[phenixbb] refinement at 3.7A

Pavel Afonine PAfonine at lbl.gov
Fri Apr 9 12:01:24 PDT 2010

Hi Vincent,

> What is low res? 

you may consider having a look at discussion about resolution limits 
presented here:
Acta Cryst. (2009). D65, 1283-1291.

These are good (in my opinion) publications about low resolution 
refinement challenges. Example of some of them:

Acta Cryst. (2006). D62, 923-932
Considerations for the refinement of low-resolution crystal structures
B. DeLaBarre and A. T. Brunger

Low-Resolution Crystallography Is Coming of Age
Structure, Volume 13, Issue 2, Pages 171-172
A. Brunger

Acta Cryst. (2009). D65, 121-127
Model-building strategies for low-resolution X-ray crystallographic data
A. M. Karmali, T. L. Blundell and N. Furnham

G.F.Schroeder, M.Levitt, A.T.Brunger,
Super-resolution biomolecular crystallography with low-resolution data.
Nature, doi:10.1038/nature08892 (2010).

> at what resolution should we use
> only rigid body refinement or can we use minimization as well?

Let's try avoid mixing unmixable first : model parameterization (rigid 
body, individual coordinates, etc.) and optimization algorithms 
(minimization, simulated annealing, etc.).

> In my case, i have a 2.7A and a 3.7A structures, 

2.7A resolution:

- coordinates:
At this resolution you definitely refine individual coordinates. 
Depending on how far your model is from the final one, you can use SA 
(if the model is still poor) or gradient-driven minimization (if the 
model quality is decent) as optimization choices.
Use NCS if available, but exercise care when selecting NCS related 
groups. Phenix.refine can select NCS groups automatically, but it is 
always a good idea to review the automatic choice manually. You can also 
try refinement without using NCS but with tighter restraints - I've seen 
cases when it works much better.
If you observe overfitting - optimize X-ray/restraints target weights 
(use optimize_wxc=true to do it automatically, or do it manually by 
playing with wxc_scale parameter). Also, it is a good idea to run a 
final refinement run (before PDB deposition) with weights optimization 
turned on.
Use local real-space refinement (fix_rotamers=true option; more details:
http://cci.lbl.gov/~afonine/rsr.pdf )

- ADPs (B-factors):
Refine individual ADPs. Use TLS (phenix.refine does combined individual 
ADP + TLS refinement, which is the best option most of the time). 
Carefully select TLS groups. Here is how I typically use TLSMD (copy 
from one of my previous posts):

1. Reset all B-factors to a average value:

phenix.pdbtools model.pdb set_b_iso=25

If model.pdb has anisotropic atoms (ANISOU records), then you need to 
convert them to isotropic: "convert_to_iso=true"

2. Refine group ADP only:

phenix.refine model_all_B_25.pdb data.mtz strategy=group_adp 

3. Submit the refined model to TLSMD

4. Discard the refined model (the one obtained at step #2).

5. Use selections for TLS groups in subsequent refinement.

If you turn on NCS restraints, they will be applied to ADP as well (to 
residual B-factors, not total).

Sometimes, very rarely, if the data is horribly incomplete and 
data-to-parameters ratio is bad, and tightening ADB restraints doesn't 
help, the group B-factor refinement (with one or two refinable isotropic 
ADP per residue) is an option to try. In my experience I had to do this 
in a refinement against neutron data.

3.7A resolution:

Try above guidelines for 2.7A first (except probably "fix_rotamers"), 
and if it doesn't work, then:

- coordinates:
Try torsion angle parameterization (constrained rigid body), and SA is 
the only working optimization protocol available in phenix.refine for 
this type of parameterization. Try secondary structure restraints if you 
see that secondary structure gets distorted during refinement.

- ADP refinement:
try group ADP refinement (in combination with TLS) instead of highly 
restrained individual.

If 3.7A structure is an analog of 2.7A structure, then you may use a 
higher resolution structure as a source of restrains for low resolution 
one (I think this option will be available soon in phenix.refine - check 
with Jeff Headd who is working on it).

If you have high-order NCS, use "thin-shells" algorithm to assign free-R 
flags (available in PHENIX). This is not done by default - you need to 
turn this flag on yourself.

Did you check for twinning?

Optimize mask calculation (main.optimize_mask=True option). This may 
give you 1-3% better R-factor at low resolution. The next version of 
PHENIX will do it automatically.

> If i do a simple rigid body refinement, I have R=32.1, Rfree=35.8. If i
> use individual site refinement, i have R=30.7, Rfree=35.8. (both with
> group ADP, TLS and NCS).

Here is how your numbers compare to the numbers for structures in PDB at 
similar resolution:

phenix.r_factor_statistics 3.7 left_offset=0.2 right_offset=0.2 n_bins=5

Histogram of Rwork for models in PDB at resolution 3.50-3.90 A:
     0.168 - 0.209      : 8
     0.209 - 0.250      : 37
     0.250 - 0.291      : 65
     0.291 - 0.332      : 36 <<< your model
     0.332 - 0.373      : 13
Histogram of Rfree for models in PDB at resolution 3.50-3.90 A:
     0.184 - 0.231      : 2
     0.231 - 0.278      : 23
     0.278 - 0.324      : 72
     0.324 - 0.371      : 51 <<< your model
     0.371 - 0.418      : 11
Histogram of Rfree-Rwork for all model in PDB at resolution 3.50-3.90 A:
     0.002 - 0.020      : 24
     0.020 - 0.038      : 42
     0.038 - 0.056      : 47 <<< your model
     0.056 - 0.075      : 35
     0.075 - 0.093      : 11
Number of structures considered: 159

Doesn't look too bad I would say.

> Basically, could you update me on what's today's dogma on what type of
> refinement we should use at what resolution?

No dogma. Try a few possible refinement strategies given the resolution 
and model quality and see which one works best. This is more robust and 
time-efficient than shaking the air with speculations about what could 
be used and why. Once you have found the best refinement protocol for 
your particular case then you can spend some time reviewing what you 
have tried and why one strategy worked better than the other. Typically, 
you need to try 3-5 refinement protocols which is not a big deal given 
the speed of today's computers.


More information about the phenixbb mailing list