Lee, Juyong

1
About BoostThreader

• Lee, Juyong
• 2009. 08. 26

2
What is BoostThreader?

• A Sequence-Structure threading program
• Published by J. Xus group
• Known to be good for hard cases
• Does not work for me

3
Lets thread!

• ???
• sequence
• protein structure
• scoring function
• algorithm

Deletion
Match
F
C
D
E
B
G
A
BAD
Good
4
Three algorithms for Alignment!
Im your father
Im Andrei Andreyevich Markov.

• Generative model
• Traditional
• Hidden Markov Chain
• Not that old
• Conditional Random Field
• Up to date
• Dynamic programming

5
Dynamic programming

• Finding the best scoring path on the alignment
  matrix

Final
Initial
The alignment!
6
Dynamic programing

• Finding the best scoring path on the alignment
  matrix

Initial
Final
The path
The alignment!
7
More about Dynamic Programming
Follow the maximum scoring path!
SEQUENCE
i i1
j j1
deletion
A ?
g Gap penalty -1
g
match
f
insertion
STRUCTURE
h
A a
F(i1, j1)
? a
h Gap penalty -1
8
??? ???? ???

• More specific!
• ??? ?? ? ?? ???? ???
• What is ???
• ??? ??? ??? ?? alignment OTL
• Squence-Structure alignment? ?? Dali server or
  TMalign ??

9
In Conventional seq.-str. alignment

• Linear sum of similarities of properties
• Functions for Match and Gap cases are only
  needed!
• Fmatch w1predicted SS real SS
  w2predicted SA real SA
  w3predicted residue depth real depth
• Fgap Opening penalty of gaps Extension
  penalty
• Only consider next step!

10
Whats different in BoostThreader?

• Dependent on the current and next step both!
• Nine scoring functions are necessary!
• Gap penalty is context-dependent
• Trained from reference alignments!
• DALI, TMalign etc
• Regression Trees are used as scoring function
• Not Linear function!

11
Regression Tree? ? ????
12
???????
Hey nature, Not all flies are not Drosophilia
13
Regression Tree!
100?? ???
Training!
1500cc? ????
???
?
5?? ?????
???
?
20?km?? ?????
???
?
?? 8???
?? 5???
?? 15???
?? 11???
14
Example in Threading
Sequence predicted properties Structure
observed properties
SS ? ????
???
?
SA ??? ????
???
?
SA ??? ????
???
?
?? 0.1 10? ?? 1?
?? 0.3 10? ?? 3?
?? 0.6 10? ?? 6?
?? 0.9 10 ??? 9?
Estimate Prob. from examples
15
Advantage of Tree

• Fast
• Interaction between variables can be easily
  considered

16
Whats really happening in BoostThreader?

• Initial Setting
• Set all F0 (u?v,seq(i),str(j)) 0
• P exp(F)
• 30 ?? ?? Sequence-Structure alignment!
• Calculate Prob. of all possible state transition!
• Probabilities of all examples!
• Forward-backward algorithm

17
All Possible Transitions?
For M?M
ABDE a b c d mmimd
AB ab
AB bc
AB cd
Generate examples!
BD ab
BD bc
BD cd
DE ab
DE bc
DE cd
18
Examples(2)
For M?I
ABDE a b c d mmimd
B- ab
B- bc
B- cd
Generate examples!
A- ab
A- bc
A- cd
D- ab
D- bc
D- cd
E- ab
E- bc
E- cd
19
Inside BoostThreader

• Examples and their probabilities
• Calculated with the current scoring functions
• Modify Scoring Functions
• ???? F? ??! F1F0 (1 P )
• ???? F? ??! F1F0 - P
• Add trees until prediction quality doesnt
  increase
• FF0F1F2F3F4F5

20
Performance
21
Summary

• BoostThreader considers Current and Next step
• Scoring function consists of Regression Trees
• Trees are trained based on Examples

22
?????!