Normally I make a series of NOE-experiments beginning with 40ms and double the time for the next experiment and ending at 640ms. This depence mainly on the mean T1-time I found for the molecule of the interesting groups. If I have a molecule lager then 500 g/M then I find often negativ or no NOE. So in this case I use the ROESY instead of NOESY, but in this case the TOCSY signals have to be eliminated instead of COSY-Signal. The mistake you normally make with sine calculation is higher than with lorentz-Gauss but the main problem is the reduction of the NOE/ROE-crosses by the overlayd COSY/TOCSY-Signals. The intensity of the NOE/ROE-signals depend with r^6, so small changes in distance make high coss-intensity changes. After I got nearly linear NOE/ROE- increase I use the highest mixing time (in the linear depence) for distance calculations. If you calculate straight lines in the linear region you find sometimes negative origins, this reveal additionl COSY/TOCSY overlays.

In the relaxation series, rule of thumb is one must use the same processing parameters through out the series! Eventually, peak heights are measured as the intensity ratio between time=0 & the series. My point is if you linear predict for the first point (t=0) and series, the intensities will get nullified and you won't overestimate. Also another point to be taken care is how much linear prediction has to be used. Also, the exponential fit and repeated points (for statical error estimation) will be a good indication about your processing parameters. If you are in doubt, try to process the spectra with or without LP and estimate the R1, R2 & ssNOE, then compare the rates & fit, if they are same or different. Then you know what to use. I never had a problem with R1, R2 & ssNOE estimation with linear prediction. I use 'LP -fb' function in NMRPipe which generally predicts twice the data size for indirect dimension or or simply the topspin processed spectra & never had any problem.

Just user

@Manfred Elend:

Thank you for the response, but I admit I don't get the point. Do you see artifacts from linear predictiion?

@Sivanandam V N:

Thanks for the response. I think it's clear that every individual spectrum of a series has to be processed absolutely the same, otherwise results would be unpredictable.

I did some comparisons between R1, R2 and NOE with and without Bruker Topspins forward LP, and with the amount of predicted points. As expected, with the same number of original data points, I have larger errors in the relaxation curves that make use of spectra with linear prediction (e.g. 200 original points plus 200 predicted points).

The mean of the rates remain approximately the same, without visible bias.

There is a slight overestimation of rates if I reduce the data points to the half and predict the other half (100 original points plus 100 predicted points) when compared with a complete set of 200 original points. (see attached plot)

And here's the plot for the first scenario (200 points vs 200 points + 200 predicted points). I don't see bias here, although I never did any quantification on the scatterplots.

I have really good experience in spectra recording and processing. I suggest you to not to use linear prediction, because in that case you got more signals i.e. crosspeaks, but some of them are not real. So, my advice is not to use it at all. I do not know what kind of machine you are using-Varian or Bruker and which software you are using for spectra processing. Sometimes, it is good to consult the user manual, so I suggest you to do it first. Bruker gives very detailed description of processing parameters for any NMR technique. For example, in Bruker's manual it is suggested to use Squared Sine-Bell function for the processing. You can use it, or it is better to try with Lorenz-to-Gauss function. It works better in most cases compared to Squared Sine-Bell function. You must keep in mind that if acquisition parameters were chosen bad for spectra recording, you are not able then to get good spectra with good processing parameters. I hope so that my advice will be helpful.

I don't use linear prediction. Some colleages do this for better results in 1D-experiment, but I didn't get good results for quantitative NMR.

Concerning to LB/GB versus sine bell process depends on your problem. Some times LB/GB will be a nice filter to noise in 2D spectra, but probably this will turn the cross peaks larger then sine bell, especially if GB parameter is less the 0.1, or 0.05 0.001 range with highest negative values of LB (-1 to -10), with LB=-.5 and GB .5 you will increase noise but the resolution will be increased a lot, test it on 1D proton spectra near of doublet or triplet. There are no good values, depends on your data (spectra) and problem to be solved, there are books about nmr processing, Try to find Peter Biegler book, very useful concepts. Linear prediction will be good sometimes and also bad sometimes, as someone here told, after use you probably can get artefacts (phantom cross peaks) the best way is test and compare with no linear prediction spectra. Other problem concern LP will be, LP will try to move to more resolution in cross peaks, if you have a cross peak which can be from two or three Amino Acids, the best way is try to have best as possible resolution, LP will try to solve the problem, but can make a function not exactly as is the separation (not enough points to good prediction) and then your final integral of those cross peaks will be really wrong, in a good relaxation curve this will be your error value or variation, the Rˆ2(fit) parameter will goes down. Sine Bell functions will turn cross peaks fine and can make poor signal to noise in the final spectra. Again, test it and compare with and without, remember also the importance in use those functions to avoid truncation at end of fid or pseudofids(2nd, Nth.. dimensions).

For R1, use integrals

For R2n only maximum height of the echo (other wise thereis rrors from field inhomogeneity (see Tiffon et al Analytica Chimica QActa, 1981, 124, 415-419)

For HOE,use integarls but remamber that in the reference soppectrum the delay should be 10 T1 (double exponential, but not a single exponential); see Canet et al, JMR-1976, 23, 36& and Harrris, JMR, 1976, 24, 449

There isno reason why LP can introduce errors in the respective peak heights

Thank you for your answers. It seems, there is quite some disagreement here. I guess, it all boils down to "it depends on your system" and "test it with and without".

I see that LP works less well with many signals and with a low S/N. Maybe I should do tests and comparisons with a protein which is a less ideal case than the examples I posted above (few signals, relatively high S/N).

Dear Martin,

to my remembering the integral of any nD NMR Signal is, from a mathematical point of view, given by the amplitude of the first datapoint S(t1=0,t2=0). This number can hardly be influenced by a FORWARD linear prediction. Therefore I'd predict a small effect (if at all) if linear prediction is in use. If you use a squared cosine with a shift this will of course reduce the amplitude of this first datapoint - but again this should not be a problem, as ANY signal is affected identically.

What we learned many years back is that for realistic 15N pulses you always will have a non-linear phase-error along the 15N dimension in 15N-CPMG type experiments which can be rather substantial. You might have a close look on signals in individual planes of your pseudo-3D, because phase-errors affect naturally the integral....we also published some work on this. Unpleasantly this phase error is dependent on the number and delay of the CPMG blocks!

Kind regards

Alfred

Thank you Alfred, I'll look into this.

Is it ok if I hijack the thread and ask what software are you using for the relaxation analysis?

I suggest you contact Anthony Vassallo who was one of the pioneers in solid-state C13 NMR. You can contact him through ResearchGate. Best of luck.

My major advice to you would be the following:

1) Make sure all the 2D spectra in the pseudo 3D are processed identically. Make sure you have good base line correction and the same baseline correction in all spectra.

2) Linear prediction can create errors in frequency (ppm). If you are using forward linear prediction only then intensity information maybe ok but its possible the predictions may still not accurately predict intensity in the predicted part of the FID which could have a _small_ effect on intensity. More likely, LP will result in peak shifts that could lead to mis-assignment.

3) How many of the R1 or R2 delay points have fully relaxed? Or another way to think about this is, how much noise is there in your estimate? You may find the errors you are seeing between LP and noLP are within your error of measurement of the parameter in the first place. Don't use points in spectra where essentially the peaks are in the noise. You can't trust these estimates. I see you have larger errors for larger R values... that is where signals are lost faster and more likely to have more noise in them.

4) You can only avoid bad truncation artifacts post-data acquisition by inducing line broadening which will possibly bleed peaks into each other. If you have 200 (complex?) points in a 15N dimension (even with TROSY) and you are still getting truncation artifacts, collect more points next time? This sounds like a nice problem to have for a 45 kDa protein... is it a sphere? ;)

Without seeing your nmrPipe scripts and data its hard to advise further. Good luck with exploring solutions.

In addition to what others have said:

1) There is always a trade off between S/N and resolution when applying a windowing function because the window function either scales down the noise in the later part of the FID where there is little signal, resulting in improved S/N, or the reverse.

2) Regarding LP, I believe some care must be taken because although the integral is the same after LP (as stated above by Alfred Ross) the maximum signal intensity is not. If LP extends a true signal, then it will add signal to the end of the FID, without adding noise. Since the integral is the same (determined by first point), and the signal is now sharper, the intensity should then go up. You might also need to adjust the order of the LP. In general LP extends the data for a specified number of signals, so if it is too low, the strongest signals will be extended and the rest will not. (On the other hand, if it is too high, noise is extended) So if you use the intensity of peaks, I would be careful with LP