Nuclear Magnetic Resonance and Artificial Intelligence

Abstract

This review explores the current applications of artificial intelligence (AI) in nuclear magnetic resonance (NMR) spectroscopy, with a particular emphasis on small molecule chemistry. Applications of AI techniques, especially machine learning (ML) and deep learning (DL) in the areas of shift prediction, spectral simulations, spectral processing, structure elucidation, mixture analysis, and metabolomics, are demonstrated. The review also shows where progress is limited.

1. Introduction

NMR spectroscopy is indispensable for the identification and structural elucidation of small compounds (e.g., metabolites). The resonance frequencies of each spin (mainly ¹³C and ¹H), also known as chemical shifts, together with their multiplicities, provide a unique fingerprint for different chemical environments within a molecule. By analyzing these shifts and peak profiles, analysts can deduce the connectivity, arrangement, and electronic environments of atoms in a molecule, making NMR an essential tool for determining the precise structure of organic and inorganic compounds. NMR spectroscopy also plays a crucial role in studying molecular dynamics and interactions. The technique can probe how molecules behave in different environments, offering insights into conformational changes, reaction mechanisms, and molecular interactions. This makes NMR invaluable not only for structural chemistry but also for understanding complex biological systems and metabolic pathways. In metabolomics, NMR is used to identify and quantify metabolites in biological samples, providing a comprehensive overview of metabolic processes [1]. By analyzing the NMR spectra of biofluids, tissues, or cells, researchers can gain insights into the metabolic state of an organism, detect biomarkers for diseases, and study the effects of drugs and other interventions [2].

Integrating artificial intelligence (AI) into NMR spectroscopy has been revolutionizing the field, enhancing the accuracy, efficiency, and scope of analyses. This review outlines this progress. Most techniques mentioned here can be classified as machine learning (ML), so we mostly use the terms AI and ML interchangeably. Deep learning (DL) is a subset of newer ML techniques. We do not delve into details of the definition of artificial intelligence, but comprise everything here which is helping to replace human expertise and input. The aim of this is to transform data handling and interpretation, enabling more complex and large-scale studies automatically. This review focuses on advancements in small molecule chemistry and the potential for future developments. Apart from the references given in the text, we point the reader to two recent special issues on the topic [3,4] and the reviews [5,6,7,8] for more materials and references. As part of this overview paper, we do not claim to cover the extensive literature exhaustively.

2. Databases and Data Standards

Large datasets are essential for training robust models capable of making accurate predictions. High-quality data that cover a broad range of chemical structures and experimental conditions enhance the model’s ability to generalize and perform well on new simulations. This is demonstrated, for example, in [9]. Correct data handling and storage practices are critical to achieving better accuracy in machine learning applications. In practice, data collections are not easily available for many problems in chemistry [10]. Apart from the simple lack of resources, available data are often difficult to access. A more formal definition of good practice is the FAIR (findable, accessible, interoperable, and reusable) data principles [11], which ensure that data are well-organized, easily accessible, and reusable by other researchers. Efficient data management systems that implement these principles can facilitate the integration of diverse datasets, enabling comprehensive analyses and more accurate predictions.

For NMR, the number of available resources has recently increased. For assigned NMR shifts, sdbs (https://sdbs.db.aist.go.jp/ (All websites in this paragraph accessed on 17 October 2024)) and nmrshiftdb2 (http://nmrshiftdb.org/), which recently celebrated its 20th anniversary [12], still seem to be the largest resources. A newly available resource is NP-MRD (https://np-mrd.org/), which holds experimental and (mostly) calculated NMR data for nearly 41,000 natural products. A bulk download is possible, but we did not manage to actually extract the data from it, mainly due to inconsistencies in the format and between files. NAPROC-13 (http://c13.usal.es) holds about 6000 natural products. nmrXiv (https://nmrxiv.org/) is part of the NFDI (Nationale Forschungsdateninfrastruktur) initiative by DFG (Deutsche Forschungsgemeinschaft) from Germany (https://www.nfdi.de/?lang=en) and is part of a landscape of research data management tools. Currently, the data in nmrXiv is very limited. Workflow4Metabolomics [13] is a platform for metabolomics workflows including NMR. It enables the saving and publication of analyses, going beyond pure analytical data.

Data standards are another important topic for machine learning, since only data that can be processed automatically are useful as training data. For example, the difficulties involved in reading the NP-MRD data [14] show the importance of data formats. In order to have fully available data, a record should contain raw data and extracted data (peaks), a full metadata record, and a complete set of spectra for a sample. The sample could be a compound, but also any other measurable substance. An NMReDATA record can hold all that information, raw data from the vendor, and annotated structure spectra in an NMReDATA file [15]. Other formats include JCAMP-DX [16] and CMLSpect [17], which hold extracted spectral data. nmrML [18], on the other hand, is focused on raw data. A consensus on which (meta)data should be reported would be important for collecting comparable data. Unfortunately, this does not exist so far. As part of NFDI4Chem, Minimum Information Metadata Standards for Chemical Investigations (MIChIs) have been developed, including NMR [19].

An alternative to using data collections is the use of artificial or computer-generated data. This includes spectra calculated using density functional theory (DFT), but also other tools or methods, depending on their suitability for the specific purpose. In any case, care needs to be taken to avoid the introduction of systematic errors from insufficiencies in the data generation.

3. Chemical Shift Prediction

One of the oldest applications of AI in conjunction with NMR is in shift prediction for compounds. This is typically used to compare measured shifts to the predicted shifts of potential structures of unknown compounds. Such comparisons can be carried out manually as part of a structure elucidation process, but can also be used during computer-aided structure elucidation (CASE) processes (see Section 6 for details). A recent example is [20], in which an optimization process guided by the similarity of the generated structures’ predicted spectra with the measured spectra was used to find good hits. Clearly, the prediction quality is a crucial factor for such applications. Quality here means primarily the closeness of the predicted shifts to the real shifts of these compounds, if they could be measured.

Therefore, a competition has been ongoing to improve prediction methods. In [21], a long term comparison was carried out between prediction quality for ¹H and ¹³C NMR. We continue this comparison in Table 1, which presents work undertaken since the publication of [21]. Figure 1 shows the results, including those from [21], in a single chart. This only contains results for small molecule solution NMR. We have also made an online version of this table and chart, which will be updated in the future. This is available at the following link: https://nmrshiftdb2.sourceforge.io/predictionhistory/history.html (accessed on 17 October 2024).

As before, the comparison is not a precise like-for-like comparison. In particular, the underlying datasets and the evaluation methods are not necessarily the same. See [21] for a thorough discussion. We still believe this is a valuable comparison.

Methods for NMR prediction can be divided into either ab initio or data-driven methods.Ab initio methods, including DFT, do not use any shift data, but calculate shifts directly from quantum mechanical models. They are known for accuracy, but are still very demanding in terms of computing power and are slow when used with common hardware. Data-driven methods, on the other hand, rely on collections of known shifts and try to use these to infer values for other compounds. These methods, including machine-learning approaches, can be very fast with standard hardware, but struggle to reach the accuracy of ab initio methods, which themselves are not AI methods, but can serve as a reference. Data-driven methods can be divided into three broad groups, namely increment-based, HOSE (Hierarchically Ordered Spherical Environment) code, and machine learning. Machine learning can be either older techniques or recently invented deep learning methods. For graph structures like molecules, graph convolutional networks and message passing are fundamental techniques. For details, see [21].

Looking at the figures for ¹³C predictions, it can be seen that recent machine learning predictions are close to what is possible using DFT. It should be noted that the prediction in 8/2002 achieving 1.0 ppm was for a single compound. Recent results using deep learning methods have achieved errors below 1 ppm. For ¹H, ref. [22] clearly achieved a historically best result with 0.1 ppm mean average error (MAE). This uses a single-solvent dataset, which up to this point has not been available. Since solvents matter significantly for ¹H predictions, it cannot be distinguished how much the result is due to the new technique or the new dataset. A very good result was also reported by [23], using a fragment-based approach. This is contrary to the recent trend towards graph neural network (GNN) approaches. In [21], it was concluded that, since the least squares linear regression for ¹H was almost horizontal, real progress was not visible. Since the least squares linear regression line is now clearly falling, this assessment has to be revised. A clear trend of improved results now is visible also for ¹H prediction.

All the above mentioned papers use a training set of a certain size (normally as big as possible). Judging a model on this alone was shown in [9,24] to give an incomplete picture. They suggested looking at the results of the model with differently sized training sets to get a full picture of the capabilities of a model. Another issue with shift prediction is the use of three-dimensional coordinates, which were used, e.g., in [25]. The generation of appropriate conformations is a problem on its own, considered in [26] using clustering techniques. In [27], AI methods were used to select the best conformer. The conformers were generated using a force field technique, and the prediction was undertaken using DFT, so the AI application was not for prediction directly, but for conformer selection.

Table 1. Historically achieved MAEs for various methods and datasets. This is a continuation of Table 1 in [21]. *** Root mean squared error instead of mean absolute error.

Date	1H MAE (ppm)	13C MAE (ppm)	Method	Training	Test Dataset	Ref.
9/2023	0.209	2.18	FullSSPrUCe (GNN)	nmrshiftdb2	nmrshiftdb2	[28]
11/2023	0.168	2.938	ComENet	GlycoNMR 80%	GlycoNMR 10%	[29]
11/2023	0.145	2.550	DimeNet++	GlycoNMR 80%	GlycoNMR 10%	[29]
11/2023	0.140	2.492	SchNet	GlycoNMR 80%	GlycoNMR 10%	[29]
11/2023	0.146	3.044	SphereNet	GlycoNMR 80%	GlycoNMR 10%	[29]
12/2023	0.138	1.79	fragment-based	COLMAR	768 COLMAR Metabolites	[23]
3/2024	0.210	2.228	GNN	nmrshiftdb2 subset	HMDB and CH-NMR-NP	[30]
5/2024	0.10	-	GNN	PROSPRE 3755 compounds	PROSPRE 272 compounds	[22]
5/2024	-	0.7	DFT	-	132 shifts	[31]
6/2024	0.185	0.944	DFT+3D GNN	nmrshiftdb2 80%	nmrshiftdb2 20%	[32]
6/2024	-	0.9 ***	GNN	2026 organic molecules	171 benzenic structures	[33]

Table 1. Historically achieved MAEs for various methods and datasets. This is a continuation of Table 1 in [21]. *** Root mean squared error instead of mean absolute error.

Date	1H MAE (ppm)	13C MAE (ppm)	Method	Training	Test Dataset	Ref.
9/2023	0.209	2.18	FullSSPrUCe (GNN)	nmrshiftdb2	nmrshiftdb2	[28]
11/2023	0.168	2.938	ComENet	GlycoNMR 80%	GlycoNMR 10%	[29]
11/2023	0.145	2.550	DimeNet++	GlycoNMR 80%	GlycoNMR 10%	[29]
11/2023	0.140	2.492	SchNet	GlycoNMR 80%	GlycoNMR 10%	[29]
11/2023	0.146	3.044	SphereNet	GlycoNMR 80%	GlycoNMR 10%	[29]
12/2023	0.138	1.79	fragment-based	COLMAR	768 COLMAR Metabolites	[23]
3/2024	0.210	2.228	GNN	nmrshiftdb2 subset	HMDB and CH-NMR-NP	[30]
5/2024	0.10	-	GNN	PROSPRE 3755 compounds	PROSPRE 272 compounds	[22]
5/2024	-	0.7	DFT	-	132 shifts	[31]
6/2024	0.185	0.944	DFT+3D GNN	nmrshiftdb2 80%	nmrshiftdb2 20%	[32]
6/2024	-	0.9 ***	GNN	2026 organic molecules	171 benzenic structures	[33]

4. Spectral Simulations

In addition to shift values, an NMR spectrum is characterized by other parameters, most prominently coupling constants, which determine the shape of the signals. These can be predicted as well, using AI techniques. ${}^{1}J_{CH}$ couplings were predicted in [28] with a precision of 0.679 Hz and ${}^{1}J_{CC}$ of 0.194 Hz. For ${}^{1}J_{CH}$, ref. [25] reported 0.870 Hz. Ab initio-calculated training data were used in all cases, since there was no big enough dataset of assigned coupling constants available. Based on predicted shifts and coupling constants, ref. [28] provided simulated spectra, fully based on machine learning.

Another option is the simulation of spectral data, including chemical shifts and coupling constants, using high-level spin dynamics. Spinach [34] is a robust computational tool for simulating NMR experiments, as it is adept at handling complex spin systems through advanced mathematical frameworks. Central to Spinach is its ability to model the behaviour of spins—nuclei or electrons possessing a magnetic moment—under various magnetic fields. These spins interact among themselves and with external magnetic fields, a phenomenon mathematically encapsulated by the Hamiltonian. The Hamiltonian provides a comprehensive representation of the total energy of the spin system, accounting for Zeeman interactions, J-coupling, and dipole–dipole interactions. The spin system’s state is described by the density matrix, which includes information about the populations and quantum coherences of the spins. The temporal evolution of this state is governed by the Liouville–von Neumann equation. Its practical workflow includes defining the spin system by specifying the spins, their interactions, and the external magnetic field; constructing the Hamiltonian to form a comprehensive energy landscape; and setting the initial state of the system. Subsequent steps involve simulating the state evolution, applying radio frequency pulses, and detecting the resultant NMR signals. Spinach leverages these mathematical techniques to predict and analyze NMR spectra, facilitating a deeper understanding of complex spin dynamics and enabling the design of more effective NMR experiments. Spinach also uses deep learning techniques, e.g., to interpret double electron–electron resonance (DEER) data [35] or to assign protein NMR signals [36].

Using full spectrum simulations, it might be possible to identify fragments and substructures of compounds [37], simulate the chemical shifts and the J-coupling values, and generate the simulated spectra to be used as training datasets. These simulated spectra can then be used for machine learning algorithms, enhancing their ability to interpret complex NMR data and automate the identification process. By leveraging both experimental and simulated NMR data, we can achieve a deeper understanding of molecular structures and dynamics, pushing the boundaries of chemical and biochemical research.

5. Spectral Processing and Peak Picking

Peak picking, a critical step in automated NMR data analysis, has been enhanced by AI techniques. AI algorithms can automatically identify and quantify peaks in complex spectra, improving the accuracy and efficiency of the process [38]. DEEP picker [38] works on two-dimensional (2D) spectra, using a convolutional neural network (CNN). For one-dimensional (1D) spectra, ref. [39] provided an automatic deconvolution of 1D spectra. It uses a combination of deep learning techniques, including CNNs and long–short-term memory (LSTM) networks. This uses synthetic data for training purposes.

An alternative to peak picking frequently used in metabolomics is binning, where the intensities of a spectrum are integrated over certain areas ("bins"). Those bins can either be of fixed width, or use so-called "intelligent" binning methods. To our knowledge, these do not apply AI or machine learning methods, but use statistical evaluations (e.g., [40] or [41]).

In [42], an algorithm to correct shift-uncertainties to better match known metabolic profiles was presented. This used a genetic algorithm combined with a least squares optimization.

6. (Pure Compound) Structure Elucidation

Computer-aided structure elucidation (CASE) systems represent a significant advancement in the field of chemical analysis by automating and enhancing the process of molecular structure determination. These systems leverage advanced algorithms, extensive spectral databases, and artificial intelligence to interpret NMR, MS, IR, and other types of spectral data, reducing the manual effort and potential for human error. CASE systems can automatically identify and label peaks in spectral data, simulate spectra from proposed structures for comparison with experimental data, and generate possible molecular structures based on spectral features. By integrating techniques like COSY, HSQC, and HMBC or 2D NMR interpretation, these systems can deduce atom connectivity and build more complete and accurate molecular structures [43,44].

For the core process of structure finding, CASE systems have historically relied on optimization and heuristics. Optimization techniques are traditionally considered part of the research field of artificial intelligence. The core challenge is to find a structure that optimally matches the constraints provided, including the measured NMR spectra. Even if restricted by, e.g., a molecular formula, testing all possible candidates is not feasible; the number of structures to test can be reduced by using heuristics like simulated annealing (e.g., SENECA [45]), genetic algorithms (e.g., GENIUS [46]), or swarm intelligence (e.g., [47]).

Key examples of CASE software include Structure Elucidator (ACD/Labs), Mnova (Mestrelab), and open-source tools like Magma [48] and Sherlock [20]. These tools offer capabilities ranging from automated data processing and peak picking to structure generation and ranking, using probabilistic scoring methods such as DP4. The integration of AI techniques, such as machine learning and deep learning, can further enhance the ability of CASE systems to handle complex datasets and predict accurate structures. An example of this is the DeepSAT system [49]. As opposed to the directed generation of compounds to find one that matches the measured spectra best, DeepSAT predicts a fingerprint and a chemical class from the spectra using machine learning techniques and searches a database for the best fit to those predicted properties. In [49], good results using this technique were reported. It should be noted that the method is restricted by the database to search in, but these do not have to be NMR databases, so a much larger range of databases can be used than with a direct NMR search.

Despite the challenges of data quality and the interpretation of complex structures, CASE systems significantly improve the efficiency, accuracy, and consistency of structure elucidation, making advanced analytical techniques accessible to a wider range of researchers and professionals in chemistry.

Bai, M. et al. (2020) [50] used ACD/Structure Elucidator software v.18 to determine the structure of four new alkaloids from the stems of Picrasma quassioides. These were elucidated through a combination of computer-aided structure elucidation software (ACD/Structure Elucidator v.18), 2D, and 1D NMR chemical shift calculations based on the GIAO (gauge-independent atomic orbital) method. The simulated NMR data aided the chemical shift assignment and ACD/Structure Elucidator yielded connectivity information for the structural formula. Comparing the experimental chemical shift values with the simulated values, the authors refined the proposed molecular structures. The GIAO simulation clarified ambiguities in the structures, especially in cases where the spectra presented overlaps or difficult-to-interpret peaks.

Natural products containing the epoxide groups were reviewed using a hybrid parametric/DFT approach called DU8+, which enabled the evaluation of more than 20 structures, including achicretin 2, guaianolide A, artanomalide B, and chloroklotzchin. CASE was shown to be crucial for overcoming difficulties in interpreting NMR data, especially in molecules with complex structural features, such as the overlap of spin–spin coupling constants and the interpretation of NOE effects. This study, by Kutateladze et al., exemplified how advanced computational techniques could significantly improve the accuracy of structure assignment in natural products, contributing to the advancement of knowledge in organic chemistry and pharmacognosy [51].

Elyashberg and Argyropoulos (2020) [52] demonstrated the application of computer-assisted structure elucidation (CASE) in determining complex natural product structures. They highlighted that modern CASE systems, such as ACD/SE, utilized NMR data to elucidate molecular structures based on molecular formulas obtained from high-resolution MS and 1D and 2D NMR spectra, including COSY, HSQC, and HMBC. The authors also noted the adoption of new experimental approaches, such as long-range correlation experiments and pure-shift methods, which enhanced the capability for structural elucidation. Caffeine, quercetin, and salicylic acid are examples of substances successfully elucidated using these computational tools. The integration of computational chemistry methods and machine learning algorithms is anticipated to further advance the accuracy and reliability of structural elucidation, solidifying the role of CASE as an indispensable tool in contemporary chemical research.

Liu et al. (2017) [53] focused on elucidating the structure of cryptospirolepine, emphasizing the importance of the CASE approach in correcting previously mischaracterized structures. The original structure, proposed in 1993 based on NMR spectroscopy data, was later found to be incorrect. Utilizing advanced MicroCryoProbe technology and 1,1-HD-ADEQUATE experiments, the authors revised the structure in 2015 and proposed possible candidate structures by comparing experimental chemical shift data with theoretical values calculated using DFT. The comparison between the original and revised structures revealed a significantly stronger correlation between the experimental and theoretical data for the correct structure, as indicated by a much lower Q value (0.122 for the revised structure, compared to 0.245 for the original). This difference underscored the effectiveness of using RDC and RCSA in structural validation, showing that the new approach not only corrected the structure but also provided a robust tool to avoid characterization errors in future studies. These findings highlighted the relevance of combining computational and experimental methods in elucidating complex molecular structures, particularly when crystallization is not feasible, as often occurs with natural compounds.

Howarth et al. (2023) [54] introduced a significant innovation in the structural elucidation of organic compounds with the development of the DP4-AI system. This automated system processes and interprets data from carbon and hydrogen nuclear magnetic resonance (NMR) spectroscopy, enabling the rapid and efficient analysis of complex molecules, such as 2,3-dihydroxy-1,4-naphthoquinone, and 1,2-dihydroxy-3,4-naphthoquinone. DP4-AI employs advanced algorithms for NMR peak selection and matches chemical shifts calculated via density functional theory (DFT) with experimental data, resulting in precise and reliable spectral assignments. With a 60-fold increase in processing speed compared to traditional methods, DP4-AI significantly reduces analysis time and the manual workload for chemists, who previously spent hours on data interpretation. This system optimizes the workflow in structural elucidation, minimizing the need for manual intervention and making the process more efficient and accessible for high-throughput analyses. DP4-AI’s capabilities are particularly valuable in research areas where speed and accuracy are critical, such as drug development and natural product characterization. The results obtained with DP4-AI underscore its potential to revolutionize organic chemistry, offering a robust tool that combines automation and computational intelligence to address the challenges of modern structural elucidation [54].

A class of AI models exploited recently for structure elucidation are transformers. A encoder–decoder architecture was used in [55]. In this, spectra are transformed to text (most prominently shift values) and structures are encoded as SMILES (simplified molecular-input line-entry system). The encoder–decoder network then learns how to transform those two representations into each other. For this, successive layers learn the context of the individual tokens in the text representations. In [55], accuracies for structure recall of 50% up to 100% (for very small molecules) were reported. An accuracy of 69.6% was reported in [56] for molecules up to 19 heavy atoms. This model predicted substructures from the spectrum and assembled them to a solution, using a transformer model for the last step and another neural network for the first.

Specifically in metabolomics, ref. [57] illustrated the use of quantum chemistry methods to identify metabolites in complex samples, addressing the challenge of characterizing the vast chemical diversity present in the metabolome. They explained how quantum calculations can be used to generate chemical properties and reference spectra for metabolites like caffeine, salicylic acid, and flavonoids, which are common in metabolomics studies. The research underscored the importance of developing in silico libraries for molecular identification, which can help overcome the limitation of the unavailability of authentic chemical standards for many compounds. They highlighted the need for free energy corrections and the inclusion of entropic contributions in the analysis of reaction mechanisms, which enhances the accuracy of predicting fragmentation patterns in mass spectrometry (MS) experiments. This approach not only aids in compound identification but also opens new avenues for designing reactions and catalysts, significantly advancing the field of metabolomics.

7. Mixtures and Metabolomics

NMR spectra can be recorded from mixtures in solution as well from single compounds. In the case of mixtures, the peaks of the individual compounds show up in one spectrum as the sum of their intensities, which reflect their respective concentrations. Since it is not clear which peaks belong together to compose a single structure, the potential search space becomes even bigger than for single mixtures. Dereplication strategies are still possible, implemented for example in MixONat [58]. COLMAR [59] uses a combination of HSQC and TOCSY spectra and queries them against a database. COLMAR NMR has emerged as a powerful analytical tool in metabolomics, natural products research, food science, environmental monitoring, and clinical diagnostics. Its application in metabolomics enables the broad profiling of biofluids such as urine [60], serum [61], and others [62,63], facilitating the identification and quantification of metabolites that reflect physiological and pathological states, including those associated with cancer, diabetes, and neurological disorders. The method’s robustness, when the user reproduce sample requirements, offers detailed spectral information with the capacity to unravel complex mixtures, offering valuable insights across scientific disciplines. Clustering methods which make use of shared shifts of peaks in 2D spectra were demonstrated in [64,65] as a tool named NMRfilter. Clustering was also used in [66], but required a separation, albeit a simple one compared to a full separation. NMRfilter is still underused by the community, but its application was shown in ref. [67,68]. SCORE-metabolite-ID [69] correlates NMR with MS data over the third dimension of the time course of a chromatographic fractionation. Similarly, DAFdiscovery [70] applies the concept of statistical heterospectroscopy (SHY) to correlate data from NMR, MS, and biological activity (or any combination of these) to pinpoint compounds of interest, mainly directed to natural products discovery.

With the emergence of AI technologies, more advanced analyses have become possible. In [37,71], a convolutional neural network was used to predict the chemical class of sub-fragments directly from 2D spectra. This demonstrates the power of deep learning technologies, since the network extracts a typical pattern from the training set, finds those in the new cases, performs fuzzy matches where needed, and performs the classification. This works for mixtures as well as for single compounds. Similar techniques (e.g., fingerprint prediction from spectra) were used in [49]. SMART-Miner [72] uses a CNN as well for the identification of components in mixtures. As opposed to using processed data, ref. [73] used raw data and identified functional groups by a combination of peak sampling and a recurrent neural network (RNN).

In metabolomics, both targeted and untargeted approaches have benefited from AI for statistical evaluation and for data interpretation. For a recent review of AI in metabolomics generally, not only in NMR, see [74]. Metabolomics stands for the comprehensive study of metabolites, which are small molecules produced or consumed in metabolic processes within a biological system, such as a cell, tissue, organ, or organism. It is used to understand the metabolic profiles and pathways involved in various physiological states, diseases, or responses to external factors like drugs or environmental changes by creating prediction models. This process is generally performed using statistical regression methods, such as logistic regression (see [75]). However, as machine learning models began to outperform statistical models in other fields, they were increasingly applied to this task as well. Recently, ref. [76] used a neural network (called a metabolomic state model) to predict multidisease risks. This is an application of untargeted metabolomics, where the complete metabolome is measured. Peak picking and binning, described in Section 5, are typically used for further evaluation. This process is referred to as metabolomic fingerprinting. In [77], machine learning was used to evaluate features from LC-MS (liquid chromatography–mass spectrometry)and NMR spectra in order to diagnose renal cell carcinoma. This can potentially replace a costly combination of imaging and invasive techniques.

In the food sector, metabolic fingerprinting is used to determine various parameters of food, including origin, freshness, species (e.g., for meat), or processing steps. This was traditionally done using statistical methods, and has been applied to a variety of food (for example [78,79,80,81]). Recent studies have claimed that using AI techniques can distinguish samples which purely statistical methods cannot deal with, e.g., the origin of fish from closely related water bodies [82]. NMR is routinely used for fingerprinting of plasma and urine samples for human metabolomics studies. Metabolomics-based methods have immense potential for early disease diagnosis and monitoring therapeutic responses [83,84]. The identification and quantification of metabolites through NMR spectroscopy and mass spectrometry has enabled the discovery of novel biomarkers with the potential for early disease diagnosis and personalized treatment [85]. An open question when discussing AI applications is the explainability and trustworthiness of results. In [86], an AI application was demonstrated which performed quantification of a limited set of analytes and showed which regions of 1D spectra had been used for this.

A very important area in metabolomic profiling is lipoprotein profiling, pioneered by [87], which traditionally relies on de-convolution and curve-fitting. Improved results using machine-learning techniques, specifically random forest and logistic regression, were shown in [88].

In contrast to untargeted metabolomics, targeted metabolomics measures a set of defined analytes. It is generally agreed upon that only a fraction of the relevant metabolites are currently known. A total of 150,000 is considered a lower limit, whereas HMDB lists about 42,000 [89]. In [90], a pipeline of deep learning techniques was used to identify metabolites in human plasma samples. This study reported the successful identification of known metabolites, but could not deal with the “metabolic dark matter”. For clinical and other purposes, a targeted analysis can identify conditions which an untargeted analysis may not show, as reported for example in [91].

8. Conclusions

This review highlights the significant role of artificial intelligence in enhancing nuclear magnetic resonance spectroscopy and metabolomics. NMR is still essential for identifying and understanding small molecules, and AI techniques, particularly machine learning, have improved the accuracy and efficiency of data analysis. By integrating AI, increasingly complex datasets can be handled. This is aided by the improved prediction of chemical shift data, and the more efficient simulation of spectra.

The adoption of FAIR data principles and advanced data management systems can further support these advancements. At the moment, data collection and accessibility is still an issue. By combining experimental and simulated data, AI-driven approaches are pushing the boundaries of NMR and metabolomics, enabling more precise compound identification and deeper insights into biological systems.