Different stats with different number of macrocycles

newer
phenix real space refine & free R...

older
Phenix Generate Table 1 Tool for...

Patrick Loll

3 May 2017 3 May '17

9:07 p.m.

Hi, I have an odd observation. I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles. OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement? In case this isn’t clear: In the 3-macrocycle job, after the first macrocycle, RMS bonds/angles = 0.005 Å/0.9 deg, R/Rfree = 0.208/0.236 (by the end of the third macrocycle, these numbers have changed to 0.009 Å/1.41 deg and 0.212/0.244) At the end of the 1-macrocycle job, RMS bonds/angles = 0.008 Å/1.4 deg, R/Rfree = 0.209/0.240 This is Phenix version 1.11.1-2575-000. Any insights welcome. Cheers, Pat --------------------------------------------------------------------------------------- Patrick J. Loll, Ph. D. Professor of Biochemistry & Molecular Biology Drexel University College of Medicine Room 10-102 New College Building 245 N. 15th St., Mailstop 497 Philadelphia, PA 19102-1192 USA (215) 762-7706 [email protected]

Show replies by date

Pavel Afonine

4 May 4 May

4:20 p.m.

Hi Patrick,

...

I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles.

it's hard to comment on this one because I don't know how you define "worse". For example, I'd call the same "R/Rfree = 0.208/0.236" and "R/Rfree = 0.209/0.240" but some may think they are different.

...

OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement?

phenix.refine may change internal strategies based on specified number of macro-cycles. So your observation is not too unexpected to me. Pavel

Patrick Loll

6:25 p.m.

Hi Pavel, I agree that the differences in the R values are not large, but I found the geometry differences compelling (RMS bonds/angles can be either 0.005 Å/0.9 deg or 0.008 Å/1.4 deg). I’m glad to hear the strategies may change based on macrocycle number (otherwise I’d question my understanding of refinement). Is there any way a user can change influence this choice of strategy? Given two models with essentially equivalent R values, I’d prefer the one with nicer geometry. Thanks, Pat

...

On 4 May 2017, at 12:20 PM, Pavel Afonine wrote:

Hi Patrick,

...
I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles.

it's hard to comment on this one because I don't know how you define "worse". For example, I'd call the same "R/Rfree = 0.208/0.236" and "R/Rfree = 0.209/0.240" but some may think they are different.

...
OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement?

phenix.refine may change internal strategies based on specified number of macro-cycles. So your observation is not too unexpected to me.

Pavel

Schnicker, Nicholas J

6:36 p.m.

Hi Pat, I believe they suggest using target weight optimization in this case. It can be specified on the command line as below: optimize_xyz_weight=true optimize_adp_weight=true I've found for a couple structures I had to use this or else the geometry was much worse. They also recommend to use this in the last round of refinement. Cheers, Nick ________________________________ From: [email protected] on behalf of Patrick Loll Sent: Thursday, May 4, 2017 1:25:21 PM To: Pavel Afonine; [email protected] Subject: Re: [phenixbb] Different stats with different number of macrocycles Hi Pavel, I agree that the differences in the R values are not large, but I found the geometry differences compelling (RMS bonds/angles can be either 0.005 Å/0.9 deg or 0.008 Å/1.4 deg). I’m glad to hear the strategies may change based on macrocycle number (otherwise I’d question my understanding of refinement). Is there any way a user can change influence this choice of strategy? Given two models with essentially equivalent R values, I’d prefer the one with nicer geometry. Thanks, Pat

...

On 4 May 2017, at 12:20 PM, Pavel Afonine wrote:

Hi Patrick,

...
I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles.

it's hard to comment on this one because I don't know how you define "worse". For example, I'd call the same "R/Rfree = 0.208/0.236" and "R/Rfree = 0.209/0.240" but some may think they are different.

...
OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement?

phenix.refine may change internal strategies based on specified number of macro-cycles. So your observation is not too unexpected to me.

Pavel

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

Phil Jeffrey

9:26 p.m.

But the variable weighting scheme that phenix.refine uses actually causes problems with this - the optimal weight(s) for cycle 1 is most assuredly not the optimal weight for cycles 2....n and it even does this with converged high resolution structures, so it's not just a result of the usual violence I do to structures when rebuilding them. I think this trend of the first cycle weight being an outlier is quite possibly the source of what Pat is seeing, and it's not clear to me why phenix.refine does this where I see a lot less of it (if any at all) in Refmac. Phil Jeffrey Princeton On 5/4/17 2:36 PM, Schnicker, Nicholas J wrote:

...

Hi Pat,

I believe they suggest using target weight optimization in this case. It can be specified on the command line as below:

optimize_xyz_weight=true optimize_adp_weight=true

I've found for a couple structures I had to use this or else the geometry was much worse. They also recommend to use this in the last round of refinement.

Cheers,

Nick

Pavel Afonine

5 May 5 May

7:03 a.m.

phenix.refine updates weight every macro-cycle. It is true that a good weight for macro-cycle 1 does not have to be a good weight for macro-cycle 2, and a good weight for mc 2 is not necessarily good for mc 3, etc. This is one of reasons why doing just one macro-cycle isn't best idea in general, unless you are experimenting with something particular. Weight calculation used in phenix.refine isn't something I invented. We use protocol implemented by Axel Brunger back in 1989 in X-plor and since then nobody came up with a better way to do this (I did spend quite some time investigating this matter). Alternative way to calculate weight is implemented in TNT (by Dale Tronrud); likely Refmac uses same idea. This is dictated by the minimizer they use, which is 2nd derivatives-based. Pavel On 5/4/17 14:26, Phil Jeffrey wrote:

...

But the variable weighting scheme that phenix.refine uses actually causes problems with this - the optimal weight(s) for cycle 1 is most assuredly not the optimal weight for cycles 2....n and it even does this with converged high resolution structures, so it's not just a result of the usual violence I do to structures when rebuilding them.

I think this trend of the first cycle weight being an outlier is quite possibly the source of what Pat is seeing, and it's not clear to me why phenix.refine does this where I see a lot less of it (if any at all) in Refmac.

Phil Jeffrey Princeton

On 5/4/17 2:36 PM, Schnicker, Nicholas J wrote:

...
Hi Pat,

I believe they suggest using target weight optimization in this case. It can be specified on the command line as below:

optimize_xyz_weight=true optimize_adp_weight=true

I've found for a couple structures I had to use this or else the geometry was much worse. They also recommend to use this in the last round of refinement.

Cheers,

Nick

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

Pavel Afonine

4 May 4 May

6:47 p.m.

Hi Patric, I guess to make refinement consisting of one macro-cycle yield the same result as 1st macro-cycle of refinement consisting of N macro-cycles (N>1), you need to set these parameters main.target=ml strategy=individual_sites+individual_adp+occupancy OR strategy=individual_sites+individual_adp+occupancy+individual_sites_real_space You can do it using GUI or command line. Normally, you need to run refinement until convergence, which is almost always more than 3-5 macro-cycles. Pavel On 5/4/17 11:25, Patrick Loll wrote:

...

Hi Pavel,

I agree that the differences in the R values are not large, but I found the geometry differences compelling (RMS bonds/angles can be either 0.005 Å/0.9 deg or 0.008 Å/1.4 deg).

I’m glad to hear the strategies may change based on macrocycle number (otherwise I’d question my understanding of refinement). Is there any way a user can change influence this choice of strategy? Given two models with essentially equivalent R values, I’d prefer the one with nicer geometry.

Thanks,

Pat

...
On 4 May 2017, at 12:20 PM, Pavel Afonine wrote:

Hi Patrick,

...
I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles. it's hard to comment on this one because I don't know how you define "worse". For example, I'd call the same "R/Rfree = 0.208/0.236" and "R/Rfree = 0.209/0.240" but some may think they are different.

...
OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement? phenix.refine may change internal strategies based on specified number of macro-cycles. So your observation is not too unexpected to me.

Pavel

Mark A. White

6:54 p.m.

Patrick, At this resolution the (RMS bonds/angles =0.005 Å/0.9 deg) are over-constrained (Bad). At high resolution we should expect values ~ 0.018 Å/3.0 deg. At your moderate resolution RMS bonds/angle values of 0.008 Å/1.4 deg are typical (Good). Getting a very low RMSD just indicates that the model was over-constrained, and is no different than having the same B-factors for every residue. It is just not physically reasonable. -- Yours sincerely, Mark A. White, Ph.D. Associate Professor of Biochemistry and Molecular Biology, Manager, Sealy Center for Structural Biology and Molecular Biophysics Macromolecular X-ray Laboratory, Basic Science Building, Room 6.658A University of Texas Medical Branch Galveston, TX 77555-0647 Tel. (409) 747-4747 Fax. (409) 747-1404 mailto://[email protected] http://xray.utmb.edu QQ: "What is earnest is not always true; on the contrary, error is often more earnest than truth." - Benjamin Disraeli -----Original Message----- From: Patrick Loll To: Pavel Afonine , [email protected] Subject: Re: [phenixbb] Different stats with different number of macrocycles Date: Thu, 4 May 2017 14:25:21 -0400 Hi Pavel, I agree that the differences in the R values are not large, but I found the geometry differences compelling (RMS bonds/angles can be either 0.005 Å/0.9 deg or 0.008 Å/1.4 deg). I’m glad to hear the strategies may change based on macrocycle number (otherwise I’d question my understanding of refinement). Is there any way a user can change influence this choice of strategy? Given two models with essentially equivalent R values, I’d prefer the one with nicer geometry. Thanks, Pat

...

On 4 May 2017, at 12:20 PM, Pavel Afonine wrote:

Hi Patrick,

...
I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles.

it's hard to comment on this one because I don't know how you define "worse". For example, I'd call the same "R/Rfree = 0.208/0.236" and "R/Rfree = 0.209/0.240" but some may think they are different.

...
OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement?

phenix.refine may change internal strategies based on specified number of macro-cycles. So your observation is not too unexpected to me.

Pavel

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

Pavel Afonine

7:02 p.m.

At resolution 2.5A I doubt you can distinguish difference of 0.005 or 0.018A in bond length. Therefore, any bond/angle rmsd is good that: a) does not exceed upper limit (which is, for example, 0.02/2.0; numbers vary based on personal opinions); AND b) corresponds to lowest Rfree and acceptable Rfree-Rwork gap. BTW, phenix.refine uses restraints for stereochemistry, not constraints. Pavel On 5/4/17 11:54, Mark A. White wrote:

...

Patrick,

At this resolution the (RMS bonds/angles =0.005 Å/0.9 deg) are over-constrained (Bad). At high resolution we should expect values ~ 0.018 Å/3.0 deg. At your moderate resolution RMS bonds/angle values of 0.008 Å/1.4 deg are typical (Good). Getting a very low RMSD just indicates that the model was over-constrained, and is no different than having the same B-factors for every residue. It is just not physically reasonable.

-- Yours sincerely,

Mark A. White, Ph.D. Associate Professor of Biochemistry and Molecular Biology, Manager, Sealy Center for Structural Biology and Molecular Biophysics Macromolecular X-ray Laboratory, Basic Science Building, Room 6.658A University of Texas Medical Branch Galveston, TX 77555-0647 Tel. (409) 747-4747 Fax. (409) 747-1404 mailto://[email protected] http://xray.utmb.edu

QQ: "What is earnest is not always true; on the contrary, error is often more earnest than truth." - Benjamin Disraeli

-----Original Message----- *From*: Patrick Loll mailto:Patrick%20Loll%20%[email protected]%3e> *To*: Pavel Afonine mailto:Pavel%20Afonine%20%[email protected]%3e>, [email protected] mailto:[email protected] *Subject*: Re: [phenixbb] Different stats with different number of macrocycles *Date*: Thu, 4 May 2017 14:25:21 -0400

Hi Pavel,

I agree that the differences in the R values are not large, but I found the geometry differences compelling (RMS bonds/angles can be either 0.005 Å/0.9 deg or 0.008 Å/1.4 deg).

I’m glad to hear the strategies may change based on macrocycle number (otherwise I’d question my understanding of refinement). Is there any way a user can change influence this choice of strategy? Given two models with essentially equivalent R values, I’d prefer the one with nicer geometry.

Thanks,

Pat

...
On 4 May 2017, at 12:20 PM, Pavel Afonine mailto:[email protected]> wrote:

Hi Patrick,

...
I am finishing a refinement at 2.5 Å, using the Phenix GUI. I performed a three-macrocycle refinement, and saw that the geometry (RMS bonds/angles) and R/Rfree all got better in the first macrocycle, and then worsened in the subsequent two macrocycles.

it's hard to comment on this one because I don't know how you define "worse". For example, I'd call the same "R/Rfree = 0.208/0.236" and "R/Rfree = 0.209/0.240" but some may think they are different.

...
OK, fine. So I repeated the refinement, except I performed only a single macrocycle (starting from the exact same input coordinates). However, the statistics after this one-macrocycle job did not match the stats seen after 1 macrocycle in the 3-macrocycle job (?!). This doesn’t make sense to me; if you’re starting from the exact same coordinates, shouldn’t the first macrocycle always wind up at the same place, regardless of whether or not the program goes on to do additional macrocycles of refinement?

phenix.refine may change internal strategies based on specified number of macro-cycles. So your observation is not too unexpected to me.

Pavel

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

2793

Age (days ago)

2795

Last active (days ago)

List overview

Download

8 comments

5 participants

participants (5)

Mark A. White
Patrick Loll
Pavel Afonine
Phil Jeffrey
Schnicker, Nicholas J

Different stats with different number of macrocycles

tags

participants (5)