1. Metabolite Imaging, In Vivo Nuclear Magnetic Resonance spectroscopy (NMR), and Clinical NMR
Different medical conditions present with distinct metabolic activities and metabolic abnormalities. For decades, metabolic alterations have been detectable using ex vivo medical NMR studies and in vivo magnetic resonance spectroscopy (MRS). The utility of MRS has led to its implementation in various clinical fields ranging from oncology to neurology. However, early clinical MRS and MRS imaging (MRSI) were limited because of low spectral resolution. This was due to the low magnetic field strength and relatively low field homogeneity of clinical MR scanners compared to ex vivo NMR instruments. Attempts to overcome these challenges have focused on improving both imaging hardware and data processing software. For instance, to enhance MR signal detection hardware, several specially designed surface coils, such as endorectal coils for prostate cancer or endovaginal coils for cervical cancer, have been developed
[1][2]. Data processing software improvements have focused on enhancing the way that MRS images can be displayed. In contrast to univariate or intensity-based imaging data (such as X-rays or CT scans), MRS and MRSI data are multivariate (i.e., all measurable metabolites) and cannot be readily interpreted through simple visual evaluations. Instead, MRS and MRSI data interpretation must rely on computer assistance, artificial intelligence (AI) or machine learning. Several analytical software packages have been developed to analyze clinical MRS data and visualize their clinical implications. These include the widely used LCModel
[3] and jMRUI
[4] programs, which can automatically identify and quantify metabolites contributing to the signals seen in MRS and MRSI spectra. These software tools are very similar in concept to other software packages (such as MagMet, Bayesil and B.I. QUANT) used for automated ex vivo NMR metabolomics. Continuing developments in MR scanner technologies, including higher magnetic field strengths and improved coil array designs, have significantly increased our ability to generate MRS images with greater spatial resolution
[5][6][7][8], as well as investigations of tissue cellular microstructures through diffusion-weighted MR spectroscopy
[9]. Thanks to these improvements, MRS and MRSI are now offering metabolomics researchers and clinicians the ability to monitor detailed metabolic changes at high spatial resolution with good sensitivity in real time, in living organisms or in live patients. For example, NMR has been applied for identifying inborn errors of metabolism in clinical settings
[10][11]. No other metabolomics technology (not LC-MS or GC-MS) offers this kind of chemical window on living systems. However, the use of MRS and MRSI in metabolomics studies has been remarkably light, and its promise remains largely unfulfilled. More widespread adoption of MRS and MRSI by the NMR metabolomics community will be key to bringing this technology into the mainstream of metabolomics studies.
2. Lipoprotein Profiling and NMR
Another important advance in NMR-based metabolomics has been the development of automated tools for lipid and lipoprotein particle (LDL, HDL, VLDL and so on) analysis
[12][13][14][15]. Lipoprotein particles are the metabolic by-products of cholesterol metabolism and consist both of proteins and cholesterol-containing lipids. As such, lipoprotein profiling is an important field of lipid metabolism and metabolomics. The first techniques for NMR-based lipoprotein analysis in serum/plasma were described in 1991 by Jim Otvos and colleagues
[16]. Otvos showed how 1D
1H NMR spectra could be rapidly and automatically deconvoluted to identify lipoprotein components and extract accurate lipoprotein concentrations. This led to the creation of LipoScience Inc. in 1994. LipoScience was initially dedicated to performing research on NMR-based cholesterol testing using plasma and serum samples. In 2005, a successful FDA-approved LDL test called LipoProfile was created
[17]. The LipoProfile NMR test is highly automated and provides 11 different measures of lipoprotein concentrations and sizes from plasma samples. In 2014, LipoScience was acquired by Labcorp, which now uses the same technology to offer comprehensive NMR-based lipoprotein profiling in many lab centers across the US and Canada. NMR-based lipoprotein profiling has become the “gold-standard” for lipoprotein measurement by clinicians because of its speed, accuracy, and the number of possible measurements
[18][19]. Indeed, NMR-based lipoprotein profiling represents one of the more successful examples of metabolomics being translated into clinical practice.
Due to the success of LipoScience, several other companies, including Bruker (IVDr Lipoprotein Subclass Analysis (B.I. Lisa)) and Nightingale, have begun to offer lipoprotein analyses of serum and plasma samples through an automated NMR spectral data collection and fitting process
[20][21]. These metabolomics profiling techniques are now available as either an on-site subscription-based service or an off-site clinical service. Both Bruker and Nightingale use spectral deconvolution or spectral fitting concepts such as the technique pioneered by LipoScience but offer more identified features or parameters. Bruker’s B.I. Lisa can measure 112 lipoprotein parameters
[22], while Nightingale’s service reports on 228 lipoprotein parameters, including 20–30 small molecules
[23]. Nightingale’s automated NMR pipeline is fast and inexpensive, which allows metabolomics to be performed at a scale unmatched by LC-MS, GC-MS, or CE-MS platforms. Indeed, Nightingale recently used its NMR platform to analyze >120,000 samples from the UK BioBank. Many large biobanks and research organizations are now turning to companies like Bruker and Nightingale to analyze tens of thousands of samples because of the high throughput, low cost, and broad metabolite coverage. The successful commercialization of NMR-based metabolomics pipelines demonstrates the tremendous potential that NMR offers for future high-throughput metabolomics, lipidomics, and lipoprotein profiling. Based on the growth in medical testing and diagnostics, it is likely that NMR-based lipoprotein profiling will soon represent the majority of samples processed by the entire metabolomics community.
3. Fluxomics and In Situ NMR
Metabolomics routinely relies on an endpoint or final state measurement of a metabolic profile. Instead, fluxomics studies the dynamic and temporal process of metabolite changes, metabolic reactions or metabolic fluxes
[24]. As a result, metabolic reaction rates may be calculated from these measured fluxes. An important advantage of NMR for a fluxomics study is the fact that sample pre-preparation is not needed. This allows one to rapidly measure metabolic reactions in situ and on a real-time scale. NMR fluxomics studies have been conducted for decades through perfused measurements of animal organs using the injection of stable NMR-active isotopes (i.e.,
13C- labeled compounds). Through these perfusion studies, a time-series of
1H,
13C, or
31P NMR spectra can be recorded, from which the intensities of the originally injected or perfused compounds and their reaction products can be quantified. The time-dependent series of peak intensities can then be used to produce reaction rates for all measurable and active metabolic pathways.
Fluxomics studies have been used in a wide range of pre-clinical and clinical metabolomics studies, including many involved in cancer
[25][26][27][28]. Recent technology developments have further enhanced the use of fluxomics
[29] by combining isotope labeling with hyperpolarized compounds. The use of hyperpolarized compounds and hyperpolarizing agents has resulted in upwards of a 1000× enhancement in NMR signals for certain compounds. Developments in high-resolution magic angle spinning (HRMAS) methods have also allowed for the mechanistic, real-time probing of cell-line metabolomics by similarly measuring isotope-labeled reactions
[30]. This method has shown superior results in the fluxomic evaluation of aerobic metabolic pathways and in differentiating between intra- and extra-cellular metabolites.
While fluxomics studies are increasingly being performed using LC-MS methods coupled to isotopic perfusion techniques, it is important to remember that NMR-based fluxomics has a key advantage over LC-MS fluxomics. NMR-based fluxomics can exploit the ability of NMR to easily localize the exact position of a given (labeled) atom in a specific molecule. The spatial localization of specific
13C or
2H isotopes incorporated within a specific metabolite can provide a clear indication of the enzymes or pathways used to generate that metabolite
[31][32]. Isotope labeling allows NMR-based fluxomics to easily link metabolites to proteins and pathways. In other words, NMR-based fluxomics offers a route to a complete, system-wide view of metabolism that is not achievable by almost any other method. Given the many strengths offered by NMR-based fluxomics, more widespread adoption of this approach by the NMR metabolomics community could lead to a closer link between metabolomics and systems biology.
4. Intact Tissue Metabolomics with HRMAS
NMR has a distinct advantage over techniques such as LC-MS for measuring metabolites in intact tissue. This is because LC-MS requires that one extract, homogenize and destroy tissues to measure their chemical composition. As a result, the tissue cannot be re-used or re-analyzed via microscopy by a pathologist. In contrast, living tissues or live biospecimens can be analyzed intact by NMR, with no need for extraction or homogenization. Indeed, NMR metabolomics studies of intact, living tissues have been conducted for many decades. However, the quality of NMR spectra collected from intact tissues tended to be quite poor, with relatively low spectral resolution and poor signal intensity. This was primarily due to tissue matrix effects leading to inhomogeneous signals and excessive line-broadening. Fortunately, these issues were resolved by the application of HRMAS to the analysis of intact tissues
[33][34]. HRMAS involves the spinning (6 to 10 kHz) of a sample at the magic angle (54.7°) to eliminate anisotropies and reduce the line-broadening effects arising from residual dipolar interactions and magnetic susceptibility variations.
High-resolution NMR spectra comparable to those measured from aqueous solutions can be obtained from biological tissues without any pre-treatment. Furthermore, HRMAS does not destroy tissue architectures or alter the spatial location of metabolites. Thus, microscope-based pathological evaluations can be conducted on the same specimens after the HRMAS measurements have been completed. This unique capability of HRMAS was critical to the development of intact tissue metabolomics and to its adoption in several clinical settings and studies
[35][36]. For instance, the presence and quantity of cancer lesions within a tissue sample analyzed by HRMAS would be unknown without a subsequent pathological evaluation of the tissues obtained from the suspected cancer patient. Accordingly, the conclusions drawn from HRMAS metabolomics studies can be clearly correlated with specific tissue pathologies. A further advantage of HRMAS NMR is its signal enhancement that allows for clinically informative metabolomics datasets to be measured on small tissue samples (<10 mg)
[34], or a minute amount (<10 mL) of scarce human biofluid
[37]. To better preserve tissue pathological architectures, various slow HRMAS methods have been proposed to ensure an accurate correlation between the metabolomics investigation and the disease pathology
[34][38]. The ability of HRMAS NMR to characterize the metabolome of intact tissues non-destructively and quantitatively, coupled with its amenability to a post-analysis pathological or microscopic examination, makes HRMAS NMR an ideal tool for clinical metabolomics (especially biopsies) and metabolically guided anatomical studies. While both are still emerging areas of metabolomics, NMR and specifically HRMAS NMR are ideally suited to address these tasks.
5. NMR Techniques for Fast Data Acquisition
NMR-based metabolomics studies commonly rely on 1D
1H NMR spectral data that can be rapidly acquired in a few minutes with maximal signal to noise. However, 1D
1H NMR spectra tend to suffer from large solvent signals that obscure relevant peaks. They may also be affected by background signals arising from large biomolecules, as well as poor resolution and peak overlap due to a combination of limited spectral resolution and peak splitting from J-coupling. Several NMR pulse sequences have been developed to address each of these issues. For example, the first increment of a 2D nuclear Overhauser effect spectroscopy (NOESY) pulse sequence with pre-saturation (i.e., 1D pr-NOESY), or pulse sequences that employ excitation sculpting or the PURGE pulse sequence, all provide efficient water suppression
[39][40][41][42]. Similarly, the Carr-Purcell-Meiboom-Gill (CPMG) or PROJECT pulse sequence can efficiently remove background signals resulting from protein contamination using a T
2 filter that relies on the large molecular-weight difference between small metabolites and large biomolecules
[39][41][43]. A diffusion ordered spectroscopy (DOSY) edited pulse sequence can achieve a similar outcome using molecular weight-dependent differences in diffusion coefficients
[44]. Of course, protein precipitation or protein filtering techniques may be a preferred alternative to removing protein contamination instead of relying on NMR pulse sequences
[45]. The complexity of a 1D
1H NMR spectrum can be reduced by using isotopically (
13C,
15N or
2H) labeled tracers or detecting alternative nuclei such as
31P
[46]. Similarly, distributing the peaks into two dimensions can also reduce the spectral overlap and complexity. However, these 2D NMR approaches tend to result in substantially longer acquisition times (hours instead of minutes).
Fortunately, several recent discoveries and developments have occurred that can substantially reduce 2D acquisition times. For example, the use of non-uniform sampling (NUS) enables a more efficient acquisition of high-resolution 2D NMR spectra with significantly shorter experimental times
[47][48]. Instead of collecting the entire data matrix for a 2D NMR spectrum, NUS sub-samples only a fraction of the matrix, leading to a sparse data set. The resulting sparsity, usually 25–50%, directly determines the reduction in acquisition time. Other advancements in pulse sequences have led to further improvements in resolution and sensitivity
[49]. For example, NUS and “pure shift” methods (see below) can be combined to yield an increased resolution along both dimensions in 2D experiments while still obtaining faster acquisition times. This is achievable because pure shift methods work independently of the chosen NUS schedule. In addition, several new pulse sequences have emerged for rapid acquisition of 2D NMR spectra, especially for
13C and
15N labeled samples (e.g., ASAP-HSQC, ALSOFAST-HSQC, CLIP-ASAP-HSQC, ASAP-/ALSOFAST-HSQC)
[50]. NUS can be combined with these methods to achieve a further reduction in acquisition times with practically no loss in resolution and sensitivity
[51]. Furthermore, Kupče et al. recently introduced the NMR by ordered acquisition (NOAH) super sequence using
1H-detection. NOAH combines two pulse sequence modules, ZZ-heteronuclear multiple bond correlation (HMBC) and ASAP-COSY, with multiplicity-edited HSQC and NOESY to obtain multiple NMR spectra from a single experimental measurement
[42]. In essence, two or more NMR pulse sequences are interleaved and simultaneously acquired during the acquisition time of a single experiment. NMR, like numerous other analytical techniques, is highly dependent on state-of-the-art computers for data processing, analysis, and storage. The use of graphics processing units (GPUs) to advance and accelerate the application of artificial intelligence to challenging NMR problems is expected to transform NMR data processing. For example, a recent proof of principle application of deep neural networks has shown great promise in the reconstruction and processing of multi-dimensional NMR spectra acquired with NUS while avoiding artifacts and distorted peak shapes and positions
[52].
Other approaches have also emerged to improve resolution or shorten acquisition times for 1D NMR. In most NMR spectra, a significant reduction in resolution occurs due to the splitting of signals into multiplets resulting from J-coupling. “Pure shift” NMR spectroscopy is a broadband decoupling method that can be used to significantly enhance the resolution and sensitivity of an NMR spectrum by removing these splitting patterns
[53][54][55]. Broadband homonuclear decoupling methods reduce multiplets to singlets through the removal of
1H-
1H J-coupling, thereby reducing peak crowding, and improving resolution. These pure shift methods have been applied to both 1D
1H NMR spectra and to the indirect
1H dimension in 2D NMR experiments. Although the application of pure shift NMR, NUS and SOFAST methods to NMR metabolomics has been relatively minimal, the potential two-to-three-fold enhancement in signal sensitivity (via pure shift NMR) or the up to 10-fold faster data collection time (via SOFAST, ASAP or NUS methods) suggests that these methods should be routinely employed by the NMR metabolomics community.
6. Hardware Sensitivity Enhancement
Recent advances in NMR instrumentation have been aimed at lowering the traditional barriers to purchasing or using NMR spectrometers. In addition to developing turnkey instrumentation designed to streamline data collection and processing, the development of benchtop NMR spectrometers built using permanent magnets has led to a wealth of new applications that would have been logistically challenging with conventional NMR
[24]. Benchtop NMR instruments (with operating frequencies ranging from 40 to 80 MHz) are less expensive and more compact and can be installed and employed at locations where NMR spectroscopy has not been practical due to physical or financial constraints. Additionally, the lower cost of operation for benchtop NMR spectrometers, which are built around permanent magnets as opposed to cryogenically cooled magnets, will enable NMR metabolomics to enter new underserved arenas that are inaccessible with current NMR technologies.
7. Databases and Software for Compound Identification
NMR continues to be the gold standard for chemical identification, and most chemistry and natural products journals require evidence of a new chemical’s presumptive structure with NMR spectral data showing its atomic or molecular connectivity. Many fundamental tools relating to NMR data processing and analysis for chemistry have been available through NMRbox
[56]. For structure elucidation, NMR has historically required substantial interpretive expertise; however, many new tools are evolving that enable a larger user base to easily obtain metabolite identification. Several substantial NMR databases are now available that allow searching of 1D and 2D NMR data for individual metabolites. These include the Human Metabolome Database (HMDB)
[57], the Madison-Qingdao Metabolomics Consortium Database (MQMCD), the Biological Magnetic Resonance Bank (BMRB)
[58][59], and more recently, the Natural Product Magnetic Resonance Database (NP-MRD)
[60]. All are encouraged to contribute to these valuable community resources to help expand their utility. In addition, algorithms have been developed to create molecular networks from 2D NMR data, allowing chemical annotation from NMR data to extend into unknown, structurally similar metabolites
[58][61]. Software tools with spectral databases are also facilitating the identification of both polar and non-polar metabolites in mixtures via both 1D NMR (e.g., Chenomx NMR Suite, B.I. QUANT, MagMet, Bayesil) and 2D NMR (e.g., COLMAR)
[62][63]. Although the quantity of data contained in today’s NMR databases for authentic compounds is growing, these spectral databases are expected to remain far from comprehensive in the foreseeable future. Fortunately, great strides are being made in the large-scale prediction of NMR spectra using quantum mechanical (QM) principles and machine learning (ML)
[64]. Indeed, the NP-MRD is one of the first examples of an NMR database to contain tens of thousands of predicted NMR spectra of known compounds derived from state-of-the-art QM and ML techniques. In the near future, it is expected that these approaches will help fill the void of authentic data. However, QM and ML methods still require more authentic (experimentally acquired) data for further validation and for improving prediction accuracy. Thus, data contributions are again encouraged from the public with an expected return on investment by expanding the accuracy and completeness of the predicted data content.