[phenixbb] solvent model in phenix.refine

Pavel Afonine pafonine at lbl.gov
Thu Dec 1 14:59:54 PST 2011

Hi Oliver,

> Hello,
> the ordered_solvent option in phenix.refine apparently uses these parameters by default:
>      min_model_peak_dist = 1.8
>      max_model_peak_dist = 6.0
>      min_peak_peak_dist = 1.8
> I was wondering about the justification for this (extremely relaxed) set of limits. As far as I know, the shortest hydrogen bonds (if defined as donor-acceptor distance) are about 2 A, and even these are considered exceptional and rare. Similarly, anything in excess of 3.5 A is quite unlikely to be significant energetically, so why use 6 A as an upper limit? Even if  max_model_peak_dist was only meant to limit the extent of the outer solvent shell, there is no such thing as max_peak_peak_dist to avoid unreasonably long individual bonds.
> My personal suggestion would be a range of about 2.3 - 3.2 A for a significant hydrogen bond, with the lower limit exemplified by charged interactions (carboxyl-water or oxonium-water), which tend to be shorter - and stronger - than average.

these are the parameters that you can cast them to match your taste.

"max_model_peak_dist=6.0" is the maximum distance that a water is 
allowed to be from the macromolecule. However, this doesn't mean that 
the water can freely float at 6A or so distance from the macromolecule, 
since the other restrictions (h_bond_max = 3.2) will make sure that this 
water is bridged to the macromolecule or another water.

Note, the parameters you referred to

       min_model_peak_dist = 1.8
       max_model_peak_dist = 6.0
       min_peak_peak_dist = 1.8

are used to initially select peaks, and the other set of parameters

     h_bond_min_mac = 1.8
     h_bond_min_sol = 1.8
     h_bond_max = 3.2

is used to make further selection.

The next level of selection is the map correlation between 2mFo-DFc and 
DFc maps computed for each water (the default threshold is 0.7 - that is 
if the map CC is lower than 0.7, the peak-candidate is discarded). There 
are bunch of other selection criteria as well.

I agree, 1.8 is a generous choice, but given the amount of (many) other 
selection criteria that phenix.refine uses to distinguish water from 
non-water, I hope in the end the noise gets filtered out.
On the other hand, we all know that we never work with the exact 
electron density but with its Fourier image that is always distorted due 
to finite resolution and data completeness (plus other data and model 
errors). These kinds of  distortions may perturb water peak position in 
both directions. The shorter distance cutoff (1.8A) is meant to account 
for this in order to not miss such waters.

> On the other hand, I do frequently observe elongated densities which, in accordance with the default parameters, are typically interpreted by phenix.refine as two water molecules about 2 A apart. So the question arises whether this is a realistic assumption, compared to a "sliding water" model in which the position of a single molecule is just ill-defined in a certain direction (due to a competition of adjacent binding sites for example)?

This kind of situations is best modeled using waters in alternative 
conformations (if I correctly understood what you mean), which is not 
implemented yet in phenix.refine - phenix.refine does not build waters 
in alternative conformations automatically). So I guess placing two 
waters is probably not ideal, but still is better than nothing. Ideally 
it would be better to assign one water altloc A and the other one altloc 
B, so their occupancies are constrained to 1.

Please let me know if you have a specific example showing that 
phenix.refine placed wrong waters, and I will be happy to have a closer 

All the best!

More information about the phenixbb mailing list