Dear all
Thanks for your help.
I attache the self rotation function from molrep. 6 fold NCS is
visible. I could not find log file for
phenix.reflection_file_converter
--use-lattice-symmetry-in-r-free-flag-generation. How do I make sure
the test set are chosen considering the NCS?
I did anisotropic ADP refinement in phenix as suggested by Robbie
Original Run TLS+isotropic ADP
Rwork 0.1828
Rfree 0.2081
RMSD bonds 0.006
RMSD angles 0.913
comments: bits of positive and negative density at the compound position
Run 1, anisotropic ADP
Rwork 0.1674
Rfree 0.2067
RMSD bonds 0.006
RMSD angles 0.869
comments: still has positive and negative density at the compound
position, but I re-modified some waters and loops, use this PDB as
input for following refinements.
Run 2, TLS+isotropic ADP+X-ray/stereochemistry weight optimization
Rwork 0.1818
Rfree 0.2074
RMSD bonds 0.017
RMSD angles 1.563
comments: Compound density improved. Bond angle deviations are mostly
at the compound.
Run 3, anisotropic ADP+X-ray/stereochemistry weight optimization
Rwork 0.1674
Rfree 0.2056
RMSD bonds 0.014
RMSD angles 1.286
comments: compound density is acceptable. Bond angle deviation are reduced.
Run 4, TLS on protein+anisotropic ADP on
compound+X-ray/stereochemistry weight optimization
Rwork 0.1921
Rfree 0.2135
RMSD bonds 0.015
RMSD angles 1.432
comments: I had not done separate TLS and ADP before. Surprised to see
it made the refinement statistics worse .
Question: 1) At 1.6 angstrom resolution, the anisotropic ADP
refinement is not justified as the Rfree did not decrease
significantly. Is my understanding correct?
2) It looks like the weight optimization improves the compound
density. Could the cif file be the problem? I input SMILES to phenix
elbow AM1 to generate the cif file.
3) For Rfree set, should I force the index to P6 regardless of Rsym,
generate test set and expand to P1? The Run 3 could be deposited. But
Rfree is not improved comparing with the TLS refinement+isotropic ADP.
Thanks for your time and reply in adance. I probably should send this
to phenix board. But my original question get answers from CCP4 board.
Best,
Zheng (Joe) Zhou
On Mon, Mar 7, 2016 at 8:26 PM, Gerard Bricogne <[log in to unmask]> wrote:
> Dear Zheng,
>
> Robbie beat me to it, and his reply is very thorough.
>
> First of all, apologies for leaving your enquiry unanswered for
> the whole of last week: we were all commandeered into reciprocal space
> during that time. Claus will shortly be replying to your TLS question
> on buster-discuss.
>
> I was going to ask you for a little more information on what you
> mean by "I used lattice symmetry to generate test set in phenix." Do
> you mean that your P1 lattice is nearly hexagonal? Or that you used a
> feature that tries to assign free reflections as closely related by a
> 6-fold symmetry axis (that you would have identified through a self-
> rotation function) as possible?
>
> Generally, the very same reason that makes it difficult to find
> an independent test set under a regime of high-order and genuine NCS
> implies that you don't need it! This type of NCS couples amplitudes
> and phases so inextricably (unlike crystallographic symmetry, that
> relates amplitudes to amplitudes and phases to phases without creating
> any interaction between amplitudes and phases) that it essentially
> protects you against phase bias: you can't get good agreement with the
> measured amplitudes without at the same time being forced towards
> phases that are a good approximation to the correct phases. This is
> where the power of this type of NCS comes from in enabling you to
> iteratively improve poor starting maps into stunnning ones, if you
> have that NCS accurately described.
>
> In other words: IF your characterisation of your NCS and of your
> space group are correct, don't worry about the small Rwork-Rfree gap:
> you are in a situation that is far away from the typical one at that
> resolution, and the statistics for structures of similar resolution,
> most of which will not have NCS or not of such a high order, should
> not intimidate you.
>
> If the difference density for the ligands was better with BUSTER,
> was mostly consistent with the NCS without that having been imposed on
> it, and was interpretable and refinable, then that would be quite a
> bit of evidence that your result is correct.
>
>
> With best wishes,
>
> Gerard.
>
> --
> On Mon, Mar 07, 2016 at 12:53:14PM +0100, Robbie Joosten wrote:
>> Hi Zheng,
>>
>> Assuming your choice of test set is not terrible, the small gap could be fine. That is, if the symmetry is really P1 and not higher (check this some more, 6-fold NCS in P1 makes people suspicious). With high NCS is becomes increasingly difficult to get an independent test set, this is why with viral capsids some authors don't even bother using a test set.
>> That said, a small R-factor gap may mean that your model is not done yet. At 1.6A you can try a lot of things such as anisotropic B-factors instead of TLS (Ethan Merritt wrote a nice paper about this, you can also try PDB_REDO to use the Hamilton test he proposes) and you can build a lot of non-protein, so go through the difference density a bit more thoroughly than normal. The average R/R-free at this resolution is 18.0/21.2 (n = 2321) in the PDB and 16.7/19.5 in PDB_REDO, so the average gap is quite a bit larger. The numbers are similar if you only take the P1 cases (n =110).
>>
>> Cheers,
>> Robbie
>>
>>
>>
>>
>> > -----Original Message-----
>> > From: CCP4 bulletin board [mailto:[log in to unmask]] On Behalf Of
>> > Zheng Zhou
>> > Sent: Monday, March 07, 2016 11:23
>> > To: [log in to unmask]
>> > Subject: [ccp4bb] Rwork Rfree gap differences
>> >
>> > Dear all,
>> >
>> > I have sent the question to the buster discussion board two weeks ago.
>> > I think I'd better consult this problem on the active ccp4 discussion board.
>> > Sorry for a bit off topic.
>> >
>> > I recently solved a P1 crystal structure, 6 fold NCS at 1.6 angstrom resolution.
>> > I used lattice symmetry to generate test set in phenix.
>> > The phenix.refine gave me Rwork/Rfree = 0.1828/0.2081. There are pieces of
>> > positive and negative density for the compounds.
>> >
>> > So I tried buster refinement with autoncs, TLS. The density for the
>> > compounds are much improved, no obvious positive density. However, It
>> > generated Rwork/Rfree = 0.1801/0.1912. I also find similar R factor examples
>> > at the similar resolution in the buster wiki:
>> >
>> > https://www.globalphasing.com/buster/wiki/index.cgi?AutoBusterExample4
>> > chawaterNCS
>> >
>> > Would the small gap between Rwrok and Rfree be an issue? I have not
>> > deposited buster refined PDB yet, would this acceptable? Any reference to
>> > help if any reviewer questions about the small gap?
>> >
>> > Thanks for your advice in advance.
>> >
>> > Best,
>> >
>> > Zheng
|