In principle, two major kinds of metabolomic analyses can be applied—targeted and untargeted—which are in detail reviewed elsewhere (Chen et al., 2007; Dettmer et al., 2007; Dunn, 2011; Monteiro et al., 2013; Zhang et al., 2016; Cuykx et al., 2018; Ghanbari and Sumner, 2018; Kind et al., 2018). While targeted analysis will focus on an a priori known number of defined metabolites, untargeted metabolomics or discovery metabolomics aims to capture all metabolomic information in a sample. In the latter, features of interest are filtered after data acquisition applying different uni- and multi-variate statistical methods followed by their identification. A schematic of an untargeted metabolomics workflow, the method of choice for biomarker search, is given in Figure 2. A wide variety of targeted and untargeted methods have already been reported in the literature for the separation and quantification of components belonging to the metabolome. However, it was found that no single analytical platform is capable of capturing all metabolomics information in a single run (Dinis-Oliveira, 2014). LC- and GC-MS, nuclear magnetic resonance (NMR) spectroscopy, and LC with electrochemical detection are all used (Ning et al., 2018), but the most widespread analytical instruments utilized are MS and NMR spectroscopy. The advantages and disadvantages of MS or NMR use within the field of metabolomics have been extensively discussed in corresponding reviews (Schlotterbeck et al., 2006; Pan and Raftery, 2007). In summary, MS shows much better sensitivity and resolution and the ability for high-throughput applications, while NMR profits from a comprehensive coverage of chemical species (Chen et al., 2007). NMR analysis certainly provides several advantages in metabolome studies, however, GC- and LC-MS platforms are more widely available in forensic toxicological laboratories (Drummer, 2007; Maurer, 2010; Peters, 2011; Meyer et al., 2014). For introducing a prepared biological sample into a mass spectrometer, GC, LC, direct injection or capillary electrophoresis can be used (Dettmer et al., 2007). In recent years, LC-MS techniques gained importance having the advantage of simpler sample preparation approaches compared to e.g., GC techniques where one- or two-step derivatization is usually mandatory. Avoiding numerous, tedious sample preparation steps can reduce overall measurement variations and will result in more reliable and comparable metabolomics data. Reversed-phase (RP) methods using C18 stationary phases in combination with mobile phases consisting of water (A) and methanol or acetonitrile (B), with additional formic acid (FA), are often the preferred choice due to their non-specific retention mechanism. Although this method is powerful in separating many—especially lipophilic metabolites such as steroids (Marcos et al., 2014) or endocannabinoids (Pastor et al., 2014), use of this single platform is non-optimal, lacking retention for polar metabolites and resolution for many apolar metabolites. Hydrophilic liquid interaction chromatography (HILIC), capillary electrophoresis, and ion-pairing reversed-phase chromatography are solutions often described in metabolomic applications to increase the separation of polar metabolites (Cuykx et al., 2018). For targeted analysis, all kind of MS devices, including triple quadrupole instruments, can generally be applied. For untargeted screening approaches, MS instruments with high-resolution mass measurements, such as time-of-flight (TOF), quadrupole TOF (qTOF) or Fourier transformation (FT) e.g., Orbitrap mass spectrometers, are preferred. Generally, some kind of MS/MS data acquisition, mostly based on data-dependent acquisition (DDA), is used to generate further MS information for identification. Data independent acquisition can increase totality of MS/MS information, but at present cannot be handled by many (commercial) software solutions for data analysis (Boxler et al., 2018a). Knowledge of accurate masses provides the basis for peak identification across different samples as it allows for calculation of empirical formulae and facilitates feature identification using online or commercially available databases such as METLIN (Guijas et al., 2018), the Human Metabolome Database (Wishart et al., 2007) (HMDB, V4.0), NIST (Linstrom and Mallard, 2001), and Lipidblast (Kind et al., 2013) as well as a priori unknown identification.

If no reference material is available for the metabolites of interest—as will be the case in untargeted metabolome studies—comparison between groups and/or conditions will be performed on the basis of relative abundances. Reproducible measurements thereby are a prerequisite for reliable data processing and analysis. It is evident, that method validation in the classical sense is impossible for untargeted metabolomics. Nevertheless, quality control (QC) is essential and needs to be implemented (Dunn et al., 2012; Cuykx et al., 2018). Different strategies—at best in combination—are generally accepted: addition of internal standards (IS) into each analyzed sample; continuous measurement of system suitability tests (SST) over the whole analytical batch containing a defined number of metabolites ideally evenly distributed over the chromatographic run; and inclusion of QC pool samples prepared from all experimental samples, hence representing the average of the data set (Dunn et al., 2012; Broadhurst et al., 2018). Further, randomization of the samples at least for analysis is an important step in untargeted metabolomics to prevent technical/instrumental biases (Dunn et al., 2012).

Typically, all samples of one metabolome experiment—ideally collected and stored under the same controlled conditions—will be measured within the same batch in order to avoid bias caused by sampling, storage or day to day variations in instrument performance. Recently, Nielsen et al. performed an untargeted metabolome experiment on a retrospective dataset collected for forensic toxicology routine analysis. There, a subgroup of samples positive for 3,4-methylenedioxymethamphetamine (MDMA) was compared to a negative control group. Only small retention time shifts were observed across all sample chromatograms, despite the fact that data was collected over a long period of time. Normalization by the NOMIS method which utilizes variability information from multiple IS compounds to find an optimal normalization factor for each individual feature (Sysi-Aho et al., 2007), also enabled the distinction between biological variation from obscuring variation related to sample preparation and ion source variation. The fact that MDMA and its phase I and II metabolites were positively identified and clearly up-regulated in MDMA users (specified by almost all applied methods) proved successful overall data-analysis. These findings suggest that analysis of retrospective data is generally possible, e.g., if samples are initially collected in the same sampling tubes, sample storage and preparation is highly standardized, and routine methods are quality controlled as is typically the case in routine forensic analysis (Nielsen et al., 2016). Mollerup et al. successfully applied a similar strategy to search for new markers/metabolites of valproic acid ingestion amenable to positive electrospray ionization (ESI) (Mollerup et al., 2019).

Metabolomics platforms generate large and complex data highlighting the need for appropriate data processing tools that allow the preparation of chromatographic and spectral data for multivariate data analysis (Katajamaa and Oresic, 2007). Often procedures like data condensation and reduction (by means of centroiding and deisotoping mass spectra), chromatographic alignment (to prevent misinterpretations due to retention time variations), filtering (for removal of noise or background signals) and peak recognition, and collection [by applying threshold windows for mass (m/z) and retention time] are used in this context. Details on different options of data processing can be found e.g., in a recent publication of Cuykx et al. (2018). To achieve minimal influence of systematic and sample biases (e.g., degree of urine dilution), it is recommended to normalize MS data either by the parameters of the whole dataset (e.g., total ion count, median ion count, etc.) or by the intensities of multiple or single ISs (Sysi-Aho et al., 2007). Data of biological origin is often skewed and commonly quantitative data are mean-centered, log-transformed, and normalized (unless absolute quantification is carried out). As every data processing step (for example filtering, scaling, peak picking, missing value imputations, and normalization) can have a significant impact on the interpretation of experimental results, it needs to be adequately described in the method sections (Yin and Xu, 2014; Alonso et al., 2015).

Different kinds of statistical tests—uni- and multi-variate—are usually performed for data interpretation. Univariate tests (t-test; ANOVA) are straightforward and compare the intensities of single features between different groups. While the interpretation of these tests is very clear, the need for multiple repetitions for hundreds of variables in metabolomic studies increases the risk for detection of false positive features (Broadhurst and Kell, 2006; Kim and van de Wiel, 2008). This should ideally be accounted for by applying false discovery corrections (e.g., Bonferroni or Benjamini–Hochberg correction). The most widely used unsupervised multivariate technique in science is principal component analysis (PCA). It projects the maximum variance of a multi-dimensional space in principal components and summarizes the data set in a limited number of components. PCA is mainly used as an exploratory technique as it is unsupervised and hence does not explicitly account for class-based separations. In contrast to this, (orthogonal) partial least-square discriminant analysis [(O)PLS-DA] is a supervised multivariate technique, which describes the greatest variance to differentiate between experimental classes, with the aim to find the metabolic patterns that are most important for the classification. Based on S-plots or variable importance in projection (VIP) values, the metabolites that have a large impact on the projection are selected. Supervised multivariate approaches are prone to overfitting to irrelevant or noisy features. Therefore, cross-validation procedures need to be carried out (Broadhurst and Kell, 2006; Gromski et al., 2014; Cuykx et al., 2018).

Knowledge on a drug's biotransformation is of major importance for a variety of questions. For urine screening approaches, particularly if the parent compound itself remains undetectable in urine specimens, effective detection of drug intake will only be possible over one or several of its (unique) metabolites. Usually, focusing on highly abundant/major metabolites will be sufficient. However, in case of common major metabolites for several structurally related compounds, other minor metabolites might be necessary to finally prove intake of a particular drug. Also, if metabolism studies are performed in other than human species (e.g., rats or mice), investigation of minor metabolites in these studies is very important, as different excretion patterns in humans are likely. Hence, a minor metabolite in an animal study could become a major metabolite for human excretion. One major challenge in urine drug metabolism studies represents the definite identification of xenobiotic metabolites considering that urine contains a large variety of chemical species. Common strategies to identify metabolites involve the use of processing filters to exclude expected metabolites. For instance, typically occurring mass differences derived from prevalent, described metabolic reactions such as oxidation (+16) or demethylation (−14) can be extracted in a targeted manner. Additionally, thorough evaluation of MS/MS data or typical MS/MS patterns can be the basis for structure elucidation of new metabolites (Kim et al., 2018). However, the major disadvantage of these strategies is their restriction to a priori known, typical changes. More individual biotransformation e.g., dealkylation, hydrolysis, peroxidation or structural rearrangements are typically not covered (Guengerich, 2001; Chen et al., 2007; Kim et al., 2018). Untargeted metabolomics techniques were used as an alternative approach to detect new (uncommon) drug metabolites. A review by Chen et al. extensively describes and discusses the use of metabolomics in drug metabolism research (Chen et al., 2007). An overview on DOAs or compounds of forensic interest applying these techniques is given in Table 1.

Steuer et al. used an untargeted metabolome approach to search for new biomarkers of gamma-hydroxybutyric acid (GHB) intake. GHB's use as a knockout drug in cases of drug facilitated crimes makes it particularly important in forensic toxicology. The detection of GHB and particularly differentiation between exogenous intake and endogenous base levels remain challenging because of its extremely short detection windows (only up to 12 h in urine) caused by a fast metabolism. In contrast to many other DOAs, for GHB no metabolite has been identified so far, that allows for longer GHB detection compared to using the parent compound itself. Analysis of urine samples collected 4.5 h after GHB or placebo intake of a randomized, double-blind, placebo-controlled crossover study in 20 men allowed identification of novel GHB metabolites GHB carnitine, GHB glycine, and GHB glutamate as exemplified in Figure 3. However, more studies addressing quantitative values, pharmacokinetics, and stability are demanded for a final conclusion on the routine applicability of these markers (Steuer et al., 2018c). Mollerup et al. performed an interesting omics-based retrospective analysis to identify potential markers of valproic acid in blood that should allow the detection of valproic acid intake using the commonly applied positive ESI-MS mode. The antiepileptic drug valproic acid represents an important compound in forensic toxicological analysis, but can only be detected with negative ionization techniques or by GC-MS (Mollerup et al., 2019). A retrospective data evaluation of routinely measured samples on a qTOF instrument in ESI positive mode were performed forming a valproic acid positive group (determined by an additional targeted valproic acid method) and a negative reference group. The authors were able to identify eight potential (indirect) targets for valproic acid (Mollerup et al., 2019).

Application of alternative screening methods not directly targeting the analyte's or its metabolite's chemical structures but e.g., aim on certain endogenous biomarkers, seem to be a desirable approach that would facilitate the complex analytical scenario to detect NPS or chemical adulteration. Different strategies have been applied so far and are summarized in Tables 2, 3: (a) search for analytical, endogenous markers after a certain stimulus such as drug intake; (b) finding markers which derive from a common drug preparation, e.g., herbals used for “spice” products or non-physiological ingredients of artificial urine products; (c) search for endogenous markers that can level out inter-individual variations; (d) search for markers of urine manipulation attempts; and (e) identification of stable/unchanged markers that can proof integrity of a urine sample. Overall, data on such approaches for actual NPS are scarce and only few studies are available on common DOAs. Such studies—particularly in humans require highly controlled conditions which are of course ethically restricted for illegal drugs in many countries.

Metabolomics combined with DOAs did not only focus on biomarkers indicating acute drug consumption but also on identification of hints for drug addiction, the severity of drug addiction or the interpretation of the severity of a given intoxication. While these studies certainly do not help in terms of new strategies for NPS detection, they can be useful in interpretation of clinical or forensic analytical results. Currently available literature focused on cocaine, crack, heroin, and MA. An overview of the analytical methods used and potential markers identified is given in Table 4.

For example, Costa et al. performed ¹H NMR-based metabolomics analysis of crack users' serum samples aiming to investigate whether drug dependency changes the endogenous profile and to further identify potential biomarkers that might be linked to brain dysfunction. The rationale of the study was the current lack of reliable diagnostic tools for crack dependency that at present mainly relies on self-reporting, medical history and physical examination. Early diagnosis should, however, result in better treatment outcomes. Serum samples of two groups were compared, crack users on the one hand against healthy individuals with similar age, gender and body mass index on the other hand. Differences were observed particularly in lactate, acylcarnitines, histidine, tyrosine, and phenylalanine which according to the author's opinion could be linked to altered brain functions. PLS-DA obtained 89.8% accuracy in differentiation of crack users from healthy controls. However, these observations must be considered cautiously due to confounding factors such as medical treatment of some crack users and the small sample size of drug users examined in general (Costa et al., 2018).

An untargeted metabolome analysis using GC-MS in rats receiving heroin for 10 days followed by a withdrawal of 4 days indicated increased myo-inositol-1 phosphate levels and decreased threonate concentrations in serum. In contrast to other biomarkers observed, these levels did not restore to baseline even after heroin withdrawal for 4 days. Therefore, these compounds were discussed as potential indicators of heroin abuse even when the consumer has been abstinent from heroin for some days. An even more sensitive marker would be the ratio of myo-inositol-1-phosphate to threonate. The drug morphine, which is also the main metabolite of heroin, interestingly did not result in the same changes as described for heroin itself. This, in the author's opinion, might potentially allow a differentiation of heroin addiction from morphine dependency (Zheng et al., 2013). However, so far, these findings have never been confirmed in larger (human) studies.

As already described in the Search for Analytical, Endogenous Biomarkers of Drug Intake, Shima et al. identified 5-oxoproline, saccharic acid, uracil, 3-HB, adipic acid, glucose, glucose 6-phosphate, fructose 1,6-bisphosphate, and TCA cycle intermediates like fumarate, as potential biomarkers helpful to assess the severity of MA-induced intoxications (Shima et al., 2011). A follow-up study on chronic MA intake and MA addiction to proof the initial findings to be specific for acute MA toxicity resulted in quite different results compared to previously found acute effects. As a conclusion, different metabolomic changes can be observed depending on the MA tolerance—either acute or chronic consumption. MA tolerance after chronic consumption could have resulted in several adaptations with no significant changes in the metabolite levels (Zaitsu et al., 2014).

The last and major group of currently available metabolome approaches linked to DOAs include studies primarily aiming to elucidate the mechanisms of drug action. While these studies were not designed to identify biomarkers, they might serve as a basis to further evaluate the general applicability of metabolome approaches for biomarker discovery. For example, they might provide first hints or estimates to assess the specificity for a particular DOA, the time frames of observed effects or confounding factors caused by other drugs or diseases. Observed metabolome changes, study design and analytical conditions of these studies are summarized in Table 5. A good summary can also be found in three recent review articles on DOAs and metabolomics (Dinis-Oliveira, 2014; Zaitsu et al., 2016; Ghanbari and Sumner, 2018). For a more detailed discussion on the mechanistic effects, the interested reader is referred to these works. The current review will just use results from the publications that appear helpful in the context of biomarker detection and evaluation of their general applicability.

As can be seen in Table 5, certain changes in endogenous compounds are detected for all studied DOAs. However, the observed changes, albeit statistically significant, appear rather small (Figure 5), particularly when they should allow for differentiation in a non-controlled setting such as drug testing of random subjects. Furthermore, very often similar compounds or in general the same pathways e.g., the energy metabolism or the TCA cycle, are affected. From a pharmacological point of view this is not surprising. As stated e.g., by Ning et al. neuronal activity is extremely energy demanding, and the brain energy supply requires oxidative metabolism of glucose in mitochondria and demands lactic acid from glycolytic processes (Ning et al., 2018). However, albeit e.g., MA and GHB act at different pharmacological targets both were shown to influence succinate concentrations (Zheng et al., 2014; Palomino-Schatzlein et al., 2017). It remains to be determined whether or not the observed changes might ever be able to specifically indicate consumption of a particular DOA/NPS or substance groups or drug use in general. Most likely, changes of single endogenous metabolites will be too unspecific for a certain drug or even drug class. Evaluating fingerprints or changes in particular metabolic pathways appear more promising, but currently lack sufficient studies to support or reject this hypothesis.

Also, other confounding factors need to be considered such as underlying diseases. For example, Mannelli et al. found elevated levels of N-methylserotonin in plasma of human opioid abusers. N-methylserotonin is an analog to serotonin and similar to the derivative bufotenine, both are known for their hallucinogenic and psychotropic effects. However, also in terms of psychiatric disorders associated with hallucinations and altered perceptions elevated levels of e.g., serotonin can be found (Takeda, 1994; Takeda et al., 1995; Mannelli et al., 2009).

The kind and state of drug consumption of course also influences the outcome on the metabolome. Addiction, withdrawal and relapse can also change the metabolome in partly similar ways as acute drug consumption (Zheng et al., 2013, 2014). In some studies, however, drug addiction/chronic drug use may change and/or eliminate the observed effects as e.g., shown for MA (Zaitsu et al., 2014) (see Metabolomics for Biomarker Search of Drug Addiction).

Well-designed controlled studies in rats can give first insights into the duration of the metabolome effects. For example, multivariate analysis (PLS-DA plots) showed that after withdrawal from MA for 2 days (after initial 5 days intake) metabolite values of those rats clustered close to the values of the control group, yet not overlapping with the control data. This is suggestive of an efficient restoration of urine metabolites to baseline levels after withdrawal (Zheng et al., 2014). In contrast to MA, metabolic changes induced by heroin recovered more slowly and were more pronounced (Zheng et al., 2013, 2014). For instance, heroin withdrawal for 4 days did neither restore elevated serum myo-inositol-1phosphate levels nor decreased serum threonate to values prior to heroin consumption. Similar findings were observed for 9-(z) hexadecenoic acid in serum and hydroxyproline in urine with little effect of heroin withdrawal on their concentration levels (Zheng et al., 2013).

The currently available knowledge is of course too preliminary to draw any final conclusion. It definitely needs to be considered that all studies mentioned here were performed for a totally different aim. An overall, critical discussion on these studies as well as the few studies performed for actual biomarker research will be provided in section Current Limitations and Discussion.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.