[phenixbb] Discrepancy between R-factors from phenix.refine vs phenix Generate "Table 1"

Jon Schuermann schuerjp at anl.gov
Fri May 31 12:15:51 PDT 2013

Sorry if this is a little long, but I saw that more than one person was 
having a similar problem.

First, I agree with everything Nat has said, but I will explain a little 
bit more. Twinning and/or pseudosymmetry can be a hairy issue to deal 
with. In some cases, NCS can be a real pain. I have seen many cases 
where just pseudosymmetry is present, but people assume it is twinned as 
well. To be fair, it is very easy to come to that conclusion, especially 
at the early stages. Let me just set the stage with a hypothetical 

You collect a 2.5A dataset that processes well in P622 with an Rsym=0.11 
(overall). The intensity statistics look normal, which I won't go into 
right now because it is complicated (even though Nat eluded to it with 
the 'L-test'). You perform MR with a search model that has 60% sequence 
identity and get a solution with 1 molecule in the AU in P6522. The 
solvent content is 0.6. You run the solution through rigid-body, 
followed by positional and B refinement with simulated annealing. The 
resulting R=0.35, Rfree= 0.40. You look at the maps and the density does 
not look good. You can see density for most of the main-chain but almost 
all of the side-chain density is not present. You go about trying to 
build in the density but the R/Rfree do not drop much. What is your next 
step? Some would go to Se-Met or heavy atom soaks to solve by SAD or 
MIR, but the results are the same... Now What? Most people would say, 
its hexagonal where merohedral twinning is possible so it must be 
twinned. You rescale the data in P6, P312, P321, and even P3 
(tetartohedral twinning) and the Rsym is about the same in all the 
subgroups with P6 and P3 being the lowest with Rsym=0.08. You perform MR 
in P65 and get a solution with 2 molecules in the AU. To 'check' if it 
is twinned, you run refinement twice from the same starting model; 
first, twin refinement specifying the 'twin law' (from XTRIAGE), and 
second regular refinement and then compare the results. You think the 
results speak for themselves... with the twin refinement, the R=0.24, 
Rfree=0.28 with the twin fraction refining to 0.49 (nearly perfect 
twinning). The maps look great! The regular refinement (without 
specifying the 'twin law') refines to R=0.30, R-free=0.32 and the maps 
look better than P6522, but not as good as the maps from the twin 
refinement. You think, 'This has to be twinned!' But here is the kicker, 
IT'S NOT TWINNED!!! How could this be???

When you input the 'twin law' in refinement and the twin fraction 
refines to >0.4, Phenix detwins using the proportionality rules. This 
method uses the model in the detwinning process which WILL introduce 
model bias in the maps, so they look great. Even if your model is NOT 
correct, the maps will still look good, because of the model bias. For 
this reason, you should calculate several 'omit maps' over different 
regions of your structure and inspect them, or look at maps calculated 
from refinement when you did NOT specify the 'twin law'. As for the 
R-factor being lower, the calculated twin R-factor is usually lower than 
the standard R-factor since the equations are NOT the same. So, a drop 
in R-factor at this early stage is not necessarily conclusive that your 
data is twinned. There are a couple of papers explaining this (which Nat 
eluded to and I believe I reference below).

I have seen many people get 'itchy finger' and try to use the twin 
refinement too early. As Nat mentioned, you don't want to run it unless 
you are at the end of model building and ready for deposition, but the 
R-factor is still too high. Then it may be necessary to run refinement 
with the 'twin law' included to see if the R-factor drops 'significantly'.

To summarize, if you suspect an issue with twinning and/or 
pseudosymmetry since the maps don't look as good as they should for a 
given resolution and the R-factor is too high, rescale/reprocess your 
data in lower symmetry SG's and rerun MR. When you get a solution, 
refine it WITHOUT specifying 'twin law' and look at the maps.
     A)If they look much better than the solution in the higher symmetry 
SG, continue your model building and refinement (WITHOUT 'twin law') 
cycles in this lower symmetry SG until the model building is complete 
and you would be ready to deposit the structure in the PDB. If the 
R-factor is still too high, then run refinement including the 'twin law' 
with the twin operator from Xtriage. If you see an appreciable decrease 
in R-factor, then you probably have twinning. If the R-factor dropped, 
but is still not low enough, then either your model is still not 
completely correct or you have something else going on as well (see end 
of B below).
    B) If the maps still do not look any better than the higher symmetry 
SG, try MR in another lower symmetry SG and repeat. I sometimes go down 
to P1, if I have enough data. If the maps still look bad (assuming you 
have a complete dataset and MR solution), then you may not have a SG 
issue at all, but something else. I would look into other issues 
including anisotropy, pseudo-translational symmetry, order-disorder, 
missing molecules, etc. I will usually scroll through the images to see 
if I see something obvious. Reprocess the data more carefully looking 
for things like extra spots not predicted, or predicted spots that 
aren't actually there, etc.

Here are a few references to papers that explain some of these issues 
more in depth with examples:
Acta Crystallogr. (2008), D64, 99-107.
Acta Crystallogr. (2012), D68, 1541-1548.
Acta Crystallogr. (2006), D62, 83-95.

I hope this helps some people...


Jonathan P. Schuermann, Ph. D.
Beamline Scientist, NE-CAT
Argonne National Laboratory, 436E
9700 S. Cass Ave.
Argonne, IL 60439

Email: schuerjp at anl.gov
Tel: (630) 252-0682

On 05/30/2013 09:26 AM, Nathaniel Echols wrote:
> On Thu, May 30, 2013 at 7:09 AM, Heather Condurso <condurso at bc.edu> wrote:
>> I have a similar issue.  My data is nearly perfectly twinned so Xtriage usually fails to find any twin law.  However when I solve the structure in the lower symmetry space group with twinning my maps look beautiful.  In the higher symmetry space group, only 3 of the 6 monomers fit into nice looking density.  I simply use generate table 1 and uncheck the box so it won't regenerate R factors.  The discrepancy between the R-factors is however much more significant than 2%.  From refine I get 18/22 and from table1 28/31.
> That's actually a little surprising that Xtriage wouldn't work - would
> you be willing to send us the data?  Keep in mind that if you have
> perfect twinning and refine with the twin law, I think the maps will
> look beautiful pretty much by definition, because of the model bias.
> (Which doesn't mean that you did it wrong - the R-factors sound
> realistic - you just need to be very careful and make lots of omit
> maps.)
> It helps to have actual examples when warning against the dangers of
> twin refinement and model bias in general, so here's one (attached).
> This map isn't beautiful, but it matches the model very closely,
> including the complete lack of sidechains.  Another Phenix user got
> this result by accident with a mostly-if-not-entirely incorrect MR
> solution - fortunately the combination of missing sidechain density
> and relatively high R-free (~40%) prompted him to email us.  I suspect
> that if we searched the PDB we'd find an actual published example
> (especially considering how many spurious ligands have been found
> there).
> (Sorry to sound like a broken record here, but this keeps coming up in
> emails and workshops - perhaps we need better validation tools.)
> -Nat
> _______________________________________________
> phenixbb mailing list
> phenixbb at phenix-online.org
> http://phenix-online.org/mailman/listinfo/phenixbb

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://phenix-online.org/pipermail/phenixbb/attachments/20130531/ce7a400e/attachment.htm>

More information about the phenixbb mailing list