Hi Nat, Would you define "significant" for me (as you see it of course)? I am NOT a X-ray crystallographer but I was intrigued by this post. I have RE-refined structures and have seen R/R (Free) go up say by 0.4 Say from 0.110 to 0.150 Would you say that is significant for high resolution structures ? Thank you, George On Mon, Oct 27, 2014 at 4:13 PM, Nathaniel Echols wrote:

On Mon, Oct 27, 2014 at 1:00 PM, Cedric wrote:

...

Hi Phenix and CCP4 community,

(sorry for the cross posting).

I was looking at a PDB file.

The http://www.rcsb.org/ website page gives the following values:

R-Value: 0.103 (work) R-Free: 0.134

The actual PDB file gives the following:

R :0.110 FREE R VALUE : 0.170

I was wondering why the difference. The structure is 1.0A resolution 2CWS.

I don't have an exact explanation, but there is some inconsistency in the way data are stored internally by the PDB. The mmCIF file is the most complete representation:

_refine.ls_R_factor_R_work 0.103 _refine.ls_R_factor_R_free 0.134 ... _pdbx_refine.R_factor_all_no_cutoff 0.1101 _pdbx_refine.R_factor_obs_no_cutoff ? _pdbx_refine.free_R_factor_no_cutoff 0.1704

So, the first set is what gets displayed on the web page, the second set ends up in the PDB header. I suspect something went awry in deposition, but only the PDB and the depositors can answer that question.

I wouldn't take the advertised statistics at face value anyway; I prefer to rely on recalculated values (and if these are significantly different, I view the structure with suspicion). In this case, phenix.model_vs_data says:

r_work(re-computed) : 0.1081 r_free(re-computed) : 0.1377

which, accounting for the precision loss in PDB format and the differences between SHELXL and Phenix, are reasonable enough, and suggest that the values on the web page are probably accurate.

-Nat

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

On 10/27/14 4:43 PM, Nathaniel Echols wrote:

On Mon, Oct 27, 2014 at 1:19 PM, George Devaniranjan mailto:[email protected]> wrote:

Would you define "significant" for me (as you see it of course)?

Pavel's definition:

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2906258/

"...the difference between R factors computed using the different methods is typically less than 0.01%."

I think this is probably a typo and it is supposed to mean "1%" or "0.01", which would have been my estimate. Certainly differences below 0.005 are hardly worth noticing, and below 0.001 is statistical noise. Differences above 0.01 are more worrisome (although not entirely unheard of).

It depends what you compare and how.. Two scenarios: 1) Compute Fcalc using FFT (Fc_fft) and direct summation (Fc_direct) and compute R-factor(Fc_fft, Fc_direct). In this case indeed typically the R-factor will be below 1% (not 0.01% !). Attached script illustrates this (to run: phenix.python run.py). Also, Table 4 in Acta Cryst. (2004). A60, 19-32. On a fast calculation of structure factors at a subatomic resolution P. V. Afonine and A. Urzhumtsev does exactly this comparison. 2) You run two identical refinements, in one you use FFT and in the other one direct summation. In this case the difference between R-factors is likely to be below 0.01%. This is because refinable parameters will absorb the differences. Pavel

On Mon, Oct 27, 2014 at 1:48 PM, Phil Jeffrey wrote:

I believe SHELXL reports R1 cut at 4 sigma F and also R1 on all data, which I've got to bet is the source of those two different values. However I've never used SHELXL for protein and don't know if the cutoff is redefined for protein data.

No, but the fact that phenix.model_vs_data comes closer to the lower values (without discarding any data) suggests that the higher values don't reflect the statistics with cutoffs applied. (As expected, model_vs_data reports slightly lower R-factors when the cutoffs are used.) -Nat