Refinement with pseudo-translation

From: Shiva Kumar
Date: 22 March 2012 22:28


Dear CCP4bb members

I have a 3.0 Å dataset which has an off-origin peak of height 36% in ‎patterson map. The peak is at fractional co-ordinates 0, 0.5, 0.5. Data has been indexed in P2(1)22(1) SG using HKL2000.  I have located all the molecules in asu (as far as I know) using Molrep with the 'locked rotation' and 'Use' PST feature. After 1 round (20 cycles) of rigid body and 1 round (10 cycles) of restrained refinement (Refmac), the R and Rfree are 49 and 53 %. Although the R factors are very high, I feel the solution might be correct because the electron density follows c-alpha trace in almost all places. To be sure, I deleted a beta strand from the structure's core and repeated the refinement and found that the electron density for the strand was still present (I have no experimental phases).  I have the following questions:

1) Are the final R factors never expected to reduce to acceptable values given the 36% off-origin peak?

2) What is the best way to settle the SG? I was considering P2(1)2(1)2(1), P2(1)22, P2(1)2(1)2 and P2(1)22(1), considering the off-origin peak is at 0,0.5,0.5. I found all the molecules in asu only using the P2(1)22(1) data in Molrep. Is this the best way to settle my SG?

3) Some CCP4bb archives advise either refining against weak and strong reflections alternatively (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1010&L=ccp4bb&D=0&1=ccp4bb&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4&P=204490) or refining against medium intensity reflections. Should I also be doing these things? If yes, then what is the best way of doing it?

Your suggestions and corrections to my interpretation of our data would be appreciated.

Regards
Shiva

----------
From: Esko Oksanen

  Dear Shiva,

  Refining against weak and strong reflections separately only works if your intensity distribution is bimodal, because the problem arises from the assumptions of the scaling protocol (that typically assumes a unimodal distribution) that leads to underestimating the sigmas of the strong reflections and overestimating those of the weak reflections. So the first thing to look at is the cumulative intensity distribution (e.g. from Truncate) and if it looks normal, this trick won't help you. If you do see a cumulative intensity distribution biased towards weak reflections, then you could try scaling the pseudo-absent reflections separately, as explained in (doi:10.1107/S0907444906031519). I would expect that in this case the R-values will remain high, since the data are correlated and the normal assumptions for the expected values of R don't really hold any more.

  HTH,

  Esko

----------
From: Garib N Murshudov

Dear Shiva

I would be careful with omit maps in the presence of pseudotranslation. You have to omit bits of molecule from PST related molecules simultaneously, otherwise you will have bias. If you look at the equation it becomes clear why.

R factors 50/53 is close to random (in the presence of solvent random R factors seem to be around 53). 

If you want you can send your data and we can try to have a look. I cannot promise anything but at least you can have a second opinion

regards

Garib

----------
From: <Herman.Schreuder


Dear Shiva,

First a word of caution. Your diffraction pattern is the convolution of the molecular transform and the lattice transform. Your molecular transform (determined by the orientation of your model) will almost certainly be right. The influence of e.g. one misidentified screw axis will not be very strong. This means that if you leave out a strand, density may come back, even with a misidentified screw axis. I would reprocess the data in P1 and generate a model (2 molecules in 1 pdb file) consisting of one molecule at (0,0,0) and one at (0,0.5,0.5) in the orientation you found already and run molrep with that model.

It could also be illuminating to look at what is behind the off-origin peak: it means that you have one molecule (say A) at (0.0.0) and one molecule (say B) at (0,0.5,0.5) which are very similar, but not identical. If they were identical, your off-origin peak would be crystallographic and an origin peak. If you look at the relative phases, the phase of molecule A would be 2pi(h*0+k*0+l*0)=0 whereas the relative phase of molecule B would be 2pi(h*0+0.5*k+0.5*l)=pi(k+l).

The total diffraction is then: Ftotal = FA (cos(0) + isin(0)) + FB (cos(pi(k+l) + isin(k+l)).
For k+l even pi(k+l)= 0 or 2pi and Ftotal = FA (1 + 0) + FB ( 1 + 0) = FA + FB (strong reflections)
For k+l odd  pi(k+l)= pi       and Ftotal = FA (1 + 0) + FB (-1 + 0) = FA - FB (weak reflections, zero if FA=FB)

This means, you could also discard all k+l reflections by reprocessing the data in the smaller unit cell and run molrep with that data. You will have to search for only one molecule and the final density will be the sum of molecules A and B. This may give you hints how the packing of your larger unit cell may look like. As I said before, these almost crystallographic NCS is extremely trickey and can easily fool automatic programs and crystallographers alike.

Good luck!
Herman

----------
From: Eleanor Dodson

You don't say how many molecules you are looking for?

I would try to first work in SG C2 pretending that the pseudo translation is 100% of the origin.

Reindex the data as k,l,h, change the SG to C2 and then redo the scaling and merging. You will lose all the reflections k+l =2n+1 in your indexing - h+k = 2n+1 after reindexing, but the MR should be simpler.

Once you have a C2 solution you can convert it back to the orthorhombic one with a certain amount of intellectual struggle! The C2 origin along your b axis is free, so you will have to

Or you can use ZANUDU from the ysbl software web site which will maybe help.

Eleannor

-- 

----------
From: Boaz Shaanan


Hi,

I would go even further down SG ladder, i.e. all the way to P1, reprocess the data in P1 (you won't lose reflections with special indices that way, the only thing to worry here is completeness, but hopefully you recorded highly redundant data) and do the MR. It worked for me and for others.

 Cheers,

                Boaz

----------
From: Boaz Shaanan


Just to add to my previous note: and then (after MR) you can go to Zanuda as Eleanor suggested, this time with the P1 data and model, in case you haven't figured it out by then.

              Boaz