Hi Jacob, I wonder what are those 'certain conditions' under which the anomalous sign is flipped for twinned data. Are they cases where the anomalous signal is weak (marginal or hardly existent) in which case using them for phasing is not a good thing (grave mistake I would think) or do you also notice this effect also for reflections with significant anomalous signal? 
Just wondering.
Cheers, Boaz


-------- Original message --------
From: "Keller, Jacob" <[log in to unmask]>
Date: 11/05/2016 16:02 (GMT+02:00)
To: [log in to unmask]
Subject: Re: [ccp4bb] Dano--Sign Convention

The second option (detwinning Dano's) is not currently possible I don't think, since Dano's are negative ~half the time, and I am not sure that Detwin can handle that. One would also have to fool the program somehow into thinking Danos were F's or I's, since Detwin takes input only as Imean, I+/-, Fmean, or F+/-. Regarding unmatched reflections, it simply omits them from the output file, which could be a problem when completeness is not great, since some (twinned) Bijvoet differences will be chucked. This might be one reason detwinning does not work so perfectly.

I did discover an interesting effect that under certain conditions, not only the magnitude but also the sign of Dano for a given reflection can change due to twinning. This should be equivalent to flipping the phase 180 deg from the true value, which is certainly bad news! But detwinning should theoretically get rid of this. This sign flip might be the major reason anomalous methods suffer more from twinning than MR?

I would suggest the way to address this issue would be to place, during HA finding and phasing, "evil twin" atoms with occupancy weighted by the twin fraction and spatially related by the twin operator (both fraction and operator known from the various tests). In other words, modelling the twinning from the beginning of the whole process.

JPK

-----Original Message-----
From: Kay Diederichs [mailto:[log in to unmask]]
Sent: Wednesday, May 11, 2016 6:00 AM
To: [log in to unmask]; Keller, Jacob
Subject: Re: Dano--Sign Convention

Hi Jacob,

I can think of several different ways of detwinning in presence of anom signal, whereas you may be assuming that there is only one way - the one implemented in CCP4 detwin?

The ways I can think of, given some reflection with index triple h and its twin mate h', and a twin fraction alpha, in hopefully obvious notation, are:
a) one could calculate I_true(+) and I_true(-) from I_obs(+),I_obs'(+) and I_obs(-),I_obs'(-), respectively, and then calculate delta_true(h) and delta_true(h') from I_true(+)-I_true(-) and I_true'(+)-I_true'(-), respectively
b) one could calculate delta_obs(h) and delta_obs(h') from I_obs(+),I_obs(-) and I_obs'(+),I_obs'(-), respectively, and then calculate delta_true(h) and delta_true(h') from delta_obs(h) and delta_obs(h')
c) for both a) and b) one has a decision to make how to proceed if one or more out of {I_obs(+),I_obs'(+), I_obs(-),I_obs'(-)} are not measured. 

I have not worked out in detail any of these possibilities, but clearly (at least that's what my mathematical intuition says ...) the results are not identical. Also, I have no idea which of these is implemented in CCP4 detwin, and whether or not the other possibilities were tested by anyone, or if some of them are clearly superior to others. These are important scientific question, and it would be worthwhile to answer them - your data may be a very good example. It would not be difficult (I think) to implement these possibilities in CCP4 detwin.

best,

Kay


You wrote:
> You are right--the absolute peak heights are in fact higher in the
> detwinned case, but I am still astonished that the Z scores are not
> better. I can also see a couple of other anomalous sites in the
> non-detwinned data which are not visible in the detwinned case.
>
> Why does it matter anyway what the absolute height is, if the Z score
> is lower? The signal has gone away, despite nominally “better fit”
> between the data and model. R differences are ~20/25 versus 12/16.
>
> JPK
>
> From: Eleanor Dodson [mailto:[log in to unmask]]
> Sent: Monday, May 09, 2016 5:04 PM
> To: Keller, Jacob
> Cc: [log in to unmask]
> Subject: Re: [ccp4bb] Dano--Sign Convention
>
> Well - you can't compare a twinned R factor to untwinned - there is
> that extra parameter - the twin factor . (It is explained in the
> REFMAC paper on twinned refinement - don't have it here at home, and
> can't remember the formula but it is always the case that twinned Rs
> are lower..)
>
> But it is funny that the signal should be reduced. Is the absolute
> peak height different?
> Looking at peak heights as sigma factors means thet the  map sigma
> levels should be comparable  for the diff protocols.
>
> I can't see how detaining could muddle the Friedel pairs - the twin
> operator is always a rotation without any inversion eg P3 h k l to k h
> -l  which does not change the hand
>
> Eleanor
>
>
>
> On 9 May 2016 at 21:55, Keller, Jacob <[log in to unmask]> wrote:
> I’ll just say what I was thinking:
>
> Is it possible that in the course of detwinning the Bijvoet sign of
> one of the twin domain’s reflections is getting flipped, such that
> when on detwins I+ and I- independently to preserve anomalous
> differences, one is really detwinning I+ of domain A versus I- of domain B?
>
> The phenomenon that I am having difficulty explaining is the following.
> I have a very good quality structure solved by MR with a twinned
> dataset. When refined against non-detwinned data, and without twin
> refinement, I get Ca++ anomalous peaks ~17 sigma. When I refine either
> against detwinned data or with twin refinement, I get anomalous peaks
> of only ~9-12 sigma. The R values with detwinning/twin refinement are
> dramatically better, but result in much weaker anomalous peaks. Why on
> earth should that be?
>
> Yes, the variance goes way up when one detwins, but how is it that
> twin refinement then can procure such improved R values without
> concomitant improvement in anomalous peak height? Isn’t something peculiar here?
>
> JPK
>
> From: Eleanor Dodson [mailto:[log in to unmask]]
> Sent: Monday, May 09, 2016 4:11 PM
> To: Keller, Jacob
> Cc: [log in to unmask]
> Subject: Re: [ccp4bb] Dano--Sign Convention
>
> Well - DANO is calculated between merged Fhkl and MergedF(-h -k -l)
>
> I trust that the software correctly groups the true symmetry
> equivalents for the Friedel pairs.
>
> Of course all twinning corrections assume that the twin fraction is
> the same for all measured reflections. This may well not be true if
> the crystal is bigger than the beam..
>
> But I don't think things are worse for friedel pairs than for all
> other reflections?
>
> Eleanor
>
> On 9 May 2016 at 19:28, Keller, Jacob <[log in to unmask]> wrote:
> Dear Crystallographers,
>
> I am curious about the sign convention of Dano as follows:
>
> The definition of Dano of hkl is F(hkl) - F(-h-k-l), so for example
>
> Dano(2,3,5) = F(2,3,5) - F(-2,-3,-5)
>
> What about mixed-sign indices, for example
>
> Dano(2,3,-5) = F(2,3,-5) - F(-2,-3,5)
>
> OR
>
> Dano(2,3,-5) = F(-2,-3,5) - F(2,3,-5)
>
> What is the convention? I think this might possibly have ramifications
> for detwinning.
>
> JPK
>
>
> *******************************************
> Jacob Pearson Keller, PhD
> Looger Lab/HHMI Janelia Research Campus
> 19700 Helix Dr, Ashburn, VA 20147
> email: [log in to unmask]
> *******************************************
>
>