On Sunday, 17 March 2013, Pavel Afonine wrote:
> Hi Ethan,
>
>
> I would place the expected resolution break-even point at more like
> > 1.2 - 1.3 A. But that's only an expectation, not a rule to rely on.
> > You should justify anisotropic refinement of a structure on the basis
> > of its own particular model and measured data. Robbie Joosten has
> > already pointed out that you can use the PDB-Redo scripts to test
> > whether individual anisotropic ADPs are justified.
> >
>
> when it comes to a point when a choice of refinement strategy cannot be
> uniquely and reliably chosen based on theoretical considerations, that
> opens a great opportunity for endless perennial discussions like this. My
> point was that if you are lucky and there isn't many options (in this case
> there are only two: iso vs aniso!) there is an easier, quicker and robust
> alternative: simply try both and that will give your THE answer. Perhaps
> not very scientific (no monster formula derived) but quick, easy and robust!
>
>
> > > - higher than 1.2A: all anisotropic (macromolelcule, water, ligands)
> > > - lower than 1.7A: all isotropic;
> > > - 1.5-1.7A is a grey area where there is only one single way to know for
> > > sure: try both (isotropic and anisotropic) and see which one works best.
> > > I realize "works best" is a broad term, but I would say Rwork, Rfree,
> > > Rfree-Rwork and values of refined anisotropic ADPs should be enough to
> > make
> > > a decision.
> >
> > Unfortunately, Rfree cannot be used reliably for this purpose.
> >
>
> Yes, Rfree cannot be used for this purpose reliably, very true. This is
> exactly why I wrote above ".. Rwork, Rfree, Rfree-Rwork *and values of
> refined anisotropic ADPs* should be enough to make a decision." My
> understanding is that your "To B.." method is one of possible ways of
> looking at the values of refined anisotropic B-factors.
That is a misunderstanding that misses the fundamental point.
Looking at the refined ADPs only helps in the artificial case that
you have a known-correct set of ADPs as a point of comparison for the
newly refined ADPS [*].
You cannot do that in the case of a new structure.
The Hamilton R test does not look at the individual ADPs or at any other
refined parameter values. It looks only at the crystallographic residuals
and the respective degrees of freedom. In this sense it is analagous
to comparing Rfree values, but as demonstrated in the paper the Hamilton
test can correctly detect an invalid model in cases where Rfree cannot.
Ethan
[*] Of course if the new ADPs are inherently unreasonable, for instance
non-positive definite, that is good reason to reject the model.
But I am assuming that a properly behaving refinement program will not
produce such an inherently unbelievable model. Instead the outcome
of refinement will be internally consistent, plausible, but wrong.
You won't learn that just by inspecting the ADPs it produces.
> All the best,
> Pavel
>
