[phenixbb] Very low resolution data set

Nathaniel Echols nechols at lbl.gov
Wed Jul 3 14:01:52 PDT 2013

On Wed, Jul 3, 2013 at 1:42 PM, Xiang Li <li338 at purdue.edu> wrote:
> I am working on a very low resolution data set (14 A). It is a designed DNA crystal and unit cell size & symmetry fit well with the assumption. When I do the refinement I found the Rfree is extreme high, ~0.49 while Rwork is about 0.4. But the model seems fits in well with the density map.

Model bias is totally pervasive at this resolution - it is very easy
to get a map that looks exactly like the model even if the model is
entirely wrong.  You can also set all amplitudes to 1000 and
re-calculate the map and it will probably still look like the model.
(At least this is what happened when I tried it at 6Å.)

Personally, I would not believe anything at this resolution without
heavy atom information - at least a difference map showing clear sites
(if you can measure anomalous scattering, even better).  MR is also
pretty difficult unless you have a huge structure with a very distinct
shape; what were the statistics from Phaser?

> 1 What is the expected Rfree and Rwork for such low resolution data? I looked up PDB bank, and there is a 11.5A crystal with Rfree=~0.4. But not sure whether there is a kind of value for Rfree to consider the refinement works or not.

It's difficult to come up with any meaningful statistics, because
these structures are usually barely refined, for obvious reasons.
(R-free in the high 30s is not uncommon for published structures,
however.)  I don't understand the math well enough to know what to
expect, but I've seen R-factors in the mid-to-low 40s for structures
refined against nonsense low-resolution data (again ~6Å).  An R-free
of 0.49 is very suspicious.

> 2 There might be overfit problem during the refinement. Currently what I did  is to optimize the weight with PHENIX gui, add restraints between basepairs and NCS restraint. I tried to manually adjust the weight, but seems the optimized weight gives the best result. Are there other things I should try?

I would stay conservative at first: stick to rigid-body and TLS
refinement if possible (I guess it depends on how much your DNA
bends).  Full refinement of coordinates and B-factors can be done
stably if very tightly refined, although this offends some purists -
but you have to be especially careful.  In any case I wouldn't worry
too much about the refinement until you're certain that the phasing is

> 3 How to confirm that the result is real?

Heavy atoms!  And maybe lots of omit maps, but I'm not sure how those
will behave either at low resolution, and you still need to be careful
about phase bias.


More information about the phenixbb mailing list