1. AMPLE – Using de novo or ab initio protein structure modelling techniques to create and enhance search models for use in Molecular Replacement
Jaclyn Bibby, Jens Thomas,
Olga Mayans and Daniel Rigden
Institute of Integrative Biology
Ronan Keegan and Martyn Winn
Collaborative Computational Project 4 (CCP4)

2. ab initio modelling of proteins for molecular replacement
AMPLE
ab initio modelling of proteins for molecular replacement
Joint development by CCP4 and the University of Liverpool
AMPLE is a comprehensive project to assess the suitability of using cheaply obtained ab initio models in molecular replacement
An additional goal of the project is to make AMPLE into an automated software tool that can be made generally available to potential users through CCP4

3. Ab initio structure prediction
Ab initio (or de novo) structure prediction is the prediction of a target structure fold based purely on its sequence information
Methods have greatly improved in recent years with the aid of the CASP experiments (Critical Assessment of Protein structure prediction)
Some examples are:
Rosetta
I-TASSER
QUARK

4. Ab initio structure prediction
1000’s of “Decoys” assembled of fragments from PDB structures
Decoys are clustered and centroid representatives of largest cluster are considered candidate fold predictions
Side chains added to selected decoys
Refinement under a more realistic physics-based force field
Initial fragment assembly stage requires relatively modest computing power
Refinement stage can require supercomputing resources

5. Ab initio structure prediction
1000’s of “Decoys” assembled of fragments from PDB structures
Decoys are clustered and centroid representatives of largest cluster are considered candidate fold predictions
Side chains added to selected decoys
Refinement under a more realistic physics-based force field
Initial fragment assembly stage requires relatively modest computing power
Refinement stage can require supercomputing resources

6. Ab initio modelling and Molecular Replacement
Combining this method with molecular replacement can be a powerful technique for solving the phase problem in cases where there are no obvious homologous structures available
Two approaches have been taken
All-atom modelling to produce single search models of maximum completeness and accuracy (Qian et. al, 2007)
Solved 1/3 of test set of 30 targets (Das et. al, 2009)
Computationally expensive
Taking cheaply obtained decoys from the initial fragment assembly step and preparing them has been shown to produce successful MR search models (Rigden et al. 2008, Caliandor 2009)

7. Molecular Replacement
Synergies
ab initio modelling
Molecular Replacement
Produces clusters of similar model structures
Works effectively with superposed ensembles approximating the target
Within and between clusters, similarity indicates accuracy  can trim inaccurate regions leaving more reliable core
May only require a partial model
Fast modelling is polyAla only
Side chains are often (partially) removed for MR

8. The AMPLE Pipeline
Uses Rosetta to perform ab initio modelling and the generation of decoy models
Can also accept models generated externally
Currently designed for < 120 residues and resolution better than 2.2 Å (but may work outside these restrictions e.g. transmembrane, coiled-coil proteins)

9. Decoy Generation Decoys generated first with Rosetta
Quark has also been used during development
Typical number of decoys required is 1000 but this can be varied
In easier cases as few as 50 decoys can be sufficient

10. Decoy Clustering
Clustering using SPICKER to identify the most likely fold for the target
A large top cluster is usually indicative of a correct prediction
A subset of decoys (max. 200) closest to the centroid of each of the largest 3 clusters are selected for further processing

11. Decoy Clustering
Each cluster is then structurally aligned using the maximum likelihood algorithm implemented in Theseus (Theobold & Wuttke, 2006)
This helps to identify structurally conserved regions
Gives a variance score which can later be used to guide truncation

12. Ensemble Truncation
We’ve found that success or failure in the molecular replacement step is highly sensitive to the accuracy of the search model
Sampling many degrees of truncation with different levels of side chain inclusion is essential
High variance regions are cut away in steps to give a set of ensemble models

13. Further processing
These truncated clusters are further processed and side chains are added to give a large set of search models for molecular replacement

14. Molecular replacement
Molecular replacement is performed using MrBUMP from the CCP4 suite which automates the procedure
Search models are processed by both Phaser and Molrep
Post molecular replacement, positioned search models are refined using Refmac5 to get an initial indication of success or failure
C-alpha tracing with SHELXE, model building with Buccaneer, ARP/wARP

15. Testing Test set of 295 small proteins (40-120 residues) from the PDB
Structure factor data also available
Resolution of 2.2 Å or better
Single molecule in the asymmetric unit
Mixture of all-α, all-β and mixed α-β secondary structure
1000 decoys generated for each case using Rosetta
Information from any homologues was excluded from the fragment generation step

16. Assessing Solutions
Initially we used Reforigin to compare solutions with the deposited structures
More stringent method: attempt to rebuild the structures
SHELXE: partial CC of >25% & average fragment length of 10 or more
Further confirmation provided through building with ARP/wARP and Buccaneer.

17. Using these guidelines, 126 successes out of 295 (~43%) were achieved
These are solutions that could be successfully traced in SHELXE
Other well positioned solutions existed but could not be traced. These may be possible to solve through manual model building

18. Results based on secondary structure type
Overall success rate: all-α = 80%; all-β = 2%; mixed α-β = 37%

19. Variance and Truncation
Variability between decoys in each cluster corresponds to their deviation from the deposited native structure
2P5K example: C-terminal region predicted as least reliably modelled portion by Theseus alignment variance score

20. Search model ensemble size/truncation

21. Running Times
Average times for complete run (decoy generation, preparation, MR and chain tracing) was 2 CPU days
A parallelised version of the code making use of Sun Grid Engine for batch farming of model generation and molecular replacement significantly speeds up the process. Results can be achieved in less than 1 hour

22. Exploiting distant homologues
a. Clustered decoy models, b. Truncated ensemble, c. Positioned MR solution, d. Shelxe c-alpha trace, e. Completed structure

23. Remodelling related NMR structures or distant homologues
Can provide AMPLE with a template for the target which could be a related NMR structure or a distant homologue
AMPLE will use Rosetta to “re-model” this template to something that should in theory be closer to the target

24. Transmembrane Proteins
Experimentally very difficult to work with/crystallise
Represent ~30% of all proteins
Make up < 3% of structures in the Protein Data Bank
Presence in the membrane constrains their shape, so they can be easier to model
MR with ab initio transmembrane models hasn't been tried yet
extra cellular
(aqueous)
transmembrane region
(hydrophobic)
intra cellular
(aqueous)
Image:

25. Selected 18 transmembrane proteins:
23 – 249 residues
1.45 – 2.5A resolution
7 clear successes
5 possible successes
223 residue structure (3GD8) could be largest ever solved with ab initio modelling.

26. Selected 18 transmembrane proteins:
23 – 249 residues
1.45 – 2.5A resolution
7 clear successes
5 possible successes
223 residue structure (3GD8) could be largest ever solved with ab initio modelling.

27. Coiled-coil targets Difficult to solve in MR even with good homologues
α-helical nature makes them suitable targets for AMPLE
Initial testing has been very promising with 80% success rate
Some novel structures have also been solved

28. AMPLE Availability Beta version available as part of CCP4 6.3.0.
Improved and more robust version to be released as part of CCP
Requires installation of several non-ccp4 packages:
Rosetta, SHELXE, Theseus, SPICKER, Maxcluster
Future versions will have a reduced number of dependencies

29. Documentation available from http://ccp4wiki.org

30. Summary
AMPLE is a pipeline designed to prepare cheaply obtained decoy models from ab initio modelling for use as search models in molecular replacement
Results show that the method works well for smaller proteins particularly those containing α-helical secondary structure
Tests were limited to structures of 120 residues in length but has worked for cases up to 250 residues
New avenues – NMR, Homolgue remodelling
Several real successes
Currently available as a beta-version in CCP

31. Acknowledgements
Jaclyn Bibby, Daniel Rigden, Jens Thomas, University of Liverpool
Olga Mayans, University of Liverpool
Martyn Winn, Daresbury Laboratory
Andrea Thorn, Tim Gruene & George Sheldrick (SHELX)
Developers of Rosetta and Quark
Refmac: Garib Mushudov, LMB-MRC Cambridge
Molrep: Alexei Vagin & Andrey Lebedev
Phaser: Randy Read, Airlie McCoy & Gabor Bunkozci
Thanks to authors of all underlying programs
Funding:
BBSRC
Support from CCP4 & the Research Complex at Harwell
Poster: MS04-12 (Rootes Building)