The documentation you quote is specifically referring to a "well-refined protein structure". When you deposit your model it will, one hopes, not contain "correctable modeling errors". What you do during refinement is a separate matter. For the very reason you describe, it is often useful to allow poor geometry at times during refinement. Simulated Annealing is the ultimate as the model can look pretty ugly at high temps. It may not be common practice, but I consider an unusual amount of movement of a residue while the geometry is relaxed and tightened again to be a flag suggesting additional scrutiny during model building. When there is a significant conformational error in the model the geometry of the atoms is often distorted, as you note. I've found that the same groups of atoms tend to stand out regardless of the tightness of the restraint. With a stronger weight the deviations are smaller on an absolute level, but they are deviations non-the-less. This brings us to Z scores. Judging a model based on rmsd Angstrom or rmsd degree is quite problematic. The method has a long history but we are well past the time when it is the best. Particularly for angles, there are some restraints which are tighter than others. In the E&H library the NCaC angle has a SD of 2.8 deg while the internal angles in the Phe ring are able 1.7 deg. A one deg deviation in one is quite different than one deg in the other. The deviations need to be normalized. Note that a 2.0 deg SD is too tight for NCaC while too loose for the Phe ring. At high resolution, where you don't need geometry restraints on the well ordered parts of your model, you will be penalized for your perfectly okay main chain bond angles. If you have an angle with a 3 SD deviation from expectation you should have a look and it doesn't matter if the standard deviation is 1 or 2 degrees. That standard deviation is from the library and not the rmsd of your model. In the end your model should be around 1 rms delta/sigma if your resolution is so high that restraints don't matter and lower if you have more typical resolutions. How low depends on the "quality" of your data and I'm not sure anyone has a measure of that that I find useful. Empirically one minimizes the free R to set the strength of the restraints. Dale Tronrud P.S. For angles particularly, it is usually observed that very high resolution models have bond angles deviations of greater than one. The Andy Karplus lab has published the view that the E&H library underestimates the quantity of angle variability. His new Conformation Dependent Library for protein main chain is normalized better and very high resolution models are scored more reasonably. (Disclaimer: I've been involved in this project.) On 12/11/2017 7:44 AM, wtempel wrote:
Hello all, the documentation https://www.phenix-online.org/documentation/faqs/refine.html states that
This is somewhat controversial, but absolute upper limits for a well-refined protein structure at high resolution are typically 0.02 for RMS(bonds) and 2.0 for RMS(angles); usually they will be significantly lower.
I understand that exceedingly high RMSDs from ideal could indicate overfitting. On the other hand, local deviations from ideal geometry may point to correctable modeling errors, and I am concerned that overly tight restraints may cause that diagnostic tool to become less sensitive, or local errors to be spread in to the model. For what bond and angle rms deviations from ideal do my colleagues on the BB aim and how have they arrived at those targets? Did my web search miss a relevant paper? Thank you in advance. Wolfram Tempel
_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb Unsubscribe: [email protected]