Hi Maddy, You are welcome! I have answered your questions below. On 21 July 2010 16:59, M Strickland <M.Strickland@xxxxxxxxxxxxx> wrote:
Dear Edward, Thank you for your help concerning running Relax earlier in the year. I have now successfully run the full_analysis.py script on two receptors from different species (ie similar structures, but different binding affinity to a peptide). I have a few questions... 1. For each run - sphere, oblate, etc. I had to remove different residues in order to help the run to finish (I generally removed those that were fluctuating between two models endlessly) and restarted the run. The runs went well past fifteen rounds if I didn't do this.
This problem has been encountered by many others. If you carefully compare the parameters of fluctuating models, you will see that there is one parameter on the boarder of insignificance that disappears and returns in a never-ending loop. This is ok, as the dynamic description of the system does not change. Models are of zero importance, only the dynamics should be of interest. So it is safe to terminate this never ending loop at any point, and any of the results from this circular loop can be used. The problem here is now caught in the newest version of the full_analysis.py script. The script will identify these loops and instantly terminate. If you would like to use this, the feature can be found in relax 1.3.5. Note that your earlier full_analysis.py script cannot be used, you will need to modify the new one.
2. Does this mean that I am affecting the final model selection by doing this? By removing all of the R2/R1/NOE values from a whole loop, I could be changing how well the model performed. For example this loop could be more condusive of a spherical model, but with the residues removed, perhaps an oblate model performs better (although it shouldn't).
If this causes a change of global diffusion model, i.e. sphere to spheroid, you will find that the Da diffusion parameter will be close to zero. Again this should not make much of a difference (though if you do see it, carefully check the different diffusion models and compare the internal dynamics). But I would not remove the data based on this endless loop. I would only recommend removing data that is found to be inconsistent. For this, please see Sebastian Morin's paper: http://dx.doi.org/10.1007/s10858-009-9381-4
3. The obvious answer to this is run all of the models with all the residue's R1/R2/NOE values left in, but this would take months. It was taking over a day for each model anyway, so in total around two weeks for the complete calculation.
One to two weeks is quite normal with relax. That is the price you pay for higher quality results. It could be changed to hours (just as in Modelfree4) by simply decreasing the precision from 1e-25 to 1e-5. You could use the new version of the script with relax 1.3.5, or you could use the current one but manually terminate the script once this loop has been encountered.
4. So - my case, I have two species of the same receptor - human was found to be oblate and chicken was found to be spherical. I would expect spherical for both, as the spherical models very well describe what is to be expected for each receptor and both receptors are very similar in structure, with chicken only containing three extra residues. Should I discount the oblate model?
Note that the relaxation data is most sensitive to the global tumbling, maybe 80% of the rates are from this and only 20% from the internal motions. Therefore a slight anisotropy or rhombicity is highly significant! Ignoring it will cause artificial motions to appear (http://dx.doi.org/10.1039/b702202f). The spherical model is very rarely encountered, even if your molecule looks roughly spherical. There are many factors which will cause the protein to no longer tumble spherically. If you have a mobile loop or a mobile terminus pointing into the solution - this will add significant anisotropy. Also note that the protein is not the only thing tumbling here. There are 2, 3 or more layers of water shell causing the correlation time to be up to twice as big as estimated from the hydrodynamic beads model of Garcia de la Torre. But if you have a non-uniform charge or hydrophobic residue distribution on the protein surface, then the water shell will be significantly large around charges or non-existent around hydrophobic patches. So encountering a perfectly spherical diffusion tensor almost never occurs - though assuming a perfectly spherical diffusion tensor occurs far more than it should :S
5. Comparing to ModelFree4 results - rigid residues have similar S2 values, but those that are more flexible have much lower S2 values in Relax than in MF4. Is this simply because Relax has models 6-9 in addition to the first 5 and so S2 values will be different? For the MF4 results I simply calculated a spherical model, models0-5, at both 600 MHz and 900 MHz. In Relax I combined them.
This is caused by many different factors. The first is that relax's optimisation is about 20 orders of magnitude more precise than Modelfree4 or Dasha. The second is model selection - this makes a huge difference for mobile parts of your molecule. The third could be because of bugs in Modelfree4 (though Art Palmer has incorporated my patches which fix most of these issues). There are a few others as well, all of which is described in full detail in my papers at http://www.nmr-relax.com/refs.html, specifically: d'Auvergne and Gooley, 2008a d'Auvergne and Gooley, 2003. So what you are seeing are well documented analysis artifacts that I have solved and placed into relax.
I've only been using dynamics programs for 6 months, so if I've missed something obvious, please let me know. In addition if you would like any files to look at - just email this address.
Please don't hesitate to ask more questions. I hope my answers have helped. Regards, Edward