[phenixbb] phenix.refine 1.3 vs. 1.4 map quality ?
PAfonine at lbl.gov
Thu Aug 6 08:11:57 PDT 2009
> About a year ago I refined a 1.6 Å RNA structure with phenix.refine
> 1.3b-rc6 and then got distracted.
I'm not sure I got the message... Was anything wrong about the maps? If
so, did you let us know (sorry, I don't remember).
> I picked it up today and
> essentially repeated the last round of refinement with 1.4-57 (and
> also 1.4-159). The maps are subtly different, but consistently
> slightly worse with 1.4.
Which maps we are looking at? I presume 2mFo-DFc (not 2Fo-Fc): regular?
filled? kicked? (see copies of previous emails to the BB explaining
these maps at the bottom of this email).
> Here are two examples: http://sage.ucsc.edu/~wgscott/mystuff/old_vs_new.pdf
May be because I'm looking at static pictures, but apart from the water
#5, the differences seem subtle indeed.
> In the second example, this would lead to deletion of one of the
> octahedrally coordinated waters on a known Mg++ ion.
Do you mean you would delete this water manually because you don't see
it in the map, or phenix.refine deletes it?
> Has the weighting for the bulk solvent mask or something like that
> been increased?
I can name two major things:
- phenix.refine outputs "filled" maps by default since December 2008 (or
may be January 2009?);
- a few weeks (a month?) ago we switched to using a faster and more
memory efficient bulk solvent mask calculation code. Plus this new code
has a minor bug fixed that existed many years in the old one.
If you send me 1.6A resolution data and your best model, I will be happy
to have a closer look.
PS> Copies of replies to BB about kicked and filled maps:
there will be a paper about it in next coming Acta D. This method was
introduced about 10-15 years ago by Dusan Turk in his program MAIN is
used since that. Here is the copy-paste from the manuscript:
"(...) An average kick map (AK map) is computed as following (Gunc(ar et
al., 2000; Turk, 2007; Praz(nikar et al., 2009): a large ensemble of
structures (several hundreds) is created where the coordinates of each
structure from the ensemble are all randomly shaken. The shake amount
(rmsd distortion introduced to coordinates) varies from 0 to 1.0 Å. Then
for each structure a map is computed ((mFobs-DFmodel)exp(i?model) or
(2mFobs-DFmodel)exp(i?model) or any other map, for example a ligand-omit
map). Finally, all maps are averaged out to produce one averaged kick
map. An AK map is expected to have less or no bias, less noise, enhance
existing signal and potentially can clear up some initially bad
I think it's time to review this subject since it is one of the most
frequently asked questions. Ironically, I think this is because we tried
to make it as clear as possible -:)
So, what kind of maps phenix.refine outputs by default? phenix.refine
outputs two types of maximum-likelihood weighted maps (or, in other
words, sigmaa-weighted maps): 2mFo-DFc and mFo-DFc.
Now, as many of you noticed, the MTZ file with map coefficients
"_map_coeffs.mtz" contains in fact four maps: 2mFo-DFc and mFo-DFc, and
"filled" 2mFo-DFc and mFo-DFc (in fact fo-fc maps should be nearly
identical; actually I have to fix it and not write identical fo-fc
maps). The first two maps are computed using original Fobs (Fo), and
the last two maps are computed using "filled Fobs", that is the original
Fobs where missing reflections are "filled" with DFc. It is well known
(I can spell a long list or references) that the data incompleteness
affects the map quality, and sometimes, certain types of data
incompleteness can *severely* distort maps.
A possible solution (in order to reduce this negative effect) is to
"model" missing Fobs somehow. One possibility is just to put in DFc in
place where Fobs is missing, or as suggested by the classics, one can
use <Fobs> taken in a resolution bin around a missing reflection. I even
tried to use the random numbers and it was also better than doing
nothing. Obviously, there is a nearly invisible line between the
benefits of "filling in" missing Fobs and introducing bias. Where this
line goes - is the subject of a research that to my knowledge is not
Anyway, this is why phenix.refine writes out "regular" and "filled"
maps: one is to give you unbiased but eventually lower quality map, and
the other one is to give you a better-looking map with a risk of being
biased. This way users have more options in exploring their maps (and
less reasons for saying that Refmac produces better-appearing maps than
phenix.refine -- see below).
I have to mention that to my knowledge REFMAC always writes "filled"
maps (those with missing Fobs substituted by DFc):
- it is mentioned in Maria Turkenburg's thesis:
- and in Refmac docs:
"Missing Data: For those reflections where the FP are missing, mFo is
set equal to dFc. (...)".
Correct me if I'm missing something.
Please let me know if I wasn't clear in my attempt to explain this. I
will update the phenix.refine documentation accordingly.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the phenixbb