Hi Martin, Please see below:
among various substantial improvements over the available old-school mf modeling packages, relax implements a different set of model-free models than e.g. Art Palmer's Modelfree4 or Martin Blackledge's TENSOR2. Neither of the above programs includes the m0 model (as relax does), where no mf parameters are determined. You described the m0 model as a special case, where none of the internal motions / model-free parameters are statistically significant.
This is correct. The key point here is that we are dealing with applied physics. Hence we don't deal with physical significance, only statistical significance. These are quite different concepts and often confused in the field of NMR.
Is this equivalent of a "failed" model-free analysis?
Most definitely not!
Does that also mean, in such a case where relax chooses m0, TENSOR or Modelfree would choose a nonsignificant model which of course wouldn't be appropriate?
If relax chooses m0, then TENSOR, Modelfree and Dasha should choose model m1 (this is assuming that the diffusion tensor and other parameters are identical in all these programs).
Would relax also choose a model by "chance" if I just left out m0 from the analysis?
The next level of statical significance if m0 is not present is m1, as this is the simplest model available with a single S2 parameter. All other models are more complex. Note that an exception is model m9, which is also only available in relax, which competes with m1 for statistical complexity.
I'm not sure what the reason for all parameters being insignificant is and what the implications are, i.e. what it means in terms of protein mobility.
The key is again that this is applied physics. There is much more happening in your system than what you can see through your relaxation data. But you can only see what you can see. Therefore a motion being statistically insignificant, with respect to your data and experimental errors, does not mean that it is not present! This is often the case with nanosecond loop motions which when using the old model-free protocols are much more often judged to be insignificant, even though proper statistics (AIC, BIC, ICOMP, etc.) says otherwise.
It cannot mean that the "m0"-residues behave like a static body (S^2 would be 1).
Statistically, yes. Physically, no. You just can't see it from the data you have. That is the meaning of this model. A good analogy is as follows - you could have a picture of an elephant but, if you only have 4 pixels in that picture, you probably won't be able to tell that your picture is of an elephant.
Does it simply mean that the supplied data and models are not sufficient for the type of motion that the protein is exhibiting? I learned that model-free works best if the motions are in the extreme narrowing limit, but it is a misconception that it is limited to this time domain. But aren't there limitations to the kind of motions that can be modeled by mf?
The information that "model-free works best if the motions are in the extreme narrowing limit" is itself a misconception! This comes from the original Lipari and Szabo model-free papers themselves. One thing they did to help them with the optimisation, back when computers were much slower, was to use the extreme narrowing limit to simplify their equations for the calculations used in the 1982 papers. Look specifically at Equation 36 in the original paper. I don't know if this is completely relevant to your question, but noise is another issue which affects the reliability of the te parameters. As te increases, so does the errors. The cause of this is that you need to look at this upside-down. The real parameter affecting your relaxation data is the diffusion rate equal to 1/(6*te). On the diffusion rate scale, the errors are relatively invariant, but if you invert this then you have hyperbolic behaviour of the te errors. Note that you won't find such information written anywhere in the literature.
If the motions approach the NMR timescale, I'd first of all expect broadened peaks (which I have a lot). From my understanding, that precludes analysis, but would not be a limitation of the mf formalism itself. These motions could be picked up by the Rex parameter ... So m0 wouldn't occur if I have amides moving around at a nanosecond timescale.
Here I'm not exactly sure what you mean. Rex broadens peaks whereas nanosecond timescale motions cause you peaks to be the strongest in the spectrum. Maybe this relates to model m9 in relax. Sometimes the very weak peaks, broadened by chemical exchange, are too noisy to extract model-free motions from. This is visible in relax as the selection of model m9. In such a case, model m0 will probably not be picked.
I'm also not sure what would happen if the internal motions approach the global correlation time ...
As I mentioned above, the errors on te become bigger and bigger. There is a point where te and tm merge, and this is governed again by noise and statistics. For example with relax I easily am able to separate a te value of 8.192 ns from a tm of 10 ns when no noise is present (see my 2008 paper http://www.nmr-relax.com/refs.html#dAuvergneGooley08a). Note you can physically have motions above tm contributing to relaxation, but that it is very hard to statistically extract this from the noise (the feasibility of te > tm is talked about in the original papers). So despite people saying te cannot be slower then tm, this is another misconception. Physically yes you can, but statistically we currently cannot.
I don't know if inaccurate or inconsistent data would favor such a behavior. We now use selective pulses in the R1 and NOE-experiments, temperature compensatio
Inaccurate/inconsistent data is not noise, but rather bias. Whereas noise shifts parameter values around randomly and governs which motions are statistically significant, bias on the other hand shifts everything in one direction. I describe this is detail in my PhD thesis (http://www.nmr-relax.com/features.html#primary_refs), specifically in section 2.1.4 and figure 2.1. The effect on a model-free analysis is generally the introduction of artificial motions (either Rex or nanosecond). Bias could probably in some cases hide motions, but more likely will result in artificial motions. Bias could also be introduced if the spherical, spheroidal, or ellipsoidal diffusion is too simplified for the system or if partial dimerisation is occurring. I hope some of these explanations help. Regards, Edward