[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
[ccp4bb]: We are having problems with newer versions of refmac5
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.ccp4.ac.uk ***
Hi all
This message is two things.
It is a plea for suggestions/insight/clues on what we might be doing
wrong in using newer refmac versions.
It is also a heads-up notice that at least in our hands refmac 5.1.x
is not always refining as well as the earlier version 5.0.36 (the one
included in CCP4 4.1.1). We see this for all newer versions we have tried
up to and including 5.1.80. I am concerned that other groups may be
suffering from similar problems, but may not have noticed.
We first encountered this problem when trying to reproduce refinement
runs of a pair of high resolution (~1A) structures. The original refinements
had been done in parallel in shelxl and refmac 5.0.36 using full aniso
treatment. The shelxl and refmac runs agreed well with each other.
The same structures fail to refine as well in refmac 5.1.x.
In our problem cases, the final model after refinement in 5.1.x
looks OK on its own. But when the restraint weights are set to yield
acceptable measures of quality (DPI, geometric residuals, anisotropy)
the R and Rfree residuals are noticeably poorer than those for the final
model from version 5.0.36. If we had not set out to
replicate an earlier refinement we might never have noticed something
was wrong.
We have tried other, simpler, test cases which do refine OK in both the
new and the old refmac5, so the problem is not universal.
But we have not been able to pin down what it is about the failing cases
that is making them fail.
General summary
------------------
Running the exact input scripts, mtz files, and starting models originally
used with refmac 5.0.36 through newer refmac5 versions causes the refinement
to blow up (atoms lost, residuals diverge, floating exceptions, all-around
bad behaviour). Some exploration of the code and new documentation clued us
in that the interpretation of weighting parameters has changed in the new
versions. We therefore systematically searched for revised weighting
parameters that would yield behaviour equivalent to our original nice runs.
At the same time we sought for other possible sources of error, either in
our input or in the newer program source code. Unfortunately after several
weeks of looking we have not been able to find any set of weighting
parameters, refinement options, or input *.cif library files that yield
results from version 5.1.x quite as nice as the ones from 5.0.36.
We have confirmed that the difference between versions 5.0.36 and 5.1.x
persists across tests
- under linux (g77 or ifc compilers) or alpha (DU4.0D)
- with or without generation of hydrogen atoms
- aniso or iso refinement of the model
- inclusion or exclusion of all ligand atoms other than metals
- varying the Shannon sampling interval for the FFT grid
The 5.1.x versions are much more sensitive to weighting parameters than
5.0.36 was, especially MATRix. This sensitivity is somewhat less on the
[64-bit] alpha than on linux/x86, which suggests to me that maybe there
is an ill-conditioned matrix somewhere.
Details
-------
Below I give a summary of 1A refinement runs of Carbonic Anhydrase I
as reported by refmac 5.0.36 and by the "best" run we were able to
obtain from version 5.1.80. The input model and *.mtz files were
identical. Both these runs included generation of hydrogens and a fully
anisotropic model.
Difference in final R values
version R Rfree
5.0.36 0.1362 0.1502
5.1.80 0.1450 0.1563
RMS Difference in final coordinates
529 CA atoms 0.0198A
2511 all atoms 0.0624A [>2x the ESU from Cruickshank DPI]
Differences we do not understand, which may or may not contribute to the
observed difference in refinement results:
- the count of VDW terms is radically different in the new version
- the count of "free" atoms is also radically different
- HBOND and period 3 torsion terms are no longer reported
- despite our best efforts to find appropriate weighting parameters,
the B restraints terms and gradient do not match very well between
the new and old runs
- notice that the ratio of X-ray gradient to Geometric gradient is
roughly 2:1 in the 'old' run and roughly 1:1 in the 'new' run.
One would think that increasing the MATRix weight in the new run
to up-weight the X-ray term would improve R at the cost of the resulting
geometry. It doesn't. An otherwise identical run of 5.1.80 with only
the MATRix weight changed to balance the two gradients yields both
worse R and Rfree and worse geometry.
- 5.1.x seems to have convergence problems. In many (most?) runs,
R and Rfree reach a minimum after 3-10 cycles, and then get worse again
before converging to something higher. By contrast, 5.0.36 converges
to the minimum and is stable there for at least the 25 cycles over
which we tracked it.
- 5.1.x reports a very small scale factor for the Babinet solvent correction
(0 in the example below; near 0 in most other runs). Does this mean that
it essentially not making a solvent correction? That right there might
explain the difference. Why would the Babinet correction be so different
between the new and old versions of refmac?
- The reported "Partial structure 1" scales are also different.
What does this refer to?
- The reported Overall figure of merit is better in 5.1.x.
How exactly is this quantity calculated?
****** REFMAC 5.0.36 ****** ****** REFMAC 5.1.80 ******
---------------------------------------------------------------------------- ------------------------------------
Restraint type N restraints Rms Delta Av(Sigma) N restraints Rms Delta Av(Sigma)
Bond distances: refined atoms 2233 0.014 0.021 2233 0.013 0.021
Bond distances: others 1942 0.001 0.020 1942 0.005 0.020
Bond angles : refined atoms 3024 2.190 1.966 3024 2.280 1.966
Bond angles : others 4590 1.059 3.000 4590 2.444 3.000
Torsion angles, period 1. refined 258 6.011 3.000 258 6.797 5.000
Torsion angles, period 3. refined 417 19.851 15.000
Chiral centres: refined atoms 317 0.125 0.200 314 0.113 0.200
Planar groups: refined atoms 2415 0.009 0.020 2415 0.008 0.020
Planar groups: others 434 0.004 0.020 434 0.004 0.020
VDW repulsions: refined atoms 489 0.374 0.300 936 0.275 0.200
VDW repulsions: others 2009 0.245 0.300 6658 0.426 0.200
VDW; torsion: refined atoms 4 0.724 0.500 640 0.108 0.200
VDW; torsion: others 5 0.984 0.500 4483 0.283 0.200
HBOND: refined atoms 273 0.172 0.500
HBOND: others 1 0.158 0.500
Metal-ion: refined atoms 6 0.401 0.500 6 0.566 0.200
VDW repulsions; symmetry: refined atoms 17 0.257 0.300 114 0.315 0.200
VDW repulsions; symmetry: others 64 0.254 0.300 276 0.374 0.200
HBOND; symmetry: refined atoms 37 0.256 0.500
M. chain bond B values: refined atoms 1297 1.785 1.000 1297 1.218 1.000
M. chain bond B values: others 522 1.650 1.000 522 1.164 1.000
M. chain angle B values: refined atoms 2110 2.698 1.500 2110 1.883 1.500
M. chain angle B values: others 1794 2.557 1.500 1794 1.804 1.500
S. chain bond B values: refined atoms 936 3.333 2.000 936 2.236 2.000
S. chain bond B values: others 1420 1.859 2.000 1420 1.273 2.000
S. chain angle B values: refined atoms 915 4.737 3.000 915 3.262 3.000
S. chain angle B values: others 2796 2.820 3.000 2796 1.953 3.000
Sphericity. Free atoms 1 0.833 8.000 337 11.448 8.000
Sphericity. Bonded atoms 2138 6.441 8.000 2174 4.558 8.000
Rigid bond restraints 2198 0.773 0.500 2233 0.517 0.500
---------------------------------------------------------------------------- ------------------------------------
Resolution limits = 15.000 1.000 15.000 1.000
Number of used reflections = 121649 121649
Percentage observed = 98.2038 98.2038
Percentage of free reflections = 3.9684 3.9684
Overall R factor = 0.1362 0.1450
Free R factor = 0.1502 0.1563
Overall weighted R factor = 0.1243 0.1308
Free weighted R factor = 0.1378 0.1443
Overall correlation coefficient = 0.9752 0.9694
Free correlation coefficient = 0.9694 0.9628
Cruickshanks DPI for coordinate error= 0.0269 0.0262
Overall figure of merit = 0.9179 0.9336
---------------------------------------------------------------------------- ------------------------------------
Norm of X_ray positional gradient 44.1 28.4
Norm of Geom. positional gradient 27.2 31.1
Norm of X_ray B-factor gradient 0.171E+04 898.
Norm of Geom. B-factor gradient 292. 300.
Product of X_ray and Geom posit. gradients -0.930E+07 -0.403E+07
Cosine of angle between them -0.581 -0.341
Product of X_ray and Geom B-fact gradients 0.258E+08 0.227E+08
Cosine of angle between them 0.003 0.005
---------------------------------------------------------------------------- ------------------------------------
Overall : scale = 1.079, B = 0.014 scale = 1.071, B = 0.003
Babinet"s bulk solvent: scale = 0.112, B = 174.346 scale = 0.000, B = 196.718
Partial structure 1: scale = 1.985, B = 31.170 scale = 0.388, B = 128.847
Overall anisotropic scale factors
B11 = 0.10 B22 = 0.12 B33 = -0.27 B12 = 0.00 B13 = -0.11 B23 = 0.00 (identical)
yours in puzzlement,
Ethan
--
Ethan A Merritt merritt@u.washington.edu
Biomolecular Structure Center Box 357742
University of Washington, Seattle, WA 98195