lastest phenix.refine reproducible problem
Dear all, I've been learning phenix refinement on a MAD dataset which is already deposited in PDB. There are something I can not get through. Basically I had two steps refinement: 1->phenix.refine data.mtz model.pdb xray_data.high_resolution=3 fix_rotamers=true strategy=individual+adp+individual_sites+tls 2->phenix.refine data.mtz model-refined.pdb xray_data.high_resolution=3 fix_rotamers=false strategy=none main.number_of_macro_cycles=1 model-refined.pdb is actually from the 1st step, so in principle the R/R-free values should be the same for 1 and 2. However in my case, I had 0.2506/0.2996 for the 1st step but 0.2503/0.2998 for the second. I compared the resulting pdb and mtz files from the two steps. PDB model were exactly the same but there were little difference in the mtz file. Why there is difference in mtz file and would it be the reason for the difference of R factors? I use the latest version of phenix, phenix-dev-1009. thanks, Wenchang
Wenchang,
I had two steps refinement:
1->phenix.refine data.mtz model.pdb xray_data.high_resolution=3 fix_rotamers=true strategy=individual+adp+individual_sites+tls
2->phenix.refine data.mtz model-refined.pdb xray_data.high_resolution=3 fix_rotamers=false strategy=none main.number_of_macro_cycles=1
may I ask why is the step #2 ?
model-refined.pdb is actually from the 1st step, so in principle the R/R-free values should be the same for 1 and 2. However in my case, I had 0.2506/0.2996 for the 1st step but 0.2503/0.2998 for the second.
Well, 0.2506/0.2996 vs 0.2503/0.2998 I would say the same. Ralf made a great comment about this sometime ago - if interested search in phenixbb archives. Something like a tiny rounding while reading out/in makes the scaling to take a route to a slightly different minimum and voila - a slightly different R. Or that same rounding makes density sampling a tiny bit different and here is tiny bit difference. I guess one can imagine a long list of such things.
I use the latest version of phenix, phenix-dev-1009.
As of today the latest is 1054, so you are 45+ days behind -:) Pavel
2012/5/9 Pavel Afonine
Wenchang,
I had two steps refinement:
1->phenix.refine data.mtz model.pdb xray_data.high_resolution=3 fix_rotamers=true strategy=individual+adp+individual_sites+tls
2->phenix.refine data.mtz model-refined.pdb xray_data.high_resolution=3 fix_rotamers=false strategy=none main.number_of_macro_cycles=1
may I ask why is the step #2 ?
Hi Pavel, thanks for your consideration. I may explain what I'm doing from the beginning. I feed my model.pdb into phenix.refine and do the refinement against different resolution from 2.00 to 3.00. There is a R/R-free value for each refinement. Then in the 2nd step I refine each model from the former refinement at high resolution limit 3.00, in order to have meaning comparison on which high resolution limit could give the best model since in this case all the models are refined at the same high resolution limit.
model-refined.pdb is actually from the 1st step, so in principle the
R/R-free values should be the same for 1 and 2. However in my case, I had 0.2506/0.2996 for the 1st step but 0.2503/0.2998 for the second.
Well, 0.2506/0.2996 vs 0.2503/0.2998 I would say the same. Ralf made a great comment about this sometime ago - if interested search in phenixbb archives.
I did a search, by heart, in order to find Ralf's great comment but I did not find it. Could you send the link if you have a way to find it?
Something like a tiny rounding while reading out/in makes the scaling to take a route to a slightly different minimum and voila - a slightly different R. Or that same rounding makes density sampling a tiny bit different and here is tiny bit difference. I guess one can imagine a long list of such things.
if reading out/in would make the slightly different R, does this mean I will get another different R if I rerun it again?
I use the latest version of phenix, phenix-dev-1009.
As of today the latest is 1054, so you are 45+ days behind -:)
Pavel
_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb
On Wed, May 9, 2012 at 2:14 PM, 周文昌
I did a search, by heart, in order to find Ralf's great comment but I did not find it. Could you send the link if you have a way to find it?
The problem is that the precision of the atomic properties in the PDB file is limited to 0.001 for coordinates and 0.01 for B-factors. Internally, the precision is much greater, but some of this is lost when writing them out. For example, the B-factor for an atom might refine to 36.0423947, but it will be written out as 36.04. The cumulative effect of this on every atom results in the slight difference in R-factors. Jeff and I found a similar "problem" with running phenix.refine from command-line arguments only versus running it from a parameter file - the map grid used for FFT structure factor calculations is spaced resolution * 1/3, but this gets written out to the parameter file as 0.3333333333, and this is enough to result in slightly different final models and R-factors. I used scare quotes around "problem", because these kinds of R-factor differences aren't meaningful anyway. -Nat
participants (3)
-
Nathaniel Echols -
Pavel Afonine -
周文昌