On Wed, Mar 31, 2010 at 8:47 AM, Dirk Kostrewa <kostr...@genzentrum.lmu.de>wrote:
> this is very interesting! From the list of changes, it appears that in > version 1.6.1, you use a similar idea to implement hydrogen bonds via DSSP > during refinement that I used to stabilize the 4.3 A refinement of Pol II in > complex with TFIIB [1]: > . . . > Could you please give more details of how this is implemented? What are > your hydrogen bond target distances and sigmas? Do you update this list > during refinement? > We're still testing this method (thus "alpha version" in Paul's email); my initial experiments indicate that it may improve R-free by up to 0.5% at moderate resolution, and generally keeps R-work and R-free closer than would otherwise be the case. It can also make R-free worse, although I think this happens less frequently. I had expected this to be resolution-dependent, but I tried it on partially built structures at either 2.25A and 3.1A and the results looked similar. Part of the problem is deciding how to filter outliers; the default behavior is to throw out any "bonds" greater than a specified threshold (e.g. 2.5A for H-O). KSDSSP (UCSF's free clone) has some quirks with respect to helix assignment which result in incorrect bonding if you take the results as-is without filtering. However, for poor models at low resolution, leaving the outliers in may be essential. All of the restraint parameters are adjustable, but I found that 1.975A for H-O was appropriate, and I left the sigma at 0.05 but 0.02 is sometimes better. We don't update the bonds during refinement, but I suppose we could - haven't done enough testing yet to know whether this is helpful. I used a simple list of additional 2.9 A target-"bond"-distances between N > and O with a target sigma of 0.05 A. This list was determined with DSSP and > a self-made Fortran95 program using a user-defined "energy-threshold" prior > to refinement and was kept constant during refinement. > Currently PHENIX uses a 3.0A distance for N-O (sigma also 0.05), but that wasn't rigorously derived. My gut feeling is that N-O "bonds" are problematic, partly because the distribution of distances (in the Richardson lab's Top500 database) appears to be bimodal - I think this results from the difference between parallel and anti-parallel sheets. We probably need to account for this when calculating the bonds. I also found that the restraints usually improved R-free more when hydrogens were used, but I haven't tried anything close to a 4.3A structure yet. At that resolution, anything that keeps the helices wound probably helps. I hadn't realized that you used similar restraints in the PolII/TFIIB paper - I'd be interested in knowing more about this, since I haven't found many useful references so far. (I couldn't find the page on the BUSTER wiki, by the way.) Personally, I think, using secondary structure hydrogen bonds should be an > option in every refinement program, especially at lower resolution!!! The > BUSTER Wiki describes the procedure that I used. For REFMAC, I haven't seen > anything similar, yet. > If you don't mind using PHENIX for this: phenix.secondary_structure_restraints model.pdb format=refmac I haven't made it as far as testing the output with REFMAC, but it's there if anyone wants to try. FYI, the program is still clumsy about switching to N-O bonds when hydrogens are absent, so the argument "h_bond_restraints.substitute_n_for_h=True" may be necessary. -Nat
