Hi Smith,

Phenix version 1.10.1 (the latest official) is the one to use, and by design it is supposed to be superior than previous versions. Actually this is what you confirm with your results: with 1.10.1 you get better model (less Ramachandran plot outliers, similar other metrics) and better model-to-map fit (CC =0.761 vs 0.687).

It is not clear to me why you find the model from 1.10.1 version unacceptable.

I suggest you to find a handful of most severe steric clashes and resolve them.

Pavel

Today I just completed another protein phenix.real_space_refine. For the following comparison runs, the initial model same, the map same, all 5 macro_cycles default  run.

For version 1.10.1, the results were as following,

Map CC (whole unit cell)： 0.672
Map CC （around atoms）： 0.761
rmsd （bonds):  0.01
rmsd (angles): 1.55
All-atom clahscore: 39.27
Ramachandran outlier: 0.98%
Ramachandran allowed: 12.71%
Ramachandran favoured: 86.31%
Rotamer outliers: 0.91%
C-beta deviations: 0

For version 1.9-1692, the results were as following,

Map CC (whole unit cell)： 0.645
Map CC （around atoms）： 0.687
rmsd （bonds):  0.0071
rmsd (angles): 1.15
All-atom clahscore:16.88
Ramachandran outlier: 2.00%
Ramachandran allowed: 6.23%
Ramachandran favoured: 91.77%
Rotamer outliers: 0.05%
C-beta deviations: 0

For the  version 1.10.1 phenix.real_space_refine, the most unacceptable was on the clashscore, the whole refine process seems did not care on the clashcore refine at all.

I am looking forward to getting an explanation from you on the causes of the difference for phenix.real_space_refine for the 2 versions, and the possible strategy we can take to eliminate the difference to get the acceptable results.

Best regards.

Smith  






On Friday, October 9, 2015 12:52 AM, Smith Lee <[email protected]> wrote:


Dear Pavel，

As I have told you, I have tried phenix.real_space_refine for 2 protein with version 1.9-1692 and version 1.10 (and 1.10.1) for many times, with cif or without cif, with or without annealing. The general feedback was that the version 1.9-1692 was much much better for the final analysis of the PDB given at the end of the run of the phenix.real_space_refine. Especially for clashing score, version 1.9-1692 can gave acceptable results (below 20 or around 20), but the clashing score given by the above 2 new versions of phenix.real_space_refine was very high (around 50 or 100). For C-beta deviations,  the output by 1.9-1692 was also much better

Before I have tried the other nightly built versions of phenix.real_space_ refine between version 1.9-1692 and version 1.10 (and 1.10.1) , all I gave acceptable output analysis (something like the Molprobity analysis at the end of the run).

Thus may I ask the reason you have version 1.10 (and 1.10.1) phenix.real_space_refine was because you have realized that the result given by version 1.9-1692 was too satisfactory?

One direct concern of me was that you have recommended not to use the old version phenix. But I cannot get the satisfactory result with the new version phenix.real_space_refine.

The other concern of me was that, I have noticed there were publications In Nature and Science which wrote they use phenix.real_space_refine to got the results in Nature and Science. The Phenix they used should be the versions earlier than version 1.10. Considering what I mentioned above, the Nature and Science results may be not so correct, if new version of phenix was considered as much better than the old version.

Thus am I the only user who communicated with you on the high clashing score given by the new version of phenix？Can you give some comments on what I mentioned above on the high clashing score given by the new version phenix.real_spac_refine?


Best regards.


Smith