Hi Mark, Sorry for the late reply. I've been stuck at home for the last 2.5 weeks looking after my 5 year old son, due to local COVID-19 outbreaks in his Kindergarten. It is difficult to do anything at the moment!
I received your data files in your private email but have not been able to even open them for now. My instinct is that this is input data related. If you were to plot the peak intensities (heights) in a HSQC, could it be that the problematic data points are those with the lowest intensity? Or those with the greatest intensity? Related to this, do you use the standard process of variance averaging across all peaks in the spectra? You can probe if the errors are underestimated by artificially increasing the errors and seeing if there is a point where relax can fit the data with no patterns left in the residuals. I hope these leeds might help. Regards, Edward On Fri, 21 Jan 2022 at 19:17, Bostock, Mark <[email protected]> wrote: > > Dear Edward, > > > Thanks for your reply - this issue occurs for a small subset of residues. > Most of the data is well fitted, hence I don't think there are any > systematic issues with the data. The errors are estimated based on repeat > planes. The data is single-scan interleaved and the cpmg pulsing frequencies > are randomised. > > > I have managed to get better fits for these residues on some occasions, > however, it's not systematic. e.g. I tested LM63 and CR72 models just on one > residue and the fit was much better. However, when I refitted all the > residues the improved fit wasn't replicated. It looks like the fit gets stuck > in a local-minimum and doesn't converge. Perhaps I could share the script and > sample data with you to see if there are any parameters I can further > optimise? > > > Many thanks again for your help, > > > Best wishes, > > > Mark > > ________________________________ > From: Edward d'Auvergne <[email protected]> > Sent: 21 January 2022 11:54:29 > To: Bostock, Mark > Cc: [email protected] > Subject: Re: [relax-users] Inconsistent relaxation dispersion fits. > > Hi Mark, > > Do you have this issue with all your data? The residuals are quite > small but probably statistically significant and there seems to be a > pattern to it, but these can sometimes be due to incorrect error > estimates and biases respectively. What techniques did you use for > temperature control and calibration > (https://www.nmr-relax.com/manual/Temperature_control_and_calibration.html)? > Improper control can lead to bias and "patterns" in residuals. And > how did you estimate the errors for each data point? If these are > out, the non-linear least squares fitting algorithms can fail. The > errors influence the curvature of the optimization space (rather than > topology) and incorrect errors can sometimes squeeze valleys in this > space creating false minima. > > Regards, > > Edward > > On Fri, 21 Jan 2022 at 00:06, Mark Bostock <[email protected]> wrote: > > > > Dear relax-users, > > > > I'm trying to fit some methyl-13C SQ CPMG data. I have a number of > > residues, which appear to have an exchange contribution, but result in poor > > fits e.g. > > > > I've tried a variety of different relaxation dispersion models (CR72 full, > > B14 full, NS CPMG 2-site expanded, IT99, TSMFK01) but the fit doesn't > > improve. I've also tried increasing the grid increment parameter from 11 to > > 21, but again this doesn't improve the fit. Very occasionaly when I have > > been testing conditions, a model has accurately fitted the data (in the > > following NS CPMG 2-site expanded) but I am unable to replicate this > > consistently. > > > > > > Any suggestions to improve the reliability of this fitting would be very > > much appreciated. > > > > Many thanks, > > > > Mark > > > > > > > > _______________________________________________ > > nmr-relax-users mailing list > > [email protected] > > https://lists.sourceforge.net/lists/listinfo/nmr-relax-users _______________________________________________ nmr-relax-users mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/nmr-relax-users
