Structure prediction

1
Structure prediction

• Why do we need structure prediction ?
• Intellectual
• Practical
• Levinthals paradox
• Anfinsens experiments
• How it will be solved
• Physics
• Computer science
• The protein design problem

2
4 Basic Levels of Protein Structure and all
information is in the sequence
3
What is homology
4
Structures change linear with ED
5
Homology

• Similar sequence - Similar structure ?

6
Examples of similar structure
7
Zones
8
Homology can be detected by Sequence Alignment

• Key aspect of sequence comparison is sequence
  alignment
• A sequence alignment maximizes the number of
  positions that are in agreement in two sequences

9
Alignments

• Local alignment
• Global alignments

Global Alignments LGPSTKDFGKISESREFDN
LNQLERSFGKINM-RLEDA Local
Alignments ----------FGKI----------
----------FGKI----------
10
Dotplots
11
Methods to align

• Optimal alignment
• Maximise similarities
• Minimize gaps
• Score of an alignment
• Score for substition
• Gap-opening and gap-extension costs
• Dynamic programming
• Finds optimal solution
• BLAST
• Heuristic, fast algorithm using indexes (hashes)?

12
When are two sequences homologous
13
Identities do not provide the best similarity
14
Statistics of Sequence scores

• Local alignments
• Follows extreme value distribution
• Scores depends on log(length)?
• E (or P-value)?
• Global alignments
• Heuristics
• Randomize sequences

15
Statistical comparison of alignment scores
16
How to improve alignments

• Use more evolutionary information
• Multiple alignments
• Profiles
• HMMS
• Profile-profile alignments
• Using additional information
• Structure
• Structural alignments

17
Multiple sequence alignments

• Computationally intensive
• Heuristic methods

18
Profiles can be used to detect distant homologs

• Extra information
• How to best use
• Different methods
• Patterns
• Evolutionary method
• Profile methods
• HMMs
• ANNs

19
PSI-BLAST in a nutshell

• With a protein sequence as query, use BLAST to
  search a protein sequence database.
• Collapse significant local alignments (those with
  E-value less than or equal to a set threshold h)
  into a multiple alignment, using the residues of
  the query sequence as alignment-column
  placeholders.
• Abstract a position-specific score matrix from
  the multiple alignment.
• Search the database with the score matrix as
  query.
• Iterate a fixed number of times, or until
  convergence.

20
Protein structure prediction (and other uses for
molecules in life in a computer)?

• Secondary structure predictions
• Homology detection
• What is homology
• Why is is related to protein structure
• How does it work
• Simulations of folding
• What is physics ?
• Realistic simulations (folding_at_home)?
• Smart simulations (rosetta_at_home)?

21
It's not that simple...

• Amino acid sequence contains all the information
  for 3D structure (experiments of Anfinsen,
  1970's)
• But, there are thousands of atoms, rotatable
  bonds, solvent and other molecules to deal
  with...

22
All the 3D information is in the sequence
23
Levinthal Paradox

• Cyrus Levinthal, Columbia University, 1968
• Levinthal's paradox
• If we have 3 rotamers per residues a 100 residue
  protein have 3100 possible conformations. To
  search all these takes longer than the time of
  the universe. But proteins fold in less than a
  second.
• Resolution Proteins have to fold through some
  directed process
• Goal is to understand the dynamics of this process

24
Old vs. New Views of folding

• Old
• Hierarchical view of protein folding
• Secondary structures form, then interact to form
  tertiary structures
• General order of events
• New
• Statistical ensembles of states
• Potential energy landscape
• Folding Funnel

25
Two alternatives for structure prediction

• Simulation of protein folding
• Folding_at_home (Erik next week)?
• Identification of lowest energy structure
• More successful (today)?
• Several layers
• Secondary structure
• 3D-structure

26
Secondary structure prediction

• AA preferences for different SS
• Pro
• Does not have a NH backbone
• No H-bonds
• Prefers Coils
• Also in N-terminal part of helices and Beta-turns
• Gly (compared with Ala)
• No sidechain on Gly (more flexible)?
• Polar groups in loops
• Additional H-bonds to backbone

27
Amino acid preferences in coil
28
Amino acid preferences in ?-Strand
29
Secondary structure preferences

• C? branched AAs prefers sheets
• Entropic cost in helices of sidechain rotations
• Hydrophobic groups prefers SS-elements
• Negatively charged residues at C-terminal end of
  helices due to dipole effect.

30
Amino acid preferences in ?-Strand
31
Amino acid preferences in ?-Helix
32
Secondary structure preferences

• C? branched AAs prefers sheets
• Entropic cost in helices of sidechain rotations
• Hydrophobic groups in SS-elements
• Polar in loops
• Negatively charged residues at C-terminal end of
  helices due to dipole effect.

33
Templates for helix, loops and sheet
34
More elaborate templates

• Key residues
• Gly in turns

35
Incorporating globular effects

• Hydrophobic lake

36
Exemple of SS predictions
37
PhD (Rost Sander, 1994)?
38
PhD-Input
39
PhD-architecture
40
PhD-predictions
41
PhD summary

• First methods with gt70 Q3
• Correct length distributions
• Much better beta strand predictions
• Good correlation between score and accuracy
• Better predictions for larger multiple sequence
  alignments

42
Threading

• A priori prediction of the Interferon fold in
  1985
• Good precdiction of helices

43
Prediction of interferon fold
44
FR methodologies
45
3D profiles
46
Threading or Fold recognition
VIFVLWGNAARQKCN LLFQTKHQHAVLACPH
47
PROSA/THREADER
48
How good is FR?

• LiveBench and CASP measure performance.
• E-values work reasonably well.
• In the real world, you might get a few percent
  more hits'' with FR compared to PSI-BLAST.
• Individual researcher vs. genome-wide analysis
• Structure information not necessary?

49
Sucess of FR
50
Alignments are not always perfect
51
Does threading really work ?

• Evolutionary methods work better
• Secondary Structure Predictions might help

52
AB-INITIO methods

• Simulate the process of folding
• Folding_at_home - MD simulations of small peptides
• Find the lowest energy structure
• Not simulate process.
• Consequences of small energy gap
• Unrealistic to model exactly
• Easier to distinguish between Correct/Incorrect
  than between Folded/non-folded.

53
How Rosetta Works

• Minimize energy in the folded state
• Uses a combination of energy formulas based on
  the likelihood of particular structures, and the
  fitness of the sequence
• Side-chains simplified to a centroid located at
  center of mass of the side-chain
• Average of observed side-chain centroids in known
  structures
• Local sequence does not decide the local
  structure, it only biases the decision
• Non-local favorable conditions
• Buried hydrophobic fragments
• Paired ß strands
• Specific side-chain interactions

54
Rosetta clustering the models

• Compare models to each other with RMSD
• Models can come from different family members
• Cutoff varied to give 80-100 members in largest
  cluster
• The largest clusters are assumed to contain the
  best structures (attractors in folding space...?)

55
Recent improvements to Rosetta

• Refinement in HR rosetta
• Make small dihedral changes
• Rebuild sidechains
• Minimize (in dihedral space)?
• Evaluate energy
• Go To 1
• 5 out 16 small proteins lt 1.5 Å

56
Physics of Rosetta

• Is Rosetta physical ?
• What is the most important terms in globular and
  local free energies ?
• How does proteins really fold ?
• What do you think ?

57
Designs

• Molten globule designs
• Regan 50

58
Deign of four-helical bundle (De Grado, 1991)
Molten Globule
59
What characterizes a molten globule

• Compact
• Good secondary structure
• Not solid
• Sidechains not packed
• No cooperative folding

60
Mayo method

• Automatic design
• Take fold (backbone)?
• Take sequence (random)?
• Mutate sequence
• Build sidechains
• Calculate energy
• Accept/reject
• Go to 3

61
Designing a non-zinc finger
62
Design of a non-zinc finger (Dahiyat and Mayo)
63
Alfabetin
64
Non-MG alfabetin shows cooperative folding
65
TOP7 (Rosetta Design)

• Novel fold
• Iterate between design and refinement
• Non molten globular behavior.

66
Iteration between Seq and Str
Sequence Structure
67
The project

• Three goals
• Learn how to develop a (binary) predictor
• Read the background literature
• Write a scientific report about your work
• Write a program that can do all of this.
• Additional goal (for top grades)
• Make a web-server
• Combine your predictor into a full system

68
Tools

• Python (or other language)
• Write scripts to do the work
• svmlight
• Used in the bioinformatics course
• Preparsed datafiles
• Annotations
• PSIBLAST needs to be parsed
• Evaluations programs should be developed.

69
The projects

• Binary classifier
• Alpha-helix/other etc..
• Surface area
• Membrane non-membrane
• Globular or membrane datasets

70
The program

• Input.
• Sequence in fasta file
• Output
• A prediction for each residue in the sequence

71
Web-server

• For top grades (A,B) a web-server should be
  developed, using the following steps.
• Learn how to use PHP
• Ask for an account on a web-server.
• Use the templates index.php available from the
  web-page.

72
The report

• The following sections
• Abstract
• Introduction
• Methods
• Results and Discussion
• Conclusions
• References
• More info May 8