[phenixbb] Implementation of French & Wilson data correction?
yamashita at castor.sci.hokudai.ac.jp
Tue Mar 1 16:45:12 PST 2011
Dear Randy and Pavel,
It's amazing to me that MLI target test case gave significantly better
reesults because I thought contributions of weak intensities to
refinement are very small.
I wonder when does MLI work very well -- e.g. in case there're many
negative intensities (measured with much noise)?
And, Did the test case where MLI gave significantly better results
give also better electron density map?
I understand MLI can give more accurate model but there're still some
problems such like I->F conversion etc..
I will read the paper introduced by Pavel -- thanks!
2011/3/1 Randy Read <rjr27 at cam.ac.uk>:
> Hi Pavel,
> Actually, when we defined the MLI target (originally called MLF2: Pannu & Read, 1996), we did two tests where we compared R-factors (computed with the same set of F values) and phase errors (measured with the mean cosine of the phase error). In one case (Streptomyces griseus trypsin, where the bovine trypsin starting model was a very poor model) the difference was very small, but in the other case (Trypanosoma brucei GAPDH) the MLI target gave significantly better results.
> I'm not sure why you say that you couldn't use R-factors (though I prefer phase errors or map correlations). You could argue that R-factors computed using F should be biased to give better results to a target based on F than on I, but that wasn't a problem in our tests.
> Clearly there's room for more exhaustive tests, but that initial result demonstrated the theoretical advantage mentioned by Keitaro (accounting for negative net intensities) plus the advantage that it's a better approximation to an ideal likelihood target (measurement errors are closer to Gaussian in measured intensities than to Gaussian in amplitudes -- assumed in our MLF target -- or complex Gaussian in complex structure factors -- assumed in the Rice likelihood target). It would also be good to bring the Rice likelihood target into the tests.
> On 1 Mar 2011, at 01:31, Pavel Afonine wrote:
>> Hi Keitaro,
>>> A little off-topic, but will phenix implement MLI (maximum likelihood
>>> intensity target)?
>> eventually. The formulas are there, someone just needs to get enough of motivation to spend a couple of days and code them thus making available in phenix.refine.
>> Once this is done, then there will be another set of troubles to deal with:
>> - what to use to compute R-factors: Iobs or Fobs? If Fobs, then in turn you will go back to the original problem of converting I to F. If Iobs, then I bet a number of people will scream that they want to see the R-factors computed the "usual way" (using Fobs, that is).
>> - Same questions about computing the maps.
>>> I heard that MLI target gives a good result because MLI handles
>>> negative intensities naturally.
>> "Good" compared to what? How you evaluate that one is better than the other in this case? Obviosuly, you wouldn't be able to use the R-fators as the criteria. I guess I know how, but I'm not aware that anyone has ever done it (another little project for someone enthusiastic).
>> Overall, I conclude quoting one of my favorite papers:
>> "The choice of reﬁnement against F**2 or F has generated more discussions than it probably warrants. (...)"
>> David Watkin
>> "Structure refinement: some background theory and
>> practical strategies"
>> J. Appl. Cryst. (2008). 41, 491–522
>> All the best!
>> phenixbb mailing list
>> phenixbb at phenix-online.org
> Randy J. Read
> Department of Haematology, University of Cambridge
> Cambridge Institute for Medical Research Tel: + 44 1223 336500
> Wellcome Trust/MRC Building Fax: + 44 1223 336827
> Hills Road E-mail: rjr27 at cam.ac.uk
> Cambridge CB2 0XY, U.K. www-structmed.cimr.cam.ac.uk
> phenixbb mailing list
> phenixbb at phenix-online.org
More information about the phenixbb