Automated phase improvement and model building with Parrot and Buccaneer

1
Automated phase improvementand model building
withParrot and Buccaneer
Kevin Cowtancowtan_at_ysbl.york.ac.uk
2
X-ray structure solution pipeline...
Data collection
Data processing
Experimental phasing
Molecular Replacement
Density Modification
Model building
Refinement
Rebuilding Validation
3
Density modification

• Density modification is a problem in combining
  information

4
Density modification

• 1. Rudimentary calculation

FFT
F, f
?(x)
ffmod
Modify ?
?mod(x)
Fmod, fmod
FFT-1
Real space
Reciprocal space
5
Density modification

• 3. Phase probability distributions

centroid
FFT
F, P(f)
?(x)
Fbest, fbest
P(f)Pexp(f),Pmod(f)
Modify ?
?mod(x)
Pmod(f)
Fmod, fmod
FFT-1
likelihood
Real space
Reciprocal space
6
Density modification
DM, SOLOMON, (CNS)

• 4. Bias reduction (gamma-correction)

centroid
FFT
F, P(f)
?(x)
Fbest, fbest
Modify ?
P(f)Pexp(f),Pmod(f)
?mod(x)
?-correct
??(x)
Pmod(f)
Fmod, fmod
FFT-1
likelihood
J.P.Abrahams
7
Density modification
PARROT

• 5. Maximum Likelihood H-L

centroid
FFT
F, P(f)
?(x)
Fbest, fbest
Modify ?
?mod(x)
?-correct
??(x)
Fmod, fmod
FFT-1
MLHL
8
Density modification

• Traditional density modification techniques
• Solvent flattening
• Histogram matching
• Non-crystallographic symmetry (NCS) averaging

9
Solvent flattening
10
Histogram matching
P(?)
Noise
True

• A technique from image processing for modifying
  the protein region.
• Noise maps have Gaussian histogram.
• Well phased maps have a skewed distribution
  sharper peaks and bigger gaps.
• Sharpen the protein density by a transform which
  matches the histogram of a well phased map.
• Useful at better than 4A.

?
11
Non-crystallographic symmetry

• If the molecule has internalsymmetry, we can
  averagetogether related regions.
• In the averaged map, thesignal-noise level is
  improved.
• If a full density modificationcalculation is
  performed,powerful phase relationshipsare
  formed.
• With 4-fold NCS, can phasefrom random!

12
Non-crystallographic symmetry

• How do you know if you have NCS?
• Cell content analysis how many monomers in ASU?
• Self-rotation function.
• Difference Pattersons (pseudo-translation only).
• How do you determine the NCS?
• From heavy atoms.
• From initial model building.
• From molecular replacement.
• From density MR (hard).
• Mask determined automatically.

13
Density modification in Parrot

• Builds on existing ideas
• DM
• Solvent flattening
• Histogram matching
• NCS averaging
• Perturbation gamma
• Solomon
• Gamma correction
• Local variance solvent mask
• Weighted averaging mask

14
Density modification in Parrot

• New developments
• MLHL phase combination
• (as used in refinement refmac, cns)
• Anisotropy correction
• Problem-specific density histograms
• (rather than a standard library)
• Pairwise-weighted NCS averaging...

15
Estimating phase probabilities

• Solution
• MLHL-type likelihood
• target function.

Perform the error estimation and phase
combination in a single step, using a likelihood
function which incorporates the experimental
phase information as a prior. This is the same
MLHL-type like likelihood refinement target used
in modern refinement software such as refmac or
cns.
16
Recent Developments

• Pairwise-weighted NCS averaging
• Average each pair of NCS related molecules
  separately with its own mask.
• Generalisation and automation of multi-domain
  averaging.

C
B
A
A
C
B
B
C
A
17
Parrot
18
Parrot Rice vs MLHL
Map correlations Comparing old and
new likelihood functions.
19
Parrot simple vs NCS averaged
Map correlations Comparing with
and without NCS averaging.
20
DM vs PARROT vs PIRATE

• residues autobuilt and sequenced
• 50 JCSG structures, 1.8-3.2A resolution

79.1
78.4
74.2
DM
PARROT
PIRATE
21
DM vs PARROT vs PIRATE

• Mean time taken
• 50 JCSG structures, 1.8-3.2A resolution

887s
10s
6s
DM
PARROT
PIRATE
22
DM vs PARROT vs PIRATE

• residues autobuilt and sequenced
• 50 JCSG structures, 1.8-3.2A resolution

79.1
78.4
74.2
DM
PARROT
PIRATE
23
DM vs PARROT vs PIRATE

• Mean time taken
• 50 JCSG structures, 1.8-3.2A resolution

887s
10s
6s
DM
PARROT
PIRATE
24
Buccaneer

• Statistical model building software based on the
  use of a reference structure to construct
  likelihood targets for protein features.
• Buccaneer-Refmac pipeline
• NCS auto-completion
• Improved sequencing

25
Buccaneer Latest

• Buccaneer 1.2
• Use of Se atoms, MR model in sequencing.
• Improved numbering of output sequences (ins/del)
• Favour more probable sidechain rotamers
• Prune clashing side chains
• Optionally fix the model in the ASU
• Performance improvements (1.5 x)
• Including 'Fast mode' (2-3 x for good maps)
• Multi-threading (not in CCP4 6.1.1)
• Buccaneer 1.3
• Molecular replacement rebuild mode
• Performance improvements, more cycles.

26
Buccaneer Method

• Compare simulated map and known model to obtain
  likelihood target, then search for this target in
  the unknown map.

Reference structure
Work structure
LLK
27
Buccaneer Method

• Compile statistics for reference map in 4A sphere
  about C? gt LLK target.

• Use mean/variance.

4A sphere about Ca also used by 'CAPRA' Ioeger et
al. (but different target function).
28
Buccaneer

• 10 stages
• Find candidate C-alpha positions
• Grow them into chain fragments
• Join and merge the fragments, resolving branches
• Link nearby N and C terminii (if possible)
• Sequence the chains (i.e. dock sequence)
• Correct insertions/deletions
• Filter based on poor density
• NCS Rebuild to complete NCS copies of chains
• Prune any remaining clashing chains
• Rebuild side chains

29
Buccaneer

• Use a likelihood function based on conserved
  density features.
• The same likelihood function is used several
  times. This makes the program very simple (lt3000
  lines), and the whole calculation works over a
  range of resolutions.

Finding, growing Look for C-alpha environment
Sequencing Look for C-beta environment
... x20
ALA
CYS
HIS
MET
THR
30
Buccaneer

• Case Study
• A difficult loop in a 2.9A map, calculated using
  real data from the JCSG.

31
Find candidate C-alpha positions
32
Grow into chain fragments
33
Join and merge chain fragments
34
Sequence the chains
35
Correct insertions/deletions
36
Prune any remaining clashing chains
37
Rebuild side chains
38
Comparison to the final model
39
Buccaneer Results

• Model completeness not very dependent on
  resolution

40
Buccaneer Results

• Model completeness dependent on initial phases

41
Buccaneer
Cycle BUCCANEER and REFMAC for most complete model
Single run of BUCCANEER only (more options) quick
assessment/advanced use
42
Buccaneer
43
Buccaneer

• What it does
• Trace protein chains (trans-peptides only)
• Link across small gaps
• Sequence
• Apply NCS
• Build side chains (roughly)
• Refine (if recycled)
• WORK AT LOW RESOLUTIONS
• 3.7A with good phases

44
Buccaneer

• What it does not do (yet)
• Cis-peptides
• Waters
• Ligands
• Loop fitting
• Tidy up the resulting model
• In other words, it is an ideal component for use
  in larger pipelines.

45
Buccaneer

• What you need to do afterwards
• Tidy up with Coot.
• Or ARP/wARP when resolution is good.
• Buccaneer/ARP/wARP betterfaster than ARP/wARP.
• Typical Coot steps
• Connect up any broken chains.
• Use density fit and rotamer analysis to check
  rotamers.
• Check Ramachandran, molprobity, etc.
• Add waters, ligands, check un-modeled blobs..
• Re-refine, examine difference maps.

46
Buccaneer Summary

• A simple, fast, easy to use (i.e. MTZ and
  sequence) method of model building which is
  robust against resolution.
• User reports for structures down to 3.7A when
  phasing is good.
• Results can be further improved by iterating with
  refinement in refmac (and in future, density
  modification).
• Proven on real world problems.

47
Achnowledgements

• Help
• JCSG data archive www.jcsg.org
• Eleanor Dodson, Paul Emsley, Randy Read,
  Clemens Vonrhein, Raj Pannu
• Funding
• The Royal Society