Hi Uma, The optimal weight is indeed resolution dependent, but hard to predict. In Refmac you can follow LLfree when you optimize the restraint weight and also keep an eye on the gap between R and R-free (it should not be too wide). Like Rob said, your geometry should be 'reasonable'. This may be a bit vague, but there is no clear target for bond/angle rmsd at a given resolution (some referees will disagree). If you look at the rmsZ values Refmac gives, the target is a bit clearer: rmsZ < 1.000. The average rmsZ does go down with resolution (i.e. lower resolution gives lower rmsZ), but an ideal value cannot be given easily (or at all). Tightening the restraints improves the effective data/parameter ratio of your model. You can also improve it by adding additional restrains (e.g. NCS restraints) or by removing parameters (e.g. changing the complexity of your B-factor model). Note that the absence of geometric outliers does not prove that your model is optimal. If you use too tight restraints you can end up hiding genuine fitting errors. Cheers, Robbie
Date: Fri, 27 Apr 2012 10:04:11 +0200 From: herman.schreu...@sanofi.com Subject: Re: [ccp4bb] Refmac and sigma value To: CCP4BB@JISCMAIL.AC.UK It all will depend on the resolution. At low resolution, relaxing the geometric restraints will allow the refinement program to tweak the model such that the difference between Fobs and Fcalc is minimized, but not that the model gets closer to the "truth". I once struggled for a long time with a 3.5Åish data set with a protein where the most important feature was a rather flexible loop. It was before maximum likelyhood methods and Rfrees and the only way I could get rid of the model bias was to use extremely tight geometric restraints. The Rfactor would go up, but suddenly the electron density maps would no longer accept incorrectly placed side chains and new features, not present in the model, would appear. So my advice: at low resolution use as tight restraints as possible and monitor with Rfree if you are going in the right direction. At high or very high resolution, you can follow what your diffraction data tells you. In fact many very high resolution structures (< 1.5 Å) have higher rmsd's for bond lenghts and angles as medium resolution structures. However, at medium or low resolution there is not enough data to justify to relax the geometric restraints too much. Best regards, Herman From: CCP4 bulletin board [mailto:CCP4BB@JISCMAIL.AC.UK] On Behalf Of Robert Nicholls Sent: Friday, April 27, 2012 9:25 AM To: CCP4BB@JISCMAIL.AC.UK Subject: Re: [ccp4bb] Refmac and sigma value Hi Uma, Altering sigma affects the strength of geometry restraints throughout the model - bonds, angles, etc. Choosing a very low sigma will cause geometry to be more tightly restrained towards "ideal" values, which is why you observe improvements in Coot validation. Note that strengthening the geometry weight causes the observations (data) to be less influential in refinement. The "risk" of this is that your model may no longer appropriately/optimally describe your data. You can assess this locally by manual inspection of the electron density, and globally by considering overall refinement statistics (as reported at the bottom of the Refmac5 log file). Ideally, you want your model to both describe the data and have reasonable geometry. Regards Rob On 26 Apr 2012, at 21:26, Uma Ratu wrote: Hi, Alex: > Which sigma do you mean? The one for automatic weight, not for Jelly-body refinement. I did not turn the "Jelly-body refinement" on. Thanks Ros On Thu, Apr 26, 2012 at 4:08 PM, aaleshin <aales...@burnham.org> wrote: Hi Uma, Which sigma do you mean? The one for Jelly-body refinement? J-B sigma=0.01 means very small fraction of the gradient will be used in each step. It is used usually with very low resolution (less then 3A) Alex On Apr 26, 2012, at 11:38 AM, Uma Ratu wrote: > > Dear All: > > I use Refmac5 to refine my structure model. > > When I set the sigma value to 0.3 (as recommended from tutorial), the > resulted model has many red-bars by coot validation (geometry, rotamer, > especially, Temp Facotr). > > I then lower the sigma value to 0.1, the resulted model is much improved by > coot validation. > > I then lower the sigma value to 0.01, the resulted model is almost perfect, > by coot validation and Molprobity. > > My question is: what is the risk for very low value sigma value? > > Thank you for your advice > > Ros
