-1)
S (a,b) )
S(i,j) = max

0 𝛾
+
S(i
+
Smith TF, Waterman MS. Identification of common molecular
subsequences.
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
𝛔(T,T) 𝛾 𝛔
S(0,0) +
(T,T) + 𝛾
S(0,1)
S(1,0) + 𝛾
𝛾
0 0
T 0
(0,1) (1,1) }
Query (x)
Align Gap in Gap in
 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)
subjec
tGap 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

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
Sequenc # possible
e length alignments
(N)
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