Let me clarify: twin *refinement* is definitely the way to go. Detwinning data *prior* to structure solution, however, always seems to worsen things, whether for MR or experimental phasing. But it is hard to imagine how detwinning makes data worse than not, if we’ve modelled twinning appropriately; any degree of detwinning should bring intensities closer to their true values, making structure solutions better. Or at least not *worse*!
I would therefore suggest that there is something about our twinning model that does not fit the data. Maybe: A. Twinned data generally have variable twin fractions across the datasets due to changes in illuminated volumes. Since we detwin between reflections measured at different rotation angles, the math would be wrong. B. Diffraction from different twin domains combines not only additively as intensities but also vectorially as structure factors. This would be more pronounced as twin domains get smaller in size. C. Something else? It is just so strange to me that twinning, which is in principle something fully understood, should ever prevent structure solution. Even a 50% twin should be no worse than halving the number of observations, right? JPK From: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] On Behalf Of Bernhard Rupp Sent: Tuesday, October 11, 2016 8:09 AM To: CCP4BB@JISCMAIL.AC.UK Subject: Re: [ccp4bb] Why Does Detwinning Not Work? Tom Terwilliger recently pointed me to a relevant discussion on the phenix bb: http://phenix-online.org/pipermail/phenixbb/2013-May/019836.html In essence, in case of detwinning, the refinement target is not a ML target anymore. With the model being used as a basis for the detwinning of Fs, the refi procedure becomes susceptible to bias. I do not know what the current state is of developing the ML target for refining against twinned (I, F) data is. Garib, Randy, perhaps? Thx, BR From: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] On Behalf Of herman.schreu...@sanofi.com<mailto:herman.schreu...@sanofi.com> Sent: Tuesday, October 11, 2016 1:23 PM To: CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK> Subject: [ccp4bb] AW: [ccp4bb] AW: [ccp4bb] Why Does Detwinning Not Work? Normally, twinned refinement is the way to go. However, if the maps don’t become clearer and without twinned refinement the R/Rfrees are already 0.25/0.29, there might be some other problem. Here I would be more cautious. Herman Von: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] Im Auftrag von Eleanor Dodson Gesendet: Dienstag, 11. Oktober 2016 13:09 An: CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK> Betreff: Re: [ccp4bb] AW: [ccp4bb] Why Does Detwinning Not Work? Search for: "experimental phasing with detwinned data" and you will find some hits. You need a low degree of twinning, and a strong sub-structure signal.. Re MR solutions - it is usually possible to solve the structure providing you get the space group right, but in my experience the maps are clearer if you do twinned refinement. I cant see any reason not to use it? Eleanor On 11 October 2016 at 10:05, <herman.schreu...@sanofi.com<mailto:herman.schreu...@sanofi.com>> wrote: Dear Jacob, I agree with Chris. In my experience for MR, twinning is just like having a space group with one extra 2-fold (or other twin operator) so MR is indeed largely immune to twinning. Also the maps, calculated with uncorrected data, often look surprisingly good, just having a higher level of random noise. I did not do the math, but my feeling is that the (greatly) inflated measurement errors (exploding as you mention near a twin fraction of 0.5) make direct phasing methods, that are often at the very limit of useable signal to noise ratio, fail after detwinning. Also, an incorrectly estimated twin fraction may introduce systematic errors, especially in case of a varying twin fraction depending on the rotation of the crystal. Probably random noise is less harmful than systematic errors. My 2 cts, Herman Von: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK>] Im Auftrag von Chris Fage Gesendet: Dienstag, 11. Oktober 2016 10:40 An: CCP4BB@JISCMAIL.AC.UK<mailto:CCP4BB@JISCMAIL.AC.UK> Betreff: Re: [ccp4bb] Why Does Detwinning Not Work? Dear Jacob, I'm not an expert on the topic, but from my experiences with twinning I can agree with you. I recently solved my second twinned structure by MR (twin fraction of 0.43, as estimated by Xtriage). Performing twin refinement in Refmac or phenix.refine dropped the R-factors, as expected, but worsened the geometry considerably without a noticeable improvement in the maps. For this reason, I opted *not* to go with the twin refinement... I don't know if others would make the same choice, though it seemed reasonable to me. Besides, my Rwork/Rfree is down to 0.25/0.29, which ain't too shabby for 2.6 A resolution. Cheers, Chris On Tue, Oct 11, 2016 at 2:15 AM, Keller, Jacob <kell...@janelia.hhmi.org<mailto:kell...@janelia.hhmi.org>> wrote: Dear Crystallographers, Based on some data sets I have looked at and anecdotal-type evidence here and there I have gotten the impression that detwinning does not help in structure solution. (Please let me know if you have a case where detwinning saved the day.) Is there a clear answer to this enigma anywhere, to anyone’s knowledge? Wouldn’t it seem that *any* detwinning would be better than *no* detwinning? I understand that the errors explode as one approaches 50% twins and does detwinning, but still, I don’t think one *loses* information by detwinning, right? Take the case of a 33% twin: since the twin-reflections are on average about half the intensity of the non-twin, and since they are generally not correlated in intensity, isn’t this like having noise added at 50% of the measured intensity? So why does detwinning make things worse generally? Is there something wrong in the assumptions underlying the detwinning algorithm, or perhaps something about the calculation that throws things off? A related sub-enigma: why is MR generally immune to twinning, but anomalous methods are susceptible? All the best, Jacob Keller ******************************************* Jacob Pearson Keller, PhD Research Scientist HHMI Janelia Research Campus / Looger lab Phone: (571)209-4000 x3159<tel:%28571%29209-4000%20x3159> Email: kell...@janelia.hhmi.org<mailto:kell...@janelia.hhmi.org> *******************************************
