[phenixbb] Problem in reducing R-factor

[Nathaniel Echols]

On Fri, Oct 11, 2013 at 6:19 PM, Ryan Spencer <rspencer at uci.edu> wrote:

>         For high resolution structures, below 1.5 Å, is there a set of
> rules
> for best practices for refinement?
>

Not really rules that I'm aware of... just general guidelines, and not
well-codified.  I've wanted for a long time for us to write some sort of
extensive how-to across all resolution ranges, but like everything else,
this has to compete with writing code and "real" papers.  As Pavel
mentioned, these resolution limits are somewhat subjective depending on
completeness, how aggressively you process the data, etc.

But, a few basic suggestions:

- you will probably not want to use real-space refinement after the first
couple of rounds
- start parameterizing conservatively, then get more aggressive as you near
the end of refinement (if nothing else, it's faster this way)
- don't use any restraints other than the default geometry restraints plus
whatever custom bonds are required by the chemistry (e.g. prosthetic groups)
- use hydrogens by the end of refinement, but riding only until you're at
subatomic resolution
- the automatic weighting is probably going to work great; weight
optimization might help for the final round, but probably not necessary
- below 1.5Å, try making non-water/non-hydrogen anisotropic (and compare to
purely isotropic); below 1.2Å, you can start thinking about making all
non-H/D anisotropic.  (Also note that heavier atoms can become anisotropic
even earlier, so in your case, refining individual anisotropic B-factors
for iodine and TLS for everything else would be a reasonable alternative.)
- you should probably be able to get rid of all geometry outliers,
including clashes - but you may see real (correct) Ramachandran or rotamer
outliers at this point
- you're inevitably going to see more blobs in the difference map that you
can't explain than usual.  don't sweat it.
- you should definitely see a fair number of alternate conformations - I
don't know what is expected statistically but I'd say at least 10% of
protein residues will have them
- pay attention to waters with close contacts to oxygen - these may be
sodium (or other) ions

I've attached a few slides that give *very approximate* suggestions -
basically just reframing what Pavel and others have said over the years.
 Although some of these are pretty clear-cut (e.g. you absolutely do not
want to use a reference model), you will always need to experiment somewhat
to find out what is best for *your* model and data.

I think based on your R-factors and statistics that you're headed in the
right direction.  I'd try to get the clashscore as close to zero as
possible, though.

         I've been toying with different parameters and found that the
> weighting didn't make a huge difference, which is probably expected, with a
> 1.3 Å dataset. Do stereochemical/ADP weights have any effect?
>

Do you mean weight optimization?  I wouldn't expect it to make much
difference.


>         For the particular data set I'm working with, I'm refining directly
> against anomalous data, more specifically against 12 Iodines. Refining the
> f' and f" has worked well, generating a beautiful density map.
>

Good idea.  Also consider looking at the anomalous residual map
(map_type=anom_residual), which may show weaker anomalous scatterers like
chloride or sulfur.  (It will be flattened around the iodines if you refine
their f'' values.)

        Refining against the model with hydrogens increased the Rfree/Rwork
> values by about 2 points. From what I've read previously I thought refining
> hydrogens in a non-"riding" model was only valid below ~1.0 Å.
>

I'm not sure what you mean here.  Did you use the riding model and it
increased R-free?  Or only when you let them refine individually?  The
default is now to switch to individual only at 0.9Å or less.


> by a few points. For water, I assume that at this resolution there is a
> combination of anisotropic (big banana looking density) and isotropic
> waters
> and that one cannot simply rely entirely on the water_update option.
>

Yes, at some point you will need to complete the waters yourself.  Those
bananas may not be simply anisotropic waters, but split sites.  You may see
other small ligands too.

        The twin law has worked great, usually dropping Rfree/Rwork values
> by 5 points or so.
>

Just remember to be careful when interpreting the effect of twin law
application - I think in this case it's probably quite reasonable, but at
lower resolution and with a worse model it can be extremely deceptive.

        I apologize if any of this is common knowledge and shows my relative
> inexperience when it comes to best practices for crystal refinement.
>

No need to apologize - it is all oral tradition anyways.  It is much better
to ask questions now than to do something blindly and regret it later.

-Nat
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20131011/7683755b/attachment-0001.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: parameterization_of_refinement.pdf
Type: application/pdf
Size: 103252 bytes
Desc: not available
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20131011/7683755b/attachment-0001.pdf>