Print

Print
Hi Alastair,

since you mentioned it... In our article "On the analysis of residual density distribution on an absolute scale":
http://www.phenix-online.org/newsletter/CCN_2012_07.pdf
one of the conclusions was that we could not reproduce pronounced features on the solvent/macromolecule border shown in Matthews '09 paper.
If you look at Figure 2 the red curve is almost flat all the way and it is around zero. One thing to keep in mind that these are averaged distributions without specifically considering surface vs buried atoms. Individual per atom curves may be different.

Pavel



On Mon, Sep 9, 2013 at 11:30 AM, Alastair Fyfe <[log in to unmask]> wrote:
thanks for the reference to the script and additional discussion. I've looked through the archives a bit but couldn't find an answer to a question that's been on my mind for a while  so my apologies if this revisits  well-trod ground. One of the potential sources of disagreement contributing to the gap could be poor modeling of scattering by the region between "ordered" and "bulk" solvent . That is, the abrupt transition from point scatterers to bulk may not adequately model regions with a greater incidence of transiently occupied scattering sites. Are there any pointers to cites/software that investigates modeling a layer of semi-structured solvent, for example as a function of distance from a molecular surface "colored" by  its hydrogen-bonding potential?

Looking at the magnitude of residual density  as a function of distance from the molecular surface (Afonine, Urzhumtsev, Adams '12, Matthews '09) seems to point to a  possible misfit in that region and some calculations I've been doing using real space correlation give similar results. Deletion of waters with poor model metrics (correlation, number of neighbors, etc.) can improve Rwork while increasing Rfree suggesting that the extra scattering is contributing meaningfully, even if poorly modeled. Softening the ordered/bulk boundary with a differentiable transition (Fenn,Schnieders,Brunger'10)  doesn't address this though their concluding discussion seems to suggest it's worth investigating. The question has been examined in the SAXS literature (Virtanen, Makowski, Sosnick, Freed '11) but I haven't found found equivalent experiments  among refinement software.


On 09/07/2013 04:54 AM, James Holton wrote:

I feel like I should point out that there is about a 20% difference between "Fcalc" and something I would call a "simulated Fobs".  Fcalc is something that refinement programs compute many times every second as they apply 100 years worth of brilliant ideas to make your model (Fcalc) match your data (Fobs) as best we know how.  Despite all this, one of the great mysteries of macromolecular structure determination is just how awful the "final" match is: R/Rfree in the 20%s or high teens at best. Small molecule structures don't have this problem.  In fact, they only recently started depositing "Fobs" in to the CSD because for the most small molecule structures "Fcalc" is more accurate than "Fobs" anyway.

This has been hashed over on this BB a number of times, so I refer the interested reader to the archives.  But there are two major considerations in turning a "pdb file" into a "simulated Fobs":
1) the solvent
    SFALL (part of the CCP4 suite) is a convenient tool for turning coordinates into maps, or structure factors, but it doesn't "do" bulk solvent unless you trick it.  I wrote a jiffy for doing this here:
http://bl831.als.lbl.gov/~jamesh/mlfsom/ano_sfall.com
download the script, make it executable, and run it with no arguments to see instructions for how to use it.  What is fascinating about this very crude bulk solvent implementation I did is that refinement programs with much more sophisticated bulk solvent implementations have a heck of a time trying to "match" it.  If you want exactly the bulk solvent you would get from phenix, use phenix.fmodel, but this will not be exactly the same as the bulk solvent you get from REFMAC.  Which one is right? Probably none of them.

2) The R-factor Gap
   One can try to simulate the R-factor gap (between Rmeas and Rfree) by adding random numbers to "Fcalc" so that it becomes 20% different from Fobs, but this is hardly a physically reasonable source of error.  If you do this enough times for the same PDB file and then "average over different crystals" you'll still end up with a dataset that will refine to R/Rfree ~ 0/0.

This is the fundamental problem with making "simulated Fobs": we actually have no good way of "modelling" whatever is causing this R-factor Gap, and therefore no good way of simulating it.  If we could simulate it, then some refinement program would quickly implement a way to model the effect, and give you R/Rfree of 0% again.  There are about as many ideas for the cause of the R-factor Gap as there are crystallographers out there, but to this day nobody has come up with a "systematic error" that, when accounted for in refinement, gives you a small-molecule-style R/Rfree for pretty much anything in the PDB.  Not even lysozyme.

-James Holton
MAD Scientist


On 9/5/2013 9:35 AM, Alastair Fyfe wrote:
Below are some links to tools for simulating Fobs data:

phenix.fake_f_obs: http://cci.lbl.gov/cctbx_sources/mmtbx/command_line/fake_f_obs.py
phenix.fmodel: http://cci.lbl.gov/cctbx_sources/mmtbx/command_line/fmodel.py
sftools (calc keyword):  http://www.ccp4.ac.uk/html/sftools.html

diffraction image simulators from James Holton
mlfsom: http://bl831.als.lbl.gov/~jamesh/mlfsom/
nearBragg: http://bl831.als.lbl.gov/~jamesh/nearBragg/
fastBragg: http://bl831.als.lbl.gov/~jamesh/fastBragg/

many thanks for the replies.
Alastair