Imagine the same structure initially solved at low resolution, and then re-solved with very high resolution data. The RMSD given by the software will be about the same, but the actual values of particular bond lengths (say) will change, with more accurate values coming from the atomic resolution structure. So we should be careful how we interpret RMSD.
Indeed, it would make an interesting statistics experiments to refine the same structure independently at low and atomic resolution to see if the RMSDs from teh low resolution data makde sense. ED Lattman ----- Original Message ----- From: Eleanor Dodson <[EMAIL PROTECTED]> Date: Wednesday, August 23, 2006 8:06 am Subject: Re: [ccp4bb]: Refmac vs Ramachandran (low resolution) To: Anastassis Perrakis <[EMAIL PROTECTED]> Cc: Clemens Vonrhein <[EMAIL PROTECTED]>, "Fred. Vellieux" <[EMAIL PROTECTED]>, Bernhard Rupp <[EMAIL PROTECTED]>, 'Mario Sanches' <[EMAIL PROTECTED]>, 'CCP4 Bulletin Board' <[email protected]> > If we find that the best estimate of RMSD bonds for structures > at > atomic resolution (where prior restraints do not have much effect) > is > 002 we can believe, cant we, that this a property of protein > structures > - ie some bonds are distorted by the protein environment and the > Xray > model illustrates this. (Bioinformaticists can and do analyse which > types of bond, in which environment, show the greates > distortions...) > And I would expect that the nature of the protein folded structure > is > more or less independent of the qualiy of the crystal it forms.. > > So doesnt it follow that you should get more or less the same RMSD > for > the geometry for structures refined at any resolution? Tassos's > example > fits in with that too.. > > All this assumes that you have built the structures correctly and > of > course you are much more likely to have made errors at 3A than at > 1A. > But tightening or loosening the restraints will not correct them.. > So > for initial builds, I often think it is sensible to release > restraints > on both B factors and bond lengths, then look for wild outliers - > they > often flag the dodgey regions. > Eleanor > > > Anastassis Perrakis wrote: > > > > > On Aug 23, 2006, at 9:59, Clemens Vonrhein wrote: > > > >> But what should be those values? I guess at 1A resolution I'd > expect a > >> > >> rmsd(bond) of 0.02 or larger. And at 3.5A I'd expect a > rmsd(bond) of > >> > >> 0.005 or smaller. But these are personal choices (same as I/sig(I) > >> > >> or Rmerge cut-offs in data-reduction). > >> > >> > >> But maybe I'm completely wrong here ... correct me please? > >> > > > > > > Although i think Clemens is very far from being wrong, as usual, > I > > cant help > > contributing another set of personal choices and a brief > justification.> > > Most important, and as Clemens explains, RMSD on bond lengths > > and standard deviation of the Engh&Huber distributions, are > different > > things. > > > > However, there must be some correlation: if (to exaggerate) a > certain > > type of bond is > > in the dictionary 1.0 +/- 0.1 A, I would doubt if it makes sense > to > > restrain it in a way that > > the deviation of the distribution of this type of bonds in a 3.5 > > A structure is 1.0 +/- 0.001. > > Thats just intuition and I have a concern that this is one of the > > things that intuition serves > > you wrong and statistics would say otherwise. But, i would be > curious > > to see why this is > > wrong/right. I will go and read the Cruickshank paper that Dirk > > pointed out to us ... > > > > I also think that beside personal preferences, Rfree would have > one or > > two things to say here. > > Should we simply aim for an RMSD between 0.005-0.020, that also > gives > > minimal Rfree ? > > > > I did a quick experiment with our current favorite nightmare at > 3.3 A > > resolution, 360 residues, no NCS. > > I am reporting the achieved RMSD, the Rfree and the number of > residues > > flagged by procheck as Rama outliers. The model I started with is > > 'towards the end' and had an initial 'original' rmsd of about > 0.016 > > when I started this test. > > > > RMSD Rfree Nres > > 0.020 31.2 31 > > 0.016 31.3 27 > > 0.010 31.4 23 > > 0.007 31.9 20 > > 0.005 32.2 22 > > > > from that I would say, that - for our structure - the middle job > is > > best: i would expect that after quickly fixing 4-5 residues in > the > > graphics and running a job that would give an RMSD about 0.015 to > be > > the best at the end. But, of course now I am tempted to take the > 0.007 > > model, look at it, and then slowly 'release it' towards the end. > ;-) > > > > Having said these I want to iterate one point that was made: > > > > Use hydrogens. It can only help and it costs nothing. > > > > And to make a new point: use TLS. > > > > The major reason that some crystals do not diffract at high > enough > > resolution, > > should be rigid body motions and TLS describes exactly that, > while being > > very 'cheap' in parameters, so use it. for the above structure is > > makes a huge > > difference. > > > > btw: i found torsional refinement as implemented in CNS not > useful > > compared with refmac5 / phenix or Buster. I was a bit surprised > by > > that, but ... > > > > A. > > > > > >
begin:vcard n:LATTMAN;EATON fn:EATON E LATTMAN tel;work:410 516 8215 org:Johns Hopkins University;Biophysics adr:;;Mergenthaler 237;Baltimore;MD;21218; version:2.1 email;internet:[EMAIL PROTECTED] title:Dean of Research, Professor of Biophysics end:vcard
