Can geometry restraints include non-ideal target values?
I see that phenix is using the same kind of simplistic statistical restraints as CCP4. Are there plans to improve the geometry restraints once the rest of phenix is stable? The problem with simple statistical restraints as used by CCP4 is that they completely ignore correlations due to interactions of multiple restraints. This is particularly bad for puckered rings like proline and nucleic acids. Nucleic acids get around this by using a fixed sugar pucker, which results in a lot of bad geometries for non-helical nucleic acids (many of which make it in to the PDB). However, if you consider the geometries "ideal value" and "ideal value esd" as just another way to write a harmonic restraint, then it is possible to make an improved restraint library. In this case, some individual geometry restraints will have target values that restrain towards non-ideal values in order to produce the correct result when all terms are combined. The problem is that these terms will give higher RMSD values, which would probably confuse software like REFMAC, which is designed to tweak restraint weights based on RMSDs. So, does phenix have any built-in restraint adjustments that is based on current rmsds? If so, is there a way to disable it? Joe Krahn
I have used 3 programs to refine RNA structures recently: CNS, Refmac and Phenix. Of these, only CNS has the temerity to enforce what it thinks is a correct sugar pucker to the extent that it will pull atoms out of the electron density. Refmac and Phenix both do a much better job of refining non-canonical sugar puckers in RNA in my experience. Joe Krahn wrote:
I see that phenix is using the same kind of simplistic statistical restraints as CCP4. Are there plans to improve the geometry restraints once the rest of phenix is stable?
However, they all use hard-wired sugar puckers. PHENIX is currently essentially the same as REFMAC. They do better than CNS only because the restraint energies are reduced. Although they won't pull atoms completely out of the density, they still produce bad geometries when the actual sugar puckers differ. I just created a C2'-endo patch for PHENIX, which was needed to avoid getting several strained angles. Joe William Scott wrote:
I have used 3 programs to refine RNA structures recently: CNS, Refmac and Phenix.
Of these, only CNS has the temerity to enforce what it thinks is a correct sugar pucker to the extent that it will pull atoms out of the electron density. Refmac and Phenix both do a much better job of refining non-canonical sugar puckers in RNA in my experience.
Joe Krahn wrote:
I see that phenix is using the same kind of simplistic statistical restraints as CCP4. Are there plans to improve the geometry restraints once the rest of phenix is stable?
phenixbb mailing list [email protected] http://www.phenix-online.org/mailman/listinfo/phenixbb
Hi Joe, Our group (Berkeley) is working with the Richardsons to analyze RNA sugar puckers and adjust the restraints for the 2' and 3' states. This should become available relatively soon. Ralf
Ralf W. Grosse-Kunstleve wrote:
Hi Joe, Our group (Berkeley) is working with the Richardsons to analyze RNA sugar puckers and adjust the restraints for the 2' and 3' states. This should become available relatively soon. Ralf
OK, but is there a plan to create restraints that correctly restrain towards any sugar pucker without needing to apply hard-wired conformation patches? I suspect not, because this is not possible with the simple statistical approach used by the current monomer libraries. If the 5-membered ring angle and dihedral restraints are defined such that they are somewhat "strained", then it is possible to end up with energy minima that automatically match the actual puckers. But, implicitly "strained" restraints mean that their target values are intentionally NOT the ideal values. A similar problem occurs with proline puckers. To illustrate the problem, if you look at angle distributions of low-resolution structures refined with REFMAC, you will find that some angle distributions do not match the ideal distributions that the ideal values were derived from. This happens because some values are correlated, but simple individual "ideal value" statistics ignore the correlations. Joe
participants (3)
-
Joe Krahn -
Ralf W. Grosse-Kunstleve -
William Scott