Typical R-factors for structures at around 1.45A are: phenix.r_factor_statistics 1.45 Histogram of Rwork for models in PDB at resolution 1.35-1.55 A: 0.092 - 0.122 : 40 0.122 - 0.152 : 344 0.152 - 0.183 : 743 0.183 - 0.213 : 722 <<< your structure 0.213 - 0.243 : 206 0.243 - 0.273 : 18 0.273 - 0.303 : 3 0.303 - 0.334 : 0 0.334 - 0.364 : 1 0.364 - 0.394 : 1 Histogram of Rfree for models in PDB at resolution 1.35-1.55 A: 0.113 - 0.148 : 18 0.148 - 0.183 : 333 0.183 - 0.217 : 982 0.217 - 0.252 : 607 <<< your structure 0.252 - 0.287 : 125 0.287 - 0.322 : 10 0.322 - 0.357 : 1 0.357 - 0.391 : 0 0.391 - 0.426 : 0 0.426 - 0.461 : 2 Histogram of Rfree-Rwork for all model in PDB at resolution 1.35-1.55 A: 0.001 - 0.011 : 82 0.011 - 0.020 : 476 0.020 - 0.030 : 666 <<< your structure 0.030 - 0.040 : 381 0.040 - 0.050 : 246 0.050 - 0.059 : 121 0.059 - 0.069 : 70 0.069 - 0.079 : 26 0.079 - 0.088 : 4 0.088 - 0.098 : 6 Number of structures considered: 2078 So, your R-factors look quite normal given the resolution. Regarding Wilson vs Average B-factors, here is the statistics over PDB: Resolution_range Wilson_B Average_B Number_of_structures 0.00 - 1.00 9.77 13.11 94 1.00 - 1.25 10.58 16.44 401 1.25 - 1.50 13.50 19.14 1050 1.50 - 1.75 17.20 21.76 3600 1.75 - 2.00 22.27 26.82 5510 2.25 - 2.50 35.70 39.42 3385 2.50 - 2.75 43.71 44.73 2844 2.75 - 3.00 53.86 51.94 1628 3.00 - 3.25 65.11 60.76 780 3.25 - 3.50 81.69 78.70 165 3.50 - 3.75 92.67 88.84 100 3.75 - 4.00 111.83 102.29 30 (*) Wilson_B is computed using phenix.model_vs_data Average_B is computed using phenix.model_vs_data from PDB file (TLS is accounted for) Structures selected such that the recomputed R-factor matches the one in PDB file header within 1%. As you see, your B-factors are quite high. Should you worry about it? I don't know. If everything else is done right and look good (as good as R-factors), then no, otherwise I would think what might be the cause for this. Although it's interesting to know why this happens. I will probably look at the histogram of Wilson B and average B around this resolution to see what people had in the past. Pavel. On 8/3/10 10:48 PM, Peter Zwart wrote:

If you did use aniso B and are stuck at 19.8% for a 'true' 1.45 A data set, I suspect that something might be off ....

did you check possibilities for twinning?

did you optimize weights properly?

P

2010/8/3 Geoffrey Feld mailto:[email protected]>

Dear PhenixBBers,

I'm working on a 1.45 A structure I solved using MR (phaser) and I'm pretty close to finishing, just plopping in waters and fixing rotamers. Rw = 19.8 Rfree= 22.8. I am a little concerned because my Wilson B is 27.00 while my average B for macromolecule is more like 43, and for solvent is 48. I have enough data to use anisotropic ADP refinement, which was a big help in bringing down the Rfree, but the average B hasn't really moved much. Should I be concerned about this? Should I try adjusting the wxu, or some other parameter?

Thanks! -- Geoffrey K. Feld

Department of Chemistry 492 Stanley Hall University of California, Berkeley

"Vigilia pretium libertatis"

_______________________________________________ phenixbb mailing list [email protected] mailto:[email protected] http://phenix-online.org/mailman/listinfo/phenixbb

-- ----------------------------------------------------------------- P.H. Zwart Research Scientist Berkeley Center for Structural Biology Lawrence Berkeley National Laboratories 1 Cyclotron Road, Berkeley, CA-94703, USA Cell: 510 289 9246 BCSB: http://bcsb.als.lbl.gov PHENIX: http://www.phenix-online.org SASTBX: http://sastbx.als.lbl.gov -----------------------------------------------------------------

_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb

Hi Gino, just a few comments:

I am aware of all approximation made in defining the Wilson B-factor and how one should not take the Wilson B-factor too literally (etc.. etc...). But, as a matter of fact, Phenix kept over-refining my B-factor to crazy values.

I would say refining and not over-refining, because nobody knows where "over" starts.

1. Was your diffraction anisotropic? If so, try the Diffraction Anisotropy Server. The server applies a negative isotropic B-factor correction to your data. In my case, corrected structure factor amplitudes resulted in a 2.5% drop in the Rfree

That server removes reflections by sigma value and in selected directions. In some cases it can easily be a few thousands of Fobs. So, it it not surprising that the R-factors dropped. You can't compare the R-factors for the datasets containing different amount of reflections.

and the average B-factor of my structure

and it is not surprising that the B-factor dropped: instead of keeping it in ATOM records you gave it away to uncle Sam, by applying negative-B correction. Just a trade, matter of where you keep these extra B-factor value.

SECOND How many macrocycles of refinement are you using? In my case Phenix kept bumping up the average B-factor of my structure by ~10A2 macrocycle. A 10 macrocycles refinement would bring the averaged B-factor to ~200A2. So don't over-refine. Stick to 2-3 cycles (provided your geometry is good, of course).

Interesting. I've never seen it before. Did you try to change the target weights for ADP refinement: manually or automatically? If you did and none helped (although I'm still not sure if help is needed) I would be interested to see the data and model.

THIRD Do you have NCS and are you using TLS? If so, try to first use NCS to refine atom positions, then do TLS refinement without imposing NCS. If you think about, it doesn't make too much sense to use ncs-restraints when splitting a structure in TLS groups.

Regarding NCS: in ADP refinement it is applied only to local ADPs, and not total ones. See recent publication regarding this in PHENIX newsletter: http://www.phenix-online.org/newsletter/CCN_2010_07.pdf see page 28. So you can safely refine everything in one go (I guess). All the best! Pavel.

Hello, I'm a little confused about Pavel's response. Perhaps I'm not understanding it correctly. I don't understand why you say this is nothing to worry about. It seems to me that it means that the refinement strategy isn't stable or optimal - the contributions from the overall Baniso and the individual B factors are not being (and prossibly can not be) separated properly - maybe they can't be because the overall anisotropy is so large that it swamps the anisotropic B or perhaps the resolution is not sufficient to refine anisotropic Bs for individual atoms. In these cases shouldn't either (1) the anisotopic B be refined and then fixed rather than be allowed to creep up or (2) anisotropic Bs for individual atoms not be refined? Can this problem be debugged by looking at the anisotropic Bs in coot? If the anisotropic scale is being incorporated improperly into the atomic anisotropic Bs then would all the thermal ellipsoids have almost the same shape and orientation and only differ in magnitude? Sue On Aug 6, 2010, at 9:15 AM, Pavel Afonine wrote:

Hi Geoffrey,

thanks for sending me the data and model - this helped me to find out what happens in your and other similar repeatedly reported cases.

Have a quick look at the total model structure factor formula that is used in all phenix.refine, phenix.model_vs_data, phenix.maps and many other similar tools:

see page 6 here:

http://www.phenix-online.org/presentations/latest/pavel_refinement_general.p...

and glance through the page 29, PHENIX Newsletter:

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

As you see there is overall anisotropic scale matrix Ucryst (see Acta Cryst. (2005). D61, 850-855 and references therein for deeper level of details). In refinement, the trace of this matrix is subtracted from it and added to individual ADPs and Bsol, making Fmodel invariant under this manipulation. Most of the time, this is a small value, but sometimes it is relatively large.

Now, let's see what happens in your particular case. The Wilson B-factor is 27. If we reset all B-factors to 27 and repeat the refinement until convergence using two scenarios:

1) we add the trace of Ucryst to individual ADPs, 2) we do not add the trace of Ucryst to individual ADPs,

we will get the following:

1) R-work = 0.1728, R-free = 0.2177 Bcryst (Ucryst reported as B) = (-10.42,-11.48,21.90,-0.00,0.00,0.00); trace/3= 0.00 ksol= 0.33 Bsol= 54.75 Average ADP = 43.56

2) R-work = 0.1744, R-free = 0.2171 Bcryst (Ucryst reported as B) = (4.42,3.35,36.58,0.00,0.00,0.00); trace/3= 14.78 ksol= 0.32 Bsol= 40.00 Average ADP = 29.03

Clearly, in case "2)" you get almost exact match of Wilson B and Average ADP (27 and 29), while in case "1)" the B-factors are higher. Note, the R-factors in both cases are "identical" (negligibly different given the resolution and the R-factor value).

So I guess everything is more or less consistent and clear. It depends where and how you keep different contributions to the total ADP, and which values you use to compute mean ADP.

Answering your very original question "Should I be concerned about this?": the answer is no. But I had a quick look at your maps: and this is what I would spend some more time: you have a lot of positive and negative peaks, some of them are very strong: larger than 6sigma! I guess you are missing some ions, and alternative conformations may need some more attention. This is normal, it just needs some care and time to be spent before your structure is ready to go (to PDB).

Pavel.

PS> I'm sending you the results of these two runs off list.

On 8/3/10 10:20 PM, Geoffrey Feld wrote:

...

Dear PhenixBBers,

I'm working on a 1.45 A structure I solved using MR (phaser) and I'm pretty close to finishing, just plopping in waters and fixing rotamers. Rw = 19.8 Rfree= 22.8. I am a little concerned because my Wilson B is 27.00 while my average B for macromolecule is more like 43, and for solvent is 48. I have enough data to use anisotropic ADP refinement, which was a big help in bringing down the Rfree, but the average B hasn't really moved much. Should I be concerned about this? Should I try adjusting the wxu, or some other parameter?

Thanks!

_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb

Dr. Sue A. Roberts Dept. of Chemistry and Biochemistry University of Arizona 1041 E. Lowell St., Tucson, AZ 85721 Phone: 520 621 8171 [email protected] http://www.biochem.arizona.edu/xray

Pavel Afonine wrote:

in summary, in that email I wrote that there are two ways of how you can handle the total B-factor and one of its components (Bcryst). You can subtract the trace of Bcryst and add it to individual ADPs (what CNS does, for example) or you can keep Bcryst as it is. Both ways do not change the Fmodel and therefore the R-factors. So I can't see how you conclude from this about refinement stability. These are purely formatting/convention things.

Strikes me that the convention should be that the B-factors associated with the atoms should be in the PDB B-factor field associated with those atoms, subject to the ability to represent the ADP model. Burying a constant isotropic offset in the PDB header - which might be of variable format depending on what program is used - seems like a good way to get the overall ADP of the model to be incorrectly assessed. This would seem to be analogous to the difference between reporting the equivalent total B-iso in the isotropic B-factor field in TLS refinement (as phenix.refine does) versus reporting just the residual B (REFMAC default behavior). i.e. all the information is in the PDB file somewhere, but it's not necessarily where you expect it to be. I rather prefer the phenix.refine method for this. Since most of the interest with Wilson B seems to be comparing Fmodel with the Fobs, it may make more sense to compare the Wilson B calculated from Fmodel with the Wilson B calculated with Fobs, which might at least cancel out some of the errors in the calculation. The underlying question seems to be "does the distribution of |Fmodel| with resolution match that of |Fobs|". Phil Jeffrey Princeton

Hi Pavel, No, I was just suggesting that the trace of B-cryst be applied, in the same way that CNS does, so that less things are hidden in the header and that the B-iso for each atom would be the same as the effective B-iso for the atom after applying B-cryst. In fact I'd assumed that was what phenix.refine did although obviously I had not thought to check. If you added the anisotropic component of B-cryst I think it would cause more confusion than it resolves, since it will generate all those per-atom ANISOU entries even if no TLS or atomic anisotropic refinement was done, and nearly double the size of a PDB file. My viewpoint is rather biased to what I historically "expect" to find in PDB file atom entries, however. Phil Pavel Afonine wrote:

Hi Phil,

Just to make sure I got your point correctly... Are you suggesting to add Bcryst to the total ADP as well (so it is always included into isotropic B in ATOM record and anisotropic b in ANISOU record), similarly to what we already do with TLS contribution?

If we do that:

- all atoms will always have ANISOU (because Bcryst is anisotropic); - do you know how to convert Bcryst obtained in reciprocal space to the one in real space (so it can be added to individual ADP) (never thought about it - am I asking something stupid....?)?

Thanks! Pavel.

On 8/6/10 12:58 PM, Phil Jeffrey wrote:

...

Pavel Afonine wrote:

...

in summary, in that email I wrote that there are two ways of how you can handle the total B-factor and one of its components (Bcryst). You can subtract the trace of Bcryst and add it to individual ADPs (what CNS does, for example) or you can keep Bcryst as it is. Both ways do not change the Fmodel and therefore the R-factors. So I can't see how you conclude from this about refinement stability. These are purely formatting/convention things.

Strikes me that the convention should be that the B-factors associated with the atoms should be in the PDB B-factor field associated with those atoms, subject to the ability to represent the ADP model. Burying a constant isotropic offset in the PDB header - which might be of variable format depending on what program is used - seems like a good way to get the overall ADP of the model to be incorrectly assessed.

This would seem to be analogous to the difference between reporting the equivalent total B-iso in the isotropic B-factor field in TLS refinement (as phenix.refine does) versus reporting just the residual B (REFMAC default behavior). i.e. all the information is in the PDB file somewhere, but it's not necessarily where you expect it to be. I rather prefer the phenix.refine method for this.

Since most of the interest with Wilson B seems to be comparing Fmodel with the Fobs, it may make more sense to compare the Wilson B calculated from Fmodel with the Wilson B calculated with Fobs, which might at least cancel out some of the errors in the calculation. The underlying question seems to be "does the distribution of |Fmodel| with resolution match that of |Fobs|".

Phil Jeffrey Princeton

_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb