Gerard,
Thanks!
Actually, I was sitting on that one for a while and debating the wisdom
of posting it. With multi-million-dollar equipment, people tend to get
sensitive about "opinions".
But, yes, I suppose I didn't answer Theresa's second question about
anomalous data collection and rad dam. My answer to it is simple:
- 2 wavelengths are better than 1 (about twice as good, actually, even
with half the exposure)
- 3 wavelengths are only marginally better than 2.
- the best wavelength-changing schedule is: as often as possible.
Preferably every image. Same goes for inverse beam. That is, what I
recommend to my users is:
image phi energy
1 0 peak-inf
2 180 peak-inf
3 1 remote
4 181 remote
5 2 peak-inf
etc.
Where "peak-inf" is halfway between the inflection and the peak. This is
the best "compromise" between maximizing f" and also maximizing the
difference in f' between the two wavelengths.
The reason for the rapid interleaving is just a fundamental principle of
science: if you are measuring a difference, don't wait too long between
the two measurements you are going to subtract. I.E. don't wait until
Sunday to do the control for an experiment you did on Wednesday. Also,
don't subtract F+ at 1 MGy from F- at 20 MGy.
However, there are always caveats. Every beamline has different design
compromises. Flux, flexibility, and speed don't always go together.
Sometimes rapid wavelength changes can overheat monochromator motors,
and sometimes inverse beam can be very slow. So, ask the beamline
scientist who runs the machine you plan to use what they recommend for
their hardware. But, if the hardware can do it, it is always better to
"change up" as rapidly as you have time for.
As for high-and low-intensity passes, I recommend doing the
low-intensity pass first. Some people have passionate opinions to the
contrary, but I say if the low-dose pass causes significant radiation
damage, then you definitely shouldn't have done a high-dose pass first!
Unless, of course, you are doing a multi-crystal strategy, then it is
okay to not get complete data from any one crystal. But even in that
case, you want a complete and relatively damage-free "refrerence"
dataset first to help you "align" the partially-complete high-dose
datasets together. So, again: short exposures first.
What I specifically recommend to my users for anomalous is to do a full
360 with 2 wavelengths (peak-inf and remote) with the shortest exposure
they think they can still process as the "first pass". Then, for a
second pass, quadruple the exposure time (or reduce the attenuation by a
factor of 4). The factor of 4 is mainly because doubling the number of
photons only increases signal/noise by a factor of 1.4, quadrupling the
number of photons doubles the signal/noise ratio. Then you keep
increasing the exposure: 1s, 4s, 16s, 32s, etc for 360-degree passes
until the crystal is clearly dead.
It is also a good idea to move the detector a bit between each "pass"
so that you are not using the same pixels over and over again. That is,
try to move your spots onto new pixels for each "pass". Every pixel has
a slightly different calibration.
When you get home, you can try mergeing all that data together and start
doing "chronological cuts" (removing the last frames) to see where the
stats are "best". I tend to look at the anomalous CC. the best test of
all is, of course, the peak height in a phased anomalous difference
Fourier, but you need phases for that.
If no chronological cut works, you can try throwing out the middle:
treat the last decent dataset as the "native" and the first dataset as
the "derivative" and do RIP. You can also try mergeing both wavelengths
together and treat it as SAD data, perhaps doing chronological cuts to
minimize rad dam. This is another reason to interleave wavelengths and
inverse beam rapidly: it allows you to "dial" the rad dam by using the
image file time stamps. So, in this way, this strategy let's you try
three methods for the price of one.
-James Holton
MAD Scientist
On 5/16/2013 10:03 AM, Gerard Bricogne wrote:
> Dear James,
>
> A week ago I wrote what I thought was a perhaps excessively long and
> overly dense message in reply to Theresa's initial query, then I thought I
> should sleep on it before sending it, and got distracted by other things.
>
> I guess you may well have used that whole week composing yours ;-) and
> reading it just now makes the temptation of sending mine irresistible. I am
> largely in agreement with you about the need to change mental habits in this
> field, and hope that the emphasis on various matters in my message below is
> sufficiently different from yours to make a distinct contribution to this
> very important discussion. Your analysis of pile-up effects goes well beyond
> anything I have ever looked at. However, in line with Theresa's initial
> question, I would say that, while I agree with you that the best strategy
> for collecting "native data" is no strategy at all, this isn't the case when
> collecting data for phasing. In that case one needs to go back and consider
> how to measure accurate differences of intensities, not just accurate
> intensities on their own. That is another subject, on which I was going to
> follow up so as to fully answer Theresa's message - but perhaps that should
> come in another installment!
>
>
> With best wishes,
>
> Gerard.
>
> --
> On Tue, May 07, 2013 at 12:04:33AM +0100, Theresa Hsu wrote:
>> Dear crystallographers
>>
>> Is there a good source/review/software to obtain tips for good data
> collection strategy using PILATUS detectors at synchrotron? Do we need to
> collect sweeps of high and low resolution data separately? For anomalous
> phasing (MAD), does the order of wavelengths used affect structure solution
> or limit radiation damage?
>>
>> Thank you.
>>
>> Theresa
> --
>
> Dear Theresa,
>
> You have had several excellent replies to your question. Perhaps I
> could venture to add a few more comments, remarks and suggestions, which can
> be summarised as follows: with a Pilatus, (1) use fine slicing, (2) use
> strategies combining low exposure with high multiplicity, and (3) use XDS!
>
> As the use of Pilatus detectors has spread widely, it has been rather
> puzzling to come across so many instances when these detectors are misused,
> sometimes on the basis of explicit expert advice that is simply misguided. A
> typical example will be to see images collected on a Pilatus 6M with an
> image width of 1 degree and an exposure time of 1 second. When you see this,
> you know that there is some erroneous thinking (or habit) behind it.
>
> When talking to various users who have ended up with such datasets, and
> with people who advocate this kind of strategy, it seems clear that a number
> of irrational concerns about fine-slicing and low-exposure+high-multiplicity
> strategies have tended to override published rational arguments in favour of
> those strategies: there is a fear that if the images being collected do not
> show spots discernible by the naked eye to the resolution limit that is
> being aimed for, the integration software will then somehow not be able to
> find those spots in order to integrate them, and the final data resolution
> will be lower than expected. Perhaps this may be of concern in relation with
> the use of some integration programs, but if you use XDS, which implements a
> full 3D approach to image integration, this is simply not the case: XDS will
> collect all the counts belonging to a given reflection, whether those counts
> are all from a spot on a single 1-degree image exposed for 1 second, or from
> 10 consecutive images of 0.1 degree width exposed for 0.1 second each, or
> from 100 images obtained by grouping together the same 10 images as
> previously collected in 10 successive passes with a 10-fold attenuated beam.
> The hallmark of the Pilatus detector is to lead to equivalent signal/noise
> ratios for the last two ways of measuring that reflection, because it is a
> photon counter and has zero readout noise: therefore the combination
> Pilatus+XDS is a powerful one.
>
> What is different between these three strategies, however, is the
> quality of the overall dataset they will produce. There is nothing new in
> what I am describing below: it is all in the references that Bob Sweet gave
> you in his reply, or is an obvious consequence of what is found in these
> references.
>
> In case 1 (1-degree, 1 second - "coarse slicing") you would presumably
> also be (mis-)advised to use a strategy aiming at collecting a complete
> dataset in the minimum number of images. These strategies used to make sense
> in the days of films, of image plates, and even of CCDs because of the image
> readout noise, but they have no place any longer in the context of Pilatus
> detectors. First of all, using 1-degree image widths can only degrade the
> precision with which 2D spots on images are lifted to 3D reciprocal space
> for indexing, and hence worsen the quality of that indexing and therefore
> the accuracy with which the spot locations will be predicted (unless you
> carefully "post-refine") - then the integration step perhaps does need to
> "hunt" for those spots locally, and needs them to be somewhat visible.
> Secondly, 1 degree is usually greater than the angular width of a typical
> reflection: the integration process will therefore pick up more background
> noise (variance) than it would have done with a smaller image width.
> Thirdly, by collecting only enough images to reach completeness you will
> have substantial radiation damage in your late images compared to the early
> ones (if you don't, it means you have under-exposed your crystal) and will
> therefore end up with internal inconsistencies in your dataset, as well as
> perhaps some extra, spurious anisotropy of diffraction limits as a result of
> having to impose increasingly stringent resolution cut-offs in the later
> images. This will affect the internal scaling of that dataset and the final
> quality of the merged data.
>
> In case 2 (0.1 degree, 0.1 second - "fine slicing") you will have a
> more precise sampling of the 3D shape of each spot, hence more accurate
> indexing and prediction of spot positions if you use a genuinely 3D
> integration program like XDS. Thanks to that increased precision, spots can
> be integrated "blind", even if they are not terribly visible in the images,
> and the same number of photons will be collected with no penalty in terms of
> noise level, thanks to the photon-counting noiseless-readout nature of the
> Pilatus detector. An improvement will be that the finely sampled 3D shape of
> the spots will be used by XDS to minimise the impact of background variance
> on the integrated intensities. On the other hand, the differential radiation
> damage between early and late images will still be the same as in case 1 if
> you have chosen one of those old-style strategies (and associated beam
> intensity setting) that aim at just about exhausting the useful lifetime of
> the crystal by the time you reach completeness.
>
> In case 3 (like case 2, but collecting n times more images with an
> n-fold attenuated beam once you have collected a few "characterisation
> images" without that attenuation to carry out the initial indexing) you
> still have the two advantages of case 2 (the same total number of photons
> will be picked up by XDS, even if the individual images are now so weak that
> you can't see anything) but you are spreading the radiation damage so thinly
> over multiple successive complete datasets that you can choose to later
> apply a cut-off on image number at the processing stage, when the statistics
> tell you that diffraction quality has become degraded beyond some critical
> level. This is much preferable to having to apply different resolution
> cut-offs to different images towards the end of a barely complete dataset,
> as in cases 1 and 2. The impact of radiation damage will be quite smoothly
> and uniformly distributed across the final unique reflections, and your
> scaling problems (as well as any spurious anisotropy in your diffraction
> limits) will be minimised.
>
>
> This is becoming quite a long message: you can see why I included a
> summary of it at the beginning! Returning to it for a conclusion: Pilatus
> detectors, fine-slicing with low-exposure and high-multiplicity strategies,
> and XDS are a unique winning combination. If fears that another integration
> program may not perform as well as XDS on fine-sliced data make you feel
> tempted to revert to old-fashioned strategies (case 1) because it supposedly
> makes no difference: resist the temptation! Switch to those Pilatus-adapted
> strategies and to XDS, and enjoy the very real difference in the results!
>
>
> With best wishes,
>
> Gerard
>
> and colleagues at Global Phasing.
>
> --
> On Tue, May 07, 2013 at 12:04:33AM +0100, Theresa Hsu wrote:
>> Dear crystallographers
>>
>> Is there a good source/review/software to obtain tips for good data
> collection strategy using PILATUS detectors at synchrotron? Do we need to
> collect sweeps of high and low resolution data separately? For anomalous
> phasing (MAD), does the order of wavelengths used affect structure solution
> or limit radiation damage?
>> Thank you.
>>
>> Theresa
|
