[phenixbb] clashscore discrepancies

Nathaniel Echols nechols at lbl.gov
Sat Apr 27 16:50:39 PDT 2013

On Sat, Apr 27, 2013 at 4:21 PM, Alex Theodossis
<alex.theodossis at monash.edu> wrote:
> Given that PHENIX is now giving me the best results and all scenarios give
> reasonable results I am less inclined to be concerned. However, I still find
> these inconsistencies rather puzzling and am curious to learn what is
> causing them.

The routines in Phenix strip out and replace hydrogens when
calculating the clash score, so I would not expect Phenix versus
Molprobity with existing H atoms to be the same in any case.  There
are many other possible explanations for the remaining discrepancies,
but the fundamental reason is that the only actual code in common
between the Phenix validation and MolProbity is Reduce and Probe (and
KiNG, but that's just for visualization).  I haven't looked at the
actual MolProbity code in years, but it's a combination of other
scripts built up over the years, and I don't think they're even
written by the same people.  With different implementations it is
extremely difficult to guarantee numerically identical output.  Even
worse, we can't even guarantee identical output for the same code
built with different compilers.  (I often get very different results
running Phenix on Windows.)

Of course the underlying statistics and overall methodology/philosophy
used is the same, which is why we tend to also refer to the various
Phenix tools as "MolProbity".  In general, unless there are specific
features in the MolProbity server which you need and which we haven't
implemented in Phenix yet, there is no point to running both tools.
Jeff has begun modernizing the server code to use CCTBX as the
backend, which should make the implementations more consistent in the
future - however, even then I would not expect the results to always
be in perfect agreement.


More information about the phenixbb mailing list