CCP4 Bulletin Board Archive: sudden drop in R/Rfree

From: Rajesh kumar
Date: 2 March 2012 15:54

Dear All,

I have a 3.3 A data for a protein whose SG is P6522. Model used was wild type structure of same protein at 2.3 A.

After molecular replacement, first three rounds of refinement the R/Rf was 26/32.8, 27.1/31.72 % and 7.35/30.88 % respectively.

In the fourth round I refined with TLS and NCS abd added water and the R/Rf dropped to 19.34/26.46. It has almost 7% difference. I also see lot of unanswerable density in the map where lot of waters were placed. Model fits to the map like a low resolution data with most of side chains don't have best density.

I was not expecting such a sudden drop in the R/Rfree and a difference is 7.2%.

I am wondering if I am in right direction. I am not sure if this usual for 3.3A data or in general any data if we consider the difference.

I appreciate your valuable suggestions.

Thanks

Raj

----------
From: Regina Kettering

Rajesh;

I am not sure that you have a high enough data:refinement parameters ratio to refine TLS. It just adds more parameters to refine that can lead to over-refinement of your model, especially at the 3.3 A.

HTH,

Regina

----------
From: Ethan Merritt

I'm afraid you've got this completely backwards.
TLS uses very few parameters, and is especially useful at low resolution.
At 3.3A I would recommend trying a TLS model _instead_ of refining
individual B factors.

NCS restraints also help a lot at low resolution.

So the drop is believable, but...

You should first worry about "lots of waters were placed".
It's there that many extra parameters have been added, perhaps leading
to over-fitting. I would not expect 3.3A data to justify placement of
more than a handful of waters at most.

If you're parameter counting, you might note that 5 water molecules add
more parameters than 1 TLS model. But the TLS model may improve the model
everywhere, whereas the waters will only suppress a few local difference
density peaks.

cheers,

Ethan

----------
From: Thomas Womack

On 2 Mar 2012, at 16:02, Regina Kettering wrote:

Rajesh;

I am not sure that you have a high enough data:refinement parameters ratio to refine TLS. It just adds more parameters to refine that can lead to over-refinement of your model, especially at the 3.3 A.

TLS only adds twenty parameters per chain; so it's a really parsimonious thing to do at low resolution.

I'd say that adding lots of waters at 3.3A (at four parameters per added water) was much more likely to be the cause of a very wide R/Rfree gap.

I'm a bit worried that a user working at low resolution on a protein with more than one chain per ASU is not using NCS from the very beginning; that's another good way of adding more restraints and effectively getting the parametersto-data ratio down (because the 'parameters' in that ratio is really 'parameters minus K * number of restraints'; there is scope for a lot of debate as to the right value of K, it clearly depends on the strength of the restraints)

If he's using the Global Phasing refinement software, I would strongly suggest that Rajesh use targetting to the initial molecular replacement result throughout the refinement, as yet a third way of adding more restraints.

Tom Womack (Global Phasing)

----------
From: Francis E Reyes
I've found the following article to be useful in defining suitable refinement strategies at a particular resolution.

1. Mueller, M., Jenni, S. & Ban, N. Strategies for crystallization and structure determination of very large macromolecular assemblies. Curr Opin Struct Biol 17, 572–579 (2007).

It can be used as a rough guide line at the start of your refinement (incomplete model, fresh off of molecular replacement) with slow increase of refinement parameters the better your model becomes (and as Rfree permits).

F

---------------------------------------------
Francis E. Reyes M.Sc.

----------
From: Rajesh kumar

Dear All,

Thanks for all the suggestions. Lot to learn when its low resolution.
I have few more details....

5th round of refinement gave R/Rfree 0.2016/0.2767 after reducing waters and fixing some outliers and difference has gone up to 7.5%

I have used Buster 2.10.0 for refinement. First round after molecular replacement I used Refmac for refinement. There after I used Buster with autoncs_noprune and sim_swap_equiv_plus as it was low resolution in 2nd and 3rd round. Fourth round I used -autoncs and -TLS options. Here are the details of NCS and TLS paprameters.

REMARK 3 SIMILARITY.

REMARK 3 NCS.

REMARK 3 NCS REPRESENTATION : RESTRAINT LSSR (-AUTONCS)

REMARK 3 TARGET RESTRAINTS.

REMARK 3 TARGET REPRESENTATION : NONE

REMARK 3 TARGET STRUCTURE : NULL

REMARK 3

REMARK 3 TLS DETAILS.

REMARK 3 NUMBER OF TLS GROUPS : 2

REMARK 3

REMARK 3 TLS GROUP : 1

REMARK 3 SET : { A|* }

REMARK 3 ORIGIN FOR THE GROUP (A): -40.7570 45.8964 0.0654

REMARK 3 T TENSOR

REMARK 3 T11: 0.6834 T22: 0.5325

REMARK 3 T33: -0.6079 T12: -0.0436

REMARK 3 T13: -0.1033 T23: 0.0313

REMARK 3 L TENSOR

REMARK 3 L11: 1.6212 L22: 0.8001

REMARK 3 L33: 6.4728 L12: 0.0460

REMARK 3 L13: 0.0051 L23: 0.7670

REMARK 3 S TENSOR

REMARK 3 S11: -0.2854 S12: 0.5434 S13: 0.0788

REMARK 3 S21: -0.5270 S22: 0.3608 S23: 0.0731

REMARK 3 S31: -0.7141 S32: 0.2459 S33: -0.0754

REMARK 3

REMARK 3 TLS GROUP : 2

REMARK 3 SET : { B|* }

REMARK 3 ORIGIN FOR THE GROUP (A): -29.9645 52.6711 27.6973

REMARK 3 T TENSOR

REMARK 3 T11: 0.5441 T22: 0.5948

REMARK 3 T33: -0.6079 T12: -0.0298

REMARK 3 T13: -0.0992 T23: -0.0084

REMARK 3 L TENSOR

REMARK 3 L11: 1.7807 L22: 1.1381

REMARK 3 L33: 4.5990 L12: 0.7938

REMARK 3 L13: 0.4959 L23: 1.3513

REMARK 3 S TENSOR

REMARK 3 S11: -0.2911 S12: -0.0941 S13: 0.1302

REMARK 3 S21: -0.3180 S22: 0.1719 S23: -0.0989

REMARK 3 S31: -1.2554 S32: 0.5321 S33: 0.1192

I deleted Lot of waters on Ethan's suggestion and now have only 12 of them. I am refining now and it takes a while for a buster run in our computer and I will let you know the results.

So, I would appreciate if you could suggest what NCS restrains are needed and where they were needed and why, if the above method is not a systematic approach use NCS after obtaining a solution. And I have same question for the TLS. I guess this would benefit many people like me who are dealing with their first low resolution data.

Many thanks to Francis for pointing good reference.

Let me know if I need to give more information.

I appreciate all the suggestions and your valuable time.

Thanks,

Raj

----------
From: Pavel Afonine
Hi,

At 3.3A I would recommend trying a TLS model _instead_ of refining
individual B factors.

may be it is implementation/software/mindset dependent, but in phenix.refine refining TLS+individual ADP or simply individual ADP(*) is a better option most of the time at low resolution.

For details see pages 30 and 31 here:

http://phenix-online.org/newsletter/CCN_2010_07.pdf

Pavel

(*) ADP = Atomic Displacement Parameters

----------
From: Ian Tickle

> (*) ADP = Atomic Displacement Parameters

or "anisotropic displacement parameters"?

See http://ww1.iucr.org/comm/cnom/adp/finrepone/finrepone.html

Section 1.5 Comments about terminology

Cheers

-- Ian

----------
From: Jacob Keller

> (*) ADP = Atomic Displacement Parameters

But aren't *isotropic* b-factors subsumed under this TLA (three-letter acronym?)

JPK

----------
From: Pavel Afonine
Thanks Ian,

I'm aware of this document. Personally I prefer "ADP = Atomic Displacement Parameters" over anything ele, because, given that Atomic Displacement Parameters can be parameterized in many different ways, it makes it easier to operate with such terms like:

- isotropic Atomic Displacement Parameters (isotropic ADP);

- anisotropic Atomic Displacement Parameters (anisotropic ADP);

- group Atomic Displacement Parameters (group ADP);

- group isotropic Atomic Displacement Parameters (group isotropic ADP; example: when refining one isotropic ADP per set of selected atoms);

- group anisotropic Atomic Displacement Parameters (group anisotropic ADP; example TLS);

- ... etc, etc...

(no intention to open another "can of worms" ! -:) )

All the best,

Pavel

----------
From: Ian Tickle

> I'm aware of this document. Personally I prefer "ADP = Atomic Displacement
> Parameters" over anything ele, because, given that Atomic Displacement
> Parameters can be parameterized in many different ways, it makes it easier
> to operate with such terms like:
>
> - isotropic Atomic Displacement Parameters (isotropic ADP);
> - anisotropic Atomic Displacement Parameters (anisotropic ADP);
> - group Atomic Displacement Parameters (group ADP);
> - group isotropic Atomic Displacement Parameters (group isotropic ADP;
> example: when refining one isotropic ADP per set of selected atoms);
> - group anisotropic Atomic Displacement Parameters (group anisotropic ADP;
> example TLS);
> - ... etc, etc...
>
> (no intention to open another "can of worms" ! -:) )

Hi Pavel, but couldn't you simplify it further by omitting the word
"atomic" throughout? In fact isn't "group (an)isotropic Atomic
Displacement Parameter" a contradiction in terms? Surely it can be
either a group parameter or an atomic parameter but not both at the
same time?

I think the worms are already out and making good their escape :).

Cheers

-- Ian

----------
From: Jacob Keller
Can't there be a "group of atoms?"

JPK

--
*******************************************
Jacob Pearson Keller

----------
From: Ian Tickle

> Can't there be a "group of atoms?"

For sure, but doesn't a given parameter either apply to a single atom
or to a group of atoms?

Cheers

-- Ian

----------
From: Joseph Cockburn
Hi Rajesh,
If you're seeing a lot of extra density coming up in the map in regions
where you previously added waters, is it possible that this extra density
corresponds to a part of your protein that you previously thought was
disordered and is thus missing from the current model? At this resolution
you wouldn't expect to see many waters.
Also, to the best of my knowledge, the relative weighting of the X-ray and
geometry terms in BUSTER is set by the program so as to produce a rmsd in
bond lengths equal to a target value. The default value of this is 0.01
Ang (I think) but you can change this using the -r option on the command
line. Using a lower value will reduce the weight on the X-ray term and may
lower the R/R-free gap.
Best wishes,
Joe

----------
From: Eleanor Dodson
Presumably your data is quite anisotropic, and low resolution, so it is quite likely that a TLS model will give much better description of the B factors than more classical refinement.

Modelling solvent at that resolotion will be tricky of course.
Elesnor

--
Professor Eleanor Dodson

----------
From: Rajesh kumar

Dear All,

Thanks for all the suggestion. Data is anisotropic. Since anisotropic correction with the UCLA server didnt give any good Molprobity scores and didnt reduce the R/Rfree at 2.1A, I am using ~3.3A data for further refinement.

Using Tom's suggestion of target restrains reduced the gap brtween R/Rfee to with in 5%. I haven't tried Pavel' suggestion but when I do I will post the results.

I used a 2.3 A MR model as target to get R/Rfree 0.2209/0.2597. When I used coot real space refine zone to fix few molprobity outliers, most of the favored residues become outliers. Almost all residues changed to become outliers. I am wondering why all the residues are moving from favored to outlier regions after restrained refinement. I thought all the residues would be fixed (fit) to the map and wouldn't change during the coot exercise, and this is what I see in the 2.3 A target/model structure ( with R/Rfree 0.1693/0.1961).

I don't know if is this common for a low resolution or I didn't understand this whole thing of refinement correctly.

Thanks for all the help,

Regards,

Raj

----------
From: Eleanor Dodson
No comments on much of your message but I cannot believe you can get better results by cutting anisotropic 2.1A data to 3.3A!!

You MUST be throwing away valuable experimental information,, remember crystallography is not about getting low R factors but about getting a model consistent with your experiment.

eleanor

CCP4 Bulletin Board Archive

Friday, 2 March 2012

sudden drop in R/Rfree

No comments:

Post a Comment

Followers