I find Lothar's comments regarding H and RSRZ excellent! I would think of it as a pretty much bug report. I hope developers at that end listen. This goes very well in line with Phoebe's comment earlier today.
Pavel On Mon, Nov 28, 2016 at 2:51 PM, Dale Tronrud <[email protected]> wrote: > On 11/28/2016 12:52 PM, [email protected] wrote: > >> I found that one can get RSRZ to go way down by loosening the geometry > >> restraints. The result is a crappy structure and I don't recommend > doing > >> that, but it does get all the atoms crammed into some sort of density. > > > > Your observation is quite interesting. I can add this: when we were > working > > with low to medium resolution structures, deleting the hydrogen atoms > from > > the model after refinement moved the very bad RSRZ statistic to about the > > average in the given resolution range! Note, no re-refinement was done > just > > a simple deletion of the riding H-atoms. I find this to be odd given the > > fact that, say the phenix developers favor the inclusion of H-atoms on > > riding positions even in cases of low resolution structures. (I assume > the > > refmac5 and BUSTER-TNT developers have also a favorable opinion about > > including H-atoms in the final model - and during refinement). > > > > In my mind, it may be tempting to delete H-atoms to improve this > statistic but > > when you use them in refinement they should be included regardless of the > > outcome of the RSRZ analysis. > > Of course, if you trick a validation statistic like this you haven't > accomplished anything. All you are saying is that one should rank RSRZ > scores with and without hydrogen atoms separately. Perhaps you should > suggest that to the PDB validation people. > > Dale Tronrud > > > >> > >> RSRZ, in my most humble of opinions, seems like one of those statistics > that > >> is far more useful in theory than reality. Particularly for > >> medium-resolution structures, the fit of each entire side chain to the > density > >> is likely to be imperfect because the density is imperfect, especially > toward > >> the tips of those side chains. > >> > >> Then again, it can be a good flag for bits of the structure worth a > second > >> look in rebuilding. > > > > The latter is certainly true. It may mean that the developers of RSRZ > > analysis need to tune it a bit to make it fully useful. > > > > L. > > > >> > >> ________________________________ > >> From: CCP4 bulletin board [[email protected]] on behalf of Matthew > >> Bratkowski [[email protected]] > >> Sent: Tuesday, November 22, 2016 10:12 AM > >> To: [email protected] > >> Subject: [ccp4bb] Calculation of RSRZ Score in PDB Validation Reports > >> > >> Hello all, > >> > >> I was wondering if anyone knew how the RSRZ score was calculated in the > >> protein data bank validation reports and how useful of a metric this > actually > >> is for structure validation? I am trying to improve this score on a > structure > >> that I am working on, but I'm not really sure where to begin. From my > >> understanding, the score is based on the number of RSRZ outliers with a > score > >>> 2. In my case, I have several residues with scores between 2 and 4, > but at > >> least by eye, fit to the electron density does not look that bad. > Hence, I > >> can't justify deleting them to try to improve the score. If the score > is just > >> based on percent of outlier residues, then for instance wouldn't a > structure > >> with say 20 residues modeled with no corresponding electron density > have the > >> same score as a structure with 20 residues with RSRZ values of say 2.5? > >> > >> > >> I was also wondering how the resolution of the structure relates to the > score? > >> Glancing through several pdb validation reports, I noticed some > structure > >> with low resolution (3.5 A or lower) with relatively high scores, while > others > >> with high resolution (2 A or higher) getting low scores. It is > reasonable to > >> assume that a structure of lower than 3.5 A would be missing several > side > >> chains and may also have some ambiguous main chain electron density, > which > >> should in theory increase the RSRZ score. While of course every > structure is > >> different and the quality of it due to the rigor of the person building > the > >> model, I was wondering if there were any general trends related to > resolution > >> and RSRZ score. > >> > >> Thanks, > >> Matt > >> > > >
