[buster-discuss] Calculating the R values for a model at differing resolution cutoffs

[det102 at uoxray.uoregon.edu]

    Since Karplus and Diederichs is now in print I can be a little more
explicit in what I'm trying to do.  I have a data set that would have
been cut at 2.6 A resolution by traditional criteria but the CC* says
that there is information present out to 2.25 A.  I want to see if a
model refined out to 2.25 A will fit the data to 2.6 A better than a
model refined just to the 2.6 A data.

    I understand your description of the issue of the error model in
the likelihood calculations but for this calculation I'm perfectly
happy to use the model which was derived using the entire data set
since I am comparing to the same data set, just a subset of it.

    I've always viewed the imperfection parameters (for lack of a better
term) as part of the model and I would prefer that they be written to
the coordinate file along the the other parameters of the model.  If
this were done they would not have to be recalculated from scratch every
time and life would be simpler for us folk who want to try new ideas
without hacking into your code.

    I am on vacation at the moment and not able to dig deeply into it,
but if refine.mtz includes <Fcalc> and Sigma(Fcalc) I could probably
calculate my R values directly from them.  I have a personal aversion
to unformatted files in general and mtz files in particular, but I could
probably overcome that if need be.

Dale Tronrud


On 6/6/2012 7:24 AM, Gerard Bricogne wrote:
> Dear Dale and other BUSTER users,
>
>       I am sorry to have left this question unanswered for so long.
>
>       The answer is "No", because the R-factors calculated by BUSTER involve
> the expectation value of the full "calculated" structure factor under the
> same error model that is used to define the likelihood function. This
> requires the maximum-likelihood estimation of the Luzzati imperfection
> B-factor for the atomic model, and hence an initial ML-scaling step that is
> affected by the scope of the experimental data that are used. If you carry
> out that step with only low-resolution data, there is no guarantee that you
> will get exactly the same result as if you had used higher-resolution data,
> then had restricted your attention to a low-resolution subset of those data.
>
>       You can compare that, roughly, to what happens with Wilson scaling: it
> works poorly and unpredictably from only low-resolution data, but much
> better if you use high-resolution data then restrict the results to their
> low-resolution portion.
>
>       I hope this is not too complicated, and not too disappointing.
> Comparing things that are intrinsically different always involves using a
> statistical model for how they differ (an "error model"), and hence some
> estimation of scaling parameters that will depend on those of that error
> model.
>
>
>       With best wishes,
>
>            Gerard.
>
> --
> On Wed, May 16, 2012 at 12:23:11PM -0700, Dale Tronrud wrote:
>> Hi,
>>
>>     I am performing refinement with data to one resolution but would like to
>> know what the R values are when calculated with only a lower resolution subset
>> of the data.  Is there an easy way (I'm lazy) to just calculate the R values
>> for an existing model with whatever resolution cutoffs I choose?
>>
>> Dale Tronrud
>> _______________________________________________
>> buster-discuss mailing list
>> buster-discuss at globalphasing.com
>> https://www.globalphasing.com/mailman/listinfo/buster-discuss
>