Metabolomics experiments generate large data sets that require specialized tools for analysis. Numerous software packages for data pre-processing and statistical analysis are available and have been reviewed elsewhere ^{40, 41} . Unfortunately, no single software exists that can simultaneously perform all of the critical steps needed for an analysis of a combined NMR and MS data set. Although the statistical techniques applied to NMR and MS data sets are largely the same, each technique requires a unique set of pre-processing tools and algorithms prior to modelling. For example, an NMR spectrum has to be Fourier-transformed and phased, whereas centroiding and de-isotoping are required in MS. Owing largely to these data type–specific processing requirements, newly developed software is almost exclusively restricted to one method or the other. Conversely, there has been minimal effort in developing tools capable of working with both NMR and MS data sets ⁴⁰ .

There are two general approaches to integrating NMR and MS data sets into a single coherent study. The first involves samples simply being independently analyzed by each method. The separate data sets then are compared in order to identify consistencies in the metabolic alterations observed by each technique. The main advantage of the approach is simplicity since it does not require any significant protocol changes. Also, the confidence of a metabolite assignment may be significantly increased if it is identified by both methods. Furthermore, a measure of internal consistency may be achieved if metabolite concentrations can be estimated by both methods. However, significant information can be lost during this process since, for example, ambiguity in peak assignments sometimes can be resolved by information from the other method. There is also a lack of statistical correlation since the data sets are independently analyzed. Although the manual curation of independent data sets is the dominant method currently used by metabolomics investigators, it also suffers from reproducibility problems due to potential biases in data interpretation (for example, metabolite assignment methods) among other issues.

The second approach to combining NMR and MS data sets is to simultaneously integrate each data set into a single statistical model using a multiblock analysis. Multiblock analysis encompasses a variety of methods that combine multiple data sets prior to conventional multivariate analysis. In addition to combining multiple instrumental data sources ⁴² , multiblock analysis has been successfully employed to combine data sets from different omics disciplines ⁴³ . Multiblock methods are preferable to independent analysis since the relative contributions of each data set still can be quantified, but importantly the larger combined data set is likely to result in models with greater predictive ability and resolving power than either method alone ⁴⁴ . However, the software tools to perform multiblock analyses are crude and often rely on custom sets of pre-processing routines using multiple software packages. The lack of integrated analysis tools and software is a major roadblock in metabolomics, especially in light of the growing interest in combining NMR and MS data sets.

NMR metabolomics investigations, especially those concerned with achieving a high confidence in metabolite identification, require two-dimensional NMR methods to resolve the overlap present in one-dimensional spectra. In general, this requires isotopic labelling with NMR-active nuclei like ¹³C and ¹⁵N because of their low natural abundance. In the last few years, dynamic nuclear polarization (DNP) has evolved from a structural biology tool in the area of solid-state NMR to have potential applications in solution-state metabolomics ⁴⁵ . In DNP, a solid, frozen metabolomics sample at about 1.5 K is polarized in the presence of microwave-irradiated free-radicals, which induces a temporary hyperpolarization in spin-active nuclei through a transfer of polarization from electrons to nuclei. The sample then needs to be rapidly melted and transferred to an NMR spectrometer to take advantage of the greatly enhanced sensitivity (>10,000-fold) ⁴⁶ . The dramatic increase in sensitivity avoids the need for isotopic labelling, especially for in vivo samples, and may permit the detection of low-abundance metabolites. Nevertheless, DNP experiments are limited by T ₁ relaxation rates, resulting in a short measurement window of the dynamically polarized samples. DNP also requires substantial hardware modifications and accessories (for example, microwave generator) to rapidly thaw and shuttle samples back and forth from the NMR spectrometer. DNP has also been applied to ¹³C-labeled metabolites that then are used as a tracer compound for in vivo imaging ⁴⁷ . This requires close proximity of the polarizer and magnetic resonance imaging spectrometer to allow for rapid transfer, dissolution, and injection of the ¹³C-labeled metabolite given the relatively short T ₁ of 30 to 40 seconds for a ¹³C-labeled carboxyl group. Despite these technical obstacles, DNP has been successfully used to monitor a single metabolite (for example, pyruvate) in living tissue (for example, heart) by magnetic resonance imaging ⁴⁸ . Besides the short measurement time, another challenge with the application of DNP to in vivo imaging is the limited number of ¹³C-labeled metabolites that can be polarized and tolerated at the concentrations needed for imaging (25 to 80 mM) and that are also a useful biological probe. In addition to pyruvate, bicarbonate, fumarate, urea, glutamine, and dehydroascorbate have been used for in vivo imaging ⁴⁷ . Despite fundamental issues of reproducibility and limitations in sample preparation, DNP protocols and technology are rapidly advancing and one day could become a routine tool for metabolomics ⁴⁹ .

Metabolomics can be used to profile an individual’s responses to a drug treatment or other medical therapy by monitoring metabolite changes in readily obtainable biofluids (for example, blood and urine). A unique advantage of metabolites as biomarkers is the likely occurrence of observing a set of multiple metabolites with distinctly different concentration changes that are correlated with a disease state or treatment response. Correspondingly, multiple metabolites, instead of a single biomarker, are expected to yield a higher sensitivity and selectivity. For example, plasma baseline levels of xanthine, 2-hydroxyvaleric acid, succinic acid, stearic acid, and fructose prior to simvastatin treatment were observed to reliably predict a good or poor response in reducing low-density lipoprotein cholesterol ⁵⁰ . The OPLS model yielded a 70% sensitivity and 79% specificity with a corresponding area under the receiver operating characteristic curve of 0.84. Thus, metabolomics can be used to predict whether a patient will respond to a drug in addition to being used as a semi-quantitative prognosis of disease progression.

For example, in a recent study of patients with tuberculosis (TB), urine samples that were collected over the course of a 6-month period became more similar to those of a non-TB control group during the course of first-line anti-TB therapy (for example, isoniazid, ethambutol, or pyrazinamide). Metabolomics has also been successfully employed to identify serum metabolic alterations associated with psoriasis ⁵¹ . Importantly, the metabolomics results were consistent with trends previously observed in genomics and proteomics studies. The metabolome changes were observed to reverse following successful corticosteroid treatment ⁵² . Interestingly, the authors identified an increased demand for glutamine, which had not been previously reported in psoriasis ^{52, 53} . Glutamine demand is directly associated with diseases characterized by increased cellular proliferation, such as in cancers. A significant alteration in β-isosterol, which is a commonly employed herbal remedy, was also observed. Thus, metabolomics may also be used to identify a patient’s use of alternative treatments outside of his or her physician’s knowledge or recommendation. In this manner, metabolomics may assist in determining whether co-administration of a complementary treatment was beneficial or detrimental to a patient’s therapeutic outcome.

Much of the data analysis approach in metabolomics has been largely borrowed from the field of chemometrics, which pioneered the application of PCA and PLS to chemical systems ⁵⁴ . Although these are powerful statistical tools, the current trend in metabolomics data analysis is evaluating the efficacy of new algorithms and statistical methods to improve group separation and metabolite identification. PCA, PLS, and OPLS are all commonly employed by metabolomics investigators, but newer approaches, including support vector machine (SVM) ⁵⁵ , random forest (RF) ⁵⁶ , and self-organizing map (SOM) ⁵⁷ algorithms, have all been recently applied to metabolomics data sets.

Despite having been formalized since 1992 ⁵⁸ , SVM has been used extensively only in the analysis of gene microarray data, particularly due to its performance on data sets characterized by a large number of variables and few samples ^{59, 60} . SVM is also able to identify non-linear relationships that violate the linearity assumptions of PCA and OPLS, making it easily generalizable. SVM has been recently applied in biomarker discovery for ovarian cancer, and a model using serum-derived LC-MS spectra was able to predict disease onset with higher accuracy than the currently accepted method of CA-125 serum monitoring ⁶¹ . A major caveat of SVM is its restriction to binary classification problems: it is able to discriminate between only two experimental groups. Simply, spectra belonging to two experimental groups are represented as points in n-dimensional space, where n corresponds to the number of observed metabolites. A hyperplane then is calculated that best separates the points from the two groups. The coefficients of the calculated hyperplane are used to determine which metabolites are most important for discriminating between the two groups. Although methods have been proposed to extend SVM to multi-class problems, they are often done by breaking down the data set into an ensemble of binary groups that oversimplifies the problem and leads to uninformative models ⁶² .

The RF algorithm is a decision tree–based method that uses random subsets of the data to construct multiple models, which then are combined to create an average model in a process known as bootstrap aggregation. In the decision tree method, samples are mapped to a target value (that is, which experimental class the sample belongs to) using a set of variable-based decision rules that separate the samples into groups corresponding to the target value. These newly formed groups can be further subdivided according to new variables, and each “branch” of separation is repeated until the samples can be fully differentiated. The major advantage of the decision tree is its imperviousness to scaling and variable normalization, an extremely common problem in metabolomics data ⁶³ . The disadvantages include an extreme propensity for overfitting and having extremely poor generalizability that severely limits its utility. RF addresses this limitation by creating an ensemble of partial decision trees that, when combined into an overall model, reduces variance and overfitting ⁶⁴ . In particular, the RF algorithm, being relatively unaffected by scaling and normalization and easily handling both large data sets and missing values, is highly adaptable to the realities of real-world data sets. A major disadvantage of RF is that the method requires extensive “tuning” of default parameters by the investigator in order to obtain the best model. Also, the resulting decision tree can be hard to visualize for large data sets ⁶⁵ . RFs have shown clinical value: they have been used to determine a set of serum protein and metabolic biomarkers in prostate cancer with higher predictive accuracy than the current prostate-specific antigen biomarker ^{66, 67} . See Gromski et al. (2015) for an excellent review of the SVM and RF algorithms that also includes comparisons with other mainstream techniques ⁶⁵ .

SOM is an approach similar to PCA that reduces multi-dimensional problems to a more easily interpretable low-dimensional grid to visualize natural clustering trends and groupings within a data set. SOM can be applied to the same tasks as PCA but without the biases toward high-variance metabolites. SOMs, like SVM, have the ability to detect non-linear relationships between detected metabolites ⁶⁸ . SOMs have been successfully applied to develop biomarkers for early-stage renal cell carcinoma as well as to predict patient response to surgical intervention with a predictive accuracy of 94.74% ⁶⁹ . In comparison with the other statistical methods, SOM has been severely limited in metabolomics because of a computationally intensive algorithm and the lack of a pre-packaged software, which has significantly diminished its accessibility to the wider research community ⁷⁰ . Nevertheless, the usage of SOMs in metabolomics is steadily rising, and comparative analyses are beginning to demonstrate that SOMs are an acceptable alternative to more traditional clustering algorithms ⁷¹ .

In addition to statistical methods applied directly to spectral profiles, identified metabolites can be used with pathway analysis ⁷² to understand metabolite interactions with known pathways or to discover mechanisms of action in pharmaceutical natural product research ⁷³ . Metabolomics data sets can generate an overwhelming and seemingly disjointed list of metabolites, which pathway analysis aims to place into a broader biological context by assigning metabolites to relevant metabolic pathways. This is done through a number of software tools that integrate putatively identified metabolites with pathway information from various databases. For example, MetaboAnalyst 3.0 is a suite of metabolomics tools ( http://www.metaboanalyst.ca), which includes modules for metabolite enrichment analysis (MSEA), metabolite pathway analysis (MetPA), and an integrated pathway analysis. The user input is typically a list of metabolites (with or without concentrations) or genes or both. MSEA provides a ranked list of potentially key metabolic pathways based on the observed number of metabolites associated with that pathway (that is, metabolite set enrichment) ⁷⁴ . MetPA combines MSEA with a pathway topology analysis to provide an overall pathway analysis to identify the metabolic pathways primarily impacted in the study ⁷² . The integrated pathway analysis combines both metabolomics and genomics data with enrichment analysis and topology analysis to again identify the pathways (in rank order) that were primarily impacted in the study ⁷⁵ . MetaboSignal ( https://bioconductor.org/) is an alternative approach to pathway analysis which employs directed graphs with network topology approaches to compute centrality measures to correlate gene-metabolite relationships through shortest-path distances ⁷⁶ . Thus, unlike the MetaboAnalyst 3.0 tools, the output of MetaboSignal is a network map of gene-metabolite connectivities. Cytoscape ( http://www.cytoscape.org/) is a generalized network interaction and visualization tool that works with a variety of data sets, including metabolomics data. Cytoscape combined with MetScape 3 ( http://metscape.ncibi.org/) ⁷⁷ can generate network maps similar to those of MetaboSignal from metabolomics or genomics data or both ⁷⁸ . MetScape uses known pathways from Kyoto Encyclopedia of Genes and Genomes (KEGG) ⁷⁹ and Edinburgh Human Metabolic Network (EHMN) ⁸⁰ databases and gene set enrichment analysis to generate these networks in order to visualize the impacted metabolic pathways. In essence, there is some significant overlap in the capabilities of MetaboAnalyst 3.0, MetaboSignal, and Cytoscape/MetScape 3. Importantly, pathway analysis provides an interaction network that may identify centralized hubs where metabolic pathways coincide or where bottlenecks may occur.

The limited connectivity or altered flow through specific metabolic nodes (that is, change in metabolic flux) may identify functionally essential biological processes ⁸¹ . These essential pathways then can be selectively targeted. By genetically or chemically restricting a potentially essential metabolic pathway, it is possible to ascertain the relevance of the pathway to a systems response to an environmental stress (that is, drug resistance) and potentially reverse or negate the response ⁸² .

Pathway analysis also allows for the integration of multi-modal omics data, such as combining gene-expression and metabolomics data to uncover gene and protein functions. For example, metabolite profiles were integrated with genome-wide screening of single-nucleotide polymorphisms (SNPs) to identify the molecular mechanism of the NAT8 and PYROXD2 genes. Briefly, SNPs were ranked according to the strength of an association with observed metabolites. The regions where these SNPs occur on the chromosomes then were screened to determine at what position in the genome the gene/protein product responsible for the mediation is stored. With this approach, it was suggested that the NAT8 and PYROXD2 genes were responsible for mediation of serum diethylamine levels ⁸³ , a novel insight for these previously under-annotated genes.

As another illustration, transcriptomic and metabolomic data from Arabidopsis thaliana were integrated to characterize the biological response resulting from the over-expression of PAP1, a gene known to cause profound accumulation of anthocyanins and to encode a MYB transcription factor regulating flavonoid biosynthesis. The authors were able to correlate the biosynthesis of cyanidin and quercetin derivatives with a specific set of upregulated genes that enabled them to identify the function of two uncharacterized proteins: a flavonoid 3-O-glucosyltransferase and anthocyanin 5-O-glucosyltransferase ⁸⁴ . Numerous tools are now available for pathway analysis of metabolomics data ^{79, 85, 86} , which will significantly improve data interpretation and simplify our understanding of biological relevance. Thus, pathway analysis is becoming a routine component of a detailed metabolomics analysis.