[phenixbb] Odd behaviour of real-space refinement

[Tristan Croll]

Hi all,


I'm just finalising the re-refinement of a 3.6 Angstrom structure. The stats in general are excellent (Rwork/Rfree = 0.219/0.260; Ramachandran outliers 0.12%; Clashscore < 2; rotamer outliers 2.1%) and I've gone through the whole structure to check and double-check fit to the density. I'm very confident that it's effectively correct.


When I run through a fairly standard refinement protocol (XYZ coords, Individual B-factors, optimizing weights, secondary structure restraints), all seems fine - the Rfree rattles around within 0.2% or so. But when I include the real-space refinement option, very strange things start to happen. Each rsrl step causes a 3-4% jump in both Rwork and Rfree, and jumps the Clashscore up to 40 or so. Rotamer outliers briefly go to zero but then increase substantially (to 6+%) in subsequent steps, and the number of Ramachandran outliers increases. I thought I'd missed some dreadful error in my structure until I looked more closely at the results. Some very nonsensical changes have been made - for example, an arginine in the middle of a beta-sheet (admittedly with no observable density associated with its sidechain, but register is well determined by surrounding residues and backbone density is clear) was flipped from its original orientation to point directly into the hydrophobic core of the domain. Somewhat surprising, to say the least.


Is this a bug in the algorithm, or is it simply that 3.6A is too low resolution for it to work reliably?


My version of Phenix is a nightly build from about 2 weeks ago. Happy to share coordinates and log files if necessary.


Thanks,


Tristan

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20150222/c26db788/attachment.htm>