Thank you very much for the responses. Below is the question I asked and th=
e=20
responses. Thank you very much. The many issues raised will help me a lot. =20
I was surprised that only one person mentioned the Bayesian software.
Patrick Hays
DEA Special Testing and Research
email: STRLLIB@aol.com
=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D=3D
The Question: We are using NMR for quantitation of organic samples on our=20
Unity 500 and I=20
would like to improve on my results (comparing one peak integral to the=20
others can vary by a few percent which is significant in my work, i.e. 99%=20
versus 96%). My difficulty appears to be in the actual integration=20
(determining when to start and stop integrals, drift correction, baseline=20
correction). I either perform oversampling (5 x SW) or set my filterband=20
(FB) to twice SW to prevent reduction in signal at the spectral ends. Are=20
there some parameters I am forgetting to consider?=20
=20
Also, does anyone have any experience using Bayes software to do quants? =20
Does it really help?
The Answers:
If you are absolutely certain that you don't have a T1 or NOE problem with
the data, Bob Dykstra taught me a long time ago that the most common error
people make when integrating is to have insufficient data points to define
their peaks. There should be at least five data points in each peak of
interest.
After that, automating the procedure (which should include setting the
integration windows) normally helps also.
-------------------
Output spectrum as ascII format. Read by some program used by HPLC which has
some special function for integration, e.g. you can chose which point to sta=
rt
and which point to end integral.
I did not use it personally, but one professor's lab in our Biochemistry dep=
t
used for better integration.
-------------------
You should be able to get better than 1%. It is important
that you are running d1 > 5*T1. Baseline correction is=20
required, and I would be quite wary of the default spline=20
fits. Any type of overlap toward the baseline requires=20
deconvolution fitting rather than standard integrals.
Other than those, things should work quite well. If things
just stay inconsistant, you might look at your amplitude=20
stability test from the automated test procedures. Problems=20
with rf power amps, attenuators, relays, and the probe can=20
effect the amplitude stability, and might through off your=20
accuracies from spectrum to spectrum (not for just doing=20
integrals differently within the same spectrum, however).
=97------------------
To answer your question about where to start and stop integration check
VNMRNews (a few years back) for the relationship between percentage of peak
area versus accuracy. I would be interested in the answers to your other
questions.
--------------------=97
We regularly do quantitative NMR, with reproducible precision in our=20
integrals to better than 0.5%.
I generally set my filter to 3x SW for the filter reasons you mention, and=20
use oversampling. I have found that using about 4X my normal data size=20
(zero-filling to 128K or more is common) makes a significant difference.
Generally, I will use about a 10-20 degree pulse, and set the relaxation=20
delay to at least 5X the largest possible T1. I have been caught in the past=20
by underestimating the range of T1 values, even for routine proton NMR. =20
The last project I worked on, we obtained reproducible integrals within +/-=20
0.2% (obtained by comparing values between like-components in a two-componen=
t=20
mixture). For that one, I used a 10 second delay, a 20 degree excitation=20
pulse, and a 128K data size covering about 9 PPM on our 500.=20
You also mention baseline, which includes good phasing. I spend a LOT of=20
time optimizing the phasing and baseline to get the best possible integrated=20
intensities. The baseline correction routine in the GE Omega software is=20
really good, but I also get excellent results using Nuts, from Acorn NMR. =20
--------------------=97
Is the relaxation delay long enough? With small organic molecules and
methyl groups, T1 can be rather long. I assume that you use 2-4
dumy scans.
--------------------=97
Assuming that you have pure reference compounds, we are developing a method,=20
using the raw FID data alone, to obtain accurate and precise quantitative=20
information on many classes of organic compounds. Working with time domain=20
data sets eliminates the effects of operator subjective data processing. =20
Currently, the method is undergoing validation. After validation, a patent=20
application and journal article will be prepared.
-------------------=97
We have had some luck writing macros using the tools in NUTS to provide our
customers with more reproducible integration results. If you would like us =
to
look at your data to see if a NUTS macro could help then respond by email=20
with a
brief description of what you want done and attach a sample data set for us =
to
work on.
-------------------=97
Yes I can see that it might be significant for evidential purposes. Have yo=
u=20
tried re running the sample a few times using the same integrals each time t=
o=20
see the variation. The baseline + phase will be critical but should be OK i=
f=20
you don't have high dynamic range samples in which case APH will be fine. =20
I'm not really very Varian (although I do have a Unity300) are there other=20
phase routines that you can use in VNMR? Some algorithms work better in som=
e=20
circumstances than in others. The filters that you have fitted to the 1H lin=
e=20
will strongly affect linearity so you need to check these, is the probe tune=
d=20
precisely on resonance. If PW90 is long use a weak flip angle and longer=20
than usual D1. The short PW will evenly cover a wider SW and the longer D1=20
will take account of uneven T1 effects. Also the tuning is important as=20
otherwise you will get uneven reflection effects from the tuning capacitors.
-----------------=97
There's quite a bit of science to this, of course, and it's pretty well
documented, but I don't know a lot of references off the top of my head.
The main two points I always worry most about are:
adequate relaxation delay (5-7X T1 for pi/2 pulse, less for Ernst angle)
broad enough integral regions (usually at least 5x linewidth)
The key for T1's is accommodating differences in T1's in the sample or
between sample components. Using too-short delays in samples with
differing T1's does cause inaccuracies in integrals -- I see this too
often.
-------------------=97
Two very important parameters are the pulse length used and the delay time
between pulse. There is an excellent article in Concepts in Magnetic
Resonance, Vol 6, number 2, page 131, by Traficante and Steward which
discusses this is great detail.
-----------------------=97
Here are some observations regarding quantitative NMR.
1. Variations in T1 from sample to sample and within the sample can result
in incorrect choices of recycle time (d1 delay before the first pulse and
acquisition time). The total recycle delay should, as a minimum, be 5 T1's o=
f
the peaks/resonances of interest with the longest T1 (this will give 99.9+%
recovery of the magnetization along the Z axis).
2. Use a 90 degree pulse that is short enough to excite all the resonances o=
f
interest. The longer the pulse the narrower the region of the spectrum that
will "see" a 90 degree pulse. If possible, calibrate a 90 degree pulse with=20
the
transmitter near, with in a 100 or so Hz, the resonance of interest but not
on the resonace of interest.
3. Select an acquisition time that is long enough to avoid any truncation of
the FID.
4. Set fn to 128k to zero fill the difference between the number of points
collected and 128k. Digitization errors when observing narrow lines can add
significanly to the errors in the measurement.
5. Carefully set rof2 and alfa. If you are using dsp=3D'r' alfa is very clos=
e=20
to the
correct value (dsp=3D'r' requires the presents of a dsp board - this is=20
standard on
INOVA - if you do not have the dsp hardware installed see the command an
parameter refernce for instructions on using software dsp, digital signal
processing and oversampling). If lp is in the single digits, after you phase=20
the
spectrum, then alfa is good enough. In the ideal case the perfect alfa will
give you an lp=3D0.=20
6. If possible add an internal reference standard of known concentration and
use this samples resonance(s) to calibrate the integration.
7. Set the gain such that you are as close to filling the ADC's as possible
(ddff(1)) but do not overflow the ADC or overload the reciever..
8. Select a sample concentration that allows you to set the gain above 36
but below it's maximum. This will minimize the noise coming from the
least significant bit of the ADC. using dsp=3D'r' will further reduce
9. Make sure you have a flat baseline. Items 5 and 7 influence the baseline
flatness.
10. Oversample - see item 5.
11. Make 3 or 5 independent measurements.
------------------------=97
Many factors affect baseline shape which of course affects frequency=20
domain integration.
=20
These effects are due completely to a fundamental mismatch between=20
form of the time-domain data and the mathematical conditions required=20
for the discrete FT to work without creating artifacts.=20
=20
Most of the time, the DFT result is useful (frequency determination of=20
well-separated peaks). Often it is adequate (integration to ~5%=20
precision). Sometimes it fails completely (frequencies and/or=20
integrals of overlapped lines, integration independent of baseline=20
artifacts).
=20
Bayesian modeling avoids all these issues as the parameters=20
(frequency, decay rate , amplitude.. or integral) are estimated from=20
time domain data. Nuisance parameters (phase, overlap, first point FID=20
distortion) are integrated out of the statistical model.
=20
See J. Magn. Reson. 98 483-500 (1992) for some details. If you use=20
VNMR, the Bayesian manual also contains useful information.
-------------------------
Things to consider when doing quantitation with NMR. Know your T1's!!! If
you wish to use 90 degree pulses and the ratio of T1(long)/T1(short)is
larger than 2 you must use a recycle time that is greater than 10T1(long).
Place the carrier in the middle of the spectral window. The digital
resolution should be set such that the narrowest line is defined by a
minimum of 6 points. The integration limits should be the same for all
lines and expressed as the same multiply of linewidths. Thus, for broad
lines the integration limits in frequency units will be larger than narrow
lines. For Lorentzian lines +/-64 linewidths (if I recall correctly)
yields 99% of the intensity of the line. If each line is integrated the
same number of linewidths the error (accuracy) is the same of each line and
will cancel. If an internal quantitation standard is used. A good choice
is pyrazine with known purity (from DSC)- it is soluble in both aqueous and
organic solvents and the resonance rarely interferes; but be prepared for
long T1's. Determine the precision by running the experiment 5-10 times on
one solution (with internal quantitation standard). The precision, if you
addressed everything else, is governed primarily by S/N of the weakest
resonance you are integrating. If you are adding a quantitation standard
you have weighing (volume) errors- run two additional solutions (one
experiment per solution) and see how close the resolutions are. If you
sample is well behaved; i.e., it is not volatile, not statically charged,
not hydroscopic, etc., you should easily achieve 1% relative error. You
notice I did not say readily!!! I hope this is helpful. =20
--------------------=97
To reduce baseline roll, I use Bruker's digital filters. This is equivalent
(in theory) to using an SW of over 200 KHz with no analog filter and selecti=
ng
only the region of interest. I then carry out an automatic (polynomial)
baseline correction. I find that doing it manually gives similar results but
takes longer. An integral over a fixed number of linewidths on either side
of the peak will theoretically yield a fixed proportion of the total integra=
l.
Of cause it is important to allow enough relaxation time between acquisition=
s.
A delay of 5T1 can cause errors of up to 0.7%. (7T1 0.1%) For very accurate
work don't forget the 13C coupling. Either ensure that no 13C coupled signal=
s
appear under the peak of interest or use inverse gated 13C decoupling.
In my experience integral accuracies of 1 or 2% can be routinely attained. =
I
have not tried really hard to improve on this but it may be possible.
--------------------=97
We often do quantitation of 13C NMR spectra, mostly with selectively
labelled compounds in biological extracts, so we have to determine the
integrals of multiplets coming from 13C-13C J-coupling. We work along
the following guidelines:
- As we typically obtain thousands of transients, saturation (or
formation of a steady-state magnetization lower than quantitative) may
become an issue. The textbook rule to avoid decreasing of the signal is
waiting 4-5 T1 of the slowest relaxing nucleus between two pulses when
employing 90 degree excitation pulses. We apply 30 degree flip angles,
and then 1 T1 is enough for magnetization to recover to more than 99% of
its pre-pulse value. It is a gain in signal/time and as we apply shorter
pulses, the excited region is larger and the excitation profile is
better. Nevertheless, you have to know the T1's of the signals you are
interested in, before you set up your experiment.
- You can gain sensitivity (of a factor of 2-3) by applying
1H-decoupling during the whole experiment and generating NOE from
protons to carbons. But the factor of this enhancement is often not
uniform in every signal. So, if you want quantitative spectra, you can
either turn on 1H-decoupling during acquisition only (it is preferred
when sensitivity or concentration is not a problem) or you can have
full-time decoupling only if you know the enhancement factors for every
signal you are interested in.
- As you want to determine areas under peaks, the peaks must be defined
with satisfactory precision. It means that you need at least three real
data points between the half-width of every signal (this can be of
course achieved by zero filling).
- You need to set the baseline as flat as possible and the phase as
correct as possible.
- NMR-lines are theoretically Lorentzian curves, which are quite broad
at their bases, and bear substantial areas even under 1-2% of peak
intensity (usually in the noise region). If you have good enough
signal/noise (in 13C this is not frequently the case) you can apply
"traditional" integration, with zero points set at +-5 halfwidths from
the center of the peak. If the S/N is not that good, or (as it is very
frequently the case, for us almost always) your spectrum is too crowded
to let you set the zero points so far from the peak without overlap,
you'd better use deconvolution for the determination of peak areas.
As an answer to your particular questions, I admit that I prefer
oversampling (for us typically 8xSW on an INOVA 400) to enlarging FB,
and unfortunately we do not have the Bayes software.
----------------------=97
I seem to remember Field And Sternhell in their book ' Analytical
NMR' claim that you cannot measure integrals to better than 2% accuracy for
various technical and physics reasons. This may be true, but I guess you
should see the difference between 99 and 96 %. Here are a few comments that
you may well have thought of - I assume you are using proton NMR, other
nuclei have their own cautions.
1. Have you measured all the T1s for your systems. It is fairly quick and
easy to get ballpark figures (perhaps use saturation recovery if T1 are
long). We have found some very long T1s in some organic systems (T1s > 60s)
and so quantitative work requires recycle time of around ten minutes per
scan for these (7 * T1), yes that really is only 6 scans an hour!. Even if
you find T1s of 5s that would mean 35s delay. For quantitative work it is
probably most efficient to use a 90 degree pulse and wait the full time -
the Ernst angle is not really relevant in this situation.
2. Have you included (or excluded) all 13C satellites from your integrals.
This will obviously make several percent difference in the result.
3. Assuming linewidths (T2* s) are pretty constant, you would probably be
best advised to ensure all integrals are the same width where possible since
lorentian NMR lines to tail off to infinity.
4. I have never found the different digital processing scenarios to have
significant effects on relative integrals - If you are worried try acquiring
with a good old fashioned analogue filter.
5. One trick you might try is to get good S/N and correct the baseline
around each peak you are interested in. Then integrate without adjusting
the integral...