Hi Hugh, I'll answer below:
This has given me more interpretable numbers. From reading all the previous correspondance (some of which has touched upon this issue), I still remain a little concerned as to the dependence of the results on the input .pdb file. I am almost certainly being overly cautious but as this is probably a common concern it'd be nice to hear your thoughts.
This is an issue which I don't think anyone has studied in depth and comprehensively before. It's also difficult to study as structural issues are likely to be a bias (directional randomness) rather than a variance (pure randomisation). For example a whole structural element could be reoriented. Or there could be domain motions which are not taken into account in the current level of theory. There is a lot of potential to develop this area of analysis in the future ;)
The Dratio that has resulted from the fitting of my apo protein is 1.55, as compared to that predicted by HydroNMR (1.75). For HydroNMR to overestimate values by this much is not itself a surprise or concern, but it is striking that the predicted Dratio for the closed complex is closer (1.43). I should emphasize that this is 2-domain protein with 2 "hinge" regions - one of which is mobile on the ms-us timescale and thus not amenable to study, and the other has fitted Rex terms. Infact about a third of the residues studied have fit Rex terms in the chosen ellipsoid model. On face value this is not actually a surprise seeing as various parts of the protein are flexible on the ms-us timescale as evidenced by line-broadening.
HydroNMR, from what I've heard, is terrible at prediction when there are domain motions. The program is also not very good at predicting the behaviour of proteins at the concentrations you have in the NMR tube. It is designed for prediction of the diffusion tensor of an isolated molecule, but your molecules are very close together in the NMR tube and this has significant consequences. The lower Dratio is understandable as you have domain motions and the core is only partly affected by the other domain. Did you perform an analysis with the two domains separately? For example as in my analysis at http://www.sciencedirect.com/science/article/pii/S0022283607007073 for the DsbA oxidoreductase.
It is possible that flexibility of the hinge regions is causing fluctuations in not only the principle diffusion tensor, but also the amide bond vectors relative to it. How could one ever test for this? Obviously the long-winded way is to fit the data with the closed complex coords and see if the X2 value is lower. But is there something more sophistocated and quick?
You could treat each domain in a separate model-free analysis. But model-free analysis assumes a static, perfectly averaged structure as the backbone of the analysis. If you do not have this, i.e. there are internal reorientations caused by the domain motions, you then have to rely on the local tm models. Though these models can easily absorb and hide motions if you have data at only 1 or 2 fields. Or alternatively you could consider developing a theory or method of analysis to handle this situation.
I have looked at residues with fitted Rex terms in the spherical model which are absent in the ellipsoid model, and also residues with ns Ts terms in the spherical model which are absent in the ellipsoid. Both of these have VERY strong dependence on bond vector relative to the principle diffusion axis, which was unsurprising based on what you have said elsewhere. Encouragingly this relationship is not so strong if you compare to the bond vectors taken from the closed structure. On the other hand, the local_tm model fit very few Rex terms which is a little concerning. I have looked if the residues with fit Rex terms in the ellipsoid model but not in the local_tm model have a bond vector dependence, which they don't. Are there any other consistency tests you would recommend?
If there are Rex terms in the spherical model but not in the ellipsoid, then these are almost guaranteed to be false motions (as described by Tjandra et al, 1995). The additional ns terms are also likely to be fake as described by Schurr 1994. Note that the local_tm models could sometimes absorb the Rex values into the local tm value as the data for these residues is usually very noisy. The only real way to determine if the Rex terms are real would be to perform some relaxation dispersion measurements, although that is not always conclusive. Data at 3 field strengths is also very powerful for determining if the Rex values are real. You must also remember that we assume that Rex in a model-free analysis is in the fast exchange limit, which is not always the case, and the only way to differentiate between quadratic fast exchange and linear slow exchange (and everything inbetween) would be to have 3 or more field strength data. Or, of course, relaxation dispersion data. I hope this helps. Regards, Edward