Sequence Alignment Algorithms

DEKM book
Notes from Dr. Bino John
and Dr. Takis Benos

1
 To Do
• Global alignment
• Local alignment
• Gaps
– Affine Gaps
– Algorithm (blackboard)
• Statistical Significance
– Notes (blackboard)
• Read up on database searches
– BLAST
– FASTA
– CS tricks: suffix tree, …
• PSSMs and Multiple Sequence Alignments

2
 Why compare sequences?
• Given a new sequence, infer its function based
on similarity to another sequence

• Find important molecular regions – conserved

across species

3
 Why compare sequences? Do more..
• Determine the evolutionary constraints at
work
• Find mutations in a population or family of
genes
• Find similar looking sequence in a database
• Find secondary/tertiary structure of a
sequence of interest – molecular modeling
using a template (homology modeling)

4
 Sequence alignment
• Are two sequences related?
– Align sequences or parts of them
– Decide if alignment is by chance or evolutionarily
linked?
• Issues:
– What sorts of alignments to consider?
– How to score an alignment and hence rank?
– Algorithm to find good alignments
– Evaluate the significance of the alignment

5
6
 Dynamic Programming
We apply dynamic programming when:
• There is only a polynomial number of
subproblems
– Align x1…xi to y1…yj

• Original problem is one of the subproblems

– Align x1…xM to y1…yN

• Each subproblem is easily solved from smaller

subproblems
7
 Global alignment
PAM
T j-1 j BLOSUM
S DNA matrix
i-1

i Mij

Mi-1, j-1 + Score(Si,Tj )

Mi,j = MAX Mi,j-1 + γ Gap penalty

Mi-1,j + γ
Needleman & Wunsch, 1970

8
9
10
11
12
 Alignment: adding scores (cntd)

Score(match) = 1
Score(mismatch) = 0
Score(gap) = 0

13
Alignment: adding scores

14
 Alignment: adding scores (cntd)

(Seq #1) A
|
Alignment:
(Seq #2) A

15
 Alignment: adding scores (cntd)

(Seq #1) T A
|
Alignment:
(Seq #2) - A

16
 Alignment: adding scores (cntd)

(Seq #1) G A A T T C A G T T A
| | | | | |
Alignment:
(Seq #2) G G A - T C - G - - A

6 matches, 1 mism., 4 gaps

17
18
19
 Local alignment
Given two sequences, S and T, find two
subsequences, s and t, whose alignment has the
highest “score” amongst all subsequence pairs.

Question: Why do we need local alignment, if we have

the global one?

20
 Local alignment: an example

EGR4_HUMAN KA [FACPVESCVRSFARSDELNRHLRIH] TGHKP [FQCRICLRNFSRSDHLTSHVRTH] TGEKP [FACDV--CGRRFARSDEKKRHSKVH]

EGR4_RAT KA [FACPVESCVRTFARSDELNRHLRIH] TGHKP [FQCRICLRNFSRSDHLTTHVRTH] TGEKP [FACDV--CGRRFARSDEKKRHSKVH]
EGR3_HUMAN RP [HACPAEGCDRRFSRSDELTRHLRIH] TGHKP [FQCRICMRSFSRSDHLTTHIRTH] TGEKP [FACEF--CGRKFARSDERKRHAKIH]
EGR3_RAT RP [HACPAEGCDRRFSRSDELTRHLRIH] TGHKP [FQCRICMRSFSRSDHLTTHIRTH] TGEKP [FACEF--CGRKFARSDERKRHAKIH]
EGR1_HUMAN RP [YACPVESCDRRFSRSDELTRHIRIH] TGQKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDI--CGRKFARSDERKRHTKIH]
EGR1_MOUSE RP [YACPVESCDRRFSRSDELTRHIRIH] TGQKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDI--CGRKFARSDERKRHTKIH]
EGR1_RAT RP [YACPVESCDRRFSRSDELTRHIRIH] TGQKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDI--CGRKFARSDERKRHTKIH]
EGR1_BRARE RP [YACPVETCDRRFSRSDELTRHIRIH] TGQKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACEI--CGRKFARSDERKRHTKIH]
EGR2_RAT RP [YPCPAEGCDRRFSRSDELTRHIRIH] TGHKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDY--CGRKFARSDERKRHTKIH]
EGR2_XENLA RP [YPCPAEGCDRRFSRSDELTRHIRIH] TGHKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDY--CGRKFARSDERKRHTKIH]
EGR2_MOUSE RP [YPCPAEGCDRRFSRSDELTRHIRIH] TGHKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDY--CGRKFARSDERKRHTKIH]
EGR2_HUMAN RP [YPCPAEGCDRRFSRSDELTRHIRIH] TGHKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDY--CGRKFARSDERKRHTKIH]
EGR2_BRARE RP [YPCPAEGCDRRFSRSDELTRHIRIH] TGHKP [FQCRICMRNFSRSDHLTTHIRTH] TGEKP [FACDF--CGRKFARSDERKRHTKIH]
MIG1_KLULA -- [-------------------------] ---RP [YVCPICQRGFHRLEHQTRHIRTH] TGERP [HACDFPGCSKRFSRSDELTRHRRIH]
MIG1_KLUMA -- [-------------------------] ---RP [YMCPICHRGFHRLEHQTRHIRTH] TGERP [HACDFPGCAKRFSRSDELTRHRRIH]
MIG1_YEAST -- [-------------------------] ---RP [HACPICHRAFHRLEHQTRHMRIH] TGEKP [HACDFPGCVKRFSRSDELTRHRRIH]
MIG2_YEAST -- [-------------------------] ---RP [FRCDTCHRGFHRLEHKKRHLRTH] TGEKP [HHCAFPGCGKSFSRSDELKRHMRTH]
[ ] :* [. * * * * * :* . *:* *] ***:* [. * * : *:**** .** : *]

21
 Local alignment (cntd)
T PAM
j-1 j
S DNA matrix BLOSUM
i-1

i Mij
0
Mi-1, j-1 + Score(Si,Tj )
Mi,j = MAX
Mi,j-1 + γ Gap penalty

Mi-1,j + γ
Smith & Waterman, 1981 Similarity Scoring Expected value:
negative for random alignments
positive for highly similar sequences 22
 The Smith-Waterman Algorithm
1. Initialization
F(0,0) = F(0,j) = F(i,0) = 0

2. Iteration
for i=1,…,M
for j=1,…,N
- calculate optimal F(i,j)
- store Ptr(i,j)

3. Termination
• Find the end of the best alignment with FOPT = max{i,j} F(i,j) and trace back OR
• Find all alignments with F(i,j) > threshold and trace back

23
Local vs. global alignment

24
Local vs. global alignment (cntd)

25
 Local alignment (cntd)
Characteristics of local alignments:
• The alignment can start/end at any point in the
matrix.

• No negative scores in the alignment.

• The mean value of the scoring matrix (e.g. PAM,

BLOSUM) should be negative, but there should be
positive scores in the scoring matrix.

26
 Scoring the gaps more accurately
• A naive model γ(n)
Gap penalty is linear to the gap length
Nature “prefers” to place gaps where other gaps exist

• Convex gap penalty function

γ(n)
γ(n+1) - γ(n) ≤ γ(n) - γ(n-1)

Time O(N2M) Space O(NM)

(assume N>M)

27
 Scoring gaps: affine gaps
• Affine gaps: a compromise between linear and convex gap
penalties

γ(n) = -d - e * (n-1) γ(n)

d: gap initiation penalty

e
e: gap extension penalty d

28
29
30
31
32
33
34
35
36
37
Database searches

38
 DNA and protein databases
• EMBL/GenBank/DDBJ database of nucleic acids

39
 DNA and protein databases
• EMBL/GenBank/DDBJ database of nucleic acids (cntd)

40
 DNA and protein databases
• SWISS-PROT & TrEMBL database of proteins

41
 DNA and protein databases
• SWISS-PROT & TrEMBL database of proteins

42
 Database searches
• Database searching consists of many pairwise alignments combined in
one search.
• It helps determining the function and the evolutionary relationships
• Heuristic algorithms are used instead of DP. Why?
• Size of SWISS-PROT + TrEMBL (Rel. 9.5):
3.9M entries or 1,276M residues.

• Exact algorithms are O(NM) fast.

• Heuristic methods can look at a small fraction of the searching space
that will include all (or most) of the high scoring pairs.

43
 BLAST algorithm
• Basic Local Alignment Search Tool - The method:

• For each “word” (of fixed-length) in the query sequence,

make a list of all neighbouring “words” that score above
some threshold.

• Scan the database for these words.

• Perform (ungapped) “hit extension” until score <

threshold.

• Stop at maximum scoring extension.

44
 BLAST algorithm (cntd)
• An example:

Query: CPICHRAFHRLEHQTRHMRIHTGEKPHAC

HMR 18 HMR
HHR HMR -2+13 HIR
HIR +1+13 .
HAR BLOSUM62 -1+13 selection .
… … .

45
 BLAST algorithm (cntd)
• An example:

Query: CPICHRAFHRLEHQTRHMRIHTGEKPHAC
H+R
Sbjct: CPLCDKAFHRLEHQTRHIRTHTGEKPHAC

46
 BLAST algorithm (cntd)
• An example:

Query: CPICHRAFHRLEHQTRHMRIHTGEKPHAC
CP+C +AFHRLEHQTR H+R HTGEKPHAC
Sbjct: CPLCDKAFHRLEHQTRHIRTHTGEKPHAC

47
 BLAST algorithm (cntd)
• The idea: a high scoring match alignment is very likely to contain a short
stretch of very high scoring matches.

• Word length: 3 (proteins) and 11 (DNA).

• HSSP: multiple HSSPs can be reported for each database entry.

• Gapped alignments: more recent BLAST versions perform gapped

alignments.

48
 BLAST flavours
Query: DNA Protein
BLASTX

BLASTN

BLASTP
TBLASTN
Database: DNA Protein

TBLASTX: DNA Query to DNA Database via translation

49
 FASTA algorithm
• The method:
• For each pair of sequences (query, subject), identify all
identical “word” matches of (fixed) length.
• Look for diagonals with many mutually supporting
“word” matches.
• The best diagonals are used to extend the word matches
to find the maximal scoring (ungapped) regions.
• Join ungapped regions, using gap costs.
• Align the two (sub)regions using full dynamic
programming techniques.

50
FASTA algorithm (cntd)

51
 FASTA algorithm (cntd)
• The idea: a high scoring match alignment is very likely to contain a short
stretch of identities.

• Word length: 2 (proteins) and 4-6 (DNA).

• HSSP: usually one (extended) gapped alignment is presented.

52
 FASTA flavours
Query: DNA Protein
FASTX3

FASTA3
FASTA3
TFASTA3
Database: DNA Protein

53