[ccp4bb] Always Modelling Anomalous Signal
Dear Crystallographers, it seems to me that on a certain level we are always throwing away (sort of) about half of our data when we merge Bijvoet pairs--why shouldn't we keep them separate, since we know that they should be a little bit different, especially in light of the higher multiplicities which are more common now? Shouldn't modelling them give better R-values, and wouldn't it just be more true? I guess a sort of proof for this is that sulfurs are almost always detectable on anomalous difference maps, implying that we are actually measuring those differences accurately enough to see them (I don't think these things can arise from model bias, as anomalous differences are not modeled.) At least maybe at the final steps of refinement...? JPK -- *** Jacob Pearson Keller Northwestern University Medical Scientist Training Program email: j-kell...@northwestern.edu ***
Re: [ccp4bb] Always Modelling Anomalous Signal
On Jan 10, 2012, at 22:00 , Jacob Keller wrote: Dear Crystallographers, it seems to me that on a certain level we are always throwing away (sort of) about half of our data when we merge Bijvoet pairs--why Who We ? shouldn't we keep them separate, since we know that they should be a little bit different, especially in light of the higher multiplicities which are more common now? Shouldn't modelling them give better R-values, and wouldn't it just be more true? I guess a sort of proof for this is that sulfurs are almost always detectable on anomalous As a matter of protocol, for the last 25 years I keep anomalous signal in my data for refinement, does not matter what ignorant annotators say. I do it for several reasons, one of them is matter of principle. difference maps, implying that we are actually measuring those differences accurately enough to see them (I don't think these things The problem is sometimes to deposit the data, which were used for the refinement. Developers refuse to add the redundant information into the deposition file, annotators refuse to accept the data file without averaged value for F or I, directors keep silent. Be strong Jacob... can arise from model bias, as anomalous differences are not modeled.) At least maybe at the final steps of refinement...? JPK -- *** Jacob Pearson Keller Northwestern University Medical Scientist Training Program email: j-kell...@northwestern.edu *** Dr Felix Frolow Professor of Structural Biology and Biotechnology Department of Molecular Microbiology and Biotechnology Tel Aviv University 69978, Israel Acta Crystallographica F, co-editor e-mail: mbfro...@post.tau.ac.il Tel: ++972-3640-8723 Fax: ++972-3640-9407 Cellular: 0547 459 608
Re: [ccp4bb] Always Modelling Anomalous Signal
Jacob, Actually the R factors including the Bijvoet pairs would be higher, because the uncertainties in F(+) and F(-) are higher than that of F(mean) by a factor of about sqrt(2). R factors will always be higher for unmerged data because averaging always reduces the uncertainty. This means that we are in effect 'cheating' by throwing away the relatively imprecise anomalous differences and getting falsely lower R factors as a result! But as you imply the model would improve if we included the anomalous data in the refinement (assuming of course that it is meaningful data). This just demonstrates that low R factors do not necessarily equate to a better model - especially if you deliberately throw away the less precise data! The model would improve (marginally) because the anomalous differences would obviously provide additional information about the anomalous scatterers and therefore increase their precision, but wouldn't affect the precision of the lighter atoms. But is imprecision in the parameters of the heavy (or heavier) atoms usually an issue? - since these have bigger real scattering factors they will be more precisely determined than the lighter atoms anyway. So I don't think you would gain very much, except maybe more truthful R factors! Cheers -- Ian On 10 January 2012 20:00, Jacob Keller j-kell...@fsm.northwestern.edu wrote: Dear Crystallographers, it seems to me that on a certain level we are always throwing away (sort of) about half of our data when we merge Bijvoet pairs--why shouldn't we keep them separate, since we know that they should be a little bit different, especially in light of the higher multiplicities which are more common now? Shouldn't modelling them give better R-values, and wouldn't it just be more true? I guess a sort of proof for this is that sulfurs are almost always detectable on anomalous difference maps, implying that we are actually measuring those differences accurately enough to see them (I don't think these things can arise from model bias, as anomalous differences are not modeled.) At least maybe at the final steps of refinement...? JPK -- *** Jacob Pearson Keller Northwestern University Medical Scientist Training Program email: j-kell...@northwestern.edu ***
Re: [ccp4bb] Always Modelling Anomalous Signal
On Tuesday, January 10, 2012 02:46:21 pm Ian Tickle wrote: Jacob, Actually the R factors including the Bijvoet pairs would be higher, because the uncertainties in F(+) and F(-) are higher than that of F(mean) by a factor of about sqrt(2). R factors will always be higher for unmerged data because averaging always reduces the uncertainty. This means that we are in effect 'cheating' by throwing away the relatively imprecise anomalous differences and getting falsely lower R factors as a result! But as you imply the model would improve if we included the anomalous data in the refinement (assuming of course that it is meaningful data). This has been an option in refmac refinement for some while now. If you are using the GUI, select the option for using SAD data directly rather than the default option using no prior phase information. Of course you must also make sure that appropriate scattering factors are picked up by the program. In my limited experience this has not led to improved refinement if the anomalous differences are due only to sulfur, but has on occasion led to noticeable improvement for true SeMet/SAD data. Ethan This just demonstrates that low R factors do not necessarily equate to a better model - especially if you deliberately throw away the less precise data! The model would improve (marginally) because the anomalous differences would obviously provide additional information about the anomalous scatterers and therefore increase their precision, but wouldn't affect the precision of the lighter atoms. But is imprecision in the parameters of the heavy (or heavier) atoms usually an issue? - since these have bigger real scattering factors they will be more precisely determined than the lighter atoms anyway. So I don't think you would gain very much, except maybe more truthful R factors! Cheers -- Ian On 10 January 2012 20:00, Jacob Keller j-kell...@fsm.northwestern.edu wrote: Dear Crystallographers, it seems to me that on a certain level we are always throwing away (sort of) about half of our data when we merge Bijvoet pairs--why shouldn't we keep them separate, since we know that they should be a little bit different, especially in light of the higher multiplicities which are more common now? Shouldn't modelling them give better R-values, and wouldn't it just be more true? I guess a sort of proof for this is that sulfurs are almost always detectable on anomalous difference maps, implying that we are actually measuring those differences accurately enough to see them (I don't think these things can arise from model bias, as anomalous differences are not modeled.) At least maybe at the final steps of refinement...? JPK -- *** Jacob Pearson Keller Northwestern University Medical Scientist Training Program email: j-kell...@northwestern.edu *** -- Ethan A Merritt Biomolecular Structure Center, K-428 Health Sciences Bldg University of Washington, Seattle 98195-7742
