On 06/14/2013 07:00 AM, John R Helliwell wrote:
> Alternatively, at poorer resolutions than that, you can monitor if the
> Cruickshank-Blow Diffraction Precision Index (DPI) improves or not as
> more data are steadily added to your model refinements.
Dear John,
unfortunately the behavior of DPIfree is less than satisfactory here -
in a couple of cases I looked at it just steadily improves with
resolution. Example I have in front of me right now takes resolution
down from 2.0A to 1.55A, and DPIfree goes down from ~0.17A to 0.09A at
almost constant pace (slows down from 0.021 A/0.1A to 0.017 A/0.1A
around 1.75A).
Notice that in this specific case I/sigI at 1.55A is ~0.4 and
CC(1/2)~0.012 (even this non-repentant big-endian couldn't argue there
is good signal there).
DPIfree is essentially proportional to Rfree * d^(2.5) (this is
assuming that No~1/d^3, Na and completeness do not change). To keep up
with resolution changes, Rfree would have to go up ~1.9 times, and
obviously that is not going to happen no matter how much weak data I
throw in.
The maximum-likelihood e.s.u. reported by Refmac makes more sense in
this particular case as it clearly slows down big time around 1.77A (see
https://plus.google.com/photos/113111298819619451614/albums/5889708830403779217).
Coincidentally, Rfree also starts going up rapidly around the same
resolution. If anyone is curious what's I/sigI is at the "breaking
point" it's ~1.5 and CC(1/2)~0.6. And to bash Rmerge a little more,
it's 112%.
So there are two questions I am very much interested in here.
a) Why is DPIfree so bad at this? Can we even believe it given it's
erratic behavior in this scenario?
b) I would normally set up a simple data mining project to see how
common this ML_esu behavior is, but there is no easily accessible source
of data processed to beyond I/sigI=2, let alone I/sigI=1 (are structural
genomics folks reading this and do they maybe have such data to mine?).
I can look into all of my own datasets, but that would be a biased
selection of several crystal forms. Perhaps others have looked into
this too, and what are your observations? Or maybe you have a dataset
processed way beyond I/sigI=1 and are willing to either share it with me
together with a final model or run refinement at a bunch of different
resolutions and report the result (I can provide bash scripts as needed).
Cheers,
Ed.
--
Oh, suddenly throwing a giraffe into a volcano to make water is crazy?
Julian, King of Lemurs
