Thanks, I do understand all of that. I gave some Rmerge and resolution
values to give some idea about errors and noise expected in the data,
and an idea for up to what resolution phases would be good. And if such
low resolution phases ever yield a meaningful model. Both measures are
flawed indicators, even though they are the most common measures of data
among us. I will definitely check Ranom (which means I should try scala).
What I was curious about is practical aspects: especially in cases in
which it really worked. And I/we have gotten quite a few responses in
MAD vs SAD, inverse beam strategies, radiation damage control, etc. The
take home message for me was that noone agrees on the best data
collection strategy, although I still have to read upon some of the case
references that were sent. Another point is the success rate of
software - be it direct methods based or Patterson based - with such
data (where anomalous signal would die at even lower resolution) at
solving the substructure. I have seen a reference where if the correct
substructure could be provided in a test case, SAD was actually
successful. In another case, they confirmed the correct selenium sites
with a platinum derivative data to further proceed with phasing. To be
honest, I can't ever get shelx to find my platinums with 6 A data :)
I would also like to hear about phase extension at low resolution (which
you have mentioned).
Overall, it appears that with such data, there are too many places for
failure.
Thanks for everyone's interest. Later, I shall come up with a nice summary.
Engin
On 5/12/09 11:26 AM, James Holton wrote:
measuring anomalous differences has nothing to do with resolution.
measuring anomalous differences has nothing to do with Rmerge.
measuring anomalous differences has EVERYTHING to do with signal and
noise. (as does measuring anything else)
If your average anomalous difference is going to be ~5%, then you need
to be able to measure a 5% change in spot intensity, yes? So, if you
take your native data, and compare the merged values of I+ and I-
(known in Scala as Ranom), and they are already more than 5%
different, then ... you are in trouble. But if Ranom for native data
is less than 5%, then you stand a chance of measuring a 5% difference.
That is, for native data, the "true" values of I+ and I- should be
"the same" (within the Bijvoet ratio for the sulfurs, which is usually
< 0.5%), so comparing I+ and I- for native data is actually a very
good way to get your expected "anomalous error". You can improve this
number by increasing redundancy, even if you reduce the exposure time
to compensate. In fact, it is a VERY good idea to do this when trying
to measure anomalous differences. Redundancy is good for anomalous,
but bad for high-res data. Long exposures and fine slicing are good
for high-res data, but bad for anomalous.
Resolution comes into play because the "anomalous error" will approach
infinity as your spot intensity approaches zero, so you will never be
able to measure anomalous differences for your highest resolution
bin. The resolution to which you CAN measure anomalous differences
(with a signal-to-noise ratio greater than one) will be the resolution
where the cumulative Ranom rises to the Bijvoet ratio (5% in your
case). That is, look for the resolution limit where the overall
"native Ranom" is 5%, and that is the resolution to which you will
probably get experimental phases.
If there is no such resolution limit (Ranom > 5% in all bins), then
MAD/SAD will not work with your current data collection method.
Higher redundancy is called for.
However, do not get too excited if this resolution limit is 6 A.
Although 6 A phases are better than no phases at all, have you ever
LOOKED at a 6 A map? It can be very hard to tell if it is protein or
not, even with perfect phases and all the right hand choices, etc.
Programs and crystallographers alike can get confused by this. I know
that there are still many structural biologists out there who "just
want to get the structure", but I remind you that you can already "get
the structure" to ~50 A resolution with other techniques. Such as gel
filtration.
The success of phase extension does depend on resolution. Although I
do not have a quantitative argument for it, the success of SAD
structure determination at worse than 4 A does seem to drop
precipitously. This could simply be correlated with the crappiness of
the crystals, but it is important to remember that SAD relies heavily
on density modification technology, such as solvent flattening and
histogram matching, etc, and these methods loose a great deal of power
as the resolution of the map decreases (and the protein-solvent
contrast becomes less clear). IMHO it is ALWAYS better to collect MAD
data, because then the dichotomous phase ambiguity is resolved
experimentally. Two wavelengths are twice as good as one, even with
the exposure time cut in half.
-James Holton
MAD Scientist
Engin Ozkan wrote:
Hi everyone,
I thought I start a new thread while it is unusually quiet on the bb.
I am pondering over the practical limitations to MAD and SAD phasing
with Se-Met at low resolution. What is the lowest resolution at which
people have solved structures "only" using phases from selenium in a
"realistic" case? Let me further qualify my question: My *realistic*
*low* resolution case is where
1. Rmerge over all resolution bins is 6-10% (i.e. your crystals are
lousy).
2. Resolution limit is worse than 3.5 Angstroms, where <I>/<sigma>
in the last resolution bin is between 1 and 3 (i.e. your crystals are
really lousy).
3. Assuming good selenium occupancy (~85%; I work with eukaryotic
expression systems, so 100% is not usually achieavable),
4. The number of selenium atoms are enough many that the
Crick-Magdoff equation would give you *at least* an average 5% change
in intensities (assuming 6 electrons contributed per selenium, based
on both absorptive and dispersive differences being at about 6 e- at
the absorption edge).
5. and specifically, no other phases and molecular replacement
solutions are available.
Obviously, I have a case very similar to what's described above, and
three years of failure with heavy atom derivatization (I am still
trying). I would be happy to hear about Se-Met cases, and data
collection strategies (2wl vs. 3wl MAD vs. SAD, etc.) and phasing
methods used in these cases, or references of them. Again, no other
partial phases, and no data cut off at 3.6 A with an I/s of 15 in the
last resolution bin. Are there any examples out there? Searching the
RCSB and PubMed did not point out to me many successful cases.
Thanks,
Engin
P.S. I would also appreciate the specific query type for searching
the PDB on the web for phasing method (MR, MAD, SAD, MIR, etc.).
They seem to have everything under the sun searchable, but I cannot
find this one.
