[phenixbb] Are sigma cutoffs for R-free reflections cheating?

Ed Pozharski epozh001 at umaryland.edu
Mon Dec 7 07:20:29 PST 2009


Pavel,

Here is an example of the SCALEPACK log-file.  My reading of it is that
in the highest resolution shell I have 25% of reflections with I/sigma
less than zero, thus negative intensities.  Again, it doesn't bother
*me*, because I use French&Wilson to convert the negative intensities to
positive amplitudes and would like to reiterate that neither
phenix.refine nor any other program can be blamed for what is
essentially user error in assigning these reflections zero amplitude.
However, Joe's point, imho, was that *if* they are assigned zero
amplitudes significant portion of the data will be excluded if FOBS>0 is
required.  They are not a few, in the dataset that I pulled this from
they are apparently >12000 out of about 100000 overall.

Shell            I/Sigma in resolution shells:
  Lower Upper      % of of reflections with I / Sigma less than
  limit limit     0     1     2     3     5    10    20   >20  total
  50.00  4.52   0.6   1.8   3.4   5.3   8.8  17.4  33.4  66.3   99.8
   4.52  3.59   1.0   2.7   5.6   8.6  14.7  28.1  50.4  49.5   99.9
   3.59  3.14   2.8   7.2  13.1  19.5  30.2  48.7  70.8  29.2   99.9
   3.14  2.85   6.8  16.4  27.9  37.0  50.4  68.9  86.1  13.8   99.9
   2.85  2.65  11.5  26.2  41.4  52.0  65.5  81.5  94.0   5.8   99.8
   2.65  2.49  15.4  35.0  52.4  63.4  75.3  89.0  97.0   2.7   99.7
   2.49  2.37  17.9  41.1  59.0  69.9  80.8  91.8  97.9   1.6   99.6
   2.37  2.26  21.1  46.6  64.9  75.0  84.9  93.6  98.5   1.0   99.5
   2.26  2.18  22.6  50.5  70.0  79.8  88.7  96.1  98.9   0.3   99.2
   2.18  2.10  25.1  56.1  76.0  84.9  92.4  97.8  99.1   0.1   99.3
 All hkl       12.4  28.3  41.2  49.4  59.0  71.1  82.4  17.2   99.7

> I was just wondering how better (and what is "better" in 
> this case) the refinement results will be if one includes the negative 
> (zero) reflections into the process of refinement, and did anyone 
> clearly demonstrate it. That's all I asked.

This is indeed the key question - how do you determine if your model is
getting "better"?  R-factors are no good here, as they will definitely
improve upon rejection of weak reflections.  I guess what most people
would accept is the following:

1.  Generate synthetic data (e.g. from molecular dynamics trajectories)
and use MLFSOM and your favorite data reduction software to obtain a
dataset.
2.  Convert negative intensities to zero amplitudes.
3.  Refine the model with and without FOBS=0 cutoff.
4.  Determine the additional error introduced by the data rejection.

My suspicion is that the result will be more complex than a simple
yes/no answer.  It's possible that rejection of negative intensities
plays significant role only in certain resolution domain, but at least
the effect should be resolution-dependent (as well as vary from dataset
to dataset).

But one thing we can be certain about.  Except for the unlikely but real
scenario you mentioned when data well beyond I/s=1 resolution cutoff is
somehow included, it is my expectation that by including the weak
reflections the model quality will *not* be decreased.  If I am right
(and that is an if), then removing the FOBS=0 cutoff does no harm and
has the potential to improve models at least on some occasions.

Cheers,

Ed.


-- 
Edwin Pozharski, PhD, Assistant Professor
University of Maryland, Baltimore
----------------------------------------------
When the Way is forgotten duty and justice appear;
Then knowledge and wisdom are born along with hypocrisy.
When harmonious relationships dissolve then respect and devotion arise;
When a nation falls to chaos then loyalty and patriotism are born.
------------------------------   / Lao Tse /





More information about the phenixbb mailing list