Hi Edward,
On 05.06.2012, at 15:48, Edward d'Auvergne wrote:
Reading my thesis
(http://www.amazon.com/Protein-Dynamics-Model-free-Analysis-Relaxation/dp/3639057627)
might be easier :P
I have to admit, I underestimated the amount of math knowledge which is
needed to understand (and trouble-shoot) model-free analysis. For a humble
biologist like me, it's quite overwhelming. But that makes it interesting,
too. :)
Now that I looked more carefully at
the results file, I can see the exact problem. I would recommend
multiplying your proton frequencies by 1e6, as they should be in Hz ;)
Great news! :D Thank you so much for the hint. I'll start relax once more
with correct frequency values.
You should see something quite different!
Now it worked – somehow. Grace files, text files etc. with the results are
now in the "final" directory. BUT: There was no diffusion model selected (->
just local T_m) which is implausible for such a small, well-defined globular
protein domain like SH3. This is reflecting the results I got from TENSOR2
where any of the diffusion models would not get accepted, until I
artificially doubled the experimental errors (which have been calculated by
relax).
Now I have a lot of relax results files in a lot of directories, but I don't
know how to look at them, e.g. extract statistical parameters. Actually, I
don't even know what to look after.
***
The auto analysis works like a perfect black box: Put data of good quality in
and a few days later the final results drop out. For systems like SH3 there
should not be a problem at all! So where could be the obvious problem?
It's probably not the sample: it is stable for several years now, there is no
oligomerisation (judging from R1/R2) or aggregation/precipitation to bee
seen.
It's unlikely it's the temperature: the methanol calibration at the 600
machine showed no differences between measurements, but the resulting data
still gave no fitting models in TENSOR2. The consistency tests give no reason
to doubt it'll be different at the 750 machine. Next week I'll hopefully know
definitvely.
The pulse programs could have errors that are not obvious. Maybe we should
compare ours to sequences from another lab where the pulse programs have
proven to be good enough.
It may well be that the NMR structure (1aey, from 1997!) does not reflect
reality well enough. At the moment relax is giving the crystal structure
(1shg, even older) with artificially added protons (Amber), and I'm hoping
that it'll yield different results. What do you think, what should be used:
crystal strutures + added protons, NMR structures (averaged or single models
from the ensembles)?
If you have any ideas what could be wrong on our side, any advice would be
very welcome.
Cheers
Martin, exhausted
P.S.: I'm testing relax --multi mpi4py with a 12-core machine right now.
Seems like it's working quite nicely!