Re: [ccp4bb] Against Method (R)
On 10-10-29 12:03 AM, Robbie Joosten wrote: Hi Bart, I agree with the building strategy you propose, but at some point it stops helping and a bit more attention to detail is needed. Reciprocal space refinement doesn't seem to do the fine details. It always surprises me how much atoms still move when you real-space refine a refined model, especially the waters. I admit this is not a fair comparison. Does the water move back to its old position if you follow up the real-space refinement with more reciprocal refinement. If so, the map may not have been a true representation of reality. Basically what I was implying is that if the required model changes "details" are such that they fall within the radius of convergence then the atoms should move to their correct positions; unless something is keeping them from moving such as an incorrectly placed side chain that causes a steric conflict. Fix the incorrect side chain and your "details" will take care of themselves. I don't imply that I can always spot an easy error to fix and sometimes end up rebuilding several different ways in the hopes that one will resolve whatever was the problem. If that doesn't happen at some point you need to give up, especially if it does not affect a functionally important region. I do think it is good practice to point out regions in the model that are problematic and have never had reviewers complain about that if it is clear you made the effort to get it as good as possible given the data. High resolution data helps, but better data makes it tempting to put too little effort in optimising the model. I've seen some horribly obvious errors in hi-res models (more than 10 sigma difference density peaks for misplaced side chains). At the same time there are quite a lot of low-res models that are exceptionally good. Can't blame the data for that, in the end each person (and supervisor) need to take responsibility for the models they produce and deposit. Same applies to sequence data bases that are full of lazy errors. If humans are involved both greatness and stupidy are likely outcomes. Bart Cheers, Robbie > Date: Thu, 28 Oct 2010 16:32:04 -0600 > From: bart.ha...@ualberta.ca > Subject: Re: [ccp4bb] Against Method (R) > To: CCP4BB@JISCMAIL.AC.UK > > On 10-10-28 04:09 PM, Ethan Merritt wrote: > > This I can answer based on experience. One can take the coordinates from a structure > > refined at near atomic resolution (~1.0A), including multiple conformations, > > partial occupancy waters, etc, and use it to calculate R factors against a lower > > resolution (say 2.5A) data set collected from an isomorphous crystal. The > > R factors from this total-rigid-body replacement will be better than anything you > > could get from refinement against the lower resolution data. In fact, refinement > > from this starting point will just make the R factors worse. > > > > What this tells us is that the crystallographic residuals can recognize a > > better model when they see one. But our refinement programs are not good > > enough to produce such a better model in the first place. Worsr, they are not > > even good enough to avoid degrading the model. > > > > That's essentially the same thing Bart said, perhaps a little more pessimistic :-) > > > > cheers, > > > > Ethan > > > > Not pessimistic at all, just realistic and perhaps even optimistic for > methods developers as apparently there is still quite a bit of progress > that can be made by improving the "search strategy" during refinement. > > During manual refinement I normally tell students not to bother about > translating/rotating/torsioning atoms by just a tiny bit to make it fit > better. Likewise there is no point in moving atoms a little bit to > correct a distorted bond or bond length. If it needed to move that > little bit the refinement program would have done it for you. Look for > discreet errors in the problematic residue or its neighbors: peptide > flips, 120 degree changes in side chain dihedrals, etc. If you can find > and fix one of those errors a lot of the stereochemical distortions and > non-ideal fit to density surrounding that residue will suddenly > disappear as well. > > The benefit of high resolution is that it is much easier to pick up and > fix such errors (or not make them in the first place) > > Bart > > -- > > > > Bart Hazes (Associate Professor) > Dept. of Medical Microbiology& Immunology > University of Alberta > 1-15 Medical Sciences Building > Edmonton, Alberta > Canada, T6G 2H7 > phone: 1-780-492-0042 > fax: 1-780-492-7521 > > ===
Re: [ccp4bb] Against Method (R)
Dear George, thanks a lot! I see the point, that in reciprocal space refinement one could refine directly against the observed intensities and sigmas. But in principle, one could do iterative real space refinement, structure factor and intensity calculation for refinement statistics and weights, calculation of an improved electron density map (but that requires Fs again ...), and so forth until some convergence criterion is met. I wonder, which refinement scheme is more efficient. The missing reflections in map calculation is something that we have to live with, and unless the data are severely incomplete, I must admit, that I don't worry too much. The twinning problem is really severe! Here, I don't see how this could be done in a clever way in real space. Interesting discussion ... Best wishes, Dirk. Am 29.10.10 10:41, schrieb George M. Sheldrick: Dear Dirk, There are good reasons why real space refinement has never become popular. With reciprocal space refinement, you refine directly against what you measured, taking the standard uncertainly of each individual intensity into account. In this context I was pleased to read in CCP4bb that REFMAC will soon be refining against intensities (like SHELXL). Then the assumptions made (e.g. no distortion of the expected intensity distribution by e.g. NCS or twinning) and even 'bugs' in (c)truncate will no longer matter. If for some reason a reflection wasn't measured, then simply leaving it out it does not invalidate a recoprocal space refinement. The same applies to reflections that are reserved for Rfree. In contrast, the electron density is only theoretically correct if all reflections between 0,0,0 and infinity are included in the Fourier summation, For a twin it is even worse, because we don't know how to partition the difference between Fo^2 and Fc^2 between the twin components. None of the attempts to work around these problems are entirely convincing. Maps and real space refinement are invaluable in the intermediate stages of model building and correction, but the final refinement should be performed in reciprocal space. Best wishes, George Prof. George M. Sheldrick FRS Dept. Structural Chemistry, University of Goettingen, Tammannstr. 4, D37077 Goettingen, Germany Tel. +49-551-39-3021 or -3068 Fax. +49-551-39-22582 On Fri, 29 Oct 2010, Dirk Kostrewa wrote: Hi Robbie, yes, the apparently larger radius of convergence in real space refinement impresses me, too. Therefore, I usually do local real space refinement after manually correcting errors, either with Moloc at lower resolution or with Coot at higher resolution, prior to reciprocal space refinement. If I recall correctly, real space refinement was introduced by Robert Diamond in the 60s long before reciprocal space refinement. In the 90s Michael Chapman tried to revive it, but without much success, as far as I know. With the fast computers today, maybe the time has come again for real space refinement ... Best regards, Dirk. Am 29.10.10 08:03, schrieb Robbie Joosten: Hi Bart, I agree with the building strategy you propose, but at some point it stops helping and a bit more attention to detail is needed. Reciprocal space refinement doesn't seem to do the fine details. It always surprises me how much atoms still move when you real-space refine a refined model, especially the waters. I admit this is not a fair comparison. High resolution data helps, but better data makes it tempting to put too little effort in optimising the model. I've seen some horribly obvious errors in hi-res models (more than 10 sigma difference density peaks for misplaced side chains). At the same time there are quite a lot of low-res models that are exceptionally good. Cheers, Robbie Date: Thu, 28 Oct 2010 16:32:04 -0600 From: bart.ha...@ualberta.ca Subject: Re: [ccp4bb] Against Method (R) To: CCP4BB@JISCMAIL.AC.UK On 10-10-28 04:09 PM, Ethan Merritt wrote: This I can answer based on experience. One can take the coordinates from a structure refined at near atomic resolution (~1.0A), including multiple conformations, partial occupancy waters, etc, and use it to calculate R factors against a lower resolution (say 2.5A) data set collected from an isomorphous crystal. The R factors from this total-rigid-body replacement will be better than anything you could get from refinement against the lower resolution data. In fact, refinement from this starting point will just make the R factors worse. What this tells us is that the crystallographic residuals can recognize a better model when they see one. But our refinement programs are not good enough to produce such a better model in the first place. Worsr, they are not even good enough to avoid degrading the model. That's essentially the same thing Bart said, perhaps a little more pessimistic :-) cheers, Ethan Not pessimistic at all, just realistic and perhaps even optimistic for me
Re: [ccp4bb] Against Method (R)
Dear Dirk, There are good reasons why real space refinement has never become popular. With reciprocal space refinement, you refine directly against what you measured, taking the standard uncertainly of each individual intensity into account. In this context I was pleased to read in CCP4bb that REFMAC will soon be refining against intensities (like SHELXL). Then the assumptions made (e.g. no distortion of the expected intensity distribution by e.g. NCS or twinning) and even 'bugs' in (c)truncate will no longer matter. If for some reason a reflection wasn't measured, then simply leaving it out it does not invalidate a recoprocal space refinement. The same applies to reflections that are reserved for Rfree. In contrast, the electron density is only theoretically correct if all reflections between 0,0,0 and infinity are included in the Fourier summation, For a twin it is even worse, because we don't know how to partition the difference between Fo^2 and Fc^2 between the twin components. None of the attempts to work around these problems are entirely convincing. Maps and real space refinement are invaluable in the intermediate stages of model building and correction, but the final refinement should be performed in reciprocal space. Best wishes, George Prof. George M. Sheldrick FRS Dept. Structural Chemistry, University of Goettingen, Tammannstr. 4, D37077 Goettingen, Germany Tel. +49-551-39-3021 or -3068 Fax. +49-551-39-22582 On Fri, 29 Oct 2010, Dirk Kostrewa wrote: > Hi Robbie, > > yes, the apparently larger radius of convergence in real space refinement > impresses me, too. Therefore, I usually do local real space refinement after > manually correcting errors, either with Moloc at lower resolution or with Coot > at higher resolution, prior to reciprocal space refinement. > > If I recall correctly, real space refinement was introduced by Robert Diamond > in the 60s long before reciprocal space refinement. In the 90s Michael Chapman > tried to revive it, but without much success, as far as I know. With the fast > computers today, maybe the time has come again for real space refinement ... > > Best regards, > > Dirk. > > Am 29.10.10 08:03, schrieb Robbie Joosten: > > Hi Bart, > > > > I agree with the building strategy you propose, but at some point it stops > > helping and a bit more attention to detail is needed. Reciprocal space > > refinement doesn't seem to do the fine details. It always surprises me how > > much atoms still move when you real-space refine a refined model, especially > > the waters. I admit this is not a fair comparison. > > > > High resolution data helps, but better data makes it tempting to put too > > little effort in optimising the model. I've seen some horribly obvious > > errors in hi-res models (more than 10 sigma difference density peaks for > > misplaced side chains). At the same time there are quite a lot of low-res > > models that are exceptionally good. > > > > Cheers, > > Robbie > > > > > Date: Thu, 28 Oct 2010 16:32:04 -0600 > > > From: bart.ha...@ualberta.ca > > > Subject: Re: [ccp4bb] Against Method (R) > > > To: CCP4BB@JISCMAIL.AC.UK > > > > > > On 10-10-28 04:09 PM, Ethan Merritt wrote: > > > > This I can answer based on experience. One can take the coordinates > > from a structure > > > > refined at near atomic resolution (~1.0A), including multiple > > conformations, > > > > partial occupancy waters, etc, and use it to calculate R factors > > against a lower > > > > resolution (say 2.5A) data set collected from an isomorphous > > crystal. The > > > > R factors from this total-rigid-body replacement will be better > > than anything you > > > > could get from refinement against the lower resolution data. In > > fact, refinement > > > > from this starting point will just make the R factors worse. > > > > > > > > What this tells us is that the crystallographic residuals can > > recognize a > > > > better model when they see one. But our refinement programs are not > > good > > > > enough to produce such a better model in the first place. Worsr, > > they are not > > > > even good enough to avoid degrading the model. > > > > > > > > That's essentially the same thing Bart said, perhaps a little more > > pessimistic :-) > > > > > > > > cheers, > > > > > > > > Ethan > > > > > > > > > > Not pessimistic at all, just realistic and perhaps even optimistic for > > > methods developers as apparent
Re: [ccp4bb] Against Method (R)
Hi Robbie, yes, the apparently larger radius of convergence in real space refinement impresses me, too. Therefore, I usually do local real space refinement after manually correcting errors, either with Moloc at lower resolution or with Coot at higher resolution, prior to reciprocal space refinement. If I recall correctly, real space refinement was introduced by Robert Diamond in the 60s long before reciprocal space refinement. In the 90s Michael Chapman tried to revive it, but without much success, as far as I know. With the fast computers today, maybe the time has come again for real space refinement ... Best regards, Dirk. Am 29.10.10 08:03, schrieb Robbie Joosten: Hi Bart, I agree with the building strategy you propose, but at some point it stops helping and a bit more attention to detail is needed. Reciprocal space refinement doesn't seem to do the fine details. It always surprises me how much atoms still move when you real-space refine a refined model, especially the waters. I admit this is not a fair comparison. High resolution data helps, but better data makes it tempting to put too little effort in optimising the model. I've seen some horribly obvious errors in hi-res models (more than 10 sigma difference density peaks for misplaced side chains). At the same time there are quite a lot of low-res models that are exceptionally good. Cheers, Robbie > Date: Thu, 28 Oct 2010 16:32:04 -0600 > From: bart.ha...@ualberta.ca > Subject: Re: [ccp4bb] Against Method (R) > To: CCP4BB@JISCMAIL.AC.UK > > On 10-10-28 04:09 PM, Ethan Merritt wrote: > > This I can answer based on experience. One can take the coordinates from a structure > > refined at near atomic resolution (~1.0A), including multiple conformations, > > partial occupancy waters, etc, and use it to calculate R factors against a lower > > resolution (say 2.5A) data set collected from an isomorphous crystal. The > > R factors from this total-rigid-body replacement will be better than anything you > > could get from refinement against the lower resolution data. In fact, refinement > > from this starting point will just make the R factors worse. > > > > What this tells us is that the crystallographic residuals can recognize a > > better model when they see one. But our refinement programs are not good > > enough to produce such a better model in the first place. Worsr, they are not > > even good enough to avoid degrading the model. > > > > That's essentially the same thing Bart said, perhaps a little more pessimistic :-) > > > > cheers, > > > > Ethan > > > > Not pessimistic at all, just realistic and perhaps even optimistic for > methods developers as apparently there is still quite a bit of progress > that can be made by improving the "search strategy" during refinement. > > During manual refinement I normally tell students not to bother about > translating/rotating/torsioning atoms by just a tiny bit to make it fit > better. Likewise there is no point in moving atoms a little bit to > correct a distorted bond or bond length. If it needed to move that > little bit the refinement program would have done it for you. Look for > discreet errors in the problematic residue or its neighbors: peptide > flips, 120 degree changes in side chain dihedrals, etc. If you can find > and fix one of those errors a lot of the stereochemical distortions and > non-ideal fit to density surrounding that residue will suddenly > disappear as well. > > The benefit of high resolution is that it is much easier to pick up and > fix such errors (or not make them in the first place) > > Bart > > -- > > > > Bart Hazes (Associate Professor) > Dept. of Medical Microbiology& Immunology > University of Alberta > 1-15 Medical Sciences Building > Edmonton, Alberta > Canada, T6G 2H7 > phone: 1-780-492-0042 > fax: 1-780-492-7521 > > -- *** Dirk Kostrewa Gene Center Munich, A5.07 Department of Biochemistry Ludwig-Maximilians-Universität München Feodor-Lynen-Str. 25 D-81377 Munich Germany Phone: +49-89-2180-76845 Fax:+49-89-2180-76999 E-mail: kostr...@genzentrum.lmu.de WWW:www.genzentrum.lmu.de ***
Re: [ccp4bb] Against Method (R)
Hi Bart, I agree with the building strategy you propose, but at some point it stops helping and a bit more attention to detail is needed. Reciprocal space refinement doesn't seem to do the fine details. It always surprises me how much atoms still move when you real-space refine a refined model, especially the waters. I admit this is not a fair comparison. High resolution data helps, but better data makes it tempting to put too little effort in optimising the model. I've seen some horribly obvious errors in hi-res models (more than 10 sigma difference density peaks for misplaced side chains). At the same time there are quite a lot of low-res models that are exceptionally good. Cheers, Robbie > Date: Thu, 28 Oct 2010 16:32:04 -0600 > From: bart.ha...@ualberta.ca > Subject: Re: [ccp4bb] Against Method (R) > To: CCP4BB@JISCMAIL.AC.UK > > On 10-10-28 04:09 PM, Ethan Merritt wrote: > > This I can answer based on experience. One can take the coordinates from a > > structure > > refined at near atomic resolution (~1.0A), including multiple conformations, > > partial occupancy waters, etc, and use it to calculate R factors against a > > lower > > resolution (say 2.5A) data set collected from an isomorphous crystal. The > > R factors from this total-rigid-body replacement will be better than > > anything you > > could get from refinement against the lower resolution data. In fact, > > refinement > > from this starting point will just make the R factors worse. > > > > What this tells us is that the crystallographic residuals can recognize a > > better model when they see one. But our refinement programs are not good > > enough to produce such a better model in the first place. Worsr, they are > > not > > even good enough to avoid degrading the model. > > > > That's essentially the same thing Bart said, perhaps a little more > > pessimistic :-) > > > > cheers, > > > > Ethan > > > > Not pessimistic at all, just realistic and perhaps even optimistic for > methods developers as apparently there is still quite a bit of progress > that can be made by improving the "search strategy" during refinement. > > During manual refinement I normally tell students not to bother about > translating/rotating/torsioning atoms by just a tiny bit to make it fit > better. Likewise there is no point in moving atoms a little bit to > correct a distorted bond or bond length. If it needed to move that > little bit the refinement program would have done it for you. Look for > discreet errors in the problematic residue or its neighbors: peptide > flips, 120 degree changes in side chain dihedrals, etc. If you can find > and fix one of those errors a lot of the stereochemical distortions and > non-ideal fit to density surrounding that residue will suddenly > disappear as well. > > The benefit of high resolution is that it is much easier to pick up and > fix such errors (or not make them in the first place) > > Bart > > -- > > > > Bart Hazes (Associate Professor) > Dept. of Medical Microbiology& Immunology > University of Alberta > 1-15 Medical Sciences Building > Edmonton, Alberta > Canada, T6G 2H7 > phone: 1-780-492-0042 > fax:1-780-492-7521 > >
Re: [ccp4bb] Against Method (R)
On 10-10-28 04:09 PM, Ethan Merritt wrote: This I can answer based on experience. One can take the coordinates from a structure refined at near atomic resolution (~1.0A), including multiple conformations, partial occupancy waters, etc, and use it to calculate R factors against a lower resolution (say 2.5A) data set collected from an isomorphous crystal. The R factors from this total-rigid-body replacement will be better than anything you could get from refinement against the lower resolution data. In fact, refinement from this starting point will just make the R factors worse. What this tells us is that the crystallographic residuals can recognize a better model when they see one. But our refinement programs are not good enough to produce such a better model in the first place. Worsr, they are not even good enough to avoid degrading the model. That's essentially the same thing Bart said, perhaps a little more pessimistic :-) cheers, Ethan Not pessimistic at all, just realistic and perhaps even optimistic for methods developers as apparently there is still quite a bit of progress that can be made by improving the "search strategy" during refinement. During manual refinement I normally tell students not to bother about translating/rotating/torsioning atoms by just a tiny bit to make it fit better. Likewise there is no point in moving atoms a little bit to correct a distorted bond or bond length. If it needed to move that little bit the refinement program would have done it for you. Look for discreet errors in the problematic residue or its neighbors: peptide flips, 120 degree changes in side chain dihedrals, etc. If you can find and fix one of those errors a lot of the stereochemical distortions and non-ideal fit to density surrounding that residue will suddenly disappear as well. The benefit of high resolution is that it is much easier to pick up and fix such errors (or not make them in the first place) Bart -- Bart Hazes (Associate Professor) Dept. of Medical Microbiology& Immunology University of Alberta 1-15 Medical Sciences Building Edmonton, Alberta Canada, T6G 2H7 phone: 1-780-492-0042 fax:1-780-492-7521
Re: [ccp4bb] Against Method (R)
You're second suggestion would be a good test because you are dealing with data from the same crystal and can thus assume the structures are identical (radiation damage excluded). So, take a highly diffracting crystal and collect a short-exposure low resolution data set and long exposure high resolution data set. Let's say with I/Sig=2 at 2.0 and 1.2 high-resolution shells. Give the data to two equally capable students to determine the structure by molecular replacement from a, let's say 30% sequence identity starting model. You could also use automated model building to be more objective and avoid becoming unpopular with your students. Proceed until each model is fully refined against its own data. Now run some more refinement, without manual rebuilding, of the lowres model versus the highres data (and perhaps some rigid body or other minimal refinement of the highres model versus the lowres data, make sure R & Rfree go down). I predict the highres model will fit the lowres data noticeably better than the lowres model did and the lowres model, even after refinement with the highres data, will not reach the same quality as the highres model. Looking at Fo-Fc maps in the latter case may give some hints as to which model errors were not recognized at 2A resolution. You'll probably find peptide flips, mis-modeled leucine and other side chains, dual conformations not recognized at 2A resolution, more realistic B values, more waters ... Bart On 10-10-28 03:49 PM, Jacob Keller wrote: So let's say I take a 0.6 Ang structure, artificially introduce noise into corresponding Fobs to make the resolution go down to 2 Ang, and refine using the 0.6 Ang model--do I actually get R's better than the artificially-inflated sigmas? Or let's say I experimentally decrease I/sigma by attenuating the beam and collect another data set--same situation? JPK - Original Message - From: Bart Hazes To: CCP4BB@JISCMAIL.AC.UK Sent: Thursday, October 28, 2010 4:13 PM Subject: Re: [ccp4bb] Against Method (R) There are many cases where people use a structure refined at high resolution as a starting molecular replacement structure for a closely related/same protein with a lower resolution data set and get substantially better R statistics than you would expect for that resolution. So one factor in the "R factor gap" is many small errors that are introduced during model building and not recognized and fixed later due to limited resolution. In a perfect world, refinement would find the global minimum but in practice all these little errors get stuck in local minima with distortions in neighboring atoms compensating for the initial error and thereby hiding their existence. Bart On 10-10-28 11:33 AM, James Holton wrote: It is important to remember that if you have Gaussian-distributed errors and you plot error bars between +1 sigma and -1 sigma (where "sigma" is the rms error), then you expect the "right" curve to miss the error bars about 30% of the time. This is just a property of the Gaussian distribution: you expect a certain small number of the errors to be large. If the curve passes within the bounds of every single one of your error bars, then your error estimates are either too big, or the errors have a non-Gaussian distribution. For example, if the noise in the data somehow had a uniform distribution (always between +1 and -1), then no data point will ever be "kicked" further than "1" away from the "right" curve. In this case, a data point more than "1" away from the curve is evidence that you either have the wrong model (curve), or there is some other kind of noise around (wrong "error model"). As someone who has spent a lot of time looking into how we measure intensities, I think I can say with some considerable amount of confidence that we are doing a pretty good job of estimating the errors. At least, they are certainly not off by an average of 40% (20% in F). You could do better than that estimating the intensities by eye! Everybody seems to have their own favorite explanation for what I call the "R factor gap": solvent, multi-confomer structures, absorption effects, etc. However, if you go through the literature (old and new) you will find countless attempts to include more sophisticated versions of each of these hypothetically "important" systematic errors, and in none of these cases has anyone ever presented a physically reasonable model that explained the observed spot intensities from a protein crystal to within experimental error. Or at least, if there is such a paper, I haven't seen it. Since there are so many possible things to "correct", what I would like to find is a structure that represents the transition between the "small molecul
Re: [ccp4bb] Against Method (R)
Bart Hazes wrote > > There are many cases where people use a structure refined at high > resolution as a starting molecular replacement structure for a closely > related/same protein with a lower resolution data set and get substantially > better R statistics than you would expect for that resolution. So one factor > in the "R factor gap" is many small errors that are introduced during model > building and not recognized and fixed later due to limited resolution. In a > perfect world, refinement would find the global minimum but in practice all > these little errors get stuck in local minima with distortions in neighboring > atoms compensating for the initial error and thereby hiding their existence. Excellent point. On Thursday, October 28, 2010 02:49:11 pm Jacob Keller wrote: > So let's say I take a 0.6 Ang structure, artificially introduce noise into > corresponding Fobs to make the resolution go down to 2 Ang, and refine using > the 0.6 Ang model--do I actually get R's better than the > artificially-inflated sigmas? > Or let's say I experimentally decrease I/sigma by attenuating the beam and > collect another data set--same situation? This I can answer based on experience. One can take the coordinates from a structure refined at near atomic resolution (~1.0A), including multiple conformations, partial occupancy waters, etc, and use it to calculate R factors against a lower resolution (say 2.5A) data set collected from an isomorphous crystal. The R factors from this total-rigid-body replacement will be better than anything you could get from refinement against the lower resolution data. In fact, refinement from this starting point will just make the R factors worse. What this tells us is that the crystallographic residuals can recognize a better model when they see one. But our refinement programs are not good enough to produce such a better model in the first place. Worsr, they are not even good enough to avoid degrading the model. That's essentially the same thing Bart said, perhaps a little more pessimistic :-) cheers, Ethan > > JPK > > - Original Message - > From: Bart Hazes > To: CCP4BB@JISCMAIL.AC.UK > Sent: Thursday, October 28, 2010 4:13 PM > Subject: Re: [ccp4bb] Against Method (R) > > > There are many cases where people use a structure refined at high > resolution as a starting molecular replacement structure for a closely > related/same protein with a lower resolution data set and get substantially > better R statistics than you would expect for that resolution. So one factor > in the "R factor gap" is many small errors that are introduced during model > building and not recognized and fixed later due to limited resolution. In a > perfect world, refinement would find the global minimum but in practice all > these little errors get stuck in local minima with distortions in neighboring > atoms compensating for the initial error and thereby hiding their existence. > > Bart > > On 10-10-28 11:33 AM, James Holton wrote: > It is important to remember that if you have Gaussian-distributed errors > and you plot error bars between +1 sigma and -1 sigma (where "sigma" is the > rms error), then you expect the "right" curve to miss the error bars about > 30% of the time. This is just a property of the Gaussian distribution: you > expect a certain small number of the errors to be large. If the curve passes > within the bounds of every single one of your error bars, then your error > estimates are either too big, or the errors have a non-Gaussian distribution. > > > For example, if the noise in the data somehow had a uniform distribution > (always between +1 and -1), then no data point will ever be "kicked" further > than "1" away from the "right" curve. In this case, a data point more than > "1" away from the curve is evidence that you either have the wrong model > (curve), or there is some other kind of noise around (wrong "error model"). > > As someone who has spent a lot of time looking into how we measure > intensities, I think I can say with some considerable amount of confidence > that we are doing a pretty good job of estimating the errors. At least, they > are certainly not off by an average of 40% (20% in F). You could do better > than that estimating the intensities by eye! > > Everybody seems to have their own favorite explanation for what I call > the "R factor gap": solvent, multi-confomer structures, absorption effects, > etc. However, if you go through the literature (old and new) you will find > countless attempts to include more sophisti
Re: [ccp4bb] Against Method (R)
re are so many possible things to "correct", what I would like to > find is a structure that represents the transition between the "small > molecule" and the "macromolecule" world. Lysozyme does not qualify! Even > the famous 0.6 A structure of lysozyme (2vb1) still has a "mean absolute > chi": <|Iobs-Icalc|/sig(I)> = 4.5. Also, the 1.4 A structure of the > tetrapeptide QQNN (2olx) is only a little better at <|chi|> = 3.5. I > realize that the "chi" I describe here is not a "standard" crystallographic > statistic, and perhaps I need a statistics lesson, but it seems to me there > ought to be a case where it is close to 1. > > -James Holton > MAD Scientist > > On Thu, Oct 28, 2010 at 9:04 AM, Jacob Keller < > j-kell...@fsm.northwestern.edu> wrote: > >> So I guess there is never a case in crystallography in which our >> models predict the data to within the errors of data collection? I >> guess the situation might be similar to fitting a Michaelis-Menten >> curve, in which the fitted line often misses the error bars of the >> individual points, but gets the overall pattern right. In that case, >> though, I don't think we say that we are inadequately modelling the >> data. I guess there the error bars are actually too small (are >> underestimated.) Maybe our intensity errors are also underestimated? >> >> JPK >> >> On Thu, Oct 28, 2010 at 9:50 AM, George M. Sheldrick >> wrote: >> > >> > Not quite. I was trying to say that for good small molecule data, R1 is >> > usally significantly less than Rmerge, but never less than the precision >> > of the experimental data measured by 0.5*/ = 0.5*Rsigma >> > (or the very similar 0.5*Rpim). >> > >> > George >> > >> > Prof. George M. Sheldrick FRS >> > Dept. Structural Chemistry, >> > University of Goettingen, >> > Tammannstr. 4, >> > D37077 Goettingen, Germany >> > Tel. +49-551-39-3021 or -3068 >> > Fax. +49-551-39-22582 >> > >> > >> > On Thu, 28 Oct 2010, Jacob Keller wrote: >> > >> >> So I guess a consequence of what you say is that since in cases where >> there is >> >> no solvent the R values are often better than the precision of the >> actual >> >> measurements (never true with macromolecular crystals involving >> solvent), >> >> perhaps our real problem might be modelling solvent? >> >> Alternatively/additionally, I wonder whether there also might be more >> >> variability molecule-to-molecule in proteins, which we may not model >> well >> >> either. >> >> >> >> JPK >> >> >> >> - Original Message - From: "George M. Sheldrick" >> >> >> >> To: >> >> Sent: Thursday, October 28, 2010 4:05 AM >> >> Subject: Re: [ccp4bb] Against Method (R) >> >> >> >> >> >> > It is instructive to look at what happens for small molecules where >> >> > there is often no solvent to worry about. They are often refined >> >> > using SHELXL, which does indeed print out the weighted R-value based >> >> > on intensities (wR2), the conventional unweighted R-value R1 (based >> >> > on F) and /, which it calls R(sigma). For well-behaved >> >> > crystals R1 is in the range 1-5% and R(merge) (based on intensities) >> >> > is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded >> >> > as the lower attainable limit for R1 and this is indeed the case in >> >> > practice (the factor 0.5 approximately converts from I to F). Rpim >> >> > gives similar results to R(sigma), both attempt to measure the >> >> > precision of the MERGED data, which are what one is refining against. >> >> > >> >> > George >> >> > >> >> > Prof. George M. Sheldrick FRS >> >> > Dept. Structural Chemistry, >> >> > University of Goettingen, >> >> > Tammannstr. 4, >> >> > D37077 Goettingen, Germany >> >> > Tel. +49-551-39-3021 or -3068 >> >> > Fax. +49-551-39-22582 >> >> > >> >> > >> >> > On Wed, 27 Oct 2010, Ed Pozharski wrote: >> >> > >> >> > > On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: >> >> > > > the errors in our measurements apparently have no >> >> > > >
Re: [ccp4bb] Against Method (R)
So let's say I take a 0.6 Ang structure, artificially introduce noise into corresponding Fobs to make the resolution go down to 2 Ang, and refine using the 0.6 Ang model--do I actually get R's better than the artificially-inflated sigmas? Or let's say I experimentally decrease I/sigma by attenuating the beam and collect another data set--same situation? JPK - Original Message - From: Bart Hazes To: CCP4BB@JISCMAIL.AC.UK Sent: Thursday, October 28, 2010 4:13 PM Subject: Re: [ccp4bb] Against Method (R) There are many cases where people use a structure refined at high resolution as a starting molecular replacement structure for a closely related/same protein with a lower resolution data set and get substantially better R statistics than you would expect for that resolution. So one factor in the "R factor gap" is many small errors that are introduced during model building and not recognized and fixed later due to limited resolution. In a perfect world, refinement would find the global minimum but in practice all these little errors get stuck in local minima with distortions in neighboring atoms compensating for the initial error and thereby hiding their existence. Bart On 10-10-28 11:33 AM, James Holton wrote: It is important to remember that if you have Gaussian-distributed errors and you plot error bars between +1 sigma and -1 sigma (where "sigma" is the rms error), then you expect the "right" curve to miss the error bars about 30% of the time. This is just a property of the Gaussian distribution: you expect a certain small number of the errors to be large. If the curve passes within the bounds of every single one of your error bars, then your error estimates are either too big, or the errors have a non-Gaussian distribution. For example, if the noise in the data somehow had a uniform distribution (always between +1 and -1), then no data point will ever be "kicked" further than "1" away from the "right" curve. In this case, a data point more than "1" away from the curve is evidence that you either have the wrong model (curve), or there is some other kind of noise around (wrong "error model"). As someone who has spent a lot of time looking into how we measure intensities, I think I can say with some considerable amount of confidence that we are doing a pretty good job of estimating the errors. At least, they are certainly not off by an average of 40% (20% in F). You could do better than that estimating the intensities by eye! Everybody seems to have their own favorite explanation for what I call the "R factor gap": solvent, multi-confomer structures, absorption effects, etc. However, if you go through the literature (old and new) you will find countless attempts to include more sophisticated versions of each of these hypothetically "important" systematic errors, and in none of these cases has anyone ever presented a physically reasonable model that explained the observed spot intensities from a protein crystal to within experimental error. Or at least, if there is such a paper, I haven't seen it. Since there are so many possible things to "correct", what I would like to find is a structure that represents the transition between the "small molecule" and the "macromolecule" world. Lysozyme does not qualify! Even the famous 0.6 A structure of lysozyme (2vb1) still has a "mean absolute chi": <|Iobs-Icalc|/sig(I)> = 4.5. Also, the 1.4 A structure of the tetrapeptide QQNN (2olx) is only a little better at <|chi|> = 3.5. I realize that the "chi" I describe here is not a "standard" crystallographic statistic, and perhaps I need a statistics lesson, but it seems to me there ought to be a case where it is close to 1. -James Holton MAD Scientist On Thu, Oct 28, 2010 at 9:04 AM, Jacob Keller wrote: So I guess there is never a case in crystallography in which our models predict the data to within the errors of data collection? I guess the situation might be similar to fitting a Michaelis-Menten curve, in which the fitted line often misses the error bars of the individual points, but gets the overall pattern right. In that case, though, I don't think we say that we are inadequately modelling the data. I guess there the error bars are actually too small (are underestimated.) Maybe our intensity errors are also underestimated? JPK On Thu, Oct 28, 2010 at 9:50 AM, George M. Sheldrick wrote: > > Not quite. I was trying to say that for good small molecule data, R1 is > usally significantly less than Rmerge, but never less than the precision > of the experimental data measured by 0.5*/ = 0.
Re: [ccp4bb] Against Method (R)
blem might be modelling solvent? >> Alternatively/additionally, I wonder whether there also might be more >> variability molecule-to-molecule in proteins, which we may not model well >> either. >> >> JPK >> >> - Original Message - From: "George M. Sheldrick" >> <gshe...@shelx.uni-ac.gwdg.de> >> To: <CCP4BB@JISCMAIL.AC.UK> >> Sent: Thursday, October 28, 2010 4:05 AM >> Subject: Re: [ccp4bb] Against Method (R) >> >> >> > It is instructive to look at what happens for small molecules where >> > there is often no solvent to worry about. They are often refined >> > using SHELXL, which does indeed print out the weighted R-value based >> > on intensities (wR2), the conventional unweighted R-value R1 (based >> > on F) and /, which it calls R(sigma). For well-behaved >> > crystals R1 is in the range 1-5% and R(merge) (based on intensities) >> > is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded >> > as the lower attainable limit for R1 and this is indeed the case in >> > practice (the factor 0.5 approximately converts from I to F). Rpim >> > gives similar results to R(sigma), both attempt to measure the >> > precision of the MERGED data, which are what one is refining against. >> > >> > George >> > >> > Prof. George M. Sheldrick FRS >> > Dept. Structural Chemistry, >> > University of Goettingen, >> > Tammannstr. 4, >> > D37077 Goettingen, Germany >> > Tel. +49-551-39-3021 or -3068 >> > Fax. +49-551-39-22582 >> > >> > >> > On Wed, 27 Oct 2010, Ed Pozharski wrote: >> > >> > > On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: >> > > > the errors in our measurements apparently have no >> > > > bearing whatsoever on the errors in our models >> > > >> > > This would mean there is no point trying to get better crystals, right? >> > > Or am I also wrong to assume that the dataset with higher I/sigma in the >> > > highest resolution shell will give me a better model? >> > > >> > > On a related point - why is Rmerge considered to be the limiting value >> > > for the R? Isn't Rmerge a poorly defined measure itself that >> > > deteriorates at least in some circumstances (e.g. increased redundancy)? >> > > Specifically, shouldn't "ideal" R approximate 0.5*/? >> > > >> > > Cheers, >> > > >> > > Ed. >> > > >> > > >> > > >> > > -- >> > > "I'd jump in myself, if I weren't so good at whistling." >> > > Julian, King of Lemurs >> > > >> > > >> >> >> *** >> Jacob Pearson Keller >> Northwestern University >> Medical Scientist Training Program >> Dallos Laboratory >> F. Searle 1-240 >> 2240 Campus Drive >> Evanston IL 60208 >> lab: 847.491.2438 >> cel: 773.608.9185 >> email: j-kell...@northwestern.edu >> *** >> >> > -- Bart Hazes (Associate Professor) Dept. of Medical Microbiology & Immunology University of Alberta 1-15 Medical Sciences Building Edmonton, Alberta Canada, T6G 2H7 phone: 1-780-492-0042 fax:1-780-492-7521
Re: [ccp4bb] Against Method (R)
It is important to remember that if you have Gaussian-distributed errors and you plot error bars between +1 sigma and -1 sigma (where "sigma" is the rms error), then you expect the "right" curve to miss the error bars about 30% of the time. This is just a property of the Gaussian distribution: you expect a certain small number of the errors to be large. If the curve passes within the bounds of every single one of your error bars, then your error estimates are either too big, or the errors have a non-Gaussian distribution. For example, if the noise in the data somehow had a uniform distribution (always between +1 and -1), then no data point will ever be "kicked" further than "1" away from the "right" curve. In this case, a data point more than "1" away from the curve is evidence that you either have the wrong model (curve), or there is some other kind of noise around (wrong "error model"). As someone who has spent a lot of time looking into how we measure intensities, I think I can say with some considerable amount of confidence that we are doing a pretty good job of estimating the errors. At least, they are certainly not off by an average of 40% (20% in F). You could do better than that estimating the intensities by eye! Everybody seems to have their own favorite explanation for what I call the "R factor gap": solvent, multi-confomer structures, absorption effects, etc. However, if you go through the literature (old and new) you will find countless attempts to include more sophisticated versions of each of these hypothetically "important" systematic errors, and in none of these cases has anyone ever presented a physically reasonable model that explained the observed spot intensities from a protein crystal to within experimental error. Or at least, if there is such a paper, I haven't seen it. Since there are so many possible things to "correct", what I would like to find is a structure that represents the transition between the "small molecule" and the "macromolecule" world. Lysozyme does not qualify! Even the famous 0.6 A structure of lysozyme (2vb1) still has a "mean absolute chi": <|Iobs-Icalc|/sig(I)> = 4.5. Also, the 1.4 A structure of the tetrapeptide QQNN (2olx) is only a little better at <|chi|> = 3.5. I realize that the "chi" I describe here is not a "standard" crystallographic statistic, and perhaps I need a statistics lesson, but it seems to me there ought to be a case where it is close to 1. -James Holton MAD Scientist On Thu, Oct 28, 2010 at 9:04 AM, Jacob Keller < j-kell...@fsm.northwestern.edu> wrote: > So I guess there is never a case in crystallography in which our > models predict the data to within the errors of data collection? I > guess the situation might be similar to fitting a Michaelis-Menten > curve, in which the fitted line often misses the error bars of the > individual points, but gets the overall pattern right. In that case, > though, I don't think we say that we are inadequately modelling the > data. I guess there the error bars are actually too small (are > underestimated.) Maybe our intensity errors are also underestimated? > > JPK > > On Thu, Oct 28, 2010 at 9:50 AM, George M. Sheldrick > wrote: > > > > Not quite. I was trying to say that for good small molecule data, R1 is > > usally significantly less than Rmerge, but never less than the precision > > of the experimental data measured by 0.5*/ = 0.5*Rsigma > > (or the very similar 0.5*Rpim). > > > > George > > > > Prof. George M. Sheldrick FRS > > Dept. Structural Chemistry, > > University of Goettingen, > > Tammannstr. 4, > > D37077 Goettingen, Germany > > Tel. +49-551-39-3021 or -3068 > > Fax. +49-551-39-22582 > > > > > > On Thu, 28 Oct 2010, Jacob Keller wrote: > > > >> So I guess a consequence of what you say is that since in cases where > there is > >> no solvent the R values are often better than the precision of the > actual > >> measurements (never true with macromolecular crystals involving > solvent), > >> perhaps our real problem might be modelling solvent? > >> Alternatively/additionally, I wonder whether there also might be more > >> variability molecule-to-molecule in proteins, which we may not model > well > >> either. > >> > >> JPK > >> > >> - Original Message - From: "George M. Sheldrick" > >> > >> To: > >> Sent: Thursday, October 28, 2010 4:05 AM > >> Subject: Re: [ccp4bb] Against Method (R) > >> > >> > >> > It is instructive to look at what happens for small molecu
Re: [ccp4bb] Against Method (R)
So I guess there is never a case in crystallography in which our models predict the data to within the errors of data collection? I guess the situation might be similar to fitting a Michaelis-Menten curve, in which the fitted line often misses the error bars of the individual points, but gets the overall pattern right. In that case, though, I don't think we say that we are inadequately modelling the data. I guess there the error bars are actually too small (are underestimated.) Maybe our intensity errors are also underestimated? JPK On Thu, Oct 28, 2010 at 9:50 AM, George M. Sheldrick wrote: > > Not quite. I was trying to say that for good small molecule data, R1 is > usally significantly less than Rmerge, but never less than the precision > of the experimental data measured by 0.5*/ = 0.5*Rsigma > (or the very similar 0.5*Rpim). > > George > > Prof. George M. Sheldrick FRS > Dept. Structural Chemistry, > University of Goettingen, > Tammannstr. 4, > D37077 Goettingen, Germany > Tel. +49-551-39-3021 or -3068 > Fax. +49-551-39-22582 > > > On Thu, 28 Oct 2010, Jacob Keller wrote: > >> So I guess a consequence of what you say is that since in cases where there >> is >> no solvent the R values are often better than the precision of the actual >> measurements (never true with macromolecular crystals involving solvent), >> perhaps our real problem might be modelling solvent? >> Alternatively/additionally, I wonder whether there also might be more >> variability molecule-to-molecule in proteins, which we may not model well >> either. >> >> JPK >> >> ----- Original Message - From: "George M. Sheldrick" >> >> To: >> Sent: Thursday, October 28, 2010 4:05 AM >> Subject: Re: [ccp4bb] Against Method (R) >> >> >> > It is instructive to look at what happens for small molecules where >> > there is often no solvent to worry about. They are often refined >> > using SHELXL, which does indeed print out the weighted R-value based >> > on intensities (wR2), the conventional unweighted R-value R1 (based >> > on F) and /, which it calls R(sigma). For well-behaved >> > crystals R1 is in the range 1-5% and R(merge) (based on intensities) >> > is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded >> > as the lower attainable limit for R1 and this is indeed the case in >> > practice (the factor 0.5 approximately converts from I to F). Rpim >> > gives similar results to R(sigma), both attempt to measure the >> > precision of the MERGED data, which are what one is refining against. >> > >> > George >> > >> > Prof. George M. Sheldrick FRS >> > Dept. Structural Chemistry, >> > University of Goettingen, >> > Tammannstr. 4, >> > D37077 Goettingen, Germany >> > Tel. +49-551-39-3021 or -3068 >> > Fax. +49-551-39-22582 >> > >> > >> > On Wed, 27 Oct 2010, Ed Pozharski wrote: >> > >> > > On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: >> > > > the errors in our measurements apparently have no >> > > > bearing whatsoever on the errors in our models >> > > >> > > This would mean there is no point trying to get better crystals, right? >> > > Or am I also wrong to assume that the dataset with higher I/sigma in the >> > > highest resolution shell will give me a better model? >> > > >> > > On a related point - why is Rmerge considered to be the limiting value >> > > for the R? Isn't Rmerge a poorly defined measure itself that >> > > deteriorates at least in some circumstances (e.g. increased redundancy)? >> > > Specifically, shouldn't "ideal" R approximate 0.5*/? >> > > >> > > Cheers, >> > > >> > > Ed. >> > > >> > > >> > > >> > > -- >> > > "I'd jump in myself, if I weren't so good at whistling." >> > > Julian, King of Lemurs >> > > >> > > >> >> >> *** >> Jacob Pearson Keller >> Northwestern University >> Medical Scientist Training Program >> Dallos Laboratory >> F. Searle 1-240 >> 2240 Campus Drive >> Evanston IL 60208 >> lab: 847.491.2438 >> cel: 773.608.9185 >> email: j-kell...@northwestern.edu >> *** >> >> >
Re: [ccp4bb] Against Method (R)
Not quite. I was trying to say that for good small molecule data, R1 is usally significantly less than Rmerge, but never less than the precision of the experimental data measured by 0.5*/ = 0.5*Rsigma (or the very similar 0.5*Rpim). George Prof. George M. Sheldrick FRS Dept. Structural Chemistry, University of Goettingen, Tammannstr. 4, D37077 Goettingen, Germany Tel. +49-551-39-3021 or -3068 Fax. +49-551-39-22582 On Thu, 28 Oct 2010, Jacob Keller wrote: > So I guess a consequence of what you say is that since in cases where there is > no solvent the R values are often better than the precision of the actual > measurements (never true with macromolecular crystals involving solvent), > perhaps our real problem might be modelling solvent? > Alternatively/additionally, I wonder whether there also might be more > variability molecule-to-molecule in proteins, which we may not model well > either. > > JPK > > - Original Message - From: "George M. Sheldrick" > > To: > Sent: Thursday, October 28, 2010 4:05 AM > Subject: Re: [ccp4bb] Against Method (R) > > > > It is instructive to look at what happens for small molecules where > > there is often no solvent to worry about. They are often refined > > using SHELXL, which does indeed print out the weighted R-value based > > on intensities (wR2), the conventional unweighted R-value R1 (based > > on F) and /, which it calls R(sigma). For well-behaved > > crystals R1 is in the range 1-5% and R(merge) (based on intensities) > > is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded > > as the lower attainable limit for R1 and this is indeed the case in > > practice (the factor 0.5 approximately converts from I to F). Rpim > > gives similar results to R(sigma), both attempt to measure the > > precision of the MERGED data, which are what one is refining against. > > > > George > > > > Prof. George M. Sheldrick FRS > > Dept. Structural Chemistry, > > University of Goettingen, > > Tammannstr. 4, > > D37077 Goettingen, Germany > > Tel. +49-551-39-3021 or -3068 > > Fax. +49-551-39-22582 > > > > > > On Wed, 27 Oct 2010, Ed Pozharski wrote: > > > > > On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: > > > > the errors in our measurements apparently have no > > > > bearing whatsoever on the errors in our models > > > > > > This would mean there is no point trying to get better crystals, right? > > > Or am I also wrong to assume that the dataset with higher I/sigma in the > > > highest resolution shell will give me a better model? > > > > > > On a related point - why is Rmerge considered to be the limiting value > > > for the R? Isn't Rmerge a poorly defined measure itself that > > > deteriorates at least in some circumstances (e.g. increased redundancy)? > > > Specifically, shouldn't "ideal" R approximate 0.5*/? > > > > > > Cheers, > > > > > > Ed. > > > > > > > > > > > > -- > > > "I'd jump in myself, if I weren't so good at whistling." > > >Julian, King of Lemurs > > > > > > > > > *** > Jacob Pearson Keller > Northwestern University > Medical Scientist Training Program > Dallos Laboratory > F. Searle 1-240 > 2240 Campus Drive > Evanston IL 60208 > lab: 847.491.2438 > cel: 773.608.9185 > email: j-kell...@northwestern.edu > *** > >
Re: [ccp4bb] Against Method (R)
In addition to bulk solvent, the other well recognized problem with macromolecular structures is the inadequate description of disorder. With small molecules, the Debye-Waller works much better because the harmonic oscillator is indeed a good model there. Note that the problem is not anisotropy (which we can model if resolution is sufficiently high), but rather anharmonic motion and multiple conformations that go undetected. On Thu, 2010-10-28 at 08:00 -0500, Jacob Keller wrote: > So I guess a consequence of what you say is that since in cases where there > is no solvent the R values are often better than the precision of the actual > measurements (never true with macromolecular crystals involving solvent), > perhaps our real problem might be modelling solvent? > Alternatively/additionally, I wonder whether there also might be more > variability molecule-to-molecule in proteins, which we may not model well > either. > > JPK > > - Original Message - > From: "George M. Sheldrick" > To: > Sent: Thursday, October 28, 2010 4:05 AM > Subject: Re: [ccp4bb] Against Method (R) > > > > It is instructive to look at what happens for small molecules where > > there is often no solvent to worry about. They are often refined > > using SHELXL, which does indeed print out the weighted R-value based > > on intensities (wR2), the conventional unweighted R-value R1 (based > > on F) and /, which it calls R(sigma). For well-behaved > > crystals R1 is in the range 1-5% and R(merge) (based on intensities) > > is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded > > as the lower attainable limit for R1 and this is indeed the case in > > practice (the factor 0.5 approximately converts from I to F). Rpim > > gives similar results to R(sigma), both attempt to measure the > > precision of the MERGED data, which are what one is refining against. > > > > George > > > > Prof. George M. Sheldrick FRS > > Dept. Structural Chemistry, > > University of Goettingen, > > Tammannstr. 4, > > D37077 Goettingen, Germany > > Tel. +49-551-39-3021 or -3068 > > Fax. +49-551-39-22582 > > > > > > On Wed, 27 Oct 2010, Ed Pozharski wrote: > > > >> On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: > >> > the errors in our measurements apparently have no > >> > bearing whatsoever on the errors in our models > >> > >> This would mean there is no point trying to get better crystals, right? > >> Or am I also wrong to assume that the dataset with higher I/sigma in the > >> highest resolution shell will give me a better model? > >> > >> On a related point - why is Rmerge considered to be the limiting value > >> for the R? Isn't Rmerge a poorly defined measure itself that > >> deteriorates at least in some circumstances (e.g. increased redundancy)? > >> Specifically, shouldn't "ideal" R approximate 0.5*/? > >> > >> Cheers, > >> > >> Ed. > >> > >> > >> > >> -- > >> "I'd jump in myself, if I weren't so good at whistling." > >>Julian, King of Lemurs > >> > >> > > > *** > Jacob Pearson Keller > Northwestern University > Medical Scientist Training Program > Dallos Laboratory > F. Searle 1-240 > 2240 Campus Drive > Evanston IL 60208 > lab: 847.491.2438 > cel: 773.608.9185 > email: j-kell...@northwestern.edu > *** -- "I'd jump in myself, if I weren't so good at whistling." Julian, King of Lemurs
Re: [ccp4bb] Against Method (R)
So I guess a consequence of what you say is that since in cases where there is no solvent the R values are often better than the precision of the actual measurements (never true with macromolecular crystals involving solvent), perhaps our real problem might be modelling solvent? Alternatively/additionally, I wonder whether there also might be more variability molecule-to-molecule in proteins, which we may not model well either. JPK - Original Message - From: "George M. Sheldrick" To: Sent: Thursday, October 28, 2010 4:05 AM Subject: Re: [ccp4bb] Against Method (R) It is instructive to look at what happens for small molecules where there is often no solvent to worry about. They are often refined using SHELXL, which does indeed print out the weighted R-value based on intensities (wR2), the conventional unweighted R-value R1 (based on F) and /, which it calls R(sigma). For well-behaved crystals R1 is in the range 1-5% and R(merge) (based on intensities) is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded as the lower attainable limit for R1 and this is indeed the case in practice (the factor 0.5 approximately converts from I to F). Rpim gives similar results to R(sigma), both attempt to measure the precision of the MERGED data, which are what one is refining against. George Prof. George M. Sheldrick FRS Dept. Structural Chemistry, University of Goettingen, Tammannstr. 4, D37077 Goettingen, Germany Tel. +49-551-39-3021 or -3068 Fax. +49-551-39-22582 On Wed, 27 Oct 2010, Ed Pozharski wrote: On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: > the errors in our measurements apparently have no > bearing whatsoever on the errors in our models This would mean there is no point trying to get better crystals, right? Or am I also wrong to assume that the dataset with higher I/sigma in the highest resolution shell will give me a better model? On a related point - why is Rmerge considered to be the limiting value for the R? Isn't Rmerge a poorly defined measure itself that deteriorates at least in some circumstances (e.g. increased redundancy)? Specifically, shouldn't "ideal" R approximate 0.5*/? Cheers, Ed. -- "I'd jump in myself, if I weren't so good at whistling." Julian, King of Lemurs *** Jacob Pearson Keller Northwestern University Medical Scientist Training Program Dallos Laboratory F. Searle 1-240 2240 Campus Drive Evanston IL 60208 lab: 847.491.2438 cel: 773.608.9185 email: j-kell...@northwestern.edu ***
Re: [ccp4bb] Against Method (R)
It is instructive to look at what happens for small molecules where there is often no solvent to worry about. They are often refined using SHELXL, which does indeed print out the weighted R-value based on intensities (wR2), the conventional unweighted R-value R1 (based on F) and /, which it calls R(sigma). For well-behaved crystals R1 is in the range 1-5% and R(merge) (based on intensities) is in the range 3-9%. As you suggest, 0.5*R(sigma) could be regarded as the lower attainable limit for R1 and this is indeed the case in practice (the factor 0.5 approximately converts from I to F). Rpim gives similar results to R(sigma), both attempt to measure the precision of the MERGED data, which are what one is refining against. George Prof. George M. Sheldrick FRS Dept. Structural Chemistry, University of Goettingen, Tammannstr. 4, D37077 Goettingen, Germany Tel. +49-551-39-3021 or -3068 Fax. +49-551-39-22582 On Wed, 27 Oct 2010, Ed Pozharski wrote: > On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: > > the errors in our measurements apparently have no > > bearing whatsoever on the errors in our models > > This would mean there is no point trying to get better crystals, right? > Or am I also wrong to assume that the dataset with higher I/sigma in the > highest resolution shell will give me a better model? > > On a related point - why is Rmerge considered to be the limiting value > for the R? Isn't Rmerge a poorly defined measure itself that > deteriorates at least in some circumstances (e.g. increased redundancy)? > Specifically, shouldn't "ideal" R approximate 0.5*/? > > Cheers, > > Ed. > > > > -- > "I'd jump in myself, if I weren't so good at whistling." >Julian, King of Lemurs > >
Re: [ccp4bb] Against Method (R)
On Wed, 27 Oct 2010, Frank von Delft wrote: So, since the experimental error is only a minor contribution to the total error, it is arguably inappropriate to use it as a weight for each hkl. I think your logic has run off the track. The experimental error is an appropriate weight for the Fobs(hkl) because that is indeed the error for that observation. This is true independent of errors in the model. If you improve the model, that does not magically change the accuracy of the data. Sorry, still missing something: In the weighted Rfactor, we're weighting by the 1/sig**2 (right?) And the reason for that is, presumably, that when we add a term (Fo-Fc) but the Fo is crap (huge sigma), we need to ensure we don't add very much of it -- so we divide the term by the huge sigma. Correct. But what if Fc also is crap? Which it patently is: it's not even within 20% of Fo, never mind vaguely within sig(Fo). Why should we not be down-weighting those terms as well? Because here we want the exact opposite. If Fc is hugely different from a well-measured Fobs then it is a sensitive indicator of a problem with the model. Why would we want to down-weight it? Consider the extreme case: if you down-weight all reflections for which Fc does not already agree with Fo, then you will always conclude that the current model, no matter what random drawer you pulled it from, is in fine shape. Or can we ignore that because, since all terms are crap, we'd simply be down-weighting the entire Rw by a lot, and we'd be doing it for the Rw of both models we're comparing, so they'd cancel out when we take the ratio Rw1/Rw2? Not sure I follow this. But if we're so happy to fudge away the huge gorilla in the room, why would we need to be religious about the little gnats on the floor (the sig(Fo))? Is there then really a difference between R1/R2 and Rw1/Rw2, for all practical purposes? Yes. That was the message of the 1970 Ford & Rollet paper that Ian provided a link for. Ethan (Of course, this is all for the ongoing case we don't know how to model the R-factor gap. And no, I haven't played with actual numbers...) phx.
Re: [ccp4bb] Against Method (R)
On Tue, 2010-10-26 at 21:16 +0100, Frank von Delft wrote: > the errors in our measurements apparently have no > bearing whatsoever on the errors in our models This would mean there is no point trying to get better crystals, right? Or am I also wrong to assume that the dataset with higher I/sigma in the highest resolution shell will give me a better model? On a related point - why is Rmerge considered to be the limiting value for the R? Isn't Rmerge a poorly defined measure itself that deteriorates at least in some circumstances (e.g. increased redundancy)? Specifically, shouldn't "ideal" R approximate 0.5*/? Cheers, Ed. -- "I'd jump in myself, if I weren't so good at whistling." Julian, King of Lemurs
Re: [ccp4bb] Against Method (R)
Yes, but what I think Frank is trying to point out is that the difference between Fobs and Fcalc in any given PDB entry is generally about 4-5 times larger than sigma(Fobs). In such situations, pretty much any standard statistical test will tell you that the model is "highly unlikely to be correct". But that's not the question we are normally asking. It is highly unlikely that any model in biology is correct, if by "correct" you mean "cannot be improved". Normally we ask the more modest question "have I improved my model today over what it was yesterday?". I am not saying that everything in the PDB is "wrong", just that the dominant source of error is a shortcoming of the models we use. Whatever this "source of error" is, it vastly overpowers the measurement error. That is, errors do not add linearly, but rather as squares, and 20%^2+5%^2 ~ 20%^2 . So, since the experimental error is only a minor contribution to the total error, it is arguably inappropriate to use it as a weight for each hkl. I think your logic has run off the track. The experimental error is an appropriate weight for the Fobs(hkl) because that is indeed the error for that observation. This is true independent of errors in the model. If you improve the model, that does not magically change the accuracy of the data. Sorry, still missing something: In the weighted Rfactor, we're weighting by the 1/sig**2 (right?) And the reason for that is, presumably, that when we add a term (Fo-Fc) but the Fo is crap (huge sigma), we need to ensure we don't add very much of it -- so we divide the term by the huge sigma. But what if Fc also is crap? Which it patently is: it's not even within 20% of Fo, never mind vaguely within sig(Fo). Why should we not be down-weighting those terms as well? Or can we ignore that because, since all terms are crap, we'd simply be down-weighting the entire Rw by a lot, and we'd be doing it for the Rw of both models we're comparing, so they'd cancel out when we take the ratio Rw1/Rw2? But if we're so happy to fudge away the huge gorilla in the room, why would we need to be religious about the little gnats on the floor (the sig(Fo))? Is there then really a difference between R1/R2 and Rw1/Rw2, for all practical purposes? (Of course, this is all for the ongoing case we don't know how to model the R-factor gap. And no, I haven't played with actual numbers...) phx.
Re: [ccp4bb] Against Method (R)
Some time ago, I computed the mean value of Rcryst(F) / Rmerge(F) across the whole PDB. This average was 4.5, and I take this as a rough estimate of |Fcalc - Fobs| / sigma(Fobs). More recently, I have been looking in more detail at deposited data, but so far the few cases where this ratio is close to 1 are all cases where sigma(Fobs) is unusually high! I think the "answer" is that we can believe structures in the PDB to "within 20% error". This is "close enough" for a few things (such as government work), but not for traditional statistics like "confidence tests". For me, it is just really bothersome that we can measure structure factors to better than 5% accuracy, but still don't know how to model them. Ethan does make a good point that sig(Fobs) is the error in the measurement, and that the model-data error is not the weight one should use in refinement, etc. However, when you are comparing one PDB entry (yours) to others (published), I still don't think that sigma(Fobs) plays any significant role. -James Holton MAD Scientist On Tue, Oct 26, 2010 at 4:45 PM, Jacob Keller < j-kell...@fsm.northwestern.edu> wrote: > - Original Message - > *From:* James Holton > *To:* CCP4BB@JISCMAIL.AC.UK > *Sent:* Tuesday, October 26, 2010 6:31 PM > *Subject:* Re: [ccp4bb] Against Method (R) > > Yes, but what I think Frank is trying to point out is that the difference > between Fobs and Fcalc in any given PDB entry is generally about 4-5 times > larger than sigma(Fobs). In such situations, pretty much any standard > statistical test will tell you that the model is "highly unlikely to be > correct". > > Wow, so what is the answer to this? Is that figure "|Fcalc - Fobs| = 4-5x > sigma" really true? How, then, do we believe structures? Are there really > good structures where this is discrepancy is not there, to "stake our > claim," so to speak? >
Re: [ccp4bb] Against Method (R)
On Tuesday, October 26, 2010 04:31:24 pm James Holton wrote: > Yes, but what I think Frank is trying to point out is that the difference > between Fobs and Fcalc in any given PDB entry is generally about 4-5 times > larger than sigma(Fobs). In such situations, pretty much any standard > statistical test will tell you that the model is "highly unlikely to be > correct". But that's not the question we are normally asking. It is highly unlikely that any model in biology is correct, if by "correct" you mean "cannot be improved". Normally we ask the more modest question "have I improved my model today over what it was yesterday?". > I am not saying that everything in the PDB is "wrong", just that the > dominant source of error is a shortcoming of the models we use. Whatever > this "source of error" is, it vastly overpowers the measurement error. That > is, errors do not add linearly, but rather as squares, and 20%^2+5%^2 ~ > 20%^2 . > > So, since the experimental error is only a minor contribution to the total > error, it is arguably inappropriate to use it as a weight for each hkl. I think your logic has run off the track. The experimental error is an appropriate weight for the Fobs(hkl) because that is indeed the error for that observation. This is true independent of errors in the model. If you improve the model, that does not magically change the accuracy of the data. Ethan > Yes, refinement does seem to work better when you use experimental sigmas, > and weighted statistics are probably "better" than no weights at all, but > the problem is that until we do have a model that can explain Fobs to within > experimental error, we will be severely limited in the kinds of conclusions > we can derive from our data. > > -James Holton > MAD Scientist > > On Tue, Oct 26, 2010 at 1:59 PM, Ethan Merritt > wrote: > > > On Tuesday, October 26, 2010 01:16:58 pm Frank von Delft wrote: > > >Um... > > > > > > * Given that the weighted Rfactor is weighted by the measurement errors > > > (1/sig^2) > > > > > > * and given that the errors in our measurements apparently have no > > > bearing whatsoever on the errors in our models (for macromolecular > > > crystals, certainly - the "R-vfactor gap") > > > > You are overlooking causality :-) > > > > Yes, the errors in state-of-the-art models are only weakly limited by the > > errors in our measurements. But that is exactly _because_ we can now > > weight > > properly by the measurement errors (1/sig^2). In my salad days, > > weighting by 1/sig^2 was a mug's game. Refinement only produced > > a reasonable model if you applied empirical corrections rather than > > statistical weights. Things have improved a bit since then, > > both on the equipment side (detectors, cryo, ...) and on the processing > > side (Maximum Likelihood, error propagation, ...). > > Now the sigmas actually mean something! > > > > > is the weighted Rfactor even vaguely relevant for anything at all? > > > > Yes, it is. It is the thing you are minimizing during refinement, > > at least to first approximation. Also, as just mentioned, it is a > > well-defined value that you can do use for statistical significance > > tests. > > > >Ethan > > > > > > > > > > phx. > > > > > > > > > > > > On 26/10/2010 20:44, Ian Tickle wrote: > > > > Indeed, see: http://scripts.iucr.org/cgi-bin/paper?a07175 . > > > > > > > > The Rfree/Rwork ratio that I referred to does strictly use the > > > > weighted ('Hamilton') R-factors, but because only the unweighted > > > > values are given in the PDB we were forced to approximate (against our > > > > better judgment!). > > > > > > > > The problem of course is that all refinement software AFAIK writes the > > > > unweighted Rwork& Rfree to the PDB header; there are no slots for the > > > > weighted values, which does indeed make doing serious statistics on > > > > the PDB entries difficult if not impossible! > > > > > > > > The unweighted crystallographic R-factor was only ever intended as a > > > > "rule of thumb", i.e. to give a rough idea of the relative quality of > > > > related structures; I hardly think the crystallographers of yesteryear > > > > ever imagined that we would be taking it so seriously now! > > > > > > > > In particular IMO it should never be used for something as critical as > > > > validation (either global or local), or for guiding refinement > > > > strategy: use the likelihood instead. > > > > > > > > Cheers > > > > > > > > -- Ian > > > > > > > > PS I've always known it as an 'R-factor', e.g. see paper referenced > > > > above, but then during my crystallographic training I used extensively > > > > software developed by both authors of the paper (i.e. Geoff Ford& the > > > > late John Rollett) in Oxford (which eventually became the 'Crystals' > > > > small-molecule package). Maybe it's a transatlantic thing ... > > > > > > > > Cheers > > > > > > > > -- Ian > > > > > > > > On Tue, Oct 26, 2010 at 7:28 PM, Ethan Merritt<
Re: [ccp4bb] Against Method (R)
- Original Message - From: James Holton To: CCP4BB@JISCMAIL.AC.UK Sent: Tuesday, October 26, 2010 6:31 PM Subject: Re: [ccp4bb] Against Method (R) Yes, but what I think Frank is trying to point out is that the difference between Fobs and Fcalc in any given PDB entry is generally about 4-5 times larger than sigma(Fobs). In such situations, pretty much any standard statistical test will tell you that the model is "highly unlikely to be correct". Wow, so what is the answer to this? Is that figure "|Fcalc - Fobs| = 4-5x sigma" really true? How, then, do we believe structures? Are there really good structures where this is discrepancy is not there, to "stake our claim," so to speak?
Re: [ccp4bb] Against Method (R)
Yes, but what I think Frank is trying to point out is that the difference between Fobs and Fcalc in any given PDB entry is generally about 4-5 times larger than sigma(Fobs). In such situations, pretty much any standard statistical test will tell you that the model is "highly unlikely to be correct". I am not saying that everything in the PDB is "wrong", just that the dominant source of error is a shortcoming of the models we use. Whatever this "source of error" is, it vastly overpowers the measurement error. That is, errors do not add linearly, but rather as squares, and 20%^2+5%^2 ~ 20%^2 . So, since the experimental error is only a minor contribution to the total error, it is arguably inappropriate to use it as a weight for each hkl. Yes, refinement does seem to work better when you use experimental sigmas, and weighted statistics are probably "better" than no weights at all, but the problem is that until we do have a model that can explain Fobs to within experimental error, we will be severely limited in the kinds of conclusions we can derive from our data. -James Holton MAD Scientist On Tue, Oct 26, 2010 at 1:59 PM, Ethan Merritt wrote: > On Tuesday, October 26, 2010 01:16:58 pm Frank von Delft wrote: > >Um... > > > > * Given that the weighted Rfactor is weighted by the measurement errors > > (1/sig^2) > > > > * and given that the errors in our measurements apparently have no > > bearing whatsoever on the errors in our models (for macromolecular > > crystals, certainly - the "R-vfactor gap") > > You are overlooking causality :-) > > Yes, the errors in state-of-the-art models are only weakly limited by the > errors in our measurements. But that is exactly _because_ we can now > weight > properly by the measurement errors (1/sig^2). In my salad days, > weighting by 1/sig^2 was a mug's game. Refinement only produced > a reasonable model if you applied empirical corrections rather than > statistical weights. Things have improved a bit since then, > both on the equipment side (detectors, cryo, ...) and on the processing > side (Maximum Likelihood, error propagation, ...). > Now the sigmas actually mean something! > > > is the weighted Rfactor even vaguely relevant for anything at all? > > Yes, it is. It is the thing you are minimizing during refinement, > at least to first approximation. Also, as just mentioned, it is a > well-defined value that you can do use for statistical significance > tests. > >Ethan > > > > > > phx. > > > > > > > > On 26/10/2010 20:44, Ian Tickle wrote: > > > Indeed, see: http://scripts.iucr.org/cgi-bin/paper?a07175 . > > > > > > The Rfree/Rwork ratio that I referred to does strictly use the > > > weighted ('Hamilton') R-factors, but because only the unweighted > > > values are given in the PDB we were forced to approximate (against our > > > better judgment!). > > > > > > The problem of course is that all refinement software AFAIK writes the > > > unweighted Rwork& Rfree to the PDB header; there are no slots for the > > > weighted values, which does indeed make doing serious statistics on > > > the PDB entries difficult if not impossible! > > > > > > The unweighted crystallographic R-factor was only ever intended as a > > > "rule of thumb", i.e. to give a rough idea of the relative quality of > > > related structures; I hardly think the crystallographers of yesteryear > > > ever imagined that we would be taking it so seriously now! > > > > > > In particular IMO it should never be used for something as critical as > > > validation (either global or local), or for guiding refinement > > > strategy: use the likelihood instead. > > > > > > Cheers > > > > > > -- Ian > > > > > > PS I've always known it as an 'R-factor', e.g. see paper referenced > > > above, but then during my crystallographic training I used extensively > > > software developed by both authors of the paper (i.e. Geoff Ford& the > > > late John Rollett) in Oxford (which eventually became the 'Crystals' > > > small-molecule package). Maybe it's a transatlantic thing ... > > > > > > Cheers > > > > > > -- Ian > > > > > > On Tue, Oct 26, 2010 at 7:28 PM, Ethan Merritt< > merr...@u.washington.edu> wrote: > > >> On Tuesday, October 26, 2010 09:46:46 am Bernhard Rupp (Hofkristallrat > a.D.) wrote: > > >>> Hi Folks, > > >>> > > >>> Please allow me a few biased reflections/opinions on the numeRology > of the > > >>> R-value (not R-factor, because it is neither a factor itself nor does > it > > >>> factor in anything but ill-posed reviewer's critique. Historically > the term > > >>> originated from small molecule crystallography, but it is only a > > >>> 'Residual-value') > > >>> > > >>> a) The R-value itself - based on the linear residuals and of apparent > > >>> intuitive meaning - is statistically peculiar to say the least. I > could not > > >>> find it in any common statistics text. So doing proper statistics > with R > > >>> becomes difficult. > > >> As WC Hamilton pointed out originally, two [p
Re: [ccp4bb] Against Method (R)
On Tuesday, October 26, 2010 01:16:58 pm Frank von Delft wrote: >Um... > > * Given that the weighted Rfactor is weighted by the measurement errors > (1/sig^2) > > * and given that the errors in our measurements apparently have no > bearing whatsoever on the errors in our models (for macromolecular > crystals, certainly - the "R-vfactor gap") You are overlooking causality :-) Yes, the errors in state-of-the-art models are only weakly limited by the errors in our measurements. But that is exactly _because_ we can now weight properly by the measurement errors (1/sig^2). In my salad days, weighting by 1/sig^2 was a mug's game. Refinement only produced a reasonable model if you applied empirical corrections rather than statistical weights. Things have improved a bit since then, both on the equipment side (detectors, cryo, ...) and on the processing side (Maximum Likelihood, error propagation, ...). Now the sigmas actually mean something! > is the weighted Rfactor even vaguely relevant for anything at all? Yes, it is. It is the thing you are minimizing during refinement, at least to first approximation. Also, as just mentioned, it is a well-defined value that you can do use for statistical significance tests. Ethan > > phx. > > > > On 26/10/2010 20:44, Ian Tickle wrote: > > Indeed, see: http://scripts.iucr.org/cgi-bin/paper?a07175 . > > > > The Rfree/Rwork ratio that I referred to does strictly use the > > weighted ('Hamilton') R-factors, but because only the unweighted > > values are given in the PDB we were forced to approximate (against our > > better judgment!). > > > > The problem of course is that all refinement software AFAIK writes the > > unweighted Rwork& Rfree to the PDB header; there are no slots for the > > weighted values, which does indeed make doing serious statistics on > > the PDB entries difficult if not impossible! > > > > The unweighted crystallographic R-factor was only ever intended as a > > "rule of thumb", i.e. to give a rough idea of the relative quality of > > related structures; I hardly think the crystallographers of yesteryear > > ever imagined that we would be taking it so seriously now! > > > > In particular IMO it should never be used for something as critical as > > validation (either global or local), or for guiding refinement > > strategy: use the likelihood instead. > > > > Cheers > > > > -- Ian > > > > PS I've always known it as an 'R-factor', e.g. see paper referenced > > above, but then during my crystallographic training I used extensively > > software developed by both authors of the paper (i.e. Geoff Ford& the > > late John Rollett) in Oxford (which eventually became the 'Crystals' > > small-molecule package). Maybe it's a transatlantic thing ... > > > > Cheers > > > > -- Ian > > > > On Tue, Oct 26, 2010 at 7:28 PM, Ethan Merritt > > wrote: > >> On Tuesday, October 26, 2010 09:46:46 am Bernhard Rupp (Hofkristallrat > >> a.D.) wrote: > >>> Hi Folks, > >>> > >>> Please allow me a few biased reflections/opinions on the numeRology of the > >>> R-value (not R-factor, because it is neither a factor itself nor does it > >>> factor in anything but ill-posed reviewer's critique. Historically the > >>> term > >>> originated from small molecule crystallography, but it is only a > >>> 'Residual-value') > >>> > >>> a) The R-value itself - based on the linear residuals and of apparent > >>> intuitive meaning - is statistically peculiar to say the least. I could > >>> not > >>> find it in any common statistics text. So doing proper statistics with R > >>> becomes difficult. > >> As WC Hamilton pointed out originally, two [properly weighted] R factors > >> can > >> be compared by taking their ratio. Significance levels can then be > >> evaluated > >> using the standard F distribution. A concise summary is given in chapter 9 > >> of Prince's book, which I highly recommend to all crystallographers. > >> > >> W C Hamilton "Significance tests on the crystallographic R factor" > >> Acta Cryst. (1965). 18, 502-510 > >> > >> Edward Prince "Mathematical Techniques in Crystallography and Materials > >> Science". Springer-Verlag, 1982. > >> > >> It is true that we normally indulge in the sloppy habit of paying attention > >> only to the unweighted R factor even though refinement programs report > >> both the weighted and unweighted versions. (shelx users excepted :-) > >> But the weighted form is there also if you want to do statistical tests. > >> > >> You are of course correct that this remains a global test, and as such > >> is of limited use in evaluating local properties of the model. > >> > >> cheers, > >> > >> Ethan > >> > >> > >> > >> > >>> b) rules of thumb (as much as they conveniently obviate the need for > >>> detailed explanations, satisfy student's desire for quick answers, and > >>> allow superficial review of manuscripts) become less valuable if they > >>> have a > >>> case-dependent large variance, topp
Re: [ccp4bb] Against Method (R)
Um... * Given that the weighted Rfactor is weighted by the measurement errors (1/sig^2) * and given that the errors in our measurements apparently have no bearing whatsoever on the errors in our models (for macromolecular crystals, certainly - the "R-vfactor gap") is the weighted Rfactor even vaguely relevant for anything at all? phx. On 26/10/2010 20:44, Ian Tickle wrote: Indeed, see: http://scripts.iucr.org/cgi-bin/paper?a07175 . The Rfree/Rwork ratio that I referred to does strictly use the weighted ('Hamilton') R-factors, but because only the unweighted values are given in the PDB we were forced to approximate (against our better judgment!). The problem of course is that all refinement software AFAIK writes the unweighted Rwork& Rfree to the PDB header; there are no slots for the weighted values, which does indeed make doing serious statistics on the PDB entries difficult if not impossible! The unweighted crystallographic R-factor was only ever intended as a "rule of thumb", i.e. to give a rough idea of the relative quality of related structures; I hardly think the crystallographers of yesteryear ever imagined that we would be taking it so seriously now! In particular IMO it should never be used for something as critical as validation (either global or local), or for guiding refinement strategy: use the likelihood instead. Cheers -- Ian PS I've always known it as an 'R-factor', e.g. see paper referenced above, but then during my crystallographic training I used extensively software developed by both authors of the paper (i.e. Geoff Ford& the late John Rollett) in Oxford (which eventually became the 'Crystals' small-molecule package). Maybe it's a transatlantic thing ... Cheers -- Ian On Tue, Oct 26, 2010 at 7:28 PM, Ethan Merritt wrote: On Tuesday, October 26, 2010 09:46:46 am Bernhard Rupp (Hofkristallrat a.D.) wrote: Hi Folks, Please allow me a few biased reflections/opinions on the numeRology of the R-value (not R-factor, because it is neither a factor itself nor does it factor in anything but ill-posed reviewer's critique. Historically the term originated from small molecule crystallography, but it is only a 'Residual-value') a) The R-value itself - based on the linear residuals and of apparent intuitive meaning - is statistically peculiar to say the least. I could not find it in any common statistics text. So doing proper statistics with R becomes difficult. As WC Hamilton pointed out originally, two [properly weighted] R factors can be compared by taking their ratio. Significance levels can then be evaluated using the standard F distribution. A concise summary is given in chapter 9 of Prince's book, which I highly recommend to all crystallographers. W C Hamilton "Significance tests on the crystallographic R factor" Acta Cryst. (1965). 18, 502-510 Edward Prince "Mathematical Techniques in Crystallography and Materials Science". Springer-Verlag, 1982. It is true that we normally indulge in the sloppy habit of paying attention only to the unweighted R factor even though refinement programs report both the weighted and unweighted versions. (shelx users excepted :-) But the weighted form is there also if you want to do statistical tests. You are of course correct that this remains a global test, and as such is of limited use in evaluating local properties of the model. cheers, Ethan b) rules of thumb (as much as they conveniently obviate the need for detailed explanations, satisfy student's desire for quick answers, and allow superficial review of manuscripts) become less valuable if they have a case-dependent large variance, topped with an unknown parent distribution. Combined with an odd statistic, that has great potential for misguidance and unnecessarily lost sleep. c) Ian has (once again) explained that for example the Rf-R depends on the exact knowledge of the restraints and their individual weighting, which we generally do not have. Caution is advised. d) The answer which model is better - which is actually what you want to know - becomes a question of model selection or hypothesis testing, which, given the obscurity of R cannot be derived with some nice plug-in method. As Ian said the models to be compared must also be based on the same and identical data. e) One measure available that is statistically at least defensible is the log-likelihood. So what you can do is form a log-likelihood ratio (or Bayes factor (there is the darn factor again, it’s a ratio)) and see where this falls - and the answers are pretty soft and, probably because of that, correspondingly realistic. This also makes - based on statistics alone - deciding between different overall parameterizations difficult. http://en.wikipedia.org/wiki/Bayes_factor f) so having said that, what really remains is that the model that fits the primary evidence (minimally biased electron density) best and is at the same time physically meaningful, is the best mode
Re: [ccp4bb] Against Method (R)
Indeed, see: http://scripts.iucr.org/cgi-bin/paper?a07175 . The Rfree/Rwork ratio that I referred to does strictly use the weighted ('Hamilton') R-factors, but because only the unweighted values are given in the PDB we were forced to approximate (against our better judgment!). The problem of course is that all refinement software AFAIK writes the unweighted Rwork & Rfree to the PDB header; there are no slots for the weighted values, which does indeed make doing serious statistics on the PDB entries difficult if not impossible! The unweighted crystallographic R-factor was only ever intended as a "rule of thumb", i.e. to give a rough idea of the relative quality of related structures; I hardly think the crystallographers of yesteryear ever imagined that we would be taking it so seriously now! In particular IMO it should never be used for something as critical as validation (either global or local), or for guiding refinement strategy: use the likelihood instead. Cheers -- Ian PS I've always known it as an 'R-factor', e.g. see paper referenced above, but then during my crystallographic training I used extensively software developed by both authors of the paper (i.e. Geoff Ford & the late John Rollett) in Oxford (which eventually became the 'Crystals' small-molecule package). Maybe it's a transatlantic thing ... Cheers -- Ian On Tue, Oct 26, 2010 at 7:28 PM, Ethan Merritt wrote: > On Tuesday, October 26, 2010 09:46:46 am Bernhard Rupp (Hofkristallrat a.D.) > wrote: >> Hi Folks, >> >> Please allow me a few biased reflections/opinions on the numeRology of the >> R-value (not R-factor, because it is neither a factor itself nor does it >> factor in anything but ill-posed reviewer's critique. Historically the term >> originated from small molecule crystallography, but it is only a >> 'Residual-value') >> >> a) The R-value itself - based on the linear residuals and of apparent >> intuitive meaning - is statistically peculiar to say the least. I could not >> find it in any common statistics text. So doing proper statistics with R >> becomes difficult. > > As WC Hamilton pointed out originally, two [properly weighted] R factors can > be compared by taking their ratio. Significance levels can then be evaluated > using the standard F distribution. A concise summary is given in chapter 9 > of Prince's book, which I highly recommend to all crystallographers. > > W C Hamilton "Significance tests on the crystallographic R factor" > Acta Cryst. (1965). 18, 502-510 > > Edward Prince "Mathematical Techniques in Crystallography and Materials > Science". Springer-Verlag, 1982. > > It is true that we normally indulge in the sloppy habit of paying attention > only to the unweighted R factor even though refinement programs report > both the weighted and unweighted versions. (shelx users excepted :-) > But the weighted form is there also if you want to do statistical tests. > > You are of course correct that this remains a global test, and as such > is of limited use in evaluating local properties of the model. > > cheers, > > Ethan > > > > >> b) rules of thumb (as much as they conveniently obviate the need for >> detailed explanations, satisfy student's desire for quick answers, and >> allow superficial review of manuscripts) become less valuable if they have a >> case-dependent large variance, topped with an unknown parent distribution. >> Combined with an odd statistic, that has great potential for misguidance and >> unnecessarily lost sleep. >> >> c) Ian has (once again) explained that for example the Rf-R depends on the >> exact knowledge of the restraints and their individual weighting, which we >> generally do not have. Caution is advised. >> >> d) The answer which model is better - which is actually what you want to >> know - becomes a question of model selection or hypothesis testing, which, >> given the obscurity of R cannot be derived with some nice plug-in method. As >> Ian said the models to be compared must also be based on the same and >> identical data. >> >> e) One measure available that is statistically at least defensible is the >> log-likelihood. So what you can do is form a log-likelihood ratio (or Bayes >> factor (there is the darn factor again, it’s a ratio)) and see where this >> falls - and the answers are pretty soft and, probably because of that, >> correspondingly realistic. This also makes - based on statistics alone - >> deciding between different overall parameterizations difficult. >> >> http://en.wikipedia.org/wiki/Bayes_factor >> >> f) so having said that, what really remains is that the model that fits the >> primary evidence (minimally biased electron density) best and is at the same >> time physically meaningful, is the best model, i. e., all plausibly >> accountable electron density (and not more) is modeled. You can convince >> yourself of this by taking the most interesting part of the model out (say a >> ligand or a binding pocket) and look at the R-value
Re: [ccp4bb] Against Method (R)
W C Hamilton "Significance tests on the crystallographic R factor" Acta Cryst. (1965). 18, 502-510 Interestingly enough, I have used the Hamilton tests in Rietveld powder refinements of small molecules/intermetallics before before. One problem were partial occupancies vs split conformations in HT superconductors. Alas, you cannot cheat there either - most of the time the results showed that numerically the differences were not significant, and one again had to resort to non-statistical plausibility arguments of references. Has anyone done Hamiltons on different protein models/parameterizations and can report? I think for global parameterization changes like NCS,TLS, etc that may in fact be interesting. BR
Re: [ccp4bb] Against Method (R)
Dear all, Augustine, "Confessions", Book 11 Chap. XIV, has it: "If no one ask of me, I know; if I wish to explain to him who asks, I know not." With best wishes, Gerard. -- On Tue, Oct 26, 2010 at 01:30:11PM -0500, Phoebe Rice wrote: > Another issue with these statistics is that the PDB insists on a single value > of "resolution" no matter how anisotropic the data. Especially in the > outermost bins, Rmerge could be ridiculously high simply because the data > only exist in one out of 3 directions. >Phoebe > > = > Phoebe A. Rice > Dept. of Biochemistry & Molecular Biology > The University of Chicago > phone 773 834 1723 > http://bmb.bsd.uchicago.edu/Faculty_and_Research/01_Faculty/01_Faculty_Alphabetically.php?faculty_id=123 > http://www.rsc.org/shop/books/2008/9780854042722.asp > > > Original message > >Date: Tue, 26 Oct 2010 09:46:46 -0700 > >From: CCP4 bulletin board (on behalf of "Bernhard > >Rupp (Hofkristallrat a.D.)" ) > >Subject: [ccp4bb] Against Method (R) > >To: CCP4BB@JISCMAIL.AC.UK > > > >Hi Folks, > > > >Please allow me a few biased reflections/opinions on the numeRology of the > >R-value (not R-factor, because it is neither a factor itself nor does it > >factor in anything but ill-posed reviewer's critique. Historically the term > >originated from small molecule crystallography, but it is only a > >'Residual-value') > > > >a) The R-value itself - based on the linear residuals and of apparent > >intuitive meaning - is statistically peculiar to say the least. I could not > >find it in any common statistics text. So doing proper statistics with R > >becomes difficult. > > > >b) rules of thumb (as much as they conveniently obviate the need for > >detailed explanations, satisfy student's desire for quick answers, and > >allow superficial review of manuscripts) become less valuable if they have a > >case-dependent large variance, topped with an unknown parent distribution. > >Combined with an odd statistic, that has great potential for misguidance and > >unnecessarily lost sleep. > > > >c) Ian has (once again) explained that for example the Rf-R depends on the > >exact knowledge of the restraints and their individual weighting, which we > >generally do not have. Caution is advised. > > > >d) The answer which model is better - which is actually what you want to > >know - becomes a question of model selection or hypothesis testing, which, > >given the obscurity of R cannot be derived with some nice plug-in method. As > >Ian said the models to be compared must also be based on the same and > >identical data. > > > >e) One measure available that is statistically at least defensible is the > >log-likelihood. So what you can do is form a log-likelihood ratio (or Bayes > >factor (there is the darn factor again, it’s a ratio)) and see where this > >falls - and the answers are pretty soft and, probably because of that, > >correspondingly realistic. This also makes - based on statistics alone - > >deciding between different overall parameterizations difficult. > > > >http://en.wikipedia.org/wiki/Bayes_factor > > > >f) so having said that, what really remains is that the model that fits the > >primary evidence (minimally biased electron density) best and is at the same > >time physically meaningful, is the best model, i. e., all plausibly > >accountable electron density (and not more) is modeled. You can convince > >yourself of this by taking the most interesting part of the model out (say a > >ligand or a binding pocket) and look at the R-values or do a model selection > >test - the result will be indecisive. Poof goes the global rule of thumb. > > > >g) in other words: global measures in general are entirely inadequate to > >judge local model quality (noted many times over already by Jones, Kleywegt, > >others, in the dark ages of crystallography when poorly restrained > >crystallographers used to passionately whack each other over the head with > >unfree R-values). > > > >Best, BR > >- > >Bernhard Rupp, Hofkristallrat a.D. > >001 (925) 209-7429 > >+43 (676) 571-0536 > >b...@ruppweb.org > >hofkristall...@gmail.com > >http://www.ruppweb.org/ > >-- > >Und wieder ein chillout-mix aus der Hofkristall-lounge > >-- -- === * * * Gerard Bricogne g...@globalphasing.com * * * * Global Phasing Ltd. * * Sheraton House, Castle Park Tel: +44-(0)1223-353033 * * Cambridge CB3 0AX, UK Fax: +44-(0)1223-366889 * * * ===
Re: [ccp4bb] Against Method (R)
On Tuesday, October 26, 2010 09:46:46 am Bernhard Rupp (Hofkristallrat a.D.) wrote: > Hi Folks, > > Please allow me a few biased reflections/opinions on the numeRology of the > R-value (not R-factor, because it is neither a factor itself nor does it > factor in anything but ill-posed reviewer's critique. Historically the term > originated from small molecule crystallography, but it is only a > 'Residual-value') > > a) The R-value itself - based on the linear residuals and of apparent > intuitive meaning - is statistically peculiar to say the least. I could not > find it in any common statistics text. So doing proper statistics with R > becomes difficult. As WC Hamilton pointed out originally, two [properly weighted] R factors can be compared by taking their ratio. Significance levels can then be evaluated using the standard F distribution. A concise summary is given in chapter 9 of Prince's book, which I highly recommend to all crystallographers. W C Hamilton "Significance tests on the crystallographic R factor" Acta Cryst. (1965). 18, 502-510 Edward Prince "Mathematical Techniques in Crystallography and Materials Science". Springer-Verlag, 1982. It is true that we normally indulge in the sloppy habit of paying attention only to the unweighted R factor even though refinement programs report both the weighted and unweighted versions. (shelx users excepted :-) But the weighted form is there also if you want to do statistical tests. You are of course correct that this remains a global test, and as such is of limited use in evaluating local properties of the model. cheers, Ethan > b) rules of thumb (as much as they conveniently obviate the need for > detailed explanations, satisfy student's desire for quick answers, and > allow superficial review of manuscripts) become less valuable if they have a > case-dependent large variance, topped with an unknown parent distribution. > Combined with an odd statistic, that has great potential for misguidance and > unnecessarily lost sleep. > > c) Ian has (once again) explained that for example the Rf-R depends on the > exact knowledge of the restraints and their individual weighting, which we > generally do not have. Caution is advised. > > d) The answer which model is better - which is actually what you want to > know - becomes a question of model selection or hypothesis testing, which, > given the obscurity of R cannot be derived with some nice plug-in method. As > Ian said the models to be compared must also be based on the same and > identical data. > > e) One measure available that is statistically at least defensible is the > log-likelihood. So what you can do is form a log-likelihood ratio (or Bayes > factor (there is the darn factor again, its a ratio)) and see where this > falls - and the answers are pretty soft and, probably because of that, > correspondingly realistic. This also makes - based on statistics alone - > deciding between different overall parameterizations difficult. > > http://en.wikipedia.org/wiki/Bayes_factor > > f) so having said that, what really remains is that the model that fits the > primary evidence (minimally biased electron density) best and is at the same > time physically meaningful, is the best model, i. e., all plausibly > accountable electron density (and not more) is modeled. You can convince > yourself of this by taking the most interesting part of the model out (say a > ligand or a binding pocket) and look at the R-values or do a model selection > test - the result will be indecisive. Poof goes the global rule of thumb. > > g) in other words: global measures in general are entirely inadequate to > judge local model quality (noted many times over already by Jones, Kleywegt, > others, in the dark ages of crystallography when poorly restrained > crystallographers used to passionately whack each other over the head with > unfree R-values). > > Best, BR > - > Bernhard Rupp, Hofkristallrat a.D. > 001 (925) 209-7429 > +43 (676) 571-0536 > b...@ruppweb.org > hofkristall...@gmail.com > http://www.ruppweb.org/ > -- > Und wieder ein chillout-mix aus der Hofkristall-lounge > -- > -- Ethan A Merritt Biomolecular Structure Center, K-428 Health Sciences Bldg University of Washington, Seattle 98195-7742
Re: [ccp4bb] Against Method (R)
Another issue with these statistics is that the PDB insists on a single value of "resolution" no matter how anisotropic the data. Especially in the outermost bins, Rmerge could be ridiculously high simply because the data only exist in one out of 3 directions. Phoebe = Phoebe A. Rice Dept. of Biochemistry & Molecular Biology The University of Chicago phone 773 834 1723 http://bmb.bsd.uchicago.edu/Faculty_and_Research/01_Faculty/01_Faculty_Alphabetically.php?faculty_id=123 http://www.rsc.org/shop/books/2008/9780854042722.asp Original message >Date: Tue, 26 Oct 2010 09:46:46 -0700 >From: CCP4 bulletin board (on behalf of "Bernhard Rupp >(Hofkristallrat a.D.)" ) >Subject: [ccp4bb] Against Method (R) >To: CCP4BB@JISCMAIL.AC.UK > >Hi Folks, > >Please allow me a few biased reflections/opinions on the numeRology of the >R-value (not R-factor, because it is neither a factor itself nor does it >factor in anything but ill-posed reviewer's critique. Historically the term >originated from small molecule crystallography, but it is only a >'Residual-value') > >a) The R-value itself - based on the linear residuals and of apparent >intuitive meaning - is statistically peculiar to say the least. I could not >find it in any common statistics text. So doing proper statistics with R >becomes difficult. > >b) rules of thumb (as much as they conveniently obviate the need for >detailed explanations, satisfy student's desire for quick answers, and >allow superficial review of manuscripts) become less valuable if they have a >case-dependent large variance, topped with an unknown parent distribution. >Combined with an odd statistic, that has great potential for misguidance and >unnecessarily lost sleep. > >c) Ian has (once again) explained that for example the Rf-R depends on the >exact knowledge of the restraints and their individual weighting, which we >generally do not have. Caution is advised. > >d) The answer which model is better - which is actually what you want to >know - becomes a question of model selection or hypothesis testing, which, >given the obscurity of R cannot be derived with some nice plug-in method. As >Ian said the models to be compared must also be based on the same and >identical data. > >e) One measure available that is statistically at least defensible is the >log-likelihood. So what you can do is form a log-likelihood ratio (or Bayes >factor (there is the darn factor again, it’s a ratio)) and see where this >falls - and the answers are pretty soft and, probably because of that, >correspondingly realistic. This also makes - based on statistics alone - >deciding between different overall parameterizations difficult. > >http://en.wikipedia.org/wiki/Bayes_factor > >f) so having said that, what really remains is that the model that fits the >primary evidence (minimally biased electron density) best and is at the same >time physically meaningful, is the best model, i. e., all plausibly >accountable electron density (and not more) is modeled. You can convince >yourself of this by taking the most interesting part of the model out (say a >ligand or a binding pocket) and look at the R-values or do a model selection >test - the result will be indecisive. Poof goes the global rule of thumb. > >g) in other words: global measures in general are entirely inadequate to >judge local model quality (noted many times over already by Jones, Kleywegt, >others, in the dark ages of crystallography when poorly restrained >crystallographers used to passionately whack each other over the head with >unfree R-values). > >Best, BR >- >Bernhard Rupp, Hofkristallrat a.D. >001 (925) 209-7429 >+43 (676) 571-0536 >b...@ruppweb.org >hofkristall...@gmail.com >http://www.ruppweb.org/ >-- >Und wieder ein chillout-mix aus der Hofkristall-lounge >--
[ccp4bb] Against Method (R)
Hi Folks, Please allow me a few biased reflections/opinions on the numeRology of the R-value (not R-factor, because it is neither a factor itself nor does it factor in anything but ill-posed reviewer's critique. Historically the term originated from small molecule crystallography, but it is only a 'Residual-value') a) The R-value itself - based on the linear residuals and of apparent intuitive meaning - is statistically peculiar to say the least. I could not find it in any common statistics text. So doing proper statistics with R becomes difficult. b) rules of thumb (as much as they conveniently obviate the need for detailed explanations, satisfy student's desire for quick answers, and allow superficial review of manuscripts) become less valuable if they have a case-dependent large variance, topped with an unknown parent distribution. Combined with an odd statistic, that has great potential for misguidance and unnecessarily lost sleep. c) Ian has (once again) explained that for example the Rf-R depends on the exact knowledge of the restraints and their individual weighting, which we generally do not have. Caution is advised. d) The answer which model is better - which is actually what you want to know - becomes a question of model selection or hypothesis testing, which, given the obscurity of R cannot be derived with some nice plug-in method. As Ian said the models to be compared must also be based on the same and identical data. e) One measure available that is statistically at least defensible is the log-likelihood. So what you can do is form a log-likelihood ratio (or Bayes factor (there is the darn factor again, its a ratio)) and see where this falls - and the answers are pretty soft and, probably because of that, correspondingly realistic. This also makes - based on statistics alone - deciding between different overall parameterizations difficult. http://en.wikipedia.org/wiki/Bayes_factor f) so having said that, what really remains is that the model that fits the primary evidence (minimally biased electron density) best and is at the same time physically meaningful, is the best model, i. e., all plausibly accountable electron density (and not more) is modeled. You can convince yourself of this by taking the most interesting part of the model out (say a ligand or a binding pocket) and look at the R-values or do a model selection test - the result will be indecisive. Poof goes the global rule of thumb. g) in other words: global measures in general are entirely inadequate to judge local model quality (noted many times over already by Jones, Kleywegt, others, in the dark ages of crystallography when poorly restrained crystallographers used to passionately whack each other over the head with unfree R-values). Best, BR - Bernhard Rupp, Hofkristallrat a.D. 001 (925) 209-7429 +43 (676) 571-0536 b...@ruppweb.org hofkristall...@gmail.com http://www.ruppweb.org/ -- Und wieder ein chillout-mix aus der Hofkristall-lounge --