On 6/10/11 6:10 PM, Ed Pozharski wrote:
> On Fri, 2011-06-10 at 12:48 -0700, Ethan Merritt wrote:
>> I don't think that combination makes any sense. Whatever anisotropic
>> is being described by the TLS parameters can also be fully described
>> by the individual anisotropic U^ij terms. So the TLS parameters are
>> entirely redundant, leading the minimization function to be poorly
>> defined.
> True, but the TLS refinement is implemented in refmac as separate step
> preceding the positional/adp refinement. So in theory the TLS step will
> go fine, and the anisotropic ADP refinement will take care of anisotropy
> additional to TLS component. But of course you are absolutely right that
>
>> I am virtually certain that refinement of individual anisotropic
>> U^ij terms cannot be justified at 1.8A. Too many parameters,
>> too few observations.
> Matt reports that TLS lowers R/Rfree compared to anisotropic ADPs
> "alone". Well that is good (TLS works!), but the question is if the TLS
> +aniso is better than TLS+iso or, better yet, if R-values decrease when
> going from Biso to Baniso.
>
> Cheers,
>
> Ed.
>
What Ed is describing above is sort of what I hoped would happen when I
first tried this: the TLS refinement would take care of the global
anisotropic atomic motions, and the aniso B refinement would take care
of any residual local ones. I realized that using the two together
could be unstable, and as several people have pointed out, the data to
parameter ratio is pretty low at this resolution to be trying this out.
That said, here are the results of the calculations Ed is wondering
about above:
Starting model: 2096 (non-H) atoms, 249 protein residues, 170 waters.
Riding hydrogens built automatically in each refinement cycle.
Data set: 23665 unique reflections, 97% complete from 25 - 1.8 A. 5% of
reflections taken for free R set. Space group C2, one molecule per
asymmetric unit.
Refinement run with the same starting model each time, and all
parameters except the TLS and B refinement left the same. I set the
number of refinement cycles so the refinements would go to convergence:
30 cycles for the non-TLS jobs, 10 TLS + 25 regular for the TLS jobs.
One TLS group covering the entire (single-domain) protein. Waters are
excluded from TLS refinement, but included in the B refinement.
Here are the R and Rfree stats for each job.
Biso alone: 0.206 / 0.253
Baniso alone: 0.182 / 0.238
TLS + Biso: 0.188 / 0.228
TLS + Baniso: 0.177 / 0.230
The TLS + Biso job gives the lowest free R and the smallest difference
between R and Rfree, so it's the winner.
Thanks to all for your input. And thanks to Mischa for answering my
original question!
- Matt
--
Matthew Franklin, Ph. D.
Senior Research Scientist
New York Structural Biology Center
89 Convent Avenue, New York, NY 10027
(646) 275-7165
|