[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[ccp4bb]: Refmac vs. cns



***  For details on how to be removed from this list visit the  ***
***          CCP4 home page http://www.ccp4.ac.uk         ***

	Since there seems to be little defense of CNS in this debate, I
thought I'd give my two cents worth.  First, there is no doubt that both
ARP/wARP and REFMAC are superb programs.  However, many statements in this
debate are of questionable validity. First, CNS (and earlier, XPLOR) were
developed to address difficult refinement problems at low to moderate
resolution (i.e. when data are very limited).  So, to claim that simulated
annealing is best suited for higher resolution refinements is untrue and
misses the declared purpose of the approach-to minimize the probability of
a poor initial model being trapped in a local minimum when there are not
enough data to guide the refinement to the global minimum.  This does not
mean that you will never have to manually correct the resulting model, or
to adjust the starting temperature for the refinement. It does mean that
you don't have to rely solely on manual rebuilding to correct potentially
major errors in an initial model that has been built into poor maps. The
value of simulated annealing for this purpose has been extensively studied
and documented in the literature.  In addition, the combination of torsion
angle refinement and simulated annealing optimization reduces the
dimensionality of the search space approximately 10-fold, vastly
decreasing the risk of overfitting by eliminating bond angles and lengths
as refineable parameters.  If you have data in the 3.5 to 2.5 A range, you
have no hope of fitting these model parameters meaningfully, no matter how
tightly they are restrained, and it is proper refinement practice to not
fit them at all.
        Secondly, the Free R was not introduced as a stopgap measure to
correct for problems in simulated annealing refinement.  It was introduced
because reporting a conventional R value as an indicator of model quality
is at best naive and at worst willfully deceptive. Again, this has been
extensively documented in the literature.
        In summary, a direct comparison of REFMAC-ARP/wARP and CNS is
complex, and should include a more careful consideration of the type of
refinement being performed, the amount of data available, and the
reliability of the initial model. A collection of anecdotes about how
these programs compare in a handful of refinements does not add
meaningfully to a proper assessment of their worth.
                             Sincerely,
                            Mark Wilson