1. 5. Ab initio modeling

2. And today… Introduction to ab initio modeling: the basic principles
Rosetta ab initio modeling protocol
Grid-based large-scale modeling & FOLDIT
I-Tasser
CASP

3. Types of structure prediction
Comparative modeling
Structural template detected from sequence similarity
Fold recognition
Structural template detected from fitness to fold (threading)
Ab initio modeling (Free Modeling)
No obvious structural template: model whole folding process….
Rosetta
I-Tasser
Similarity to known structure

4. Basic Ab Initio Rosetta protocol
Select fragments consistent with local sequence preferences
Assemble fragments into models with native-like global properties
Identify the best model from the population of decoys
Figures adapted from Charlie Strauss;
Protein structure prediction using ROSETTA
Rohl et al (2004) Methods in Enzymology, 383:66

5. 1. Select fragments: local sampling
Fragment libraries
fragments for each trimer and nonamers
Recent improvement was obtained by using fragments of additional sizes:
For a helix: length 5-19 & 3-12
For b sheet: length 4-10 & 3-7
Selected from PDB
< 2.5Å resolution & < 50% seq id
Ranked by sequence similarity and similarity of predicted and known secondary structure
Discard improbable conformations

6. 2. Create compact decoys using fragment assembly
Advantages of approach
Fragment library approximates Gibbs sampling
Fragments allow an accurate, but implicit, representation of the potential energy surface for local interactions.
Computer power can be invested in optimization of global features (e.g. compactness)
local
global

7. Low-resolution step Structure Representation:
Equilibrium bonds and angles (Engh & Huber 1991)
Centroid: average location of center of mass of side-chain
(Centroid | aa, f,)
No modeling of side chains
Fast

8. Low-resolution parameters
Senv - burial preference (number of neighbors)
Spair - preferred amino acid pairs (e.g. cys-cys, glu-arg, etc)
Sss + SHS - sheet and helix-sheet geometries
compactness of structure
Scb
Svdw no clashes
Srgyr globular structure
small vs. large
radius of gyration (Rgyr)

9. 2. Create compact decoys using fragment assembly
MC search with simulated annealing – start with extended conformation
28K-36K random 9-mer fragment insertions (from top25 fragments)
XK Only vdw score (until all f,y have changed)
2K Add strand pairing score (0.3 weight)
20K Compactness: Increase pairing score + add Cb and Rgyr : ±local strand pairing weight
6K/4K Full strand pairing; Full centroid function
3-mer fragment insertions
8K gunn-type (select among least perturbing fragments)

10. Further local refinement strategies
Local moves: how to perturb the backbone with minimal effect on remote regions
1. random torsion angle perturbation (helix - 0o,strand <2o, rest < 3o)
Small move - random fi,yi pair
Shear move - Dyi-1, -Dfi compensatory movements, move peptide plane
2. selection of globally non-perturbing fragments
Chuck move – fragments that minimize atom msd
Gunn move – fragments that minimize Dy, Df

11. Further local refinement strategies
Local moves: how to perturb the backbone with minimal effect on remote regions
3. adjacent f-y variation to offset global effect of fragment insertion
Wobble move – fast analytical gradient calculation
Crank shaft - combination of several wobble moves
Smaller moves are accepted with higher frequency
wobble
crank shaft
Fig. 2. Modified ‘‘crank’’ fragment insertion into 1 dan. (A) Superposition of the protein conformations preceding (black) and following (blue) insertion of a nine-residue fragment. The fragment insertion window is shown in red. The portion of the chain unperturbed by insertion is shown in gray. (B) Superposition of the protein conformations preceding (blue) and following (green) optimization of angles at a wobble site (cyan) adjacent to the insertion window. (C) Superposition of the protein conformations preceding (green) and following (magenta) optimization of angles at a second wobble site (orange) nonadjacent to the insertion window. (D) Superposition of the original (black) and final (magenta) conformations.
Before insertion
after insertion
insert
No changes
Final conformation

12. Global sampling Fragment exchange Local moves Initial global changes
Further refinement
Movie by Jens Meiler

13. 3. Identify best structure
Generate decoy population ( )
Filter to correct sampling biases
Cluster analysis identifies broadest minimum
Fullatom refinement will identify lowest energy minimum

14. High-resolution step: parameters
VdW
12-6 Lennard Jones
linear repulsion
Cutoff within Å
Solvation (Lazaridis-Karplus)
Hydrogen bonds
rij +
Weak pair potential
Electrostatic interactions
p-p, p-+
Backbone torsions (rama score) -
polar
polar N H O C d  

15. High-resolution refinement of models
MCM protocol:
120 steps of small & shear moves
Random perturbation of 5/10 torsions angles (2-3o)
Side chain optimization: rotamer trial (each 10 steps full repacking)
minimization
steps 1-60: gradually ramp up vdw repulsive
steps : add side
chain minimization
vdW repulsive
Small backbone moves and MCM
Side chain optimization
Backbone
optimization
Side chain optimization+ minimization
Backbone
optimization

16. First atom-resolution model
Target 0281 CASP6
Topology sampled by ab initio trajectory of homolog sequence (rmsd=2.2Å)
Full atom refinement reduces rmsd to 1.5Å
Side chain packing accurately recovered

17. Atom-resolution Ab Initio (I)
Challenge:
Sample near-native conformations (<~2.5A)
Approach:
Model set of homologs → diverse population samples basin of attraction
Example: exposed Leucine
Models starting from extended conf
Models starting from native conf
Toward high-resolution de novo structure prediction.
Bradley et al (2005) Science 309:1868

18. Low-resolution homolog folding improves prediction
Collect 50 homologs
(psi-blast 2 rounds; 60% non-redundant)
For each
create 2000 low-resolution models
cluster, retain large clusters (n>5), and select 500 models
Thread query sequence back onto ~20-30K models
Proceed to fullatom refinement: evaluate also homolog sequences (2 rounds of MCM protocol) … … …

19. Atom-resolution Ab Initio (II)
Hox-B1
Ubiquitin
Step1: low resolution
model homologs
Step2: atom resolution
models
Energy-based model selection
Results:
11/16 proteins of length <88 residues are modeled within <5Å

20. Sampling of b sheet topologies
Fold-tree representation of protein allows tailored optimization

21. How can we improve? (1) More computer time
BOINC – donate idle time of many home computers for Rosetta runs
Tera=1012
strong desktop: ~ gigaflop (109)

22. More computer time – is sampling an issue?
Perform very long runs on the grid (>106 decoys)
3 categories
(a) Near-native lowest energy model (<3.5Å) ✔
(b) Problem with sampling
(E near-native structures <<E decoys)
(c) Problem with energy function (E near-native structures >E decoys)
Why don’t we sample these conformations (b) ?????
Sampling bottlenecks in de novo protein structure prediction
Kim et al (2007) JMB 393:249

23. “linchpin features” are rarely sampled
Describe models as feature vectors
Identify native features not sampled in low-energy models
O: w=cis
E: left-handed strand
G: left-handed helix
B: right-handed strand
A: right handed helix
Torsion bins
Residue position
Position 23 never samples native helix conformation
 simulations never succeed
Native torsion bin
Frequently sampled Native torsion bin
Enforcement of native-like value for feature
 Some simulations now succeed
Never sampled Native torsion bin
Sampling bottlenecks in de novo protein structure prediction Kim et al (2007) JMB 393:249

24. Examples for “linchpin features”
Near active site
Regions that form late in folding
Irregular b strand pairing (mostly in edge strands)

25. How can we improve? More brains
FOLDIT – folding game
donate idle time of
many brains to improve
structure prediction
Now as Android application!
“win the Nobel prize by just playing a game”
Look also for “black belt” foldit lessons

26. Foldit Players: human spatial reasoning
Explore also strategy space: new search algorithms
Excel in solving problems where substantial backbone rearrangement is needed to bury hydrophobic residue
Challenge: Formulate problem as game
Easy to understand to non-scientists
Competition/Collaborations
Native structure
Starting structure
Foldit Model
Predicting protein structures with a multiplayer online game Cooper et al (2010) Nature 466:756

27. Example1: help in structure determination
Solved structure
Starting model
Figure 2 M-PMV retroviral protease structure improvement by the Foldit Contenders Group. (a) Progress of structure refinement over the first 16 d of game play. The x axis shows progression in time, and the y axis shows the Phaser log-likelihood (LLG) of each model in a near-native orientation. To identify a solution as correct by molecular replacement using Phaser, the model must have an LLG better than the best random models. The distribution of these best random predictions is indicated by the intensity of the pale blue band. (Because almost all the models are too poor to allow correct placement in the unit cell, Phaser LLGs are calculated after optimal superposition of each model onto the solved crystal structure and rigid-body optimization.)
(b) Starting from a quite inaccurate NMR model (red), Foldit player spvincent generated a model (yellow) considerably more similar to the later determined crystal structure (blue) in the β-strand region. (c) Starting from spvincent’s model, Foldit player grabhorn generated a model (magenta) considerably closer to the crystal structure with notable improvement of side-chain conformations in the hydrophobic core.
(d) Foldit player mimi made additional improvements (in the loop region at the top left) and generated a model (green) of sufficient accuracy to provide an unambiguous molecular replacement solution which allowed rapid determination of the ultimate crystal structure (blue).
LLG: log likelihood of a model: useful models must have better LLG than best random models (in shade)
Nature Structure and Molecular Biology 2011

28. Example 2: Foldit Puzzle #986875
Foldit detects better structures,
… using trajectories that visit high energy structures on the way
Native structure
Starting structure
Foldit Model
Predicting protein structures with a multiplayer online game Cooper et al (2010) Nature 466:756

29. Algorithm discovery by Foldit players
Added ability to create, edit, share and rate “recipes” (each player can create its own “cookbook”)
Evaluated what strategies evolve and how they spread among players
Top Players
Used at different stages in during modeling
Main strategies
All Players
Algorithm discovery by protein folding game players
Khatib et al (2012) PNAS 108:18949

30. Algorithm discovery by Foldit players
Many new recipes evolve from “Blue Fuse”
“Blue Fuse” and “Quake” are most popular
Algorithm discovery by protein folding game players
Khatib et al (2012) PNAS 108:18949

31. Foldit players detect algorithms that are similar to those used in Rosetta
Foldit “Blue Fuse”:
very similar to new Rosetta protocol “Fast Relax”
(repeated decrease/increase of repulsive term)
Comparable efficiency for short runs
Algorithm discovery by protein folding game players
Khatib et al (2012) PNAS 108:18949

32. ab initio modeling – summary:
Highly accurate
Computationally expensive (~150 CPU hours/protein)
Server of Rosetta
Good alternative: I-Tasser
Protocol developed by Zhang and Skolnick
Based on threading of parts of sequence onto parts of known structures
Very efficient and accurate (~5 CPU hours/protein)
Server of iTasser
Roy, Kukucural, Zhang (2010) I-TASSER: a unified platform for automated protein structure and function prediction. Nat Protoc. 5:725

33. I-Tasser Iterative Threading Assembly Refinement (Zhang, & Skolnick)
Separate training of protocol for: easy/ medium/ hard targets

34. i-Tasser (Zhang & Skolnick)
Threading:
Create profile:
Psiblast -> sequence profile
Psipred -> secondary structure profile
LOMETS: Metaserver for threading (FUGUE, HHSEARCH, MUSTER, PROSPECT, PPA, SP3 & SPARKS)
Excise aligned structure elements from top-scoring templates for next step
(20/30/50, depending on difficulty of target)
Wu, Solnick, Zhang (2007) Ab initio modeling of small proteins by iterative TASSER simulations. BMC Biology 5:17
Roy, Kukucural, Zhang (2010) I-TASSER: a unified platform for automated protein structure and function prediction. Nat Protoc. 5:725

35. i-Tasser (Zhang & Skolnick)

36. Tasser: Schematic representation of polypeptide chain in on- and off-lattice Ca model
Structure assembly -
efficient modeling:
2 points/residue (Ca + SG)
on-lattice ab initio for unaligned regions
off-lattice for aligned regions
Schematic representation of a piece of polypeptide chain in the on- and off-lattice CAS model. Each residue is described by its Cα and side chain center of mass (SG). Whereas Cα values (white) of unaligned residues are confined to the underlying cubic lattice system with a lattice space of 0.87 Å, Cα values (yellow) of aligned residues are excised from templates and traced off-lattice. SG values (red) are always off-lattice and determined by using a two-rotamer approximation (9).
Zhang Y., Skolnick J. PNAS 2004;101:
©2004 by National Academy of Sciences

37. i-Tasser (Zhang & Skolnick)
Monte Carlo Search by replica exchange
Exchange between simulations at different temperatures: better sampling
Scoring function: separately trained for easy, medium and hard targets
Secondary structure (PSIPRED & SAM)
Statistical terms: backbone hydrogen bonds; hydrophobicity and Ca/side chain correlations
Spatial restraints from threading templates
Sequence-based contact predictions (SVM)
(and accessible surface area prediction; NN)

38. i-Tasser (Zhang) Example for improvement over template
Constraints from threading; contact prediction are located at different sites and complement each other

39. i-Tasser (Zhang) Clustering
additional iteration of MC simulation starting from cluster centers
Final model created by optimizing hydrogen bonds

40. Contact-assisted structure prediction
ab initio restricted to small (100aa), single domain proteins
+ information about contacts
-> dramatic increase of scope (… 500aa)
Info from:
Contact prediction (bioinfo)
Experiments: e.g. NMR chemical shifts, mutagenesis, etc
Contacts may assist in
Determination of Topology:
Filter fragments
Find fragment pairs
Refinement of Topology:
Refine structure by imposing constraints
Assessment on CASP10 of Rosetta ab initio modeling: one reliable non-local contact every <12aa> needed for reliable modeling
Kim et al. (2013) One contact for every twelve residues allows robust and accurate topology-level protein structure modeling Proteins 82:208

41. Contact-assisted structure prediction
Flowchart of protocol
Topology determination: from partial threading (SPARKS, Rosetta)
Topology refinement:
RosettaCM recombination protocol (next week)
Kim et al. (2013) One contact for every twelve residues allows robust and accurate topology-level protein structure modeling Proteins 82:208

42. Contact-assisted structure prediction
Improved models for large structures using contacts:
native
Ab initio
Assisted ab initio
Kim et al. (2013) One contact for every twelve residues allows robust and accurate topology-level protein structure modeling Proteins 82:208

43. CASP Identification of “winner strategies”:
Rosetta in CASP4-6
iTasser in CASP7 & CASP8
servers
improved combination of multiple templates in CASP9
CASP10: refinement with MD
CASP11: contact prediction methods & contact-assisted modeling
Double-blind structure prediction experiment
allows assessment of different approaches
every 2 years; summer 2014: CASP11
Steady improvement of methodology
Categories:
Template based modeling (TBM)
Free modeling (FM)
Refinement of initial models
New: prediction of contacts, unstructured regions, ligand binding sites
Proteins special issue vol:82, S2

44. CASP Around 130 targets in last rounds
Until CASP 9: Target difficulty decreases
CASP10:
131 domains
(20 free modeling)
Targets now more difficult than previous CASPs
Proteins special issue vol:79, S10
Kryshtafovych et al.(2011). Proteins 79:S196–207

45. Measure of performance
Compare predicted to solved structure
superimpose short fragments (length n=3,5,7 residues; iteratively)
find maximal superimposed part N, where
N Ca atom pairs are within xÅ
4 thresholds: x=1.0, 2.0, 4.0, 8.0
GDT_TS = ¼ (N1+N2+N4+N8)

46. CASP4 ab initio summary
18 newly solved structures predicted prior to publication of structure.
none recognized by sequence similarity
none with close structural homologs
Rosetta
Independently assessed scoring:
2=“Well Above Average”, 1=“okay”, 0=“lousy”

47. Improvement over the years
Improvement in each round
CASP7: in difficult region
CASP8: accuracy in template-based modeling (few difficult cases)
CASP9: intermediate difficulty targets
CASP10: refinement using MD (Michael Feig)

48. Free Modeling with Rosetta in CASP8

49. Free Modeling with Rosetta in CASP9
Server model: predicts kinked helix
Only model with 4 beta strands (most predictions: all helical protein)
model
best template
Kinch et al. (2011). Proteins, 79:S59–73

50. Free Modeling with Rosetta in CASP11
best template

51. Rosetta in CASP8: modification of fragment size improves prediction
longer fragments for alpha helical proteins (5-19; 3-12)
shorter fragments for beta sheets (4-10; 3-7)

52. FM with ITasser

53. Improved automatic servers
increased contribution of automatic servers
predictions of mostly similar quality
improve now also difficult targets
Human server +

54. CASP10: Foldit platform joins the game for coopetition
Start from Foldit models; proceed with different approaches
Joint forces produce best model

55. Summary steady improvement of structure prediction over the years
impressing quality of current ab initio modeling
efficient combination of appropriate sampling strategies and a tailored energy function
models now often better than template
automatic servers outperform now also FM