Affine Gap Penalty Function

BCB 597
Fall, 2007
Multiple Base Insertion/Deletion
Mutation Event
 Constant Gap Penalty
• Each gap character is penalized a constant
amount: γ
• Not a good gap penalty for our mutation event
model
• What if we had a more general gap penalty
function?
 General Gap Penalty Function
• Gap penalty function: W (g )
• Each entry in the dynamic programming table
now takes O(n) time to fill.
S[0,0]  0 • O(n3) running time
S[i,0]  W (i ) • Cannot directly apply space saving
S[0, j ]  W ( j ) • O(nm) space

S [i  1, j  1]   (i  1, j  1)
 i
S[i, j ]  max  max S [i  k , j ]  W (k ) 
k 1
 j
 max S [i, j  k ]  W (k ) 
 k 1
 Affine Gap Penalty
• W(i)=g+hi (for i >= 1)
• g: gap opening penalty
• h: gap extension penalty
• The ratio between g and h affects how much
we view the size of the gap.
– Small g, Large h: size more important
– Large g, Small h: size less important
 Remember our “What if?”
• E(AA, BA) = E(AA[0,i], BA [0,i]) + E(AA[i+1,L], BA[i+1,L])
 Alignment Decomposition
No Longer Behaves
optimal
AGCTAGAGACCAG---TCT-GAGGTAGA
AGCTAGAGACCAGCTATCTAGAGGTAGA

optimal AGCTAGAGACCAG- --TCT-GAGGTAGA not optimal

optimal AGCTAGAGACCAG- TCT---GAGGTAGA optimal

  1,   1, g  3, h  1
However, this was a sufficient condition, not a necessary condition!
 Amortize the cost of opening
AGCTAAGGAA
• Consider filling out the dynamic AGCTAAGGAC
programming table >
• The cost of g is eventually spread AGCTAAGGA-
out over the length of the gap AGCTAAGGAC
once the gap is closed.
• While gap opening might initially AGCTAAGGAAA
cause the alignment score to be AGCTAAGGACT
less than a mismatch. >
• However, the one time loss might AGCTAAGGA--
eventually become “worth it.” AGCTAAGGACT
• We cannot predict ahead to
know how much the gap opening AGCTAAGGAAAA
will cost per gap character, or we AGCTAAGGACTT
would directly know if the gap <
opening will be worth it. AGCTAAGGA---
AGCTAAGGACTT
 Three Dynamic Programming Tables
• Instead of looking to the future, we simple store
the best possible alignments that end in gaps, in
case they might eventually become “worth it.”
• Table GA: Stores the best alignment, under the
restriction that the last character in AA be aligned
with a gap.
• Table GB: Stores the best alignment score, under
the restriction that the last character in BA be
aligned with a gap.
• Table S: Stores the best alignment score with no
restrictions.
 Three Recursions

 G A [i  1, j ]  h
G A [i, j ]  max 
G A [0, j ]   S [i  1, j ]  g  h
GB [i,0]    GB [i, j  1]  h
S [0, j ]  g  jh
GB [i, j ]  max 
S [i, j  1]  g  h
S [i,0]  g  ih
S [i  1, j  1]   (i  1, j  1)
S [0,0]  0 
S[i, j ]  max  G A [i, j ]

 GB [i, j ]
 Traceback
• The path can now travel
• Start in table S
• If S[i,j] = GA[i,j] or GB[i,j], jump to the
corresponding table, without making a
“move.”
• Paths through GA must move up
• Paths through GB must move left
 Semiglobal Alignment

 G A [i  1, j ]  h
G A [i, j ]  max 
G A [0, j ]   S [i  1, j ]  g  h
GB [i,0]    GB [i, j  1]  h
S [0, j ]  0
GB [i, j ]  max 
S [i, j  1]  g  h
S [i,0]  0
S [i  1, j  1]   (i  1, j  1)
S [0,0]  0 
S[i, j ]  max  G A [i, j ]

 GB [i, j ]
 Local Alignment
• The key to Local Alignment is that the aligning
region must still start and end with a matching
character.
• This means that a local alignment path must
start and end in table S.
• Thus finding local alignment is very similar to
local alignment for constant gap penalty.
 Local Alignment
 G A [i  1, j ]  h
G A [i, j ]  max 
G A [0, j ]   S [i  1, j ]  g  h
GB [i,0]    GB [i, j  1]  h
GB [i, j ]  max 
S [0, j ]  0 S [i, j  1]  g  h
S [i,0]  0 
S [0,0]  0 S [i  1, j  1]   (i  1, j  1)

S[i, j ]  max  G A [i, j ]
 GB [i, j ]

 0
Applying Space Saving To Affine Gap
Alignment
 Finding the Intersection

 S[i, j ]  S[i  1, j ]  g  h j j
 G [i, j ]  G [i  1, j ]  g  h j j
 A A

max  S[i, j ]  G A [i  1, j ]  h j j
 G [i, j ]  S[i  1, j ]  h
j
j j
 A

S[i, j ]  S [i  1, j  1]   (i, j ) j  j 1
Diagonal Intersection

 (i, j )
…
g+2h
g+h

g+3h
g+2h
g+h

…
g+3h
g+2h
g+h
…

g+h
g+2h
…
Vertical Intersection

gh
…
2h
h

g+3h
g+2h
g+h

…
g+3h
g+2h
g+h
…

h
2h
…