[ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Hi, I have a preliminary question. I understand Rfree reflection sets are never used during automatic refinement, but, when generating the real space density maps, do we have to exclude Rfree columns? Any references will also be greatly appreciated! Best Regards, Hailiang
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
I meant :when generating the real space density maps, do we have to exclude Rfree reflections? > Hi, > > I have a preliminary question. I understand Rfree reflection sets are > never used during automatic refinement, but, when generating the real > space density maps, do we have to exclude Rfree columns? Any references > will also be greatly appreciated! > > Best Regards, Hailiang >
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: > I have a preliminary question. I understand Rfree reflection sets are > never used during automatic refinement, but, when generating the real > space density maps, do we have to exclude Rfree columns? Any references > will also be greatly appreciated! > If you are going to run overall real-space refinement on the structure, you should absolutely exclude the test set reflections from the map. If you are only going to run local refinement of small parts of the model in Coot or equivalent, it's debatable - in practice, I think most people/programs leave them in. -Nat
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Thanks Nat! I am not doing real space refinement. Actually I am only using the maps for manual model building/adjustments. In this case, if some Rfree reflections have strong scattering intensities, removing them may lead to featureless density maps. However, if we just leave them in, do you think we may have the so-called model-bias issue? Hailiang > On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: > >> I have a preliminary question. I understand Rfree reflection sets are >> never used during automatic refinement, but, when generating the real >> space density maps, do we have to exclude Rfree columns? Any references >> will also be greatly appreciated! >> > > If you are going to run overall real-space refinement on the structure, > you > should absolutely exclude the test set reflections from the map. If you > are > only going to run local refinement of small parts of the model in Coot or > equivalent, it's debatable - in practice, I think most people/programs > leave > them in. > > -Nat >
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
It should be remembered that refining in real space is equivalent to refinement in the reciprocal space (through Parseval's theorem). If you want to do consistent refinement then you need to use exactly same reflections for free and working set. If you do not use the same set of reflections for real and reciprocal space refinements then you may get very interesting results. Garib On 23 May 2011, at 21:17, Hailiang Zhang wrote: > Thanks Nat! I am not doing real space refinement. Actually I am only using > the maps for manual model building/adjustments. In this case, if some > Rfree reflections have strong scattering intensities, removing them may > lead to featureless density maps. However, if we just leave them in, do > you think we may have the so-called model-bias issue? > > Hailiang > > >> On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: >> >>> I have a preliminary question. I understand Rfree reflection sets are >>> never used during automatic refinement, but, when generating the real >>> space density maps, do we have to exclude Rfree columns? Any references >>> will also be greatly appreciated! >>> >> >> If you are going to run overall real-space refinement on the structure, >> you >> should absolutely exclude the test set reflections from the map. If you >> are >> only going to run local refinement of small parts of the model in Coot or >> equivalent, it's debatable - in practice, I think most people/programs >> leave >> them in. >> >> -Nat >>
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
On Monday, May 23, 2011 01:17:45 pm Hailiang Zhang wrote: > Thanks Nat! I am not doing real space refinement. Actually I am only using > the maps for manual model building/adjustments. In this case, if some > Rfree reflections have strong scattering intensities, removing them may > lead to featureless density maps. However, if we just leave them in, do > you think we may have the so-called model-bias issue? > > Hailiang In theory: On the one hand, if you include Fobs for the Rfree reflections and then use the map for fitting, you are contaminating your Rfree set. -> bias On the other hand, if you omit the Rfree reflections altogether then you suffer from possible Fourier truncation effects -> bias Third alternative: Include Fcalc values for the Rfree reflections rather than Fobs. -> avoids both classes of bias In practice: For manual fitting I doubt any of this makes very much difference. cheers, Ethan -- Ethan A Merritt Biomolecular Structure Center, K-428 Health Sciences Bldg University of Washington, Seattle 98195-7742
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
On Mon, May 23, 2011 at 1:17 PM, wrote: > Thanks Nat! I am not doing real space refinement. Actually I am only using > the maps for manual model building/adjustments. In this case, if some > Rfree reflections have strong scattering intensities, removing them may > lead to featureless density maps. However, if we just leave them in, do > you think we may have the so-called model-bias issue? I've always seen "model bias" used to describe phases, which don't really involve R-free. I suppose there is theoretically some slight risk of biasing your R-free value if the test set is not excluded from the maps during manual rebuilding, but it's minimal compared to what happens during refinement. The real danger is in global optimization algorithms (gradient minimization, and especially simulated annealing), which are so powerful that they can occasionally make a completely incorrect model - say, traced backwards - look reasonable (as judged by the R-factor and fit to density) after many rounds of refinement. It is very difficult to do this with modern refinement programs, regardless of what reflections are used in the maps. -Nat
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Thanks Garib, but my task was not real space refinement (just manual model building/adjustment). Following is my previous post. Thanks! """I am not doing real space refinement. Actually I am only using the maps for manual model building/adjustments. In this case, if some Rfree reflections have strong scattering intensities, removing them may lead to featureless density maps. However, if we just leave them in, do you think we may have the so-called model-bias issue?""" Hailiang > It should be remembered that refining in real space is equivalent to > refinement in the reciprocal space (through Parseval's theorem). If you > want to do consistent refinement then you need to use exactly same > reflections for free and working set. If you do not use the same set of > reflections for real and reciprocal space refinements then you may get > very interesting results. > > Garib > > > > > On 23 May 2011, at 21:17, Hailiang Zhang wrote: > >> Thanks Nat! I am not doing real space refinement. Actually I am only >> using >> the maps for manual model building/adjustments. In this case, if some >> Rfree reflections have strong scattering intensities, removing them may >> lead to featureless density maps. However, if we just leave them in, do >> you think we may have the so-called model-bias issue? >> >> Hailiang >> >> >>> On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang >>> wrote: >>> I have a preliminary question. I understand Rfree reflection sets are never used during automatic refinement, but, when generating the real space density maps, do we have to exclude Rfree columns? Any references will also be greatly appreciated! >>> >>> If you are going to run overall real-space refinement on the structure, >>> you >>> should absolutely exclude the test set reflections from the map. If >>> you >>> are >>> only going to run local refinement of small parts of the model in Coot >>> or >>> equivalent, it's debatable - in practice, I think most people/programs >>> leave >>> them in. >>> >>> -Nat >>> > > >
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
It does not matter. By fitting manually you are doing manual minimisation. The same treatment is applied. You are trying to optimise fit of the model into the electron density. I did these tests few years back and results were as expected. Independent on minimisation tools (manual, automatic, partial, full: real space, reciprocal space) exactly same "model bias" is added. As I said the reason is Parseval's theorem (with some adjustments in case when you are using maximum likelihood refinement) Garib On 23 May 2011, at 21:45, Hailiang Zhang wrote: > Thanks Garib, but my task was not real space refinement (just manual model > building/adjustment). Following is my previous post. Thanks! > > """I am not doing real space refinement. Actually I am only using the > maps for manual model building/adjustments. In this case, if some Rfree > reflections have strong scattering intensities, removing them may lead to > featureless density maps. However, if we just leave them in, do you think > we may have the so-called model-bias issue?""" > > Hailiang > >> It should be remembered that refining in real space is equivalent to >> refinement in the reciprocal space (through Parseval's theorem). If you >> want to do consistent refinement then you need to use exactly same >> reflections for free and working set. If you do not use the same set of >> reflections for real and reciprocal space refinements then you may get >> very interesting results. >> >> Garib >> >> >> >> >> On 23 May 2011, at 21:17, Hailiang Zhang wrote: >> >>> Thanks Nat! I am not doing real space refinement. Actually I am only >>> using >>> the maps for manual model building/adjustments. In this case, if some >>> Rfree reflections have strong scattering intensities, removing them may >>> lead to featureless density maps. However, if we just leave them in, do >>> you think we may have the so-called model-bias issue? >>> >>> Hailiang >>> >>> On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: > I have a preliminary question. I understand Rfree reflection sets are > never used during automatic refinement, but, when generating the real > space density maps, do we have to exclude Rfree columns? Any > references > will also be greatly appreciated! > If you are going to run overall real-space refinement on the structure, you should absolutely exclude the test set reflections from the map. If you are only going to run local refinement of small parts of the model in Coot or equivalent, it's debatable - in practice, I think most people/programs leave them in. -Nat >> >> >>
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Hailiang, r-free reflections should not participate in refinement, regardless whether it is real or reciprocal space one, done with machine driven minimizers or your hands moving atoms in Coot. Period. The issue of correcting the map appearance for missing data (resolution or completeness) is relevant but different. Removing the data is noticeable, but most of the time putting aside test set is not critical for map appearance (given that reflections are selected randomly and do not exceed a reasonable fraction of available data); and when it is critical, the cross validation should be done differently anyway. So the answer to your question is: every time you compute a map not just to enjoy its appearance but to use it to improve your model, do not include test flagged reflections into it. Pavel. On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: > Hi, > > I have a preliminary question. I understand Rfree reflection sets are > never used during automatic refinement, but, when generating the real > space density maps, do we have to exclude Rfree columns? Any references > will also be greatly appreciated! > > Best Regards, Hailiang >
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Dear all, I'm very interested in this topic. I have a question about the default behaviors in output reflection files of each refinement softwares. Are test set reflections excluded from the columns for calculating electron density maps? I found in phenix.refine documentation the option electron_density_maps.exclude_free_r_reflections was equal to False by default. (Does this option affect real space refinement in phenix.refine?) And I don't think Coot excludes test set reflections when opening MTZ file... because there's no option to specify the flag number, right? Thanks in advance, Keitaro 2011/5/24 Pavel Afonine : > Hailiang, > r-free reflections should not participate in refinement, regardless whether > it is real or reciprocal space one, done with machine driven minimizers or > your hands moving atoms in Coot. Period. The issue of correcting the > map appearance for missing data (resolution or completeness) is relevant but > different. Removing the data is noticeable, but most of the time putting > aside test set is not critical for map appearance (given that reflections > are selected randomly and do not exceed a reasonable fraction of available > data); and when it is critical, the cross validation should be done > differently anyway. > So the answer to your question is: every time you compute a map not just to > enjoy its appearance but to use it to improve your model, do not include > test flagged reflections into it. > Pavel. > > On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: >> >> Hi, >> >> I have a preliminary question. I understand Rfree reflection sets are >> never used during automatic refinement, but, when generating the real >> space density maps, do we have to exclude Rfree columns? Any references >> will also be greatly appreciated! >> >> Best Regards, Hailiang > >
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Dear Keitaro As far as I know different program behave differently. REFMAC by default replaces structure factors of the excluded reflections with their expected values (as a first approximation) that is equal to DFc, where D reflects errors in coordinates. It seems to be a balance between avoiding model bias in map calculations (due to use of FC) and bias in Rfree (due to use of free reflections in minimisation) and noise introduction due to missing reflections. As far as I recall coot uses map coefficients produced by the refinement programs. In case of refmac it uses FWT, PHWT (that has free reflection replaced with their expected values) and DELFWT PHDELWT (that does not have free reflections). In refmac (I am sure in phenix also) there are keywords to turn this option on/off. I hope it helps. regards Garib On 24 May 2011, at 01:02, Keitaro Yamashita wrote: > Dear all, > > I'm very interested in this topic. > I have a question about the default behaviors in output reflection > files of each refinement softwares. > Are test set reflections excluded from the columns for calculating > electron density maps? > > I found in phenix.refine documentation the option > electron_density_maps.exclude_free_r_reflections was equal to False by > default. > (Does this option affect real space refinement in phenix.refine?) > > And I don't think Coot excludes test set reflections when opening MTZ > file... because there's no option to specify the flag number, right? > > Thanks in advance, > > Keitaro > > > 2011/5/24 Pavel Afonine : >> Hailiang, >> r-free reflections should not participate in refinement, regardless whether >> it is real or reciprocal space one, done with machine driven minimizers or >> your hands moving atoms in Coot. Period. The issue of correcting the >> map appearance for missing data (resolution or completeness) is relevant but >> different. Removing the data is noticeable, but most of the time putting >> aside test set is not critical for map appearance (given that reflections >> are selected randomly and do not exceed a reasonable fraction of available >> data); and when it is critical, the cross validation should be done >> differently anyway. >> So the answer to your question is: every time you compute a map not just to >> enjoy its appearance but to use it to improve your model, do not include >> test flagged reflections into it. >> Pavel. >> >> On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: >>> >>> Hi, >>> >>> I have a preliminary question. I understand Rfree reflection sets are >>> never used during automatic refinement, but, when generating the real >>> space density maps, do we have to exclude Rfree columns? Any references >>> will also be greatly appreciated! >>> >>> Best Regards, Hailiang >> >>
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
The default output for REFMAC Missing Data: For those reflections where the FP are missing, mFo is set equal to dFc. Hence the terms become FWT=dFC and DELFWT=0.0. the Rfree reflections are counted as "missing" hence there shouldnt be any bias intoroduced towards those Fobs assigned as free I dont think.. Of course all maps use the PhiCalc so there is inevitably bias towards to current model.. eleanor On 05/23/2011 11:03 PM, Pavel Afonine wrote: Hailiang, r-free reflections should not participate in refinement, regardless whether it is real or reciprocal space one, done with machine driven minimizers or your hands moving atoms in Coot. Period. The issue of correcting the map appearance for missing data (resolution or completeness) is relevant but different. Removing the data is noticeable, but most of the time putting aside test set is not critical for map appearance (given that reflections are selected randomly and do not exceed a reasonable fraction of available data); and when it is critical, the cross validation should be done differently anyway. So the answer to your question is: every time you compute a map not just to enjoy its appearance but to use it to improve your model, do not include test flagged reflections into it. Pavel. On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: Hi, I have a preliminary question. I understand Rfree reflection sets are never used during automatic refinement, but, when generating the real space density maps, do we have to exclude Rfree columns? Any references will also be greatly appreciated! Best Regards, Hailiang
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
If you run coot with the FWT/PHWT columns from refmac, you are fine because refmac omits the observed F and substitutes a calculated F, giving a map which is rather less biased by omission of the free set while not contaminating the free set. The buccaneer/refmac pipeline does the same. On Tue, 24 May 2011 09:02:20 +0900, Keitaro Yamashita wrote: > Dear all, > > I'm very interested in this topic. > I have a question about the default behaviors in output reflection > files of each refinement softwares. > Are test set reflections excluded from the columns for calculating > electron density maps? > > I found in phenix.refine documentation the option > electron_density_maps.exclude_free_r_reflections was equal to False by > default. > (Does this option affect real space refinement in phenix.refine?) > > And I don't think Coot excludes test set reflections when opening MTZ > file... because there's no option to specify the flag number, right? > > Thanks in advance, > > Keitaro > > > 2011/5/24 Pavel Afonine : >> Hailiang, >> r-free reflections should not participate in >> refinement, regardless whether >> it is real or reciprocal space one, done with machine driven >> minimizers or >> your hands moving atoms in Coot. Period. The issue of correcting the >> map appearance for missing data (resolution or completeness) is >> relevant but >> different. Removing the data is noticeable, but most of the time putting >> aside test set is not critical for map appearance (given that >> reflections >> are selected randomly and do not exceed a reasonable fraction of >> available >> data); and when it is critical, the cross validation should be done >> differently anyway. >> So the answer to your question is: every time you compute a map not just >> to >> enjoy its appearance but to use it to improve your model, do not >> include >> test flagged reflections into it. >> Pavel. >> >> On Mon, May 23, 2011 at 1:02 PM, Hailiang Zhang wrote: >>> >>> Hi, >>> >>> I have a preliminary question. I understand Rfree reflection sets are >>> never used during automatic refinement, but, when generating the real >>> space density maps, do we have to exclude Rfree columns? Any references >>> will also be greatly appreciated! >>> >>> Best Regards, Hailiang >> >>
Re: [ccp4bb] do we have to exclude Rfree columns when generating the real space density maps?
Thank you for replies. I understand that real space refinement using maps generated by REFMAC doesn't affect cross validation. I found the documentation of REFMAC about this topic. http://www.ysbl.york.ac.uk/refmac/data/refmac_keywords.html#mapcalc Oops, I should have found this earlier. Thanks again, Keitaro
