Nuclear Magnetic Resonance Spectroscopy

The vast majority of NMR spectroscopic experiments are carried out in solution to identify the structures of small chemical molecules, including natural products, but a wealth of high–resolution applications is found in other areas, such as determination of the three–dimensional structures of proteins and the analyses of complex biological mixtures (e.g. biofluids). In addition, much effort has been devoted to solid–state NMR spectroscopy in which special techniques have to be used to overcome very broad NMR peaks and to recover useful chemical information. One application in this area of interest to pharmaceutical scientists is the determination of polymorphic forms of drugs in the solid state.

Finally, NMR spectra can be obtained from living humans and animals and in vivoNMR or magnetic resonance spectroscopy (MRS) has found a use in disease diagnosis. The same technology and principles lie behind magnetic resonance imaging (MRI), which is now widely available in hospitals for clinical diagnosis.

The phenomenon of NMR arises because the positively charged nuclei of certain atoms possess a quantised property called spin, denoted by a spin quantum numberI. This spin is associated with a nuclear magnetic moment, also quantised, such that in a magnetic field it is possible for the nuclear magnetic moment to take up various orientations with respect to the field. Each orientation is associated with a discrete energy state and in the presence of the magnetic field these states have different energies. I can have an integer or half–integer value, including zero. For nuclei with I = 0, there is no magnetic moment, which is the case when both the atomic number and the atomic weight are even, such as ¹²C and ¹⁶O. In general, there are 2I + 1 energy states or levels and so for the simplest magnetic nuclei, with I = ½, there are just two levels. As a consequence of the differing energies of the states, and because of the Boltzmann distribution, the populations of spins in the states are not equal and an excess of nuclear spins will occur in the lower level. It is possible to induce transitions of nuclear spins between these levels by applying an oscillating frequency field; for commercially available NMR magnets, these transitions are in the radiofrequency region of the electromagnetic spectrum. There is a linear relationship between the magnitude of the nuclear magnetic moment and the observation frequency of the NMR phenomenon for a given magnetic field strength. There is also a linear relationship for a given nucleus between observation frequency and magnetic field strength. As is shown below, not all nuclei of a given atomic isotope have the same resonance frequency (known as the Larmor frequency), and hence the NMR phenomenon gives rise to a range of resonance frequencies that correspond to peaks in an NMR spectrum.

Many good textbooks describe the theory of NMR spectroscopy in more detail, a selection of which are listed in Further reading at the end of this chapter.

Important NMR-active nuclei

NMR spectroscopy is used mostly to identify molecular structures, usually as pure compounds, although, because it is such a high–resolution technique, it is possible to characterise individual components in chemical and biological mixtures. However, sometimes it is preferable to carry out simple solid–phase extractions before measuring NMR spectra, and it is also possible to couple high–performance liquid chromatography (HPLC) directly to NMR spectrometers (see below). One example of this use has been to characterise components in combinatorial chemistry syntheses. Of course, it is increasingly difficult to identify minor components (e.g. less than 5 mol %) in mixtures, in which case concentration by solid–phase extraction may improve detection. Nevertheless, it is possible to quantify lower levels of impurities given a clearly resolved NMR resonance for the impurity. Each of the two natural–abundance ¹³C satellite signals of the main component (ca. 0.55 mol %) provides a useful quantification standard. Thus, impurity detection limits depend on the resolution of a suitable NMR resonance, but a detection level of 0.1 mol % is not unreasonable.

Almost all spectrometers used for chemical structural studies are now based on a superconducting magnet with an operating frequency for ¹H observation that ranges from 200 to 900 MHz. Both resolution and detection are improved with increased magnetic field strength. For routine chemical studies in which compound availability is not limited, 200, 250, 270 and 300 MHz observation is quite usual. Increasingly, laboratories are using 400 MHz NMR spectroscopy as a standard for chemical studies, but 500 MHz and 600 MHz systems are now relatively common (in 2003). Higher observation frequencies are available (700, 750, 800 and 900 MHz) and are used for specialised research studies, mainly in the field of protein–structure investigation and for complex biological mixtures such as biofluids. Historically, and still generally, a sample for high–resolution NMR spectroscopy is placed in an expensive, precision–machined 5 mm glass tube sealed with a plastic cap, which is inserted into the detector (known as the probe) of an NMR spectrometer. Other tube diameters are also used, including 10 mm for larger samples and 3 mm when sample quantity is limited. More recently, robotic systems to inject samples directly into a flow probe using individual sample vials or 96–well plates have been developed.

This amplitude–against–time signal is treated as a series of decaying cosine waves and is converted into an amplitude–against–frequency spectrum by the mathematical process of Fourier transformation. Multiple scans are usually co–added to improve the signal–to–noise ratio (SNR), since this is proportional to the square root of the number of scans. The resultant FID is usually multiplied by some function to improve SNR or reduce the spectral line widths, and hence increase resolution. For example, it is possible to enhance SNR by multiplying the FID by a negative exponential, exp(–At), where t is the acquisition time and A is a positive constant, but this also has the effect of increasing line widths. Alternatively, it is possible to enhance resolution by multiplying by a positive exponential, exp(+At), but this has a large deleterious impact on SNR. In addition, if the acquisition time is too short, the FID is truncated before it has decayed into the noise, and artefacts are seen as oscillations on each side of the NMR peaks. A good compromise to overcome truncation artefacts and to improve resolution without too large a compromise on SNR is to carry out a Lorentzian–Gaussian transformation. This is achieved by multiplying by a positive exponential and a negative exponential squared, exp(+At – Bt²). The value of A is chosen to match the experimental line width and B is selected to give a Gaussian line shape of the desired reduced width. Since the delays in the electronics are frequency dependent, not all peaks will appear phased (upright). This has to be corrected either manually or automatically.

A number of standard system tests are used to define the performance of NMR spectrometers to meet the specifications guaranteed by the manufacturers.

A good line shape is an important requirement of modern NMR spectroscopy. The line–shape test is often called the ‘hump’ test and comprises an examination of the ¹H NMR peak from 10% chloroform in acetone–d₆, degassed and sealed. In addition to an acceptable value for the peak width at half height (50%), a good line shape is also defined by the width at the height of the ¹³C satellites (0.55%) and at one–fifth of this height (0.11%). NMR peaks should have a Lorentzian line shape and therefore the widths at 0.55% height and 0.11% height should be 13.5 and 30 times the width at 50% height; deviations from these values should be avoided. The spectrum is measured in both spinning and non–spinning conditions.

The standard resolution test for ¹H NMR spectroscopy is the measurement of the spectrum of a sample of 10% 1,2–dichlorobenzene in acetone–d₆, degassed and sealed. This molecule gives a complex second–order spectrum, but the eighth line from the left is used and the peak width at half height is measured at a spinning rate of 20 Hz. No weighting function is applied to the FID before Fourier transformation and a typical value for a 200 MHz spectrometer is around 0.07 Hz.

The sensitivity test for ¹H NMR spectroscopy is based on a degassed and sealed sample of 0.1% ethylbenzene in CDCl₃. A single 90° pulse is used with a spectral width of 10 ppm and a 1 Hz line–broadening function is applied to the FID. The region of the spectrum between δ3 and δ5 is expanded vertically, typically by a factor of 16, to show the full extent of the noise band, and this is compared to the height of the peak from the methylene group. The SNR is defined as the peak height multiplied by 2.5 divided by the peak–to–peak noise value in that specified region.

For quantitative measurements, pharmacopoeial requirements are to verify the repeatability of the integrator response using a 5% solution of ethylbenzene in CDCl₃. Five successive scans of the protons of the phenyl and ethyl groups are carried out and the mean of the values obtained. None of the values should differ by more than 2.5% from the mean.

There are two main tests of ¹³C NMR performance. The American Society for Testing and Materials (ASTM) approach uses a solution of 60% benzene–d₆ in 1,4–dioxane, degassed and sealed, and tests only the ¹³C performance. The other test uses a solution of 10% ethylbenzene in CDCl₃, degassed and sealed, and tests both the ¹³C sensitivity and the ¹H decoupling efficiency. For the ASTM test, a single 90° pulse is applied to the sample with the ¹H decoupler switched off. The FID is zero–filled to 65 536 (64k) points, weighted by a line broadening function of 3.5 Hz, and the full spectrum is plotted. The region between δ_C80 and δ_C120 is expanded vertically by a factor of four to allow good visualisation of the noise band, and this height is compared with that from the triplet signal from the C₆D₆. When using the ethylbenzene test sample, a single 90° pulse is applied to the sample, the FID is zero–filled to 65 536 (64k) points, a line broadening of 0.3 Hz is applied and the full spectrum is plotted. The noise height is taken, as before, from the region between δ_C80 and δ_C120 and compared to the peak height of the aromatic CH ¹³C NMR signals.

One–dimensional NMR spectroscopy gives rise to one frequency axis. As mentioned above, this is achieved by Fourier transformation of a FID obtained as the response after a radiofrequency pulse or pulse sequence. By incorporating a second incremented time period into an NMR pulse sequence, it is possible to carry out Fourier transformation with respect to both time periods to give two frequency axes that result in two–dimensional NMR spectra. This spreads out the NMR spectrum in a variety of ways, and so makes interpretation simpler. Extension to the use of three or more time periods leads to the concept of three–dimensional (or more) FTNMR spectroscopy.

All two–dimensional NMR experiments can be considered to have a preparation period, which might simply be a relaxation delay or could be some way of preparing the spin system for further manipulation. Next, the evolution period contains the first variable time period, t₁, which is incremented in equal time steps. Then it is possible to include a mixing period to allow the various spins to interact in some non–equilibrium condition. However, not all pulse sequences use this period. This is followed by the detection of the FID during the second time period, t₂.

To see how this works in practice, it is possible to consider one of the simplest two–dimensional pulse sequences. This is correlation spectroscopy (or COSY), which uses two radiofrequency pulses with an incrementable time delay between them. The time–domain data comprise a series of FIDs for a range of inter–pulse delays. After the first Fourier transformation, the data consist of an array of spectra on a frequency axis, each of which corresponds to a given inter–pulse delay, but with the phase of the peaks modulated as a result of the various values of t₁. After a second Fourier transformation, the data array comprises NMR intensity as a function of two frequency axes and is shown as a three–dimensional plot. This is quite difficult to interpret, especially if it is complicated by many signals. The data are visualised more easily as a contour plot ‘looking down’ on the peaks. In the COSY spectrum the peaks of the normal spectrum appear along the diagonal and the presence of off–diagonal cross–peaks indicates two nuclei that have a common indirect spin–spin (J) coupling. Hence, this experiment correlates nuclei that are spin–coupled together and are thus close together in bond terms. The frequency axes are labelled F1 and F2, deriving from t₁ and t₂, respectively.

There are many two–dimensional NMR pulse sequences in the literature, most of which are for very specialised applications and are used only by their inventors. However, a few very useful experiments are used widely for molecular structural studies. Two–dimensional NMR experiments can be classified into two types – resolved and correlation. The former resolve a particular type of NMR parameter into the second dimension. For high–resolution solution–state studies, this type is restricted largely to the J-resolved experiment (JRES), which separates the chemical shifts and spin–coupling multiplets into the two dimensions. This can be carried out in a homonuclear mode for ¹H chemical shifts resolved from ¹H–¹H couplings and in a heteronuclear mode in which, usually, ¹³C chemical shifts are resolved from ¹H–¹³C couplings.

Most two-dimensional NMR experiments are of the correlation type, which includes the COSY experiment that provides a correlation between ¹H chemical shifts that have a direct J coupling. The total correlation spectroscopy (TOCSY) pulse sequence provides a correlation as off–diagonal peaks, also based on the J coupling, but now along an unbroken chain of couplings. The nuclear Overhauser enhancement spectroscopy (NOESY) experiment provides a correlation between ¹H chemical shifts connected by an NOE. This experiment can also be used to investigate spin systems of molecules undergoing chemical exchange.

It is possible to obtain heteronuclear correlation two-dimensional NMR experiments, usually the correlation of ¹³C and ¹H chemical shifts, although the use of other nuclei such as ¹⁵N or ³¹P is possible. Direct detection of the heteronucleus is possible and historically was the main approach. However, this is usually a low–sensitivity experiment because of the lower inherent sensitivity, or natural abundance, or both, of most heteronuclei compared to ¹H. To overcome this problem, experiments have been devised that use indirect detection of the heteronuclei attached to the protons (see the next section on ¹³C NMR spectroscopy).

A summary of the principal useful two-dimensional NMR pulse sequences is given in Table 4. This also lists the typical length of time required for data acquisition and the quantity of material required for each experiment.

Table 4. Two–dimensional NMR experiments useful for small molecules

Directly observed ¹³C NMR spectra have now been surpassed largely by a technique that involves indirect ¹³C detection through ¹H NMR spectroscopy. This has come about through the advent of inverse–geometry NMR probes and the development of NMR pulse sequences that allow transfer of magnetisation between ¹³C and ¹H. In an inverse experiment, usually conducted in the two–dimensional mode, ¹H magnetisation is generated, and then transferred to ¹³C , the ¹H and ¹³C spins are allowed to interact for a variable time based upon the one–bond ¹H–¹³C coupling constant of about 125 to 160 Hz, and then the magnetisation is transferred back to ¹H for detection. The resultant two–dimensional NMR spectrum has ¹H chemical shifts on the detection axis (F2), ¹³C chemical shifts on the indirect axis (F1) and a peak at the intersection of the ¹H and ¹³C chemical shifts for CH, CH₂ and CH₃ groups. A number of pulse sequences achieve this correlation, the two best known being heteronuclear multiple quantum coherence (HMQC) and heteronuclear single quantum coherence (HSQC). Quaternary carbons do not appear in such a spectrum and have to be detected in a separate experiment based on a longer range ¹H–¹³C spin coupling, usually that over three bonds with a typical value of 6 to 10 Hz. For long–range coupling correlation, the experiment usually used is heteronuclear multiple bond correlation (HMBC). Full details of these methods are given Croasmun and Carlson (1994) especially, and in the books listed in Further reading. Such inverse–detected correlations have rendered the conventional ¹³C-detected two–dimensional methods such as HETCOR obsolete.

¹³C has a natural abundance of only around 1.1%, so ¹³C–¹³C correlation spectroscopy has to rely on the observation of only 0.01% of the sample, and at the same time suppress the ¹³C resonances of all singly labelled molecules that are 100 times larger. However, such a method does exist, given the rather ridiculous acronym INADEQUATE, and provides a correlation between carbons separated by one bond, the ¹³C–¹³C coupling constant in this case being around 40 Hz. Typical sample requirements for two–dimensional NMR studies that involve ¹³C are given in Table 4.

The various types of NMR spectra often used to confirm a molecular structure are illustrated by the example of the non–steroidal anti–inflammatory drug ibuprofen dissolved in dimethyl sulfoxide–d₆ (DMSO–d₆). Fig 1 shows the 600 MHz ¹H NMR spectrum of this substance; the molecular structure and numbering system is also shown, along with the assignments of the various NMR signals.

Thus the two doublets at δ0.85 and δ1.33 arise from the methyl groups with typical chemical shifts, that from CH₃(12,13) being of double intensity, with both showing spin coupling to a single proton. The resonance at δ1.80 comprises nine lines in a binomial intensity pattern because it arises from CH(11), which is coupled to eight adjacent hydrogens. The 1:3:3:1 quartet at δ3.61 is from CH(7). The doublet at δ2.40 is the resonance from CH₂(10) and arises from coupling to CH(11) exclusively. The small multiplet at δ2.49 is from the residual DMSO–d₅ in the DMSO–d₆ solvent and comprises a 1:2:3:2:1 quintet because the residual CHD₂ proton is coupled to two deuterium nuclei that have a spin quantum number of unity, each giving rise to a 1:1:1 splitting, and the two splittings overlap to give the resultant multiplet. Finally, the aromatic protons give rise to the multiplets at δ7.09 and δ7.17. These are not simple doublets, as can be seen by the partially resolved lines in the wings of the main peaks. These patterns arise as a consequence of the symmetry of the aromatic ring, with H2/H6 being chemically equivalent, but not magnetically equivalent (e.g. H2 will have a different spin coupling to H3 and H5). The broad peak near δ3.3 arises from the residual water in the sample and solvent.

The 600 MHz ¹H–¹H COSY spectrum of ibuprofen is shown in Fig 2. Dotted lines are used to indicate the off–diagonal peaks that connect two chemical shifts arising from spin–spin coupled protons. For example, the peak at δ0.85 from CH₃(12,13) is connected by an off–diagonal peak to the chemical shift at δ1.80 arising from CH(11), which in turn is connected to the peak at the chemical shift of δ2.40 corresponding to CH₂(10).

The 125 MHz ¹³C NMR spectrum of ibuprofen in DMSO–d₆ is given in Fig 3. Since all spin–spin couplings between protons and ¹³C are removed by decoupling, each non–equivalent carbon gives a single resonance peak. The assignments of the peaks are also given on Fig 3. Thus, the carboxyl carbon appears at 175.3 ppm and the two quaternary aromatic carbons are at 139.6 and 138.4 ppm; the assignments to C1 and C4, respectively, are based on a two–dimensional experiment, known as HMBC. This correlates ¹³C and ¹H peaks via long–range spin–spin couplings (see the section on ¹³C NMR spectroscopy). Next come the other aromatic carbons, C2/6 at 129.0 ppm and C3/C5 at 127.0 ppm. These peaks are more intense because each arises from two equivalent carbons, and also, because of the decoupling of the ¹H–¹³C spin–spin interactions, each carbon experiences an NOE that boosts the intensity of its NMR resonance. The resonances from CH(7) and CH₂(10) appear very close together, the assignments again being based on an HMBC spectrum. The 1:3:6:7:6:3:1 septet at 39.5 ppm arises from the carbons in the solvent DMSO–d₆ and the splitting is caused by ¹³C–²H spin–spin coupling, which is not affected by decoupling the ¹³C–¹H interactions. The resonance for CH(11) is at 29.6 ppm and the methyl carbons appear at 22.1 ppm [CH₃(12,13)] and 18.4 ppm [CH₃(8)], one being twice as large as the other, as expected.

To illustrate the use of indirectly detected ¹H–¹³C two–dimensional NMR spectroscopy, the ¹H–¹³C HSQC spectrum of ibuprofen is shown in Fig 4. This provides the connectivity between ¹³C nuclei and directly attached ¹H nuclei. Hence, given the partial assignment of both the ¹H and ¹³C NMR spectra based on inspection, it is often possible to complete the assignment of both spectra using this type of data. The two–dimensional contour plot shows a number of peaks, each of which occurs at the intersection of the ¹H and ¹³C NMR chemical shifts of a given CH, CH₂ or CH₃ group. Quaternary carbons do not appear in this spectrum because they have no directly attached hydrogens to provide the one–bond ¹H–¹³C spin–spin coupling used for the correlation. For ease of interpretation, the ¹H and ¹³C one–dimensional NMR spectra, assigned earlier in Figs 1 and 3, are plotted along the appropriate axes.

Quantification is possible using NMR spectroscopy. In this case, it is necessary to ensure that, for the summation of successive scans, the spin system is in equilibrium at the start of each scan. This is achieved by waiting five times the longest T₁ for the signals of interest between scans. This ensures that 99.3% of the signal is recovered. This is reasonable for ¹H NMR spectroscopy, in which T₁ values are usually less than 4 to 5 s, but for ¹³C NMR spectroscopy (see above), T₁ can be much longer. As shown for ¹³C NMR spectroscopy, the peak areas are, in general, not related quantitatively to concentration, and steps have to be taken to overcome this. This can be achieved by adding a relaxation agent and/or by modifying the pulse sequence to remove the NOE, which can affect signal intensity.

Quantification is carried out by measuring peak areas in the usual fashion, either by adding a known amount of a pure internal standard or by the method of standard addition. For a peak from an analyte with relative molecular mass M_rA, arising from N_A equivalent nuclei having an area I_A, and for a corresponding peak from an internal standard with relative molecular mass M_rS, from N_S nuclei with area I_S, the mass of the analyte in the sample is given by:

The two enantiomers of a chiral compound give identical NMR spectra in a non–chiral solvent because the magnetic environments of the nuclei are identical in both isomers. However, addition of a chiral paramagnetic shift reagent or use of a chiral solvent allows resolution of the NMR spectra of the two forms and hence integration of peak areas and measurement of the enantiomer proportions.

The identification of components in complex mixtures has been rendered much more efficient by the ability to couple HPLC directly with NMR spectroscopy. The on–line coupling of HPLC with MS has already proved to be of immense value, and the complementary structural capability of NMR spectroscopy has meant that complete molecular identification can now be very fast. Most applications to date have focused on the identification of drug metabolites in biofluids and similar matrices, but there are also applications in drug impurity analysis and drug degradation profiles. The general scheme for HPLC–NMR–MS is shown in Fig 5, which indicates that the typical configuration is to have the HPLC apparatus situated on the bench as normal. After the column and UV (possibly diode–array) detection the eluate is split, usually in the ratio of 95:5, to the NMR and mass spectrometers, respectively. The length of the capillary lines to these spectrometers can be adjusted so that either the NMR spectrometer or MS detects the eluate first. For example, the MS could be put first to identify characteristic ions, such as those relating to the isotope patterns of chlorine or bromine. Alternatively, the NMR spectrometer could be put first, for example set up to pick up well–characterised fingerprints, such as the presence of a ¹⁹F NMR signal.

Currently, five main options can be employed for HPLC–NMR, using either isocratic or gradient elution. The simplest of these is continuous–flow detection, in which NMR spectra are collected in real time as chromatographic peaks elute, but this is only practical when using ¹H or ¹⁹F NMR for detection, unless isotopically enriched compounds are available. Where continuous–flow NMR detection is used for gradient elution, the NMR resonance positions of the solvent peaks shift with the changing solvent composition and, for effective solvent suppression, these solvent resonance frequencies must be determined as the chromatographic run proceeds.

If there is a method to detect retention time on–line (UV, MS or radioactivity), stop–flow HPLC–NMR can be carried out using all the usual techniques available for high–resolution NMR spectroscopy. In particular, these include two–dimensional NMR experiments. In practice, it is possible to acquire NMR data on a number of peaks in a chromatogram using a series of stops during elution without on–column diffusion causing an unacceptable loss of chromatographic resolution.

There are two further categories of stop–flow experiment. First, and an increasingly popular approach, fractions eluted from the column can be stored in capillary loops for later off–line NMR study (‘peak picking’). Second, the flow can be halted at short intervals during the passage of the eluting peak through the NMR flow cell (‘time–slicing’), in a manner analogous to the use of a diode–array UV detector, to obtain spectra from various portions of the peak. This allows chromatographic peak purity to be estimated. Time slicing is most useful if the separation is poor, or if the compounds under study have weak or no UV chromophores, which makes it difficult to determine the retention times.

Fully automated analysis is now possible and, in this mode, automatic detection of UV peaks in the chromatogram based on predetermined time–windows or peak intensities is allowed under software control. The successful detection of each UV peak triggers the system to stop the flow at an appropriate time to isolate the peak in the NMR flow probe. This automatic NMR operation includes field homogeneity optimisation, setting and optimisation of all NMR acquisition parameters and the predefinition of the resultant SNR required in the spectrum. The measurement of two–dimensional NMR spectra can also be performed. With currently available commercial software, the automated run can be halted at any time with reversion to manual control if desired.

Usually, for reversed–phase HPLC separations, a gradient eluent system of acetonitrile and water is used. To reduce the solvent suppression demands, the aqueous phase is made up of D₂O, but to save on cost any organic solvents are used in their conventional protio forms. In some cases, it is preferable to use deuterated acetonitrile, for example, as this aids the detection of NMR peaks near the methyl peak of the solvent. The increase in cost is outweighed easily by the time saving in carrying out the study and the improvement in the chance of success. Apart from this, it is not usually necessary to make any compromises in a desired chromatographic procedure to accommodate the various types of HPLC–NMR experiment. The powerful structural elucidation capabilities of NMR spectroscopy often ensure that complete chromatographic separation is not necessary for full characterisation of the peak.

In the case of HPLC–NMR–MS experiments, there are some additional considerations. So far, the principal MS ionisation method used has been electrospray in either positive or negative ion mode (using either single quadrupole or ion–trap mass spectrometers), which places further constraints on the chromatographic solvent systems. When using HPLC–NMR the chromatography is often developed off–line from the NMR using non–deuterated solvents. It is not always simply a matter of replacing non–deuterated solvents with deuterated solvents to reproduce the chromatography for HPLC–NMR or HPLC–NMR–MS, as this can give rise to changes in retention times. For this reason it is standard practice to carry out an initial chromatographic run with a small injection volume (e.g. 10 μL) and then scale–up (e.g. 50 μL) for stop–flow NMR, once optimum conditions have been established. By this means it is often possible to acquire valuable MS data that can be used to guide the selection of peaks for study by NMR spectroscopy. It is also possible to mix the eluent just prior to the mass spectrometer with a non–deuterated solvent to back–exchange any deuterium atoms in exchangeable situations (e.g. NH and OH groups) for hydrogens. Comparison of the MS data with and without D₂O present enables the number of exchangeable hydrogens in any compound to be determined. In addition, MS can be used to search for particular diagnostic groups or fragments, such as an increase in m/z 16 for Phase I hydroxylated drug metabolites or an increase of m/z 196 for a glucuronide.

At present, a number of different mass spectrometers have been used in HPLC–NMR–MS studies. The simplest approach is to use a quadrupole mass spectrometer that gives integer mass values for ions. Employment of an ion–trap design allows fragmentation of the initial ions to be carried out in tandem experiments, such as MS–MS, MS–MS–MS (i.e. MSⁿ), etc. The use of time–of–flight or FT mass spectrometers provides very high resolution that allows good predictions of empirical molecular formulae from accurate ion masses, including those from MSⁿ experiments. Finally, some studies using inductively coupled plasma MS have been carried out. This technique provides no molecular information, as the sample is reduced to atomic form in the plasma, but it yields atom–specific chromatographic detection, which can be valuable for identifying the retention times of molecules of interest, such as those that contain heavy atoms or halogens such as chlorine and bromine.

By far the largest body of work to date using HPLC–NMR and HPLC–NMR–MS is in the field of drug metabolism, in which the methodology has been used extensively to identify metabolites in studies from clinical trials that involve human subjects, in the investigation of model drugs in animals in vivo and also in in vitro systems, such as liver microsome incubations. Applications to compounds used clinically include paracetamol, ibuprofen, flurbiprofen, naproxen, antipyrine, tolfenamic acid, and studies on many compounds currently in development in which early information on metabolic products can be used to aid drug design.

Many marketed drugs are either natural products or are modifications of such substances. Hence considerable effort has been spent in isolating and characterising chemicals from natural sources so that they can be tested in a variety of biological screens. Often it is necessary to carry out laborious extraction and purification steps, so the advent of directly coupled HPLC–NMR–MS has been explored as an alternative technique for natural product identification. The use of HPLC–NMR, and other combined techniques such as HPLC–MS–MS, to identify natural products from plant sources has been reviewed by Wolfender et al. (1998; also see Lindon et al. 2000).

NMR spectroscopy has been incorporated into a number of tests in the 2002 edition of the British Pharmacopoeia. These include the assay for buserelin, a peptide analogue [5–oxoPro-His-Trp-Ser-Tyr-D-Ser- (tert-Bu)-Leu-Arg-Pro-NHEt]. In this case the ¹H NMR spectrum of the sample is compared visually to that of a reference NMR spectrum from the European Pharmacopoeia and must be ‘qualitatively similar’. In another approach, the test for low molecular weight heparins relies on the use of a 75 MHz ¹³C NMR spectrum, which is compared with that from a reference sample of the material. A similar approach is used to identify the antibiotic tobramycin, but the nucleus of observation is not specified.

Poloxamer is a synthetic block copolymer of ethylene oxide and propylene oxide and thus contains oxyethylene units (–O–CH₂–CH₂–O–) and oxypropylene units (–O–CH.CH₃–CH₂–O–) units. The characterisation of poloxamer includes the use  of ¹H NMR spectroscopy to determine the ratio of the two types of sub–unit, based on integration of the ¹H NMR resonances at δ1.08 from the methyl group of the oxypropylene units and at δ3.2 from methylene groups from both types of sub–unit.

High resolution one- and multidimensional NMR spectra measured on biological samples, such as biofluids, can be extremely complex, and contain many thousands of resonances. In many cases, visual inspection of such spectra releases only a small percentage of the information available in the data. For this reason there has been a move towards the use of computer–based methods to extract the maximum information from such complex spectra. Although statistical methods have been applied for many years to NMR data, the first truly concerted application of multivariate statistics to high–resolution NMR spectra in the biomedical area was the classification of ¹H NMR spectra from rat urine samples according to the type of organ toxin administered to the animals. This provided a fast non–invasive way to assess candidate drug toxicity (see Lindon et al. 2000).

In the post–genomic era, there is much emphasis on the study of the consequences of changes in gene expression (functional genomics), by examination of either the protein complement (proteomics) or the small molecule metabolite composition of a biological system (metabonomics) (Nicholson et al. 2002). The term metabonomics has been defined as ‘the quantitative measurement of the dynamic multiparametric metabolic response of living systems to pathophysiological stimuli or genetic modification’, to describe the study of the dynamic and time–dependent profile of metabolites within an organism and how they are altered by some biological process. Currently, the only effective way to extract meaningful metabonomic information is to combine high–resolution NMR spectroscopy of biofluids with multivariate statistics methods, often known as pattern recognition (PR) Lindon et al. 2001.

Genetic, disease or foreign compound effects can result, in some cases, in changes in gene expression and/or changes in protein production, but all will, in principle, cause alterations to biochemical regulation and control. Thus, all drug–induced and other pathophysiological perturbations result at least in disturbances in the ratios and concentrations, binding or fluxes of endogenous biochemicals. In body fluids, metabolites are in dynamic equilibrium with those inside cells and tissues, and consequently abnormal cellular processes in tissues of the whole organism after a toxic insult or during a disease process are reflected in altered biofluid compositions.

In all cases the problem involves the detection and identification of low amounts of analytes in a very complex matrix with many potential interferences, all obscured by inter–subject biological variation. High–resolution ¹H NMR spectroscopy is particularly appropriate for investigating abnormal body fluid compositions (such as blood plasma, urine or bile), as a wide range of metabolites can be quantified simultaneously with no sample preparation and without preselection.

To illustrate the complexity of biofluid NMR spectra, Fig 6 shows the 800 MHz ¹H NMR spectrum of a control human urine sample, with successive horizontal and vertical expansions. It has been shown that each biofluid has its own characteristic fingerprint. NMR spectroscopy may also be used effectively to screen for abnormal metabolite profiles in tissue extracts or cell suspensions. More recently, it has been shown that the same approach can be used to investigate the metabolic composition of intact tissues using a technique called high–resolution magic angle spinning (MAS) ¹H NMR spectroscopy.

Computer–based PR and expert system approaches have to be used to interpret the NMR data on biofluids and tissues obtained in various experimental toxicity or disease states. Often these include the use of principal components analysis (PCA) to classify the samples according to their NMR spectral profiles. Each spectrum can be reduced to a set of intensity–based numbers that are characteristic of that spectrum, and these descriptors are used as input for the PCA. Principal components (PCs) are new variables created from linear combinations of the starting spectral descriptors, such that each PC is uncorrelated with all other PCs, and the first PC contains the largest part of the variance of the data set, with subsequent PCs containing correspondingly smaller amounts of variance. Thus, a plot of the first two or three PCs gives the ‘best’ representation in terms of variation in the data set in two or three dimensions. Such PC maps can be used to visualise inherent clustering behaviour of samples, for example marked according to toxic mechanism. The position of a NMR spectrum from a biofluid or tissue sample on a map is determined purely by its metabolic profile, as opposed to any other independent knowledge of the class of the sample. In this simple metabonomic approach, a urine sample from an animal treated with, for example, a compound of toxic effect is mapped against a database of NMR spectra of samples from animals in a range of toxic classes and the proximity of the sample to other types of sample is determined.

However, in the real world, biochemical data are more complex as effects caused by genetic modification, disease, nutrition, lifestyle, drug therapy or drug toxicity develop and recover in real time, and hence there can be complex, time–related changes in the NMR spectra of biofluids. Also, it is more rigorous to compare the NMR-derived metabolic profile in the original multidimensional NMR metabonomic descriptor space. Hence, to develop automatic classification methods, it is necessary to develop more complex expert systems. Here, a ‘training set’ of NMR metabonomic data is used to construct a mathematical model that predicts correctly the class of each sample. This training set is tested with independent data (a ‘validation set’) to determine the robustness of the computer–based model. These models can be based on a range of different mathematical procedures, such as PCs, artificial neural networks and rule induction, and several types can be embedded in a single expert system. In general, the methods allow the quantitative description of the statistical boundaries that characterise each class of sample in terms of their metabolic profiles. Some methods also allow a level of probability to be placed on the classification and can distinguish samples that belong to none of the classes in the training set. Using such systems, a sample can be classified as belonging to a single class, to multiple classes or to no class. By building an exhaustive series of models it is possible to use such methods to provide classification probabilities for a wide range of sample types.