[phenixbb] Simulated anneal
nechols at lbl.gov
Tue Nov 13 17:50:37 PST 2012
On Tue, Nov 13, 2012 at 5:33 PM, Qiang Chen <chen at red.dfci.harvard.edu>wrote:
> I have a 3.05Å structure. I refine it to Rwork=25.2 Rfree=28.81. I use
> Molprobity to check the structure. The clash score is 100th and Molprobity
> score is 99th. I was suggested to run simulated anneal to help reduce
> Rwork and Rfree. I tried Cartesian and Torsion and different seed. The
> best result I got is: torsion anneal, start_temp=2500,final_temp=300,
> random_seed=4250146. The Rwork:25.8, Rfree=31.2. I checked the new model
> and found that some places in a domain were not correctly traced. The
> density for that domain is not good and have several flexible loops. My
> question is: does simulated anneal always work and reduce Rwork/Rfree?
No, especially when the structure is nearly finished.
> At the final stage of refinement, does it necessary to use simulated anneal
Definitely not necessary. Simulated annealing is a good way to improve a
model that fits the data poorly, or to remove model bias, but at the end of
refinement you should be much more conservative.
There are several other potential ways to reduce the R-free - such as using
TLS, or optimizing the weights. But you already have excellent geometry
and your R-free and R-free/R-work gap are both reasonable for that
resolution, so it's possible that further refinement won't help.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the phenixbb