[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: refmac for partial p-ala trace / refinements
*** For details on how to be removed from this list visit the ***
*** CCP4 home page http://www.dl.ac.uk/CCP/CCP4/main.html ***
again, as for a number of previuos cases, I would agree with Dirk's
Kostrewa comments and would even emphasise his scepticism. I think it is
worthy to understand what you are doing when you are solving a structure
and I would like to give some comments for young researchers.
There are no miracles, specially in optimisation procedures - refinement
is a praticular case of it. As in any optimisation, you have :
a) a criterion to be optimised and
b) the procedure to search for the (GLOBAL) minimum - if we believe that
this minimum corresponds to the right model.
There is NO method for global minimum search for an ARBITRARY function
(except a systematic checking of its values for ALL possible values of its
argument). Crystallographers use now a mixture of random methods with local
minimisation (gradient based) which work in general well for OUR PARTICULAR
PROBLEMS but they CANNOT garatee to find a global minimum when starting
from a quite wrong (or POOR !! or INCOMPLETE !!) model. But - we do not
have a better procedure for a time being.
ML is a minimisation CRITERION, and NOT a minimisation method. One can
compare it with the least-square criterion. It works in the same way except
that instead of fitting calculated magnitudes to the observed one it
suggests to fit them to some MODIFIED VALUES, and these modifications are
extimated through the model quality and completeness (see, for example,
explanations in the last issue of the CCP4 Newsletters). If the model is
incomplete, it is wrong to fit it to experimental amplitudes (as it is the
case in the LS refinement) and ML knows how to introduce such corrections.
Such corrections are done in resolution shells. If your model does not fit
at all your data at a given resolution shell, the ML puts the "corrected
experimental magnitudes" in this shell equal to zero. IF in this case they
still participate in optimisation, this is even dangerous to use these
reflections. IF in this case they are simply removed from the refinement
procedure, this means exactly "old good scheme" of refinement when one
increases the resolution slowly. (I do not know exactly which of these
alternatives is programmed.)
Including of all data is good when your have a high-resolution data set
(which rather mean that your crystal is good, your density is good, and
your model is supposed to be quite good and complete originally - without
or with some dummy-atoms as it in ARP).