Re: [phenixbb] Third issue of Computation Crystallography Newsletter - Volume 2, Number 2
Hi, just scanned through the article on automatic weight adjustment. If I had to summarise figures 1 and 2, I'd have to conclude that the latest weight optimization only occasionally produces better results than not doing it (fig 2); and that therefore the old optimization was considerably worse than doing nothing at all. Or have I misinterpreted it? Cheers phx On 28/07/2011 19:59, Nigel Moriarty wrote:
Dear Colleagues,
I am pleased to announce the publication of the third issue of the Computational Crystallography Newsletter:
http://www.phenix-online.org/newsletter/
A listing of the articles and short communications is given below. Please note that the newsletter accepts articles of a general nature of interest to all crystallographers. Please send any articles to me at [email protected] noting that there is a Word Template on the website to streamline production.
Articles --------
Improved target weight optimization in phenix.refine Mite-y lysozyme crystal and structure
Short communications --------------------
A lightweight, versatile framework for visualizing reciprocal-space data An extremely fast spotfinder for real-time beamline applications Hints for running phenix.mr_rosetta
Cheers Nigel
Hi Frank, thanks for comments!
If I had to summarise figures 1 and 2, I'd have to conclude that the latest weight optimization only occasionally produces better results than not doing it (fig 2);
not quite. If it did this then I wouldn't write a paper -;) I'm not sure I understand what made you to make this conclusion? Could you explain, please?
and that therefore the old optimization was considerably worse than doing nothing at all.
Both weight optimizations are better than not doing weight optimization. The new procedure is an improvement over the old one. See text of the paper for details. Pavel.
Hi Frank, thanks for prompting me to have another look at this! Yes, a quick section reiterating the results shown in both figures would be definitely helpful. Ok, I'm doing it now (better late than never!): Figure 1 (old version vs new version), going from left to right, top to bottom: - the new procedure always produces smaller Rfree-Rwork gap (except two outliers out of ~110 cases). In a few dozens cases this means going from Rfree-Rwork ~10-14% down to 8% and less. Which is I believe an improvement. - Molprobity clashscore gets much lower in all cases, and in many cases it get down from clashscore>150-200 to less than 50. Again, I take this as an improvement. - Ramachandran plot: always less outliers, and in a few dozens cases it get down from 10-25% to less than 5%. It's good too! - Always less Rotamer outliers. Not bad. - <Bij> for bonded atoms gets always smaller. In many cases it goes down from ridiculously high ~30-50A**2 to less than 20 or so. There are outliers which, as pointed out in the text, need investigation. - Rfree difference wise (the last graphic at the bottom), almost all the differences stay within "equally good range of values", which is defined in the text. Some bad outliers, are 1) inoptimal NCS groups selection (which was done automatically using the old procedure, not the new one that does it in torsion angle space), 2) not using TLS even though this was used originally, and other issues that needs an investigation. There are improvements too. Figure 2 (with weight optimization vs without weight optimization): going from left to right, top to bottom: - Rfree-Rfree gets much better in a few dozens of cases, or stays in a "equally good range of values". - clashscore either stays the same or gets smaller, which is good. Same for Ramachandran and Rotamer outliers. - Except two outliers (out of total ~110 cases), Rfree(opt)-Rfree(no opt) is consistently smaller, with a number of cases having significant improvement (over ~2% or so). Thanks again for your comments, and prompting me to spell out a proper analysis of the presented pictures. Pavel. On 7/29/11 1:21 AM, Frank von Delft wrote:
Hi, just scanned through the article on automatic weight adjustment.
If I had to summarise figures 1 and 2, I'd have to conclude that the latest weight optimization only occasionally produces better results than not doing it (fig 2); and that therefore the old optimization was considerably worse than doing nothing at all.
Or have I misinterpreted it? Cheers phx
On 28/07/2011 19:59, Nigel Moriarty wrote:
Dear Colleagues,
I am pleased to announce the publication of the third issue of the Computational Crystallography Newsletter:
http://www.phenix-online.org/newsletter/
A listing of the articles and short communications is given below. Please note that the newsletter accepts articles of a general nature of interest to all crystallographers. Please send any articles to me at [email protected] noting that there is a Word Template on the website to streamline production.
Articles --------
Improved target weight optimization in phenix.refine Mite-y lysozyme crystal and structure
Short communications --------------------
A lightweight, versatile framework for visualizing reciprocal-space data An extremely fast spotfinder for real-time beamline applications Hints for running phenix.mr_rosetta
Cheers Nigel
_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb
Thanks for the elaboration. All I meant was that figure 2 did not show improvements (points pretty much on the y=x line), while figure 1 showed huge improvements (points below the y=x line). I should have added that in the past I've had puzzled questions around here about phenix results at lower resolutions, and I always wondered whether weight optimization may be an issue. So maybe what I was looking for was a figure like figure 2, but of the old optimization. Cheers Frank On 29/07/2011 17:41, Pavel Afonine wrote:
Hi Frank,
thanks for prompting me to have another look at this! Yes, a quick section reiterating the results shown in both figures would be definitely helpful. Ok, I'm doing it now (better late than never!):
Figure 1 (old version vs new version), going from left to right, top to bottom:
- the new procedure always produces smaller Rfree-Rwork gap (except two outliers out of ~110 cases). In a few dozens cases this means going from Rfree-Rwork ~10-14% down to 8% and less. Which is I believe an improvement.
- Molprobity clashscore gets much lower in all cases, and in many cases it get down from clashscore>150-200 to less than 50. Again, I take this as an improvement.
- Ramachandran plot: always less outliers, and in a few dozens cases it get down from 10-25% to less than 5%. It's good too!
- Always less Rotamer outliers. Not bad.
-<Bij> for bonded atoms gets always smaller. In many cases it goes down from ridiculously high ~30-50A**2 to less than 20 or so. There are outliers which, as pointed out in the text, need investigation.
- Rfree difference wise (the last graphic at the bottom), almost all the differences stay within "equally good range of values", which is defined in the text. Some bad outliers, are 1) inoptimal NCS groups selection (which was done automatically using the old procedure, not the new one that does it in torsion angle space), 2) not using TLS even though this was used originally, and other issues that needs an investigation. There are improvements too.
Figure 2 (with weight optimization vs without weight optimization): going from left to right, top to bottom:
- Rfree-Rfree gets much better in a few dozens of cases, or stays in a "equally good range of values".
- clashscore either stays the same or gets smaller, which is good. Same for Ramachandran and Rotamer outliers.
- Except two outliers (out of total ~110 cases), Rfree(opt)-Rfree(no opt) is consistently smaller, with a number of cases having significant improvement (over ~2% or so).
Thanks again for your comments, and prompting me to spell out a proper analysis of the presented pictures.
Pavel.
On 7/29/11 1:21 AM, Frank von Delft wrote:
Hi, just scanned through the article on automatic weight adjustment.
If I had to summarise figures 1 and 2, I'd have to conclude that the latest weight optimization only occasionally produces better results than not doing it (fig 2); and that therefore the old optimization was considerably worse than doing nothing at all.
Or have I misinterpreted it? Cheers phx
On 28/07/2011 19:59, Nigel Moriarty wrote:
Dear Colleagues,
I am pleased to announce the publication of the third issue of the Computational Crystallography Newsletter:
http://www.phenix-online.org/newsletter/
A listing of the articles and short communications is given below. Please note that the newsletter accepts articles of a general nature of interest to all crystallographers. Please send any articles to me at [email protected] noting that there is a Word Template on the website to streamline production.
Articles --------
Improved target weight optimization in phenix.refine Mite-y lysozyme crystal and structure
Short communications --------------------
A lightweight, versatile framework for visualizing reciprocal-space data An extremely fast spotfinder for real-time beamline applications Hints for running phenix.mr_rosetta
Cheers Nigel
_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb
phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb
participants (2)
-
Frank von Delft -
Pavel Afonine