Hi,
The model selection techniques are independent of the program used.
So you can use AIC model selection with both Modelfree4 and relax.
Which model selection technique did you use with Modelfree4? Did you
use the independent FASTModelfree interface to Modelfree4? This
separate program has the hypothesis testing model selection built in
and is unavoidable. If you have used hypothesis testing with either
program, then the more complex models are disfavoured as under-fitting
occurs, and hence this will cause there to be less m5 models selected.
Therefore I doubt that model selection is causing the difference
between the two programs as you are seeing more m5 using Modelfree4.
The true cause is highly likely to be different diffusion tensors
within the two programs and problem of using the isotropic diffusion
tensor. Using this tensor when Brownian rotational diffusion is
non-isotropic is not good (I can almost guarantee that every protein
will be non-isotropic, if proper model selection is used (by proper I
mean from the statistical field of model selection)). If a vector
points along the long axis of the diffusion tensor, its correlation
time will be underestimated and this will result in artificial Rex
values will appear (see Tjandra et al., 1996). If the vector is
perpendicular then artificial nanosecond motions (model m5 and higher)
will appear (see Schurr et al., 1994). The refs are below.
As for the te value being on the nanosecond timescale, there is
nothing in the theory which stops this! There are no mathematical or
physical limits on this. The only limits are statistical - forced by
experimental noise - and which vary depending on the exact
experimental noise, the amount of data collected, processing bias,
etc. The limits can be maybe ~10 ps as a lower bound (this is highly
variable and AIC model selection will show you this significance (the
Mandel et al., 1995 hypothesis testing significantly increases the
lower bound)). The upper bound is approximately the global
correlation time, tm in the Lipari and Szabo notation, again purely
because of noise. The decoupling approximation has no effect with
these bounds, its effect is on the accuracy of the correlation time
parameter itself. This is easily tested using synthetic data. Just
use an isotropic diffusion with say tm = 10 ns and then create a
motion of ts = 2 ns with an S2s value of 0.7 and a fast internal
motion with tf = 10 ps and S2f = 0.9. The fast internal motion in
this case is insignificant because of experimental noise (add 5%
noise, optimise the model-free models, and do model selection (AIC or
hypothesis testing)). The S2 value will incorporate the slow and fast
order parameters - S2 = S2f*S2s. The tf parameter will not be able to
be extracted because of noise. Hence the optimised te parameter will
be very close to the ts parameter! This is all mathematical
modelling, and there is nothing wrong with that. I hope this
description helps in understanding why te has no problem with being on
the nanosecond timescale.
Regards,
Edward
Schurr, J. M., H. P. Babcock, and B. S. Fujimoto: 1994, 'A test of the
model-free formulas. Effects of anisotropic rotational diffusion and
dimerization.'. J. Magn. Reson. B 105(3), 211–224.
Tjandra, N., P. Wingfield, S. Stahl, and A. Bax: 1996, 'Anisotropic
rotational diffusion of perdeuterated HIV protease from N-15 NMR
relaxation measurements at two magnetic'. J. Biomol. NMR 8(3),
273–284.
On 8/14/07, Hongyan Li <hylichem@xxxxxxxxxxxx> wrote:
Dear Edward,
Sorry for the confusion. I tried to compare the results obtained from
Modelfree4
and relax using the isotropic model (since I cannot get Modelfree4 works on
axial symetric model at moment). 't-rex-sim.agr' corresponds to results from
Modelfree4 and 't-rex-relax.agr' corresponds to the results from relax.
Using
Modelfree4 there are more residues fitted to M5 where using relax, more
residues were fitted by M2 and M4, I supposed that is due to different
criteria
on model selections. However, one thing I don't understand is that te
extracted
from M2 and M4 should be on a scale of several hundreds ps (fast) instead of
several thousands ps (slow). In this regards, Modelfree4 is more resonable
and
there seems some problem in terms of model elimination and model selection
for
the Relax.
Best wishes,
Hongyan
Quoting Edward d'Auvergne <edward.dauvergne@xxxxxxxxx>:
Hi,
Sorry, I'm not exactly sure what the graphs correspond to. Is
't-rex-sim.agr' Modelfree4 using the prolate (or oblate) spheroid
(this is axially symmetric anisotropic Brownian rotational diffusion)?
And is 't-rex-relax.agr' the results from relax using the spheroid
tensor? Have you used constraints on Da in relax to isolate the
oblate and prolate spheroids? Also how many iterations of the
model-free optimisation; model elimination; model selection; and
global minimisation (the optimisation of the model-free parameters of
all spin systems together with the diffusion parameters) have you
used? What is the input data and do you have data at more that one
field strength?
I'll try to answer some of your questions, but without more
information these may not be the answers you are after. The first
thing which is a little worrying is that in 't-rex-sim.agr' there are
many ts values between 6 to 8 nanoseconds. Unless you are working
with an unfolded protein or a system that is far from globular, this
is a very strong indication that the diffusion tensor is significantly
underestimated. How did you determine the initial diffusion tensor in
the analyses? Did you use the full_analysis.py script when using
relax (which requires data at minimally 2 field strengths)? The
errors on the Modelfree ts results are also worrying. This, to me,
looks like that there has been failures in the MC simulations causing
very similar errors on all the high ts values. Did you use an upper
limit of 10 ns in Modelfree?
Another worry is that you obtained similar results from relax using
the spherical and spheroidal diffusion! How many iterations of
model-free analysis did you use? And how did you determine the
initial diffusion tensor? As for the te values in the nanosecond
range, this is perfectly normal. This is modelling slow internal
motions. Model m5 was designed for this purpose, but if the fast
internal motion is close to insignificant due to experimental noise,
then model m2 is perfectly capable in modelling the slow motion. Also
if you set the range of the y-axis in all the correlation time graphs
from 0 to 10 ns, then you can see that the results from Modelfree4 are
more worrying. For the correlation time results, it is better to make
two graphs - one for fast motions up to 200 or so picoseconds and one
for slow motions from 200 ps up. Don't forget that what you are doing
is modelling. The models don't care what the underlying true dynamics
are - they will model that motion as best as they can. So classifying
the dynamics based on which model is selected is at best distracting
or at worst misleading. It's the results that matter, not the model.
I hope this answers some of your questions.
Regards,
Edward
Re: relax and modelfree