Re: pA = 0.5 problems.Re: pA = 0.5 problems.Hi, Great, the delta_omega value should be different when clustering. For the high values, you should check the fitted curves to see if you see any issues. It is likely that this is actually the values at the minimum, and that noise or some experimental bias is pushing the values too high. But if that's what the data says, then there is nothing you can do about it. Adding constraints to prevent such situations is just an ugly hack - the end result will be meaningless anyway as it is not the minimum. The best is to report the value with its Monte Carlo simulation error. You will probably find that the errors for these high delta_omega values will also be large and hence the large values are not statistically meaningful. Regards, Edward On 30 April 2014 15:25, Troels Emtekær Linnet <tlinnet@xxxxxxxxxxxxx> wrote:
Hi Edward. For 68 residues, i get this: # Parameter description: The population for state A. # mol_name res_num res_name spin_num spin_name value error None 10 G None N 0.995627205128479 None # Parameter description: The exchange rate. # mol_name res_num res_name spin_num spin_name value error None 10 G None N 843.138159024003 None # Parameter description: The chemical shift difference between states A and B (in ppm). # # mol_name res_num res_name spin_num spin_name value error None 10 G None N 5.66240190353847 None None 11 D None N 7.24018919066445 None None 15 Q None N 1.27388427761583 None None 16 G None N 1.70637956735682 None None 37 G None N 1.39439146332152 None None 41 G None N 2.00489971256184 None None 42 L None N 1.14631824779138 None None 43 H None N 4.11812700604344 None None 46 H None N 6.74721119884288 None None 47 V None N 18.6941069719393 None None 49 E None N 7.41204467390198 None None 50 E None N 6.73571806460759 None None 51 E None N 2.6906237906698 None None 53 N None N 3.43636822580998 None None 54 T None N 1.72434404944155 None None 56 G None N 7.38662030427091 None None 57 C None N 1.88126168471543 None None 58 T None N 6.61594197477923 None None 61 G None N 3.42205122284923 None None 67 L None N 4.00714078384803 None None 68 S None N 3.02933093965657 None None 70 K None N 2.65894254799687 None None 72 G None N 4.01752138022632 None None 73 G None N 3.10502419263122 None None 75 K None N 5.52331683531287 None None 78 E None N 2.39121460031728 None None 79 R None N 2.95565292785431 None None 80 H None N 10.6521951761457 None None 81 V None N 6.46552900214463 None None 82 G None N 5.48378904252769 None None 85 G None N 4.72783895083071 None None 86 N None N 2.2535643167938 None None 87 V None N 3.42430152185329 None None 102 S None N 1.33719888517455 None None 103 V None N 1.78945522230369 None None 104 I None N 2.1193021535956 None None 105 S None N 1.20023816089299 None None 111 A None N 3.68849791596676 None None 112 I None N 1.92921136977377 None None 115 R None N 2.1336531230742 None None 118 V None N 1.1301287075642 None None 121 E None N 1.68619193009267 None None 123 A None N 4.91019478151119 None None 126 L None N 7.6255827307843 None None 127 G None N 4.89765215595432 None None 128 K None N 2.26502557102985 None None 129 G None N 1.79003350167683 None None 130 G None N 1.74650398353974 None None 131 N None N 4.91476102864345 None None 133 E None N 1.02559032422555 None None 134 S None N 0.842131709855722 None None 135 T None N 9.20627022843478 None None 137 T None N 8.00007213116674 None None 138 G None N 1.9412902050166 None None 139 N None N 6.51160366265863 None None 140 A None N 8.89216425477085 None None 141 G None N 2.354941400505 None None 142 S None N 3.50895251891688 None None 143 R None N 2.65884864234097 None None 146 C None N 2.92485233744021 None None 147 G None N 4.71130879214043 None So dw is moving fine. But we do though think that dw has high values. There is a 18 ppm and 10 ppm in there. Now trying with ShereKhan. Best Troels 2014-04-30 10:45 GMT+02:00 Edward d'Auvergne <edward@xxxxxxxxxxxxx>:Hi, I should expand on the statistics a bit more. Maybe using AIC would clarify the noise vs. real data components. Here is a short table: Set Chi2 k AIC Individual 32.97 10 52.97 Cluster 48.79 8 64.79 So even using AIC, the individual fit is better. Statistically it is not that you are just fitting more noise in the non-clustered fit. That is significant! One thing I noticed is that dw is the same for both spins in the clustered fit. Could you check if this is the case for other clustering cases? It must be different for each spin. Maybe there is an important bug there. Regards, Edward On 30 April 2014 10:16, Edward d'Auvergne <edward@xxxxxxxxxxxxx> wrote:I tried to generate sherekhan output, but since I have time_T2 of 0.04 and 0.06, for the two fields, I cannot generate the input files for ShereKhan.ShereKhan should support this, and it would be a good test for relax. The second line of the input file has this time. Was it that relax could not create the input files rather than ShereKhan not handling this?My problem origins from that I would like to compare results from Igor Pro script. Yet, another software solution.Have you run the Igor Pro script to compare to relax? With the same input data, all software solutions should give the same result. This is important - you need to determine if the issue is with relax or with the data itself. It is best to first assume that the problem is with relax, then see if other software produces a different result (the more comparisons here the better). Maybe relax is not handling the two different times correctly. Otherwise if everything has the pA = 0.5 problem then the solution, if one exists, will be very different.I now got the expected pA values of 0.97 if I did a cluster of two residues.This could indicate that the pA = 0.5 issue is in the data itself, probably due to noise. You should confirm this by comparing to other software though. Comparing to the 'NS CPMG 2-site expanded' might also be useful.If I do an initial Grid inc of 21, use relax_disp.set_grid_r20_from_min_r2eff(force=False) I get this.As I mentioned before (http://thread.gmane.org/gmane.science.nmr.relax.scm/20597/focus=5390), maybe it would be better to shorten this user function name as it is a little misleading - it is about custom value setting and not the grid search, despite it being useful for the later.:10@N GRID r2600=20.28 r2500=18.48 dw=1.0 pA=0.900 kex=2000.80 chi2=28.28 spin_id=:10@N resi=10 resn=G :10@N MIN r2600=19.64 r2500=17.88 dw=0.7 pA=0.500 kex=2665.16 chi2=14.61 spin_id=:10@N resi=10 resn=G :10@N Clust r2600=18.43 r2500=16.98 dw=2.7 pA=0.972 kex=3831.77 chi2=48.79 spin_id=:10@N resi=10 resn=G :11@N GRID r2600=19.54 r2500=17.96 dw=1.0 pA=0.825 kex=3500.65 chi2=47.22 spin_id=:11@N resi=11 resn=D :11@N MIN r2600=14.98 r2500=15.08 dw=1.6 pA=0.760 kex=6687.15 chi2=18.36 spin_id=:11@N resi=11 resn=D :11@N Clust r2600=18.19 r2500=17.31 dw=2.7 pA=0.972 kex=3831.77 chi2=48.79 spin_id=:11@N resi=11 resn=DIf you sum the chi-squared values, which is possible as these are all the same model, then you can compare the individual fits and the clustered fit. The individual fit total chi-squared value is 32.97. The cluster value is 48.79. This is very important - the individual fit is much, much better. You should make a plot of the fitted curves for both and compare. Note that a better fit does not mean a better result, as you are fitting both a data component and noise component. So the better fit might be due to the noise component. This is why clustering exists.Ideally, I would like to cluster 68 residues. But as you can see, if several of my residues start out with dw/pA far from the Clustered result, this minimisation takes hilarious long time.I can see how this would be a problem for you mass screening exercises. This will probably require a lot of investigation on your part to solve, as I have not seen any solution published in the literature. Though if you could find a solution in the literature, that would probably save you a lot of time. You could also ask others in the field. If you remember (http://thread.gmane.org/gmane.science.nmr.relax.devel/4647/focus=4648), you changed the parameter averaging to the parameter median for the clustering. So maybe that is having an effect. Anyway, you need to first compare to other software or models and see if there is a problem in relax first, before trying to invent a solution. Regards, Edward