Challenges in Identifying the Dark Molecules of Life

Abstract

Metabolomics is the study of the metabolome, the collection of small molecules in living organisms, cells, tissues, and biofluids. Technological advances in mass spectrometry, liquid- and gas-phase separations, nuclear magnetic resonance spectroscopy, and big data analytics have now made it possible to study metabolism at an omics or systems level. The significance of this burgeoning scientific field cannot be overstated: It impacts disciplines ranging from biomedicine to plant science. Despite these advances, the central bottleneck in metabolomics remains the identification of key metabolites that play a class-discriminant role. Because metabolites do not follow a molecular alphabet as proteins and nucleic acids do, their identification is much more time consuming, with a high failure rate. In this review, we critically discuss the state-of-the-art in metabolite identification with specific applications in metabolomics and how technologies such as mass spectrometry, ion mobility, chromatography, and nuclear magnetic resonance currently contribute to this challenging task.

INTRODUCTION

Metabolomics is the newest omics field focused on the examination of metabolites in complex systems, with the goal of identifying pathway alterations that correlate with the onset and progression of specific processes, such as disease (1, 2). The metabolome is typically defined as the collection of small molecules in a given biological system that roughly falls under the ∼1,500-Da molecular weight window. Metabolites in the metabolome include endogenous molecules that are biosynthesized in primary metabolism, specialized secondary metabolite signaling molecules, lifestyle or environmental exposure molecules (the exposome), and molecules originating from the microbial community associated with the organism under study (the microbiome). Metabolomics can be either targeted to detect and quantify a set of known metabolites, or nontargeted, where the emphasis is to detect as many compounds as possible, even if associated with unknown chemical species. The metabolome, by definition, spans a vast diversity of chemistries, including lipids, sugars, amino acids, steroids, and a whole array of molecule types. These species exist in large concentration ranges that can be as high as millimolars and as low as femtomolars, suggesting that no single analytical method in existence today is able to detect, identify, and quantify all present species (3). Nuclear magnetic resonance (NMR) and liquid chromatography mass spectrometry (LC-MS) are the main platforms used for metabolomics, each with its own advantages and disadvantages in terms of sensitivity and peak capacity. These techniques are used individually in most studies but can also be combined to better yield metabolome coverage and enable more accurate metabolite identity annotation (4–6).

Despite rapid technological advances in both NMR and LC-MS, metabolite identification still remains the undisputed bottleneck in nontargeted metabolomics experiments. Only a small fraction, less than 2–10%, of the detected compounds in a nontargeted metabolomics study can be annotated reliably, with the chemical identity of most species detected remaining unknown (7). This vast chemical space has been described as the metabolome dark matter (8) and continues to be the focus of intense scientific research, in terms of new hardware and methods for improved separations and structural identification, but also new databases, libraries, and algorithms that can predict some of the analyte’s molecular properties.

Much effort has been directed toward defining the confidence levels and minimum reporting standards associated with metabolite identification in any given metabolomics study. As early as 2007, Sumner et al. (9) proposed a four-level metabolite identification confidence scheme that divides metabolites into (a) identified compounds, (b) putatively annotated compounds, (c) putatively characterized compounds, and (d) unknown compounds. In level a, identified compounds are those that include at least two orthogonal molecular characterization methods that match a chemical standard [e.g., retention time (RT) and high-resolution mass spectrum, RT and NMR resonances, elemental formula and tandem MS (MS/MS) spectra, ¹H and/or ¹³C NMR, and two-dimensional (2D) NMR spectra] (9). Generation of more than two pieces of such orthogonal metabolite characterization data is widely seen as providing good evidence in metabolite annotation, but it is not always possible due to a number of factors, including low signal-to-noise ratios, unavailability of chemical standards, and a lack of database coverage. Putatively annotated compounds are those that are identified by matching to literature or database information, but for which no chemical standard can be obtained for comparison purposes. Putatively characterized compounds are those for which only similarities to a given family of compounds can be established, but further information is not available [e.g., a glycerophospholipid characterized by the presence of an m/z 184 fragment in MS/MS experiments corresponding to the phosphocholine headgroup; or the neutral loss of 141 Da corresponding to the phosphatidylethanolamine headgroup (10, 11)].

Despite its usefulness, this four-level scheme lacks granularity and detail, so here we propose to further expand it through a score card (Table 1) that further refines the assignment of metabolite identification confidence in nontargeted metabolomics experiments using a points system. This system builds on advances in metabolomics instrumentation and algorithms described below, such as RT prediction, use of ion mobility (IM) collision cross sections (CCS), and prediction of MS/MS fragmentation and NMR spectra.

Table 1.

Proposed “score card” approach to assign metabolite identity confidence levels. In this approach, points are added as various orthogonal methods of molecular characterization are successfully applied. A higher score indicates higher certainty on metabolite identity.

Qualifier		Explanation	Points assigned
(i) High resolution MS*	a. Monoisotopic peak m/z match within 5–10 ppm.	Simplest metric for assigning a tentative identity. Should be used with caution, as high ambiguity exists in ID assignment based solely on m/z.	5
	b. Monoisotopic peak m/z match within 1–5ppm.	Higher mass accuracy reduces the number of elemental formula candidates, helping in reducing the number of tentative IDs.	10
	c. Mass-to-charge ratio match within 1–2 ppm and isotopic ratio match with proposed elemental formula.	Relative abundance information found in isotopic structure can further reduce number of plausible elemental formula candidates.1	15
	d. More than one matched ESI adduct or in-source fragment (e.g. [M-H]−, [M+Cl]−).	The match of elemental formulae to more than one spectral feature (e.g. [M+H]+ and [2M+H]+) results in increased confidence.	5
	e. Mass-to-charge ratio match within 1–2 ppm and X+1, X+2 ion isotopic fine structure (e.g. high field Orbitrap, FT ICR).	Ultrahigh resolution MS measurements allow readout of elemental formulae directly from mass spectrum.	20
		Sub-Total Max	25
(ii) MSN	Match of MS2 fragmentation spectrum to experimental, database, or literature spectra.	Tandem MS experiments, although not 100% conclusive, greatly increase identification confidence.	10 total, proportional to # fragments matched
	Match of expected fragment ion ratios	Relative ratios of fragment ion pairs can discriminate between closely related, isobaric, species	5
	Match of experimental MS2 fragmentation spectrum to in silico predictions	If database entries are not available, or not match is found, some degree of success can be obtained by predicting CID fragmentation via software tools (e.g. Thermo Mass Frontier, CFM-ID, MetFrag, Mass Fragment)	5
	Manual interpretation of MS2 fragmentation spectrum consistent with proposed structure.	Manual interpretation of tandem MS spectra is sometimes necessary in the absence of database matches or unsuccessful in silico prediction.	10
	Higher than MS2-level match to MSN database (such as mzcloud)	Emerging databases now include MSN information.	5
		Sub-Total Max	20
(iii) Chromatography	Retention time match (within expected window) between candidate ID and chemical standard.	Chromatographic retention times are excellent qualifiers for increasing ID certainty. Analytes should elute outside of the dead volume window.	10 (+5 if spiked in sample)
	In absence of standard, retention time may be matched to predicted value.	QSRR tools to predict retention time for small metabolites can be used to reinforce ID confidence.	10
		Sub-Total Max	20
(iv) Ion mobility	Database CCS matches to experimental CCS.	Ion mobility is an emerging technique in terms of metabolite ID. Databases are being created to facilitate identification using CCS.	10
	Predicted CCS matches experimental CCS	Molecular descriptor-based prediction of CCS is possible.	5
	IM filtering post precursor ion selection prior to MS2	Better matches of experimental MS2data to database data can be obtained by IM separation prior to CID, mitigating precursor ion co-selection.	5
		Sub-Total Max	15
(v) NMR	1D NMR match (e.g. 1H or 13C)	Matches to COLMAR or any other query interfaces	15
	2D NMR match	Matches to COLMAR or any other query interfaces/databases.	20
‘		Sub-Total Max	20

^*Only ia, ib or ic should be applied.

(1) Pluskal, T.; Uehara, T.; Yanagida, M. Anal. Chem. 2012, 84, 4396–4403.

An example of how to apply the scoring approach described in Table 1 would be as follows: If an identity is proposed for an unknown metabolite by (a) matching the [M+H]⁺ and [M+Na]⁺ monoisotopic adduct ions (5 points) with an average 2.5-ppm error (10 points), followed by (b) an MS² (10 points) and MS³ (5 points) match to the mzCloud database, (c) its experimentally measured CCS matched within tolerance to a database value (10 points), and (d) its ¹H NMR spectrum matched to a database entry (15 points), a confidence score of (10 + 5) + (10 + 5) + (10) + (15) = 55 points could be assigned. Further improving identification confidence would require (a) increasing mass accuracy to 1–2 ppm for more than one adduct ion (5 points) while employing ultrahigh-resolution MS experiments to examine isotopic fine structure (20 points); (b) manually interpreting (5 points) the database-matched product ion spectrum to verify that observed fragment ions are compatible with known fragmentation pathway; (c) matching chromatographic RT to a chemical standard (10 points); (d) and performing 2D instead of one-dimensional (1D) NMR experiments (20 points). This would yield a more confident score of (5 + 20) + (10 + 10 + 5) + (10) + (10) + (20) = 90 points in total. By applying these more specific metabolite identification techniques to unknown species and combining the information they provide, a better confidence score is obtained. Strengths and limitations of each of these techniques, as applied to metabolomics, are discussed in the sections below.

CHROMATOGRAPHIC SEPARATIONS

Historically, chromatographic separations coupled to MS have offered one of the most versatile platforms for complex sample analysis and metabolite identification in nontargeted metabolomics studies (12). LC-MS is by far the most popular MS-based hyphenated technique in metabolomics due to its sensitivity, selectivity, reproducibility, and versatility for analyzing small molecules with a wide range of different physicochemical properties in biological samples (13, 14). Gas chromatography (GC)-MS, a predecessor to LC-MS, continues to be the obvious choice for fingerprinting volatile and low molecular weight compounds (15, 16). Analysis of less volatile compounds by GC-MS, however, requires additional preparation steps such as lyophilization, followed by chemical derivatization to increase thermal stability and volatility prior to analysis (17). To a lesser extent, capillary electrophoresis has also been utilized in metabolomics studies involving highly charged and polar ionogenic metabolites in small-volume biological samples (13, 18). Capillary electrophoresis is attractive in the sense that it separates compounds based on their intrinsic electrophoretic mobility, a parameter that reflects the charge and size of the analyte (18).

LC-MS offers the widest coverage and most efficient separation of complex metabolomes, with much enhanced peak capacity compared to MS alone (13). LC and MS are coupled by means of soft ionization techniques; electrospray ionization (ESI) (19–21) is the most common. ESI generally produces protonated or deprotonated gas-phase ions that can be used for accurate predictions of metabolite elemental formulae (22). In addition to in-source fragment ions, various adduct ions and multiply charged species that complicate data interpretation are also produced, leading to higher false-positive rates in metabolite identification (23). GC, on the other hand, is coupled to MS by means of electron ionization sources operated under vacuum and standardized at 70 eV, causing predictable fragmentation and rearrangement reactions that lead to highly reproducible mass spectra for comparison with libraries (16).

Reverse phase (RP) and hydrophilic interaction chromatography (HILIC) are by far the most commonly used chromatographic methods for LC-MS-based metabolomics (13, 24). In RPLC-MS, analytes are typically eluted by an aqueous-based mobile phase under a gradient of increasing organic solvent content from a hydrophobic stationary phase. In HILIC, analytes are eluted in order of increasing hydrophilicity from a hydrophilic stationary phase as the polarity of the mobile phase is increased by increasing the aqueous content.

Traditionally, chromatographic identification in metabolomics by either RP or HILIC approaches is best performed by RT matching with authentic chemical standards. However, not all metabolites are commercially available, and purchasing authentic standards for all possible metabolite candidates in a nontargeted study is incredibly costly and inefficient. Indeed, there are fewer reference standards available than the number of peaks that is detected in LC-MS experiments of biological samples (25–27). Although chemical synthesis can be attempted for specific cases of high-value unknowns, this degree of effort is seldom warranted.

GC-MS capillary columns are highly reproducible, which facilitates the compilation of standardized retention indices (28, 29) in libraries or public databases for compound identification (16, 30). In particular, Fiehn and collaborators (30) have greatly contributed to the metabolomics field by building FiehnLib libraries comprising mass spectra and retention indices from quadrupole and time-of-flight (ToF) GC-MS data. In contrast, it is difficult to catalog RT information from LC experiments in libraries due to the lack of procedural standardization and instrumentation involved, with varying LC pumps and injectors, columns, run temperatures, solvent gradients, mobile phase pH, and flow rates. In addition, column aging, temperature changes, MS detector drift, and other analytical factors may further compound RT variance in nontargeted LC-MS metabolomics experiments. Matrix effects caused by differences in biofluids matrix compositions can also lead to RT shifts. Spiking experiments with chemical standards can mitigate these effects (31), but in cases where chemical standards are not available, RT window prediction can complement metabolite identification efforts.

Different attempts have been reported in the literature (31–36) to integrate RT window prediction into the metabolite annotation process, most of which rely on quantitative structure–retention relationship (QSRR) modeling (37). The generalization of such predictive models depends on the application domain investigated, because using overly restricted domains for model building also leads to poor prediction in independent test sets (38). Molecular descriptors (39–42) also play a significant role in defining such application domains (38, 43, 44), as well as in defining the multiple algorithmic options and combinations of adjustable parameters involved in QSRR model building (42, 45). In general, RT factors (RF in the equation below, where T₀ represents the chromatographic column void time) are calculated to allow comparison between different chromatographic systems (32, 36). Other studies utilize retention indices (28, 29), which are measures of relative RT based on reference compounds that elute immediately prior to and immediately following the analyte of interest (30, 45–48).

$$RF=\frac{RT-T_0}{T_0}.$$

Cao et al. (31), for example, conducted QSRR modeling based on theoretical molecular descriptors and experimental RTs of 93 authentic compounds analyzed with HILIC LC-MS. A predictive QSRR model based on a random forest algorithm achieved high predictive accuracy, with mean and median absolute errors of 0.52 min and 0.34 min (5.1% and 3.2%), respectively. These authors applied this model to annotate features (RT, m/z) of perennial ryegrass (Lolium perenne) samples, significantly reducing the number of false-positive metabolite annotations that would be obtained if only accurate masses were considered. Predicted RTs were validated in this study using either authentic compounds or ion fragmentation patterns (31). Among the molecular descriptors utilized to build QSRR models, the partition coefficient (XLogP) was found to be the most relevant predictor in agreement with the HILIC model previously reported by Creek et al. (32). These authors converted the RT of 120 metabolite standards to retention factors that were input into a multiple linear regression model. The optimal QSRR model used six physicochemical variables and showed good predictive ability (cross-validated R² = 0.82 and mean square error = 0.14) for RTs of metabolites with MW <400. Availability of predicted RTs translated into the removal of 40% of the false metabolite identifications based on accurate mass alone (32). The model was evaluated to putatively identify 690 metabolites in extracts of the protozoan parasite Trypanosoma brucei. Model limitations were associated with the applicability to only low-MW metabolites, since larger compounds were poorly predicted, most likely due to errors associated with predicted log D, the octanol-water partition coefficient calculated at a pH of 3.5. Based on their QSRR model, the authors produced a template file that allows users to calculate predicted RTs for a database of metabolites based on experimental RTs measured for standards.

QSRR modeling has also been applied to RPLC data. Bruderer et al. (34) evaluated RT prediction for a metabolomics database with 532 human metabolites. The authors built a model with only 16 compounds using logD2 (calculated based on log P and pKa) and the molecular volume as molecular descriptors. The developed model was evaluated for two different RP C18 columns and two pH conditions (pH = 3.0 and 8.0 for positive and negative ESI modes, respectively), achieving good prediction accuracy for a time window below 4 min (34). In addition, RT prediction combined with the data-independent acquisition (DIA) method known as sequential windowed acquisition of all theoretical fragment ion mass spectra (SWATH) aided in annotating isobaric metabolites found in human urine, which increased identification confidence and reduced the number of false positives (34).

A different RT prediction model for RPLC was built by Wolfer et al. (44) based on 442 authentic standards, including fatty acids, nucleosides, sterols, sphingolipids, lipids, vitamins, cofactors, amino acids, aromatic biogenic acids, carbohydrates, catechols, neurotransmitters, and other metabolites to cover a large variety of polarity and chemical topology. The authors combined random forest and support vector regression models with 97 computationally determined descriptors derived from chemical structures. The model was further tested on an external validation set of 111 known compounds, predicting RTs with an average 13% error that reduced by 77% the number of incorrect candidates. The models were made available through a user-friendly interface to be integrated into existing workflows.

Identification of isobaric lipids is challenging even with high-resolution mass spectrometers. Aicheler et al. (33) developed an RT prediction model based on machine learning approaches that enabled the improved assignment of lipid structures and automated annotation of lipidomics data obtained by RPLC–high-resolution MS. A support vector regression model was built based on 201 lipids originating from mouse adipose tissue, including molecular structural features. The cross-validated model achieved good correlation (R = 0.989) between predicted and experimental RT in test samples (Figure 1) and allowed filtering out of more than half of the potential identifications, while retaining 95% of the correct candidates.

Figure 1.

Comparison of predicted and experimental retention times of the initial model. The model was trained on 50 data points and evaluated for 151 validation analytes. The depicted test lipids are distinguished by lipid class. The listed R² was computed from the test lipids. Reproduced from Ref. 33 with permission from the American Chemical Society ©2015.

Most QSRR approaches have been successful at predicting RT with good accuracy. The main limitations of these strategies, however, are related to the lack of large and sufficiently diverse metabolite training sets for building predictive models that could cover the thousands of metabolites found in biological samples. In addition, some published models have lacked comprehensive external validation, thus risking overfitting (36). Further drawbacks are associated with the type of stationary phase material used (36, 43) that determines, together with the sample preparation protocol, the metabolome fraction that can be effectively resolved.

Alternative efforts have involved the development of tools that allow crowd sourcing of RT information across laboratories and chromatographic systems such as retention projection (49) and direct mapping (50) for GC-MS and LC-MS systems, respectively. The retention projection methodology for GC-MS data, for example, was shown to be threefold more accurate than retention indexing across five different laboratories under identical experimental conditions. This made it easier to account for unintentional differences between the various GC systems, such as temperature calibration errors, flow rate nonidealities, or variance in column dimensions (49). RT mapping for LC-MS data, developed under the name of PredRet, was done pairwise between LC systems using the same chromatographic method and provided higher accuracy than QSRR (50). However, QSRR models are able to a priori predict the RT of any given metabolite structure, whereas PredRet can only predict RT for compounds already in the database, i.e., those for which the RT has been previously determined in a comparable chromatographic system (50).

MASS SPECTROMETRY

MS/MS experiments select a precursor ion for activation and fragmentation due to ion-neutral collisions with the gas filling the collision cell. The combination of chromatographic separations and MS/MS is one of the most powerful approaches to metabolite and natural product identification (51). In nontargeted metabolomics, these tandem mass spectra can be obtained in either a data-dependent acquisition (DDA) or DIA fashion (52). DIA approaches, such as MS^E, benefit from larger precursor ion coverage (53), but they suffer from difficulties in matching precursor ions to product ions when chromatographic overlap is substantial. DDA, on the other hand, preferentially targets the most abundant precursor ions, resulting in poorer sampling of lower abundant precursor species. Repeated analysis can somewhat improve precursor ion coverage, but this improvement is typically only marginal due to redundant precursor ion sampling. Advanced precursor ion selection algorithms used in proteomics (54) could also benefit metabolomic studies. Broeckling et al. (55), for example, have proposed an alternative method for enabling the comprehensive MS/MS coverage of complex samples via data set–dependent acquisition. With this approach, real-time feedback between data processing and data acquisition was achieved using a combination of R, ProteoWizard, XCMS, and WRENS software, yielding a threefold improvement in the number of peaks mapped by MS/MS. Elaborate DDA approaches that combine with higher-energy collision dissociation have also been reported (56).

Targeted metabolomics approaches are different from nontargeted approaches in the sense that they typically rely on chemical standards with known structures, using triple quadrupole mass spectrometers in multiple reaction monitoring (MRM) mode, or hybrid quadrupole ToF/Orbitrap analyzers in SWATH or parallel reaction monitoring modes (57). Approaches that bridge the targeted and nontargeted methods, however, have been reported in the literature. Chen et al. (58) recently developed a hybrid method utilizing DIA for targeted quantitative metabolomics experiments. In this method, a sequentially stepped targeted MS/MS scan is used for improving coverage. In this type of scan, multiple product ion scans are acquired for all ions in the examined m/z ranges, selecting them as the chosen precursor ions. These scans are then followed by scheduled MRM scans for numerous ion pairs that are used for quantitation. Ferreira et al. (59) have also reported an approach that bridges classical targeted and nontargeted methods. This approach, named MRM profiling, makes use of numerous product-precursor ion transitions that are developed in a supervised fashion. By combining neutral loss and precursor ion scans, a list of more than 1,000 transitions is built from a pooled sample. Relative abundances of all product ions targeted in these MRM transitions are then measured for all samples and used to build univariate and multivariate diagnostic models for a specific condition, such as polycystic ovarian syndrome (60).

Precursor ion coselection is an underappreciated issue that commonly hinders metabolite identification in LC-MS/MS metabolomics, particularly with lipids. Most high-resolution mass spectrometers use a low-resolution quadrupole mass analyzer for mass selection prior to MS² fragmentation. The selection window for the precursor ions is typically limited to 0.5–3 Da. When a high number of isobaric species chromatographically coelute, these are inevitably coselected and cofragmented, yielding a product ion spectrum that is a composite of all product ions yielded from the initial precursors in that window. One way to mitigate these interferences is to perform collision-induced dissociation (CID) experiments a posteriori from IM separations. Damen et al. (61) recently reported an ultraperformance LC (UPLC) method for the separation of closely related lipid molecular species using a stationary phase incorporating charged surface hybrid technology. The chromatographic method showed excellent RT reproducibility [intraassay relative standard deviation <0.385% and <0.451% for 20- and 10-min gradients, respectively (N = 5)]. The UPLC system was coupled to a hybrid quadrupole ToF mass spectrometer, equipped with a traveling wave ion mobility (TWIM) cell. Despite the use of a quadrupole mass analyzer for precursor ion selection, separations in the TWIM cell followed by transfer cell ion activation enabled the acquisition of cleaner low- and high-energy DIA MS/MS spectra that were more useful in terms of metabolite identification. Another approach to prevent precursor ion coselection is through the use of different variations of stored waveform inverse Fourier transform (SWIFT) ion excitation in Fourier transform ion cyclotron resonance (FT-ICR) (62). As early as 1995, O’Connor & McLafferty (63) reported on high resolution ion isolation using a capacitively coupled FT-ICR open cell. In these experiments, isotopic peaks of ubiquitin (8.6 kDa) and carbonic anhydrase (29 kDa) were isolated by SWIFT with an order of magnitude higher isolation power than previously reported in the literature. Another approach yielding high-resolution ion selection in FT-ICR MS is known as correlated harmonic excitation fields (64). With this approach, de Koning et al. (64) were able to achieve a resolution of ∼50,000 in the separation of deuterated toluene isotopes. Routine implementation of high-resolution precursor ion selection approaches such as those described above in a metabolomics context could significantly improve the odds of correct unknown identification via more selective CID experiments and higher confidence MS/MS database matching.

Although MS¹ information coupled with local network enhancement analysis can be used for tentative metabolite annotation (65), high-quality tandem mass spectral libraries are becoming essential for more confident identification (66). A growing number of these libraries are currently available, including the Human Metabolome Database (HMDB) (67, 68), METLIN (69, 70), MassBank of North America (MoNA; http://mona.ftehnlab.ucdavis.edu/), LipidBlast (71), mzCloud (72), LIPID MAPS Structure Database (64, 73), Manchester Metabolomics Database (MMD) (23), and many others (74, 75). More general databases, such as PubChem (65, 76) and ChemSpider (65, 77), can also be incredibly useful for MS-based identification of unknowns. However, metabolite identification through mass spectral library searches is far from an automated task, and the analyst is typically forced to manually search each individual database and manually curate the obtained matches (if any) to ensure that differences in the type of mass spectrometer used and the collision energy employed are considered. Along these lines, Stein and coworkers (78) have proposed an approach for creating high-quality ESI tandem mass spectral libraries. The procedure involved the acquisition of tandem mass spectra for all major precursor ions in a direct infusion experiment. This was followed by assigning spectra to clusters and creating a consensus spectrum. Filtering through intensity-based constrains for cluster membership was then applied, together with peak testing, noise reduction, and examination by an experienced human evaluator, yielding a library of >9,000 compounds with ∼230,000 spectra.

When MS/MS database matches are not found, prediction of such spectra in silico can be advantageous to provide an added level of confidence to the proposed metabolite identity. To this purpose, Wolf and coworkers (79, 80) described the popular package named MetFrag, where a candidate list of metabolite identities is first obtained by searches of precursor ion masses, followed by ranking based on the agreement between in silico fragmentation spectra and experimental data. Initial evaluation of MetFrag showed that it was able to rank most of the correct compounds in the top three candidates produced by KEGG queries, producing better results than commercial software. In related work, Duhrkop et al. (81) described an approach named CFM-ID that combines computation and comparison of elemental formulae fragmentation trees with machine learning techniques. They reported a 2.5-fold increase in correct identifications compared with state-of-the-art methods when searching PubChem. Li’s group (82) described a web interface (https://www.MyCompoundID.org) for metabolite identification based on both MS and MS/MS data for compounds in the HMDB and metabolites derived from these through in silico metabolic reactions. Fragmentation prediction for specific metabolite families, such as lipids, has certainly benefited subfields such as lipidomics (83–85), where chemical structures follow combinatorial rules. Prediction of electron impact spectra has also been achieved with a significant degree of success (86) based on an approach previously used for ESI data (87).

Combined prediction of various molecular properties, including retention indices, energy required to fragment 50% of a selected precursor ion, IM drift time, and CID spectrum, has been proposed by Grant and coworkers (88) through a package known as MolFind, but follow-up validation of such an approach through comparison with large experimental data sets has not been reported. Similar motivation led Hu et al. (89) to attempt the simultaneous prediction of both RTs and fragmentation patterns with the goal of identifying micropollutants.

ION MOBILITY SPECTROMETRY

In order to increase confidence in the identification of prioritized features after MS analysis, additional structural techniques, including infrared spectroscopy, NMR spectroscopy, and more recently, IM spectrometry (IMS), are often incorporated into nontargeted studies (90, 91) and are taken into account in our proposed identity confidence scoring scheme (Table 1). IMS is a gas-phase separation technique in which analytes are separated based on their rotationally averaged surface area or CCS. Briefly, IMS separations are conducted as analyte migration through an inert buffer gas under the influence of an applied electric field. Although specifics of gas composition, pressure, and applied field strength vary depending on the instrument configuration, interactions between these forces drive ion motion and separation in the IMS cell (92, 93). Because IMS distinguishes analytes based on their structural size in the gas phase, it is orthogonal to MS to a great extent, and it provides capabilities such as isomeric separations that are not possible with only the mass dimension (94, 95). Furthermore, IMS measurements occur on a millisecond timescale, which is readily nested into traditional LC and GC-MS workflows (96). Descriptions of various IMS platforms and configurations are available in several in-depth review articles (74, 97, 98) and are beyond the scope of this article.

Although growth in MS databases has steadily continued for several decades, CCS databases are still in their infancy. Since the commercialization of IMS-MS instrumentation in 2006 (73, 76), several studies have been devoted to collecting CCS values on a larger scale. These efforts have made CCS values available to the public, while also creating databases useful for nontargeted metabolomics experiments (77, 91, 99–101). Despite these advances, obtaining reproducible CCS values across various instrument platforms still remains challenging owing to several key factors that include variations in instrument design, experimental parameters, and calibration protocols. For example, only drift tube and differential mobility instruments can empirically measure CCS from first principles, meaning that no calibrants are needed for the values they produce. IMS platforms based on TWIM spectroscopy and trapped ion mobility spectroscopy (TIMS) instruments, however, must be calibrated with ions of known mobility before reliable CCS values can be generated (102–104). Also, the specific manner in which each instrument is calibrated can greatly influence the reproducibility of the reported CCS values. Several studies have previously described the challenges of choosing the correct calibrant ions for traveling wave devices with similar structural characteristics as the species being studied (105, 106). For example, Gelb et al. (107) observed that improper calibration of TWIMS with calibrant ions of a different chemical class and charge state could produce CCS measurements with an error in excess of 4% compared to using proper calibration protocols. Even after proper calibrant ions are chosen for a specific instrument and experiment, the optimal mathematical calibration procedure remains a topic of debate for several platforms (108, 109). As IMS-MS instrumentation continues to improve in terms of generating reliable and reproducible CCS values, slight variations in calibration procedures become increasingly critical. For example, a recent interlaboratory study from Paglia et al. (77) characterized TWIMS reproducibility to typically less than 3% relative standard deviation, suggesting that errors in CCS calibration could lead to errors in calculated CCS that are larger than the instrument’s own reproducibility. In a similar fashion, a recent interlaboratory study by Stow et al. (100) demonstrated that new advancements in drift tube technology resulted in reproducibility of typically less than 1% relative standard deviation for CCS measurements between laboratories because no calibration was required.

Despite challenges in calibration procedures, once a reproducible CCS value is measured for a given analyte, matching a molecular feature to a library entry is a straightforward process when the analyte m/z and CCS are compared to entries based on analytical standards. If there is a structural match within a certain mass error and CCS tolerance, the molecule is considered a match. However, if a database search generates no matches, further work is needed. CCS databases are typically generated based solely on commercially available analytical standards. Unfortunately, the availability of such standards limits the size of databases, as many compounds either cannot be isolated or they are simply too expensive to obtain. For these molecules, CCS matching may only be feasible against predicted values generated by computational methods (110–112). Such computational approaches have shown promise in generating CCS values, usually with <2% agreement with experimental values (113). As more CCS values are published for the various IMS platforms, it is expected that predicted values will be able to fill the gap of unavailable standards for the identification of unknown metabolites.

IMS-MS also allows identification of metabolites through monitoring the placement of the analyte’s observed CCS as a function of the measured m/z. For example, lipid molecules are typically characterized by their headgroup, length of the fatty acid tails, and the number of double bonds (114). These characteristics make the three-dimensional gas-phase structure of lipids quite rigid and, as a consequence, lipids have, on average, larger CCS values than peptides, carbohydrates, and nucleotides with similar masses, as shown in Figure 2 (99, 115). Several studies have noted these resulting mass/CCS ratios and have generated analytical trend lines describing the relationship between analyte mass and CCS values for a wide range of biomolecules (116, 117). In fact, if both the mass and CCS of an unknown analyte are accurately known, it may be possible to classify an unknown metabolite into a tentative biological class, after including other related factors such as mass defect and isotope ratio pattern. It is also worth noting that while there is significant overlap for several zones of the illustrated mass/CCS relationships in Figure 2, further advances in IMS resolving power and selectivity continue to increase the likelihood of producing better-resolved features in the CCS dimension.

Figure 2.

IMS-MS plots showing the regions occupied by (A) lipids and peptides; (B) subclasses of antibiotics; (C) compounds of various densities; and (D) corticosteroid and nonsteroidal anti-inflammatory drugs (NSAIDS). Structures are shown for: cephalexin a cephalosporin antibiotic, benzalkonium C12 an amphiphilic ammonium, clioquinol an antifungal drug, ibuprofen a common NSAID, and cortisone a common corticosteroid. Reproduced from Ref. 100 with permission from the American Chemical Society ©2017.

NUCLEAR MAGNETIC RESONANCE

NMR spectroscopy exploits the quantum mechanical interactions of atomic nuclei with an external magnetic field. These interactions arise because some nuclei have an intrinsic type of angular momentum called spin. In metabolomics applications, the most common nucleus to measure is ¹H, with some applications using ¹³C (e.g., 118–121), ³¹P (e.g., 122), or other nuclei. All of these biological nuclei are spin 1/2, which means that when they are in an external magnetic field, they can adopt two energy levels separated by the resonance frequency (ω₀) that is proportional to the magnetic field strength (B₀), according to the equation

$${\omega }_0=-\gamma B_0,$$

where γ is the gyromagnetic ratio, which is a physical constant for a specific nucleus. Thus, ¹H resonates at 600 MHz in a 14.1-Tesla (T) magnet and 900 MHz in a 21.1-T magnet. Because of the quantum mechanical underpinnings of NMR spectroscopy, it provides atom-specific information, which makes it the method of choice for the structural characterization of unknown molecules. It can be quantitative, with the integrated value of each NMR proportional to the number of nuclei and thus the concentration of the molecule. NMR spectroscopy is also nondestructive and highly reproducible because the sample never comes into direct contact with the instrument. But all of these significant strengths come at a cost of overall sensitivity. Because it is a resonance phenomenon, NMR spectroscopy has a fundamental sensitivity that is limited by the Boltzmann equation

$$\frac{N_{\text{up}}}{N_{\text{down}}}=e^{-\left(\frac{\Delta E}{k_{B}T}\right)},$$

where N_up and N_down represent the number of nuclear spins in the upper and lower energy levels, ΔE is the energy gap between levels, k_B is the Boltzmann constant, and T is the absolute temperature. For ¹H at 600 MHz and room temperature, if the number of spins in the upper energy state is 1 million, there are only 1 million + 96 spins in the lower state, so only a small fraction of the sample contributes to the NMR signal. But the low energies associated with NMR spectroscopy also allow its noninvasive application in living systems through magnetic resonance imaging.

ONE-DIMENSIONAL NUCLEAR MAGNETIC RESONANCE

NMR metabolomics applications are typically done without chromatography or significant sample extraction steps. Therefore, the measured signals represent a complex mixture of metabolites in a sample. With modern spectrometers and probes, the practical lower limit of detection is about 10 µM in a 550-µL sample. The most common experiment in NMR metabolomics is a 1D ¹H spectrum, which can have hundreds to sometimes thousands of overlapping peaks. These can be matched to databases with standard spectra of known metabolites. The most important public databases with NMR spectral libraries are the Biological Magnetic Resonance Data Bank (BMRB) (123) and HMDB (124). The primary difficulty in using these databases is that 1D NMR spectra can be heavily overlapped and, thus, there are almost always uncertainties in peak assignment using exclusively 1D methods.

TWO-DIMENSIONAL NUCLEAR MAGNETIC RESONANCE

Two-dimensional NMR offers significant advantages over 1D NMR. It not only reduces resonance overlap by spreading the signal into a second dimension, but it also can provide extra information about chemical bonding between nuclei. The drawback of 2D NMR is the length of time required for each experiment, so these are typically only used for pooled samples rather than every sample in a study. However, new approaches are improving the speed of 2D methods (125). One of the most useful 2D experiments in metabolomics is heteronuclear single quantum correlation (HSQC). The 2D HSQC experiment correlates ¹H with ¹³C (or less common in metabolomics, ¹⁵N) that are covalently bonded. Each pair of bonded nuclei give a single peak in a 2D HSQC, and this provides a useful fingerprint of a mixture. Edison & Schroeder (126) wrote a more complete description of 2D NMR experiments and their interpretation.

NUCLEAR MAGNETIC RESONANCE DEREPLICATION

Before the difficult step of unknown compound identification, it is important to first recognize and assign peaks that are known and in databases. This is called dereplication. There are several approaches to this, as both freely available and commercial packages. One of the most popular commercial software is Chenomx, which allows users to fit a library of reference standards to 1D ¹H experimental metabolomics data. This is an excellent visualization tool but suffers from the problems mentioned above about 1D NMR and peak overlap. At least one study reports inconsistent results with Chenomx using the same data set and multiple analysts (127). Bruker Corporation offers a spectral database that includes a wide range of pH values for many common metabolites, and this can be used with both 1D and 2D data in its Assure software. The 2D data add confidence in annotation, but the cost of this solution may be beyond the budget of many labs.

Brüschweiler’s laboratory has developed a suite of free web-based tools called COLMAR (complex mixture analysis by NMR; http://spin.ccic.ohio-state.edu/index.php/colmar). COLMAR has several functionalities that are useful for metabolomics. One of the simplest ways to use COLMAR is the ¹H-¹³C-HSQC query, which takes a peak list from an HSQC spectrum and finds database matches using a combination of BMRB, HMDB, and internally curated data. HSQC matches are useful but can also be prone to misinterpretation, because they only report on a ¹H–¹³C pair and do not include correlations between peaks. COLMAR adds to the HSQC query by allowing the addition of TOCSY (total correlation spectroscopy) or HSQC-TOCSY. The TOCSY experiments provide correlations between coupled ¹H spins. Adding both HSQC and TOCSY or HSQC-TOCSY data significantly improves the confidence of dereplication of known metabolites in a mixture.

COMPOUND ISOLATION

The most straightforward approach to NMR-based compound identification is to use a natural products-like strategy that involves purification of the unknown molecule. The purification steps are typically activity guided, essentially using the desired activity as a detector for the compound of interest. For example, the identification of the mating pheromone for Caenorhabditis elegans involved a series of low-resolution fractionations followed by assays of male-specific attraction (128). Once a fraction is sufficiently pure, both 2D NMR (and MS) data can be obtained and analyzed. There are several advantages of this strategy. First, the focus is on the compound of interest (i.e., the one with the desired biological activity or discriminating power). Second, the limits of detection are defined by the assay and not the NMR spectrometer. This is very important, because unknown molecules at concentrations lower than NMR detection limits can be concentrated and identified if sufficient material is available for the bioassay (129). Finally, the pure (or semipure) compound provides a straightforward way to relate NMR and MS data, which is important for a more reliable identification (see Table 1).

Although it is common to use some type of LC system for fractionation, it is not always best to start with analytical chromatography. It is often simpler to start with simpler solid-phase extraction (SPE) steps, which can be done on a larger scale than LC. Orthogonal SPE (e.g., C18 followed by ion exchange) can be quite effective at quickly simplifying mixtures, even with just a few fractions from each step. If necessary, the crude material from SPE fractionation can then be purified further using LC. This approach was used for the isolation of C. elegans (128) and Panagrellus redivivus (130) mating pheromones (and many other activity-guided fractionation studies).

Alternatively, the assay can be NMR or MS spectra in order to isolate an unknown peak of interest (131). In most metabolomics applications, important NMR or MS features are first determined from statistical analysis. If these features do not match databases, the same purification steps described for activity-guided fraction can be used, and fractions can be screened for the feature(s) of interest. Chemical fractionation is time consuming and not always possible without sample degradation (132).

DIFFERENTIAL ANALYSIS BY TWO-DIMENSIONAL NUCLEAR MAGNETIC RESONANCE SPECTROSCOPY

Frank Schroeder’s lab has developed a technique called differential analysis by 2D NMR spectroscopy (DANS) (133). A review of DANS and other approaches to NMR mixture analysis provides an additional overview (134). Briefly, DANS compares unfractionated high-resolution 2D NMR data sets (COSY) of two different genetic strains of organism, e.g., wild-type and a mutant of interest. In DANS, the data are manually overlaid, subtracted, and adjusted so that peaks common to both spectra will cancel, while peaks unique to one of the genotypes are retained. The DANS strategy does not include integration of MS data, but it does provide a very helpful overview of the major metabolic differences between the two genotypes and also then yields 2D NMR data that can be used to either partially or fully determine the structure of unknowns.

THE SUMMIT APPROACH

The Brüschweiler lab recently developed a powerful approach to link NMR and high-resolution MS data in metabolomics called structures of unknown metabolomic mixture components by MS/NMR (SUMMIT) (135). SUMMIT links MS and NMR data through computation (Figure 3). Chemical purification can be used but is not a requirement. The general idea is that both high-resolution MS and NMR data are collected on the same sample. The high-resolution MS can be obtained through either chromatography or direct infusion, e.g., with an FT-ICR instrument (136). Starting with a feature of interest from the MS data, it is possible to obtain a molecular formula directly from the intact precursor ion provided there is sufficiently high mass resolution.

Figure 3.

The SUMMIT approach is a powerful way to link high-resolution MS and NMR data. Starting from an accurate mass measurement and molecular formula, NMR chemical shift calculations are done on all potential structures consistent with that formula. The computed NMR chemical shifts are then compared to experimental NMR data, and the best match is the most likely structure. Figure from Ref. 136 with permission from the American Chemical Society ©2015.

Once a reliable molecular formula is known, it is possible to enumerate all structures that are consistent with that formula, e.g., by searching the ChemSpider database (http://www.chemspider.com). The difficulty with this step is that the number of possible molecules grows substantially with molecular weight. For example, C₄H₆O₅ (e.g., maleic acid) yields 35 structures from ChemSpider. In contrast, ChemSpider yields 1,023 results for C₂₁H₃₀O₃. Therefore, it is desirable for the number of possible structures to be reduced by other data, e.g., through association with a specific metabolic pathway through a metabolite-genome-wide association study.

The next step is to calculate the NMR chemical shifts of all possible structures from MS data. Calculations of NMR chemical shifts have become quite reliable with high-level ab initio or density functional quantum mechanical calculations (137). However, even semiempirical-based methods can provide reasonable results (135). The computed NMR chemical shifts are then compared against experimental NMR data for the closest match. This conceptually simple step can be complicated when the NMR data are from a complex and unfractionated metabolomics mixture, so a modification of this strategy would be to fractionate the NMR sample using similar chromatography to the LC-MS. By doing this step, the overall approach becomes similar to natural products fractionation described above with the additional step of using computational chemistry to determine the best structure rather than through traditional analysis.

CONCLUSIONS AND OUTLOOK

Despite the limitations associated with each of the strategies described here, all of them have significantly contributed to addressing the most challenging problem in nontargeted metabolomics studies, which is to know the unknowns (27, 69, 138). Integration of the information produced by such advanced assays, however, is still largely lacking, thus preventing identification of metabolites in a high-throughput fashion. Expected advances in metabolomics informatics pipelines are expected to propel the field to a more mature stage, in a similar fashion to what has occurred in other omics fields such as genomics, transcriptomics and proteomics.