On Tue, Oct 5, 2010 at 11:53 AM, Mark A Saper
If I take my well-refined structure as output by phenix (.pdb) and use it directly on the Molprobity web site, the clash score is much lower: 2.75 for 39 total bumps. Many of these involve residues or waters with high B factors. If I first remove all hydrogens from the phenix-refined .pdb file with a text editor, and upload it to the Molprobity web site, then add hydrogens, and then calculate clash score, the score is now 8.69 or 123 bumps for 14150 atoms. This is *identical* to what phenix calculates. Therefore, why doesn't phenix feed the refined coordinates directly into the Molprobity (probe) calculations? Has this change been made in one of the nightly builds?
There is a discrepancy in hydrogen bond lengths in Reduce versus the monomer library - I believe Reduce is using the distance to the center of the nucleus instead of the center of the electron cloud. phenix.refine will regularize the hydrogen geometry before refining the non-hydrogen coordinates, so it is correcting the bond lengths to be appropriate for X-ray refinement. It looks like they're getting stripped off and replaced when we run the clashscore calculations, so you end up with the longer bonds and thus more clashes. I think the consensus among the group was that these are more appropriate for interpreting (instead of refining) the model, but I'll double-check this with everyone. The GUI could at least be more explicit about how the calculations are done, however. (By the way, a clashscore of 8.69 is pretty decent anyway.) -Nat