The indexing ambiguities do not include anomalous pair confusion because
there is no way to rotate the lattice to make every h,k,l overlap with
-h,-k,-l. I.E. you can't rotate your left hand to superimpose it on
your right. The only way to mix those up is to change the sign of some
detector geometry parameter (I.E. looking in a mirror).
That said, anomalous differences tend to be very weak and noisy in all
but the most exotic cases of macromolecular diffraction. Twinning makes
this worse because you are (to a first order approximation) averaging
DANO(h,k,l) with DANO(k,h,-l) and the result will tend to be closer to
zero than either one taken individually. However, the biggest source of
error in LCLS datasets at the moment is partiality. Basically, you only
get one shot per crystal, you can't rotate it appreciably in the 70 fs
exposure time, the beam is a laser so there is essentially no divergence
or dispersion, and the crystals are so small as to be one mosaic domain
each, so there is no "mosaic spread". The "3D profile" of the spots is
therefore dominated by the finite size of the crystal itself (Sherrer
broadening). We were actually worried for a while that we wouldn't see
any spots at all at LCLS!
So, everything is a partial, and we currently don't have postrefinement
software that can model the shape of each crystal and give us a
partiality. At least, not in a reasonable amount of time. If we spent
30 s on each of the 3 million images, we would still be processing them
for a few more years. So, for the first run, it was decided to jut
average out the partiality errors. For example, unknown partiality
means that each spot is measured with 100% error (at best), but if you
have 700 of them, then the expected error of the average is ~3%. John
Spence called this a "Monte Carlo integration", and it turned out to be
a really good idea. We measured the error of the average by splitting
the images into two heaps and comparing the merged datasets that
resulted from each heap. I proposed calling this "R-internal" for
internal agreement, since a traditional Rmerge does not really apply.
However, I admit that for the PDB deposition I entered R-internal as
"Rmerge". Technically, R-internal is exactly what an Rmerge used to be:
the R-factor between data from different crystals.
Personally, I think "the way" to crack this "twin problem" is to scale
all the data and look at the partial intensity histograms for each
spot. In situations where the "true" values of h,k,l and k,h,-l have
radically different intensities, there will be a bimodal distribution,
and that will allow us to re-index the ~700 images that contained a spot
from one of those two hkls. Which group to flip (the bright ones or the
dim ones) is an interesting question, but probably the dim ones, since
they are the least consistent with the average intensity. Might need to
try both. After re-mergeing and re-scaling, there will be another hkl
with the strongest bimodal distribution, and then you iterate. That's
the idea anyway.
-James Holton
MAD Scientist
On 2/10/2011 6:32 AM, Jacob Keller wrote:
Would it be true that the anomalous differences could not be measured
in these types of datasets, because one would not know which
Friedel/Bivoet reflection one is measuring in a given frame? Perhaps,
given anomalous signal, there would be a way to tease out which
orientation one was looking at from the correlations of the
signs/magnitudes of anomalous-scattering-induced deviations from the
mean intensities (derived from the whole dataset) for all of the
relections observed in each frame? I guess this might also detwin the
data?
JPK
On Thu, Feb 10, 2011 at 7:17 AM, Anastassis Perrakis<a.perra...@nki.nl> wrote:
Anyway, I thought that was a cool idea, but like so many other cool
things, it had to be cut from the Nature paper. Admittedly, the problem has
not actually been solved yet. This is why we used REFMAC in TWIN mode.
Is that a hint on the:
a. wisdom of the editor
b. wisdom of 'the third referee'
c. wisdom of the dogma 'five years of eight eight lifes in 2000 words'
d. All of the above
;-)
A.