Dear Pavel, yes, such an exact prediction of ordered water molecules might be very helpful. I was sure that somebody else had this idea already. I was playing around with a few datasets truncated a low resolution (3.5 - 4.0 A) and then compared Rwork/Rfree using an input model with and without water molecules. Clearly the water molecules had a large contribution in the refinement of these artificially truncated datasets. Sascha pointed me to an example in your paper from 2002: Lunin, V.Y., Afonine, P. & Urzhumtsev, A.G. (2002) "Likelihood-based refinement. 1. Irremovable model errors.". Acta Cryst., A58, 270-282. I had a look into the literature to get an idea and found several programs evaluating the inner shell water molecules and some programs predicting water positions. I had a try only on a few programs. I found that a nice summary is given in the publication on an approach called WaterDock: Ross GA, Morris GM, Biggin PC (2012) "Rapid and accurate prediction and scoring of water molecules in protein binding sites." PLoS One 7(3):e32036. But before analyzing many structures and see whether it might work in general, my aim is much simpler. I have high resolution structures of with water molecules and try to implement the ordered water molecules into the refinement of a protein complex at low resolution. My approach was maybe a bit of naive so far but I am sure there is good way to do that. Best wishes, Guenter

Hello,

I tried this idea back in 2004. In a nutshell: using all (or categorized subset of) structures in PDB we can learn about distribution of structured water and given this knowledge we can build an a priori contribution of scattering arising from such water to the scattering of any given new structure or a structure at low resolution (where the water is not visible in maps).

Either I did not spend enough time on this or the idea wasn't viable, but one way or another this did not work in my hands. I think it may be worth revisiting this 10 years later! Perhaps I would do it better now than back then!

All the best, Pavel

On 11/16/14 2:19 PM, Nathaniel Echols wrote:

...

I will leave it to others to debate the wisdom of this strategy, but to answer the purely technical question:

On Sun, Nov 16, 2014 at 2:06 PM, Guenter Fritz mailto:[email protected]> wrote:

Is it possible to use protein and water atoms from the reference models to generate restraints for the low resolution refinement?

I don't think so. You'll probably find it easier to refine the atoms separately, i.e. one run with reference model and the individual sites selection set to "not resname HOH", followed by a run with harmonic restraints on waters and selection "resname HOH". Alternately, you could try applying harmonic restraints to the entire model, although I suspect that the waters and protein require different weights (or sigmas).

-Nat

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

Well, we happily restrain (these days) low resolution structures to high resolution structures. I see no philosophical difference. On 25/11/2014 05:58, Pavel Afonine wrote:

Hi Guenter,

while I clearly understand your motivations, I don't feel very comfortable with placing explicit atoms that are not supported by the data.

The fact that those atoms are present in high-resolution structure does not mean that they are also present in low-resolution structure. You can argue that adding these waters improves Rfree and you may think of it as an improvement. However, as a counterargument one can say that R-factor is a global metric that is unlikely to be sensitive to adding/removing just one single molecule. Therefore, while adding bulk of "structured" waters may be an improvement in general this still does not mean that all the waters you add are true and good ones. Say what if 70% of them are good and 30% are rubbish? In this case still Rfree may improve because you add more good water than bad, but adding bad ones is counterproductive anyway and introduces model bias and thus must be avoided.

All the best, Pavel

On 11/17/14 1:33 AM, Guenter Fritz wrote:

...

Dear Pavel,

yes, such an exact prediction of ordered water molecules might be very helpful. I was sure that somebody else had this idea already. I was playing around with a few datasets truncated a low resolution (3.5 - 4.0 A) and then compared Rwork/Rfree using an input model with and without water molecules. Clearly the water molecules had a large contribution in the refinement of these artificially truncated datasets. Sascha pointed me to an example in your paper from 2002:

Lunin, V.Y., Afonine, P. & Urzhumtsev, A.G. (2002) "Likelihood-based refinement. 1. Irremovable model errors.". Acta Cryst., A58, 270-282.

I had a look into the literature to get an idea and found several programs evaluating the inner shell water molecules and some programs predicting water positions. I had a try only on a few programs. I found that a nice summary is given in the publication on an approach called WaterDock:

Ross GA, Morris GM, Biggin PC (2012) "Rapid and accurate prediction and scoring of water molecules in protein binding sites." PLoS One 7(3):e32036.

But before analyzing many structures and see whether it might work in general, my aim is much simpler. I have high resolution structures of with water molecules and try to implement the ordered water molecules into the refinement of a protein complex at low resolution. My approach was maybe a bit of naive so far but I am sure there is good way to do that.

Best wishes, Guenter

...

Hello,

I tried this idea back in 2004. In a nutshell: using all (or categorized subset of) structures in PDB we can learn about distribution of structured water and given this knowledge we can build an a priori contribution of scattering arising from such water to the scattering of any given new structure or a structure at low resolution (where the water is not visible in maps).

Either I did not spend enough time on this or the idea wasn't viable, but one way or another this did not work in my hands. I think it may be worth revisiting this 10 years later! Perhaps I would do it better now than back then!

All the best, Pavel

On 11/16/14 2:19 PM, Nathaniel Echols wrote:

...

I will leave it to others to debate the wisdom of this strategy, but to answer the purely technical question:

On Sun, Nov 16, 2014 at 2:06 PM, Guenter Fritz mailto:[email protected]> wrote:

Is it possible to use protein and water atoms from the reference models to generate restraints for the low resolution refinement?

I don't think so. You'll probably find it easier to refine the atoms separately, i.e. one run with reference model and the individual sites selection set to "not resname HOH", followed by a run with harmonic restraints on waters and selection "resname HOH". Alternately, you could try applying harmonic restraints to the entire model, although I suspect that the waters and protein require different weights (or sigmas).

-Nat

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

Okay, so could one build them and restrain their positions and occupancies to those of the high res structure? The point is: if you've seen a 1st shell water in conserved 20 structures, the chances of it not being there are in fact rather slim - which is the identical assumption we make to justify NCS. On 25/11/2014 07:11, Pavel Afonine wrote:

No, this is different. Note: you restrain to high-res structure, not add it to the model. Pavel

On 11/24/14 10:54 PM, Frank von Delft wrote:

...

Well, we happily restrain (these days) low resolution structures to high resolution structures. I see no philosophical difference.

On 25/11/2014 05:58, Pavel Afonine wrote:

...

Hi Guenter,

while I clearly understand your motivations, I don't feel very comfortable with placing explicit atoms that are not supported by the data.

The fact that those atoms are present in high-resolution structure does not mean that they are also present in low-resolution structure. You can argue that adding these waters improves Rfree and you may think of it as an improvement. However, as a counterargument one can say that R-factor is a global metric that is unlikely to be sensitive to adding/removing just one single molecule. Therefore, while adding bulk of "structured" waters may be an improvement in general this still does not mean that all the waters you add are true and good ones. Say what if 70% of them are good and 30% are rubbish? In this case still Rfree may improve because you add more good water than bad, but adding bad ones is counterproductive anyway and introduces model bias and thus must be avoided.

All the best, Pavel

On 11/17/14 1:33 AM, Guenter Fritz wrote:

...

Dear Pavel,

yes, such an exact prediction of ordered water molecules might be very helpful. I was sure that somebody else had this idea already. I was playing around with a few datasets truncated a low resolution (3.5 - 4.0 A) and then compared Rwork/Rfree using an input model with and without water molecules. Clearly the water molecules had a large contribution in the refinement of these artificially truncated datasets. Sascha pointed me to an example in your paper from 2002:

Lunin, V.Y., Afonine, P. & Urzhumtsev, A.G. (2002) "Likelihood-based refinement. 1. Irremovable model errors.". Acta Cryst., A58, 270-282.

I had a look into the literature to get an idea and found several programs evaluating the inner shell water molecules and some programs predicting water positions. I had a try only on a few programs. I found that a nice summary is given in the publication on an approach called WaterDock:

Ross GA, Morris GM, Biggin PC (2012) "Rapid and accurate prediction and scoring of water molecules in protein binding sites." PLoS One 7(3):e32036.

But before analyzing many structures and see whether it might work in general, my aim is much simpler. I have high resolution structures of with water molecules and try to implement the ordered water molecules into the refinement of a protein complex at low resolution. My approach was maybe a bit of naive so far but I am sure there is good way to do that.

Best wishes, Guenter

...

Hello,

I tried this idea back in 2004. In a nutshell: using all (or categorized subset of) structures in PDB we can learn about distribution of structured water and given this knowledge we can build an a priori contribution of scattering arising from such water to the scattering of any given new structure or a structure at low resolution (where the water is not visible in maps).

Either I did not spend enough time on this or the idea wasn't viable, but one way or another this did not work in my hands. I think it may be worth revisiting this 10 years later! Perhaps I would do it better now than back then!

All the best, Pavel

On 11/16/14 2:19 PM, Nathaniel Echols wrote:

...

I will leave it to others to debate the wisdom of this strategy, but to answer the purely technical question:

On Sun, Nov 16, 2014 at 2:06 PM, Guenter Fritz mailto:[email protected]> wrote:

Is it possible to use protein and water atoms from the reference models to generate restraints for the low resolution refinement?

I don't think so. You'll probably find it easier to refine the atoms separately, i.e. one run with reference model and the individual sites selection set to "not resname HOH", followed by a run with harmonic restraints on waters and selection "resname HOH". Alternately, you could try applying harmonic restraints to the entire model, although I suspect that the waters and protein require different weights (or sigmas).

-Nat

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

Hi Pavel, yes, I see your point and I fully agree. My motivation is simple: I hoped to get better phases from a model including water molecules, i.e. better maps for those parts of the model, which are still not very well defined. I feel at low resolution we fight for every tiny little bit of information that could improve the model. So I found it an attractive idea to use structured water. It is somewhat analogous to the concept to add restraints by the use of reference models derived from structures determined seperately , although these parts might look slightly different in the complex. Finally to evaluate which water molecules might be there or not could be judged by running parallel refinement runs with several dozens of models including different water molecules. Good luck, this is not a big deal nowadays. But all in all, so far my trials in one specific case yielded only a moderate improvement. Best wishes, Guenter

Hi Guenter,

while I clearly understand your motivations, I don't feel very comfortable with placing explicit atoms that are not supported by the data.

The fact that those atoms are present in high-resolution structure does not mean that they are also present in low-resolution structure. You can argue that adding these waters improves Rfree and you may think of it as an improvement. However, as a counterargument one can say that R-factor is a global metric that is unlikely to be sensitive to adding/removing just one single molecule. Therefore, while adding bulk of "structured" waters may be an improvement in general this still does not mean that all the waters you add are true and good ones. Say what if 70% of them are good and 30% are rubbish? In this case still Rfree may improve because you add more good water than bad, but adding bad ones is counterproductive anyway and introduces model bias and thus must be avoided.

All the best, Pavel

On 11/17/14 1:33 AM, Guenter Fritz wrote:

...

Dear Pavel,

yes, such an exact prediction of ordered water molecules might be very helpful. I was sure that somebody else had this idea already. I was playing around with a few datasets truncated a low resolution (3.5 - 4.0 A) and then compared Rwork/Rfree using an input model with and without water molecules. Clearly the water molecules had a large contribution in the refinement of these artificially truncated datasets. Sascha pointed me to an example in your paper from 2002:

Lunin, V.Y., Afonine, P. & Urzhumtsev, A.G. (2002) "Likelihood-based refinement. 1. Irremovable model errors.". Acta Cryst., A58, 270-282.

I had a look into the literature to get an idea and found several programs evaluating the inner shell water molecules and some programs predicting water positions. I had a try only on a few programs. I found that a nice summary is given in the publication on an approach called WaterDock:

Ross GA, Morris GM, Biggin PC (2012) "Rapid and accurate prediction and scoring of water molecules in protein binding sites." PLoS One 7(3):e32036.

But before analyzing many structures and see whether it might work in general, my aim is much simpler. I have high resolution structures of with water molecules and try to implement the ordered water molecules into the refinement of a protein complex at low resolution. My approach was maybe a bit of naive so far but I am sure there is good way to do that.

Best wishes, Guenter

...

Hello,

I tried this idea back in 2004. In a nutshell: using all (or categorized subset of) structures in PDB we can learn about distribution of structured water and given this knowledge we can build an a priori contribution of scattering arising from such water to the scattering of any given new structure or a structure at low resolution (where the water is not visible in maps).

Either I did not spend enough time on this or the idea wasn't viable, but one way or another this did not work in my hands. I think it may be worth revisiting this 10 years later! Perhaps I would do it better now than back then!

All the best, Pavel

On 11/16/14 2:19 PM, Nathaniel Echols wrote:

...

I will leave it to others to debate the wisdom of this strategy, but to answer the purely technical question:

On Sun, Nov 16, 2014 at 2:06 PM, Guenter Fritz mailto:[email protected]> wrote:

Is it possible to use protein and water atoms from the reference models to generate restraints for the low resolution refinement?

I don't think so. You'll probably find it easier to refine the atoms separately, i.e. one run with reference model and the individual sites selection set to "not resname HOH", followed by a run with harmonic restraints on waters and selection "resname HOH". Alternately, you could try applying harmonic restraints to the entire model, although I suspect that the waters and protein require different weights (or sigmas).

-Nat