Diffusion model Round Machine Time Time_for_opt/
=============== ===== ============= ======= =============
local_tm --- Pelican 15h00
--------------------------------------------------------------------
sphere MII init Pelican 0h10
1 Pelican 42h00 31h00
2 Pelican 23h00 21h00
3 Pelican 18h00 16h00
4 Pelican
prolate MIII init Hibou 0h20
1 Hibou 76h00 74h00
2 Hibou 44h00 39h00
3 Hibou
oblate MIV init Hibou 0h20
1 Hibou 76h00 75h00
2 Hibou 44h00 40h00
3 Hibou
ellipsoid MV init Pigeon 1h00
1 Pigeon 75h00 74h00
2 Pigeon 31h00 27h00
3 Pigeon 20h00 17h00
4 Pigeon
--------------------------------------------------------------------
Those times are quite long! They're not completely unreasonable
considering you have 220 residues and data at 3 field strengths.
So, it seems the behaviour is normal with the processing times getting
shorter as the optimization proceeds... Maybe the first runs were long
because of my errors on the NOEs being too small (I think they could be
too underestimated because, with the traditional approach used in
ModelFree -- the approach from Mandel et al., 1995 -- I can select only
~20% of residues for model 1 (and almost no residues in subsequent
models) under an axially symmetric diffusion tensor...
That percentage of model m1 being selected is quite normal, maybe just
a little too high because of the deliberate under-fitting of that
ANOVA/hypothesis testing approach. The underestimation of the R1, R2,
and NOE errors could however explain the long calculation times. The
more the errors are underestimated, the more convoluted and difficult
the model-free space is to optimise in. Are you experimentally
determining the NOE errors by measuring the base plane noise or by
running both the saturated and reference NOE spectra twice? I would
value the errors of equal or greater importance than the relaxation
data itself! Under or overestimation of the errors will influence
both optimisation and model selection, and hence the final dynamic
picture of the molecule. High quality errors are very important.
Also, the fitted values seem reasonable.
That's a good sign. Despite the amount of time this is taking, you
may end up with good results, even if it does take two weeks. Unless
there is something subtly wrong causing the slow down. Have you
modified the 'full_analysis.py' script much? Are you sure your
relaxation data is being read in correctly with the correct field
strength values? How did you determine your R1 and R2 values? Were
the R1 and R2 errors determined using Monte Carlo simulations?
Anyway, it seems I'll get optimized models soon and then I could look at
them and maybe exclude models 6 to 9 if I find them irrelevant... Let's
see !
I would strongly recommend in keeping those models! You have the data
to fully characterise these and dropping them will induce
under-fitting if you have a system which exhibits large interesting
motions. It doesn't seem like this is the slow point anyway.
A question, however... Is it normal that the X2 in the opt directories
as the same value for every residue (for example in the file
ellipsoid/round_3/opt/results.bz2) ? Is it the global X2 ?
Yes. The 'opt' directory is the last part of the new protocol where
the global model - the diffusion model together with all model-free
models of all residues - is optimised. It is a single optimisation of
the single global chi-squared value.
Edward
Re: full_analysis.py