Doug Raiford

1
Protein Structure Searches

• Doug Raiford
• Lesson 18

2
Problem definition

• Given a protein conformation can we find other
  structurally similar proteins?
• Might have a database of structures (like the PDB)

3
If have a predicted and known

• Can do a simple RMSD to compare the two
  conformations
• Know precisely which aas compare to which

4
What about if not identical sequences?

• Must map aas from one to aas in the other
• How might you do this?
• Sequence similarity
• MSAs

5
Have we seen before?

• 3D PSSM
• Sequence alignment integrated with 3D alignment
• Stored in profile (position specific similarity
  profile)
• Gens 1D profiles first (MSAs)
• Then uses a structural alignment program (SAP) to
  augment profiles with structural similarity

6
SAP (structural alignment program)

• Aligning secondary structures

7
How?

• What do you think of when you hear that you will
  need to align two things?
• Dynamic programming

a a a T ß ß ß
a
a
ß
ß
ß
ß
8
Scoring

• Three components
• AA similarity (substitution matrix)
• Local structure
• E.g. both aas members of alpha helix
• Solvent exposure

Are the associated AAs similar, sequence wise
(i.e. both glycines)?
Are they both in a similar local structure?
Are they both buried or both exposed to solvent?
9
Benefits

• SAP (structure alignment) allows a profile to be
  influenced by secondary structure
• Useful to 3D PSSM in thatthreading decisions
  (whichaas match to a profile)
• Homology based protein conformation enhancedby
  making better decisions on where to insert
  gaps/varying length loops

10
Another already seen

• PFAM
• Have Markov Models for protein families
• Sequences that match models have high probability
  of matching conformation
• Even though not comparing structures (query to
  target)
• are matching a sequence to its most probable
  structure

Pfam
HMMR
11
What about similar structure in an alternative
way?

• Cant really align
• How else might it work?

12
Dali (distance matrix alignment)

• How might two distance matrices look?
• All pair wise distances from each aa to all other
  aas
• If identical proteins the matrices would be
  almost identical

Low distance region in matrix if parallel
13
How turn into a similarity score?

• Find optimum set of similar sub-structures
• Even if in different 1D locations
• Find amino acid equivalence
• Once have equivalence can easily compare
  structure similarity
• E.g. with RMSD

14
Approach

• Break matrix into a bunch of overlapping
  sub-matrices
• Do an all pair wise comparison
• Sub-matrices are merged that naturally extend
• Must find pairings of sub-matrices that yield
  best overall score

1 2 3 4 5 6 7
1
2
3
4
5
6
7
1 2 3 4 5 6 7
1
2
3
4
5
6
7
15
How optimize choice of pairings
1 2 3 4 5 6 7
1
2
3
4
5
6
7

• Monte Carlo approach
• Randomly generate pairings
• Calculate overall similarity
• Multiple solutions in parallel
• Slowly improve each by randomly altering pairings
  (like a random search)
• Have some probability of keeping a solution that
  is worse than previous

1 2 3 4 5 6 7
1
2
3
4
5
6
7
16
Once have aa associations

• Can determine similarity
• How?

17
Have to minimize aa distances

• Must perturb XYZ (translation), pitch, and yaw
  (rotation) of one of the proteins minimizing RMSD
• Like linear regression
• Cant do until know which aas are associated

18
Have to minimize aa distances

• Some numeric methods start by fixing between 2
  and 4 amino acids
• Some short cuts
• Center of gravity is the average of all vectors
• Translate
• ave(p1) ave(p2)
• Singular value decomposition to rotate (Like
  Eivenvectors)

19
(No Transcript)
20
Score more complex so

• Requires double dynamic programming
• If nxm matrix then n times m different matrices
  generated pinning return path to each aa pair
• Used to generate a position specific scoring
  which is then used in aa similarity scoring
• Reduces the constraint that two particular aas
  are equivalent

a a a T
a
a
ß
a a a T
a
a
ß
a a a T
a
a
ß
a a a T
a
a
ß

a a a T
a
a
ß