Hi Martin,
Please see below:
Until now I used relax only on 9 - 12 kDa proteins and probed backbone
relaxation. With an 8-core Intel workstation it took me only 2 days (if
not even less!) to do a model-free calculation.
The calculation time is highly dependent on the system being studied.
If the optimisation space is quite complicated, something which
appears to be independent of molecule size, then it can take much
longer. Did you run relax with Gary Thompson's multi-processor
framework to take advantage of all your CPU cores?
Now I tried a big protein–one with >240 assigned residues. It took 2 days
and 23 rounds to find a optimized spherical diffusion model, and since
yesterday it churned out 35 prolate diffusion models!
This is quite possible. I would highly recommend you create plots of
the progression of optimisation such as in:
d'Auvergne, E. J. and Gooley, P. R. (2008). Optimisation of NMR
dynamic models II. A new methodology for the dual optimisation of the
model-free parameters and the Brownian rotational diffusion tensor. J.
Biomol. NMR, 40(2), 121-133.
(http://www.nmr-relax.com/refs.html#dAuvergneGooley08b or
http://dx.doi.org/10.1007/s10858-007-9213-3).
You can obtain the data for the plots by manually opening the results
files in the 'opt' directories and then manually creating the graphs.
I understand that with increasing number of spins also computation time
increases, but how many rounds are "normal" and how does computation time
scale in respect to number of analyzed spins? Is it indicative of data
quality if it takes too long to compute?
I don't know if you could ever define "normal". For some systems, two
rounds are sufficient. For others, a huge number of rounds is needed.
This is a complex combined optimisation + model selection problem,
hence you have to traverse multiple optimisation spaces to find the
solution. I suggest reading:
d'Auvergne E. J., Gooley P. R. (2007). Set theory formulation of the
model-free problem and the diffusion seeded model-free paradigm. Mol.
Biosyst., 3(7), 483-494.
(http://www.nmr-relax.com/refs.html#dAuvergneGooley07, or
http://dx.doi.org/10.1039/b702202f).
This, together with the above 2008b paper, will explain the problem in
full detail. You really should create the plots of the 2008b paper to
see if you are slowly circling around the 'universal solution',
getting closer and closer, or if something else is happening. There
is a lot of scope still in advancing the model-free analysis to
improve the search for this solution. It could be that you are in
almost perpetual motion orbiting around two solutions, sliding in and
out of different optimisation spaces or universes, one day colliding
with one of them (see below).
The protein in the analysis is in a trimeric complex (where only one
component is labeled at a time, simply to reduce overlap), I guess relax
should model the correct correlation time and tensor for the whole
multimer, and also find the correct center of mass? Shouldn't also the
tensor parameters of any of the multimer components be identical?
If the complex exhibits domain motions, then this could explain the
long optimisation times (though not definitively). I am currently
working on a new theory that I'm calling the frame order theory to
investigate such a problem but, until this is developed, you have to
use the current techniques. If there is inter-domain or inter-subunit
motion, then the current model-free theory is an approximation and the
correlation times found will be something between the inter-domain and
global tumbling motions. This could explain the large number of
rounds required as the multi-universe space could be quite complex and
possibly have multiple solutions. I don't think anyone has
investigated such a space in the approximation of one diffusion tensor
when two overlapping tensors of vastly different time scales are
present. Again there is a lot of scope for advancing this area of NMR
(and biophysics in general)! Note that the approximation of one
tensor when multiple are present could result in artificial motions -
i.e. Rex or nanosecond motions being observed across the entire
system, with a possible XH vector orientational dependence. I hope
this description is not too abstract.
Regards,
Edward
Re: How many rounds does it take?