Review of Mass Spectrometry–Based Metabolomics in Cancer Research

Metabolomics is a new promising “Omics-discipline” in systems biology, claiming to investigate the entire set of metabolites present in a biologic system. It is an analytic approach to detect metabolites and determine their concentrations (1). Metabolomics is interdisciplinary, driven by basic sciences (analytic biochemistry, biology) and bioinformatics together with epidemiology and clinical research. The term “metabolome” was introduced in 1998 as the total metabolite content of a biologic sample (2), whereas the terms for the discipline were coined subsequently: metabonomics in 1999 by Nicholson and colleagues (3); metabolomics in 2000 by Fiehn and colleagues (4). The numbers of metabolites present in the human metabolome is estimated to lie within the range of 10⁴ to 10⁵, with the Human Metabolome Database currently containing about 15,000 entries (5). Variance in the fraction measured can be attributed largely to different analytic and detection methods for metabolites: Mass spectrometry (MS)-based techniques have the specific advantages of being more sensitive and therefore superior in terms of metabolic coverage, compared with nuclear magnetic resonance (NMR; ref. 6). Metabolomics experiments can be subdivided into targeted and untargeted analyses: Targeted studies aim to accurately determine concentrations of a limited and predefined subset of a few metabolites of a predefined class or pathway (commonly tens to hundred), whereas untargeted analyses use a more global approach to cover as many metabolites that can be detected by a given method (7, 8). The human metabolome can be interpreted as being the most downstream endpoint of cellular phenotype, influenced by changes in the proteome or genome (9, 10). It is hypothesized to carry more information on the actual phenotype than the latter two more upstream biochemical levels. Metabolomics is therefore a prime candidate for biomarker development.

Proteomics and genomics have been first developed as standard tools in cancer research and translational medicine, with regard to biomarker development for early detection, classification of tumors, measurement of therapy response, and patient prognosis (11). Both proteomics and genomics have been extensively reviewed with regard to cancer research (12–14). In a recent review, it was noted that metabolomics can be used as a promising approach in clinic1al applications (15). Metabolomic information is of high value because metabolic reactions mirror the function of the cell more directly (9).

This report aims to systematically review the results of recent metabolomic studies with respect to its use in epidemiology and translational oncology. In addition, we give a brief overview of the currently applied MS-based methods and potential future applications.

Three databases (PubMed, Embase, and Web of Knowledge) were searched for the keywords (“metabolomics” and “cancer”) or (“metabonomics” and “cancer”) in all fields (i.e., title, journal, abstract) from 1998 through November 2012. The Web of Knowledge search was further refined by excluding research areas of low relevance (see Fig. 1 and Supplementary Data for search strategy).

Only primary research studies investigating either human tissues or body fluids were included. Any study using human in vitro models or animal models was excluded. Furthermore, we included only studies employing mass spectrometry; approaches using solely NMR were excluded because of the narrow metabolic coverage. We included only studies with more than 10 metabolites measured. A distinction was made between studies using an untargeted approach (comprehensive metabolomics or metabolic fingerprinting) or analyses focusing on metabolites of a component class, i.e., only lipids or amino acids (metabolic profiling; refs. 8, 15). English language and an available abstract were further inclusion criteria.

Studies with irrelevant titles were excluded within a preselection step. Duplicate findings were removed and studies underwent a first screen for the above inclusion criteria, based on their abstract. Full text articles of studies that passed all the inclusion criteria were further reviewed in detail. A flowchart of the study selection process is given in Fig. 1.

After a detailed review of the full text articles, the following data were extracted from each study, if provided:

• Number of cancer patients, diseased controls, and healthy controls included

• Type of cancer

• Number of metabolites investigated (distinction between untargeted metabolomics and targeted metabolic profiling)

• Type of biospecimen investigated

• Platform used for the analytic assessment of the metabolome

• Significantly altered metabolites in patients cancer compared to other groups

• Study type, grouped as follows: (a) case–control comparison, (b) therapy response, (c) patient prognosis, (d) method development, (e) tissue profiling or (d) others

Data extraction was carried out by three independent researchers (D.B. Liesenfeld, N. Habermann and R.W. Owen) to avoid author bias. On the basis of the extracted data, specific indicators were assigned to the studies:

• N: At least n > 100 patients with cancer were included in the study (with an unspecified number of total study subjects)

• V: Two independent study populations (i.e., a discovery and validation set) were monitored

• T: Patients were monitored over time (either prospectively or retrospectively)

• T +: In addition to T, samples were collected repeatedly over time.

To develop a metabolic map of patients with cancer, each study that employed an untargeted profiling approach was screened for reported alterations of identified metabolites. Significantly altered metabolites were extracted and subjected to further analysis. A network was created with Cytoscape (16) using the MetScape v2.33 (17) plugin with “homo sapiens” as reference species. MetScape creates networks based on reaction from pathway information of the Kyoto Encyclopedia of Genes and Genomes (KEGG; ref. 18). Constructed networks can be found in Fig. 4. The Cytoscape file is available from the authors upon request.

In total, 106 studies were reviewed in detail. The descriptive information, i.e., the cancer type studied, sample type used, and type of study, are summarized in Fig. 2.

Colorectal cancer was investigated most often. Interestingly, cancers of the urogenital tract were ranked second, presumably because they were hypothesized to have more intensive and direct contact with the urinary matrix. In total, studies involving 21 different types of cancer were reviewed.

Conventional clinical sample types were studied most commonly (urine, followed by serum and plasma). Twenty-two studies extracted metabolites from tumor tissues, investigating the metabolome of cancer cells directly. Seven studies used less common samples with direct contact to tumor tissue, i.e., cerebrospinal fluid (glioblastoma; refs. 19, 20), exhaled air (lung cancer; ref. 21), saliva (oral cancer; ref. 22), fecal extracts (23), or volatile signatures derived from skin (melanoma; ref. 24). Wedge and colleagues compared plasma and serum and described a generally high overlap in metabolites (78%–87%) indicating that overall discriminatory abilities of the two biospecimens are comparable. They concluded that the choice for either plasma or serum is led clinically, rather than analytically (25).

The number of patients investigated is shown in Fig. 3. Only 15 of 106 studies (14%) investigated samples from more than 100 patients with cancer. Seventy-seven studies (73%) included healthy individuals as controls. Twenty-eight (26%) used additional controls, i.e., patients with benign tumors or other organ-specific conditions (i.e., hepatitis for hepatocellular cancer control) to enhance the specificity of the evaluated biomarkers. Studies investigating tumor tissue commonly used nontumorous (“normal”) adjacent tissue specimens as matched control samples.

A general definition and overview of the current instrumentation used in MS-based metabolomics can be found in the review by Dettmer and colleagues (26). Liquid (LC) and gas chromatography (GC) were equally often used for metabolite separation involving 61 and 59 studies, respectively. Few studies employed other techniques, such as capillary electrophoresis-MS, direct injection-MS (27, 28), or matrix-assisted laser desorption/ionization-MS (MALDI-MS; ref. 29). Accurate mass methodologies such as time of flight (TOF) MS dominate the untargeted metabolomic setups with 60 of 106 studies (57%). Other instruments, i.e., GC-MS with single quadrupole, tandem LC-MS with triple-quadrupole (QqQ), or quadrupole-ion traps (QTRAP) are also frequently used. Figure 3B gives an overview of the analytic setups used in the reviewed studies.

The distribution of different study types is shown in Fig. 2. The majority of studies, n = 77 (73%) examined the use of blood or urine with the intent for future early diagnosis of cancer or a cross-sectional biomarker development. These studies used metabolomics to discriminate cancer cases from healthy individuals. Studies employing this common approach are not discussed in detail here. Most studies used sample sizes of pilot character and did not have independent validation. Eighteen of 106 studies (17%) studied the metabolome of patients in a discovery and a separate validation set. Out of these studies, only 7 had patient numbers more than 100 (see Table 1). In the following, we will discuss these higher quality studies and will provide a brief overview of other studies with unusual biospecimen types and aims (tumor tissue profiling, therapy response, and prognosis).

Three of the studies with separate validation investigated colorectal cancer (30–32). Nishiumi and colleagues and Ritchie and colleagues investigated serum of patients with colorectal cancer in comparison with controls, and both achieved very good discrimination abilities with an area under the receiver operating characteristic curve (AUROC) of >0.90. Ritchie and colleagues further employed a targeted LC-MS approach focusing on polyunsaturated fatty acids (PUFA; ref. 30). Although levels of PUFAs were found to be lower in colorectal cancer, the discrimination abilities of only this targeted subset of markers was still adequate with an AUROC > 0.85 (30). The analysis of urine of patients with colorectal cancer by Cheng and colleagues revealed even better AUROC values > 0.99 with significantly lower tryptophan downstream metabolites in colorectal cancer (tryptophan, kynurenine, 5-hydroxytryptophan, indoleacetate, indole; ref. 32). Budczies and colleagues characterized the metabolome of breast cancer tissue and normal breast tissue (33) using a more invasive approach. A metabolic map of the breast cancer metabolome showed marked increases in purine and glycerolipid metabolism (33). Maeda and colleagues measured the amino acid concentrations of a total of 303 patients (34). An AUROC of 0.82 indicated a limited performance of a targeted metabolomics study design, which used amino acids as the only markers (34). Chen and colleagues identified a previously unknown metabolite as 27-nor-5β-cholestane-3,7,12,24,25 pentol glucuronide (CPG) as a biomarker for ovarian cancer (35). The authors showed in their validation in a cohort of more than 600 individuals (>200 patients with ovarian cancer) that the AUROC of the single metabolite alone was 0.74 with an accuracy comparable with the standard protein marker cancer antigen 12-5 (CA12-5; ref. 35). However, all of these studies still require marker validation in prospectively collected (nondiseased) individuals to prove that they are able to detect cancers at a very early stage.

Abaffy and colleagues collected volatile signatures directly from skin of patients with melanoma, whereas Gaspar and colleagues investigated exhaled air from patients with lung cancer. Both samples have direct tumor contact and are still noninvasive. An AUROC of 0.94 (melanoma study) and a 100% discriminating ability (lung cancer study) show, that choosing such unusual biospecimens may be particularly useful (21, 24). It should be noted, however, that sample sizes in the latter two studies were < 20.

Twenty-one studies were categorized as tissue profiling studies, investigating the metabolome of tumor tissue and, therefore, cancer cells directly. These studies indicate a possible use of metabolomics in a pathologic assessment: i.e., Brown and colleagues investigated 7 prostate and kidney cancer biopsies comparing metabolic profiles to regular histopathologic examination (36). The authors noted that metabolomic analysis yielded additional information, distinguishing aggressive versus mild forms of cancer (36). Two studies by Denkert and colleagues investigated metabolomic profiles of colorectal and ovarian cancer tissues (37, 38). Both studies showed a very high accuracy for the correct classification of tumors (88% and 95%, respectively). Two studies investigating lung (39) and gastric cancer tissue (40) were able to discriminate between tumors graded at an early stage (i.e., T1, T2) versus late stage (i.e., T3, T4). Lin and colleagues used a similar approach for renal cell carcinoma (RCC) but in this case serum instead of tissue was investigated (41). These results indicate that the process of tumor grading can be aided by metabolomics, but discrimination between different tumor stages is not always successful (42, 43).

Twelve studies (11%) monitored the impact of a therapeutic intervention (i.e., chemo/-radiotherapy or surgery) on the metabolome. Jiye and colleagues investigated the metabolome of chronic myeloid leukemia patients treated (n = 33)/not treated (n = 26) with the tyrosine kinase-inhibitor imatinib (44). The authors showed that the metabolome of patients sensitive to chemotherapy changed, whereas the metabolome of resistant patients was similar to that of untreated controls. Saylor and colleagues investigated the effects of androgen deprivation therapy on patients with prostate cancer. As expected, steroid levels were decreased, but also other metabolic effects of the therapy were observed: for example, an increase in bile acids and a reduction in lipid oxidation (45). Wibom and colleagues used an unusual approach, measuring the metabolic changes in catheter-microdialysates from patients with glioblastoma undergoing radiotherapy. The authors interpreted an increase in glutamate and glutamine after treatment as predictive for tumor proliferation and cell damage. Furthermore, a cross-link towards epigenetic mechanisms such as global DNA-hypomethylation was provided: S-methyl-cysteine, an end product of demethylation of methylated nucleosides, decreased with treatment. The authors concluded that the demethylation process could be hampered by radiation, proposing a mechanistic explanation as to why combining radiotherapy with temozolomide (an alkylating drug) works so efficiently (20). Cai and colleagues monitored metabolic changes in plasma of patients with lung cancer treated with radiation, but did not identify possible biomarkers that could distinguish treated from untreated individuals (46). Some studies investigated the effect of surgical intervention on the metabolome (47–51). Most authors report a restoration of the metabolome towards that of healthy controls after surgery, whereas only a few note a clustering into a third group (50, 51). In contrast, a study led by Silva and colleagues found no differences in pre- and postsurgical serum of patients with ovarian cancer (52).

Eight studies (8%) assessed associations of the metabolome with patient prognosis (i.e. cancer recurrence, metastasis, and survival). The group of Farshidfar and colleagues used a retrospective approach, profiling the serum of patients with colorectal cancer with locoregional recurrence, liver metastases, or extra hepatic metastases. These three groups were distinguished by metabolites from the galactose and glutamine/glutamate metabolic pathway indicating an alteration of liver metabolism during metastasis (53). Locasale and colleagues prospectively followed a group of 10 patients with glioblastoma of different stages, by investigating the metabolome of cerebrospinal fluid during lumbar puncture (19). The group was able to discriminate between newly diagnosed patients with cancer and those with recurrent disease, indicating a possible role of tryptophan metabolites (19). Wedge and colleagues described that lower levels of glycerophosphatidylcholines, erythritol, and hexadecanoic acid in plasma and serum were associated with poor survival in patients with lung cancer (25). However, no data were provided with respect to patients' tumor stage as a possible confounder and only 29 patients were included (25). One of the few prospective studies by Ye and colleagues followed 19 patients with hepatocellular carcinoma (stage II and III) prospectively and developed a classification model for recurrent disease, using lactic acid and acotinic acid as potential biomarkers. The results of this small pilot study were promising with only 1 of 7 recurring patients being misclassified (47). Tang and colleagues investigated serum of patients with nasopharyngeal cancer (n = 49). Patients were followed prospectively and a set of 4 metabolic markers (kynurenine, two n-acetylglucosamines, and hydroxyphenylpyruvate) increased gradually from hyperplasia toward cancer. This set of biomarkers was further validated within a subsequent targeted analysis. Furthermore, these markers correlated with tumor size and recurrence. In addition, their levels normalized in patients who underwent radiotherapy (54). Asiago and colleagues developed a discrimination model for recurrence in breast cancer patients (n = 56), where 86% sensitivity and 84% specificity was attained and 55% of the recurrent patients were detected 13 months before clinical diagnosis (55).

GC-MS is favored for its robustness and its low susceptibility to MS-matrix interferences such as ion suppression (56). In addition, well established electron impact databases facilitate the automated identification of metabolites. On the other hand, fewer metabolites are detected by GC-MS as compared to LC-MS. This is mainly attributed to the derivatization process necessary for GC analysis and the limit in molecular weight of metabolites analyzed by this technique. Therefore, more metabolites are typically covered by LC-MS methods. As a drawback, LC-MS–specific electrospray ionization is more susceptible to matrix interferences such as adduct formation or ion suppression (56). Fourteen studies employed both GC and LC-MS analyses.

Preprocessing is essential to analyze and interpret metabolomics data. This step includes the import/conversion of raw data files, the detection of signals (peak picking), the assignment of single ions to the same metabolite (deconvolution), the integration and alignment of chromatographic peaks, and eventually different methods for baseline correction, normalization, and smoothing (57). Extensive preprocessing software is available, both free-of-charge and commercially (57). Both, the process of preprocessing and available software have been extensively discussed previously (58–60). Appropriate data normalization techniques are critical and the type of procedure is mainly determined by the type of sample or methodology. A special concern arises for urinary metabolomics, where metabolite levels change depending on water intake and kidney function. Studies reviewed in this report generally normalize urine to creatinine levels (51, 61–66), but other types of normalization are also common (normalization to osmolarity or the mean/total peak area of the total ion chromatogram). Because normalization methods may influence the results of a study (67, 68), we provide the type of normalization used in studies monitoring urinary metabolite levels in Table 1.

Data analysis includes both uni- and multivariate statistical approaches, such as principal component analysis (PCA), some type of discriminant analysis (PLS-DA) or clustering analysis, but also machine-learning techniques such as support vector machines. A good review of metabolomics data analysis was provided by Danielsson and colleagues (69).

Although guidelines for preprocessing and data analysis of large scale human metabolomics studies have been published recently (70, 71), there is still substantial diversity in the studies reviewed here, with more than 10 different software programs used for preprocessing and for data analysis. PCA, PLS-DA, and clustering techniques are most common for studies reviewed here, but some authors favor support vector machines (28, 37, 38, 63, 72–77) or algorithms developed in-house (77). This heterogeneity reflects in part the state of metabolomics as a relatively young discipline, where fully standardized protocols are just beginning to be developed.

One hundred and six studies investigated the metabolome of patients with cancer in varying matrices. A comprehensive list of altered metabolites was extracted (see Supplementary Table S1). In total, 390 metabolites were reported as being significantly altered in patients with cancer, compared with controls. Markers are more likely to be relevant, if the metabolite is repeatedly reported by multiple investigators. Supplementary Table S1 lists the metabolites as described in the publications. However one should be careful with some of these annotations which may only be tentative, in particular for uncommon metabolites for which chemical standards may not be available. Furthermore, no standards exist thus far for reporting metabolite identification in metabolomics studies and some of the annotations proposed might be revised in future studies (78). Pathways most commonly different between cases and controls are presented in Fig. 4 and the Supplementary Data. The figures only represent the frequency of reported alterations and not their direction of change, because of the high heterogeneity within studies (i.e., different types of cancer, types of samples, and instrumentation used). Therefore, inconsistencies within the direction of change exist. Tryptophan, for example, the most commonly reported metabolite, was found in lower concentrations in the serum of patients with hepatocellular carcinoma, colorectal cancer, and RCC (41, 43, 79). In contrast, it was upregulated in urine of recurring HCC (47), breast cancer patients (62), and in bladder cancer tissues (80). In a recent report, Ng and colleagues also reviewed and discussed these types of discrepancies depending on cancer types (81). Using MetScape we identified the following pathways as potential hallmarks of the cancer metabolome.

Seven of the 10 most frequently reported metabolites altered in cancer were amino acids. This outlines the importance of an altered protein synthesis in cancer metabolism. Interestingly, essential and aromatic amino acids such as tryptophan, phenylalanine, tyrosine, and their downstream metabolites were frequently reported (refs. 41, 43, 47, 62, 79, 80; see Supplementary Table S1). For example, many catabolites of tryptophan, i.e., several indole derivatives, nicotinuric acid, and especially kynurenine were altered in various cancer types (54, 62, 73). Chen and colleagues concluded that the unusual increases of kynurenine and nicotinuric acid in urine of patients with breast cancer might be a consequence of elevated estrogen levels (62), because of the hormonal regulation of tryptophan oxygenase, the initial enzyme catabolizing tryptophan (62). Conversely, Cheng and colleagues reported lower kynurenine concentrations in the urine of patients with colorectal cancer, compared with controls (32)—this is consistent with opposing effects of estrogen on colorectal and breast carcinogenesis, as noted in the epidemiology of these diseases (82, 83). Recently, kynurenine has been described as an endogenous ligand of the human aryl hydrocarbon receptor, providing a mechanism for important carcinogenic characteristics, such as tumor cell survival, motility, and immune suppression (84). Downstream metabolites of phenylalanine, such as hippurate or p-cresol, were frequently altered (32, 43, 72, 85). The latter metabolite has been controversial in the literature: Cheng and colleagues reported a decrease in urine of patients with colorectal cancer compared with controls (32), whereas Qui and colleagues observed an opposite trend (43). The above mentioned metabolites are produced by the gut microbiota. A number of other metabolites most likely of microbial origin have also been described in various publications (Supplementary Table S1). These metabolites may play a role in cancer etiology, especially regarding colorectal cancer pathogenesis, as discussed in more detail elsewhere (86–88). Other members of the pathway, such as phenylacetylglutamine, a downstream metabolite of phenylacetic acid and glutamine were higher in urine of patients with colorectal cancer. Phenylacetylglutamine has been found to be cancer suppressive by initiating apoptosis. Its upregulation may reflect the organism's response to the presence of a tumor (65).

Alanine, aspartate, and glutamate are also frequently reported to be changed and link multiple pathways in cancer metabolism: i.e., they fulfill anaplerotic reactions for the tricarboxylic acid (TCA) cycle, providing an alternative energy source for cancer cells, which predominantly use energy produced by glycolysis rather than the TCA cycle and oxidative phosphorylation. This effect can be observed by the metabolic pathway analysis (lactate is in the top 10 metabolites altered) and has long been known as the Warburg effect (89).

Alterations in arginine and proline metabolism are furthermore linked to the urea cycle that has also been discussed in the review by Ng and colleagues (81). Impaired amino acid metabolism is therefore not only a consequence of protein synthesis. An increased use of amino acids for energy production in cancers may explain the excretion of ammonia and therefore the impairment of the urea cycle. Because of the importance of amino acids in cancer metabolism, several targeted studies, measuring their exact concentrations, have been conducted (34, 90, 91). One large scale study pooled more than 900 patients with different cancer types together. Despite this, an AUROC close to 0.75 indicates a limited value of amino acids as sole biomarkers (91).

The respective amines of aspartate and glutamate (asparagine and glutamine) are important amino donors. Glutamine's role as an energy source for cancer cells, entering the TCA cycle via α-ketoglutarate has been previously discussed by several authors (81, 92). Both amines, being nitrogen donors, can furthermore link amino acid to nucleoside synthesis (see Fig. 4). Some studies also investigated this impaired nucleoside metabolism, focusing on several modified nucleosides excreted in urine of patients with breast cancer (75, 93). A targeted study led by Yoo and colleagues showed that a decrease in downstream metabolites hypoxanthine and xanthine were suitable markers for non-Hodgkin lymphoma with AUROC values of 0.85 and 0.83, respectively (29).

Another hallmark of cancer is impaired lipid metabolism: the chemical group of lysophosphatidylcholines and lysoethanolamines were prime markers of this pathway. Changes in fatty acid profiles and various carnitines were also frequently reported by multiple groups, indicating increased membrane synthesis and cellular turnover (94–98).

Myo-inositol, which was increased in various types of cancers (48, 64, 99), might provide a link to the phosphoinositide-3-kinase (PI3K) pathway in cancer. This indicates that well-known cancer pathways, such as PI3K, not only show their effects in the proteome, but also in the metabolome.

Metabolites that are exclusively synthesized within specific organs/tissues can be considered as good candidates for markers of the respective cancer, because of their specificity. This becomes evident in studies with a focus on hepatocellular carcinoma (liver) and adrenocortical cancer (adrenal glands):

Because of the liver's prime role in human metabolism, metabolic markers of liver disease, e.g., bilirubin, have a long clinical history. Therefore, most hepatocellular carcinoma studies included diseased controls (liver cirrhosis, hepatitis) for the development of biomarkers (23, 100–103). Liver-specific bile metabolites have been reported very frequently, impacting both bile acid and taurine metabolism. The decrease in bile acids is in line with a loss of the organ's function in hepatocellular carcinoma. Moreover, taurine is considered as an important antioxidant (104) highlighting the role of oxidative stress in cancer. Cao and colleagues also reported a decrease in bile acids and an increase in lysophosphatidylcholines (23). Tan and colleagues showed that an increase of three markers, namely taurocholic acid, lysophosphatidylcholine 22:5, and lysophosphoethanolamine 16:0, were enough to discriminate hepatocellular carcinoma from patients with hepatitis B or cirrhosis with a sensitivity of 87.5% and a specificity of 72.3% (102).

Adrenocortical cancer is also characterized by specific metabolite patterns, because of the adrenal gland's function which influences the steroidal metabolome. Arlt and colleagues studied urinary steroid levels and showed that they are of high value in discriminating patients with cancer from patients with adenomas. Sensitivity and specificity in the range of 90% could be attained (105). Similar observations have been made by Kotlowska and colleagues (106).

One challenge of metabolomics is to elucidate the importance of altered metabolites in a biologic system. Different theoretical options exist:

• A metabolite reflects a change in the genome, epigenome, transcriptome, or proteome. For example, a mutation affecting an enzyme's activity may translate into a change in metabolite levels.

• A metabolite can be a driver, i.e., the altered metabolite levels may inhibit an enzyme or activate a receptor (see kynurenine; ref. 84) or have epigenetic effects.

The question, whether or not a metabolite can be considered (i) or (ii) is often difficult to answer and would most often require additional experiments. Ultimately, it is the entire biologic system which is changed: Elevated levels of 2-hydroxyglutarate, for example, are the consequence of mutated isocitratedehydrogenase(s) (i), but the metabolite itself may promote epigenetically-driven pathogenesis (ii; ref. 107).

Another option to assess the functional role of metabolites is to embed them into pathways and their alterations. This question is not easily addressed by analyses of urine or plasma/serum levels, but tissue analyses are more promising. For example, Budczies and colleagues constructed a metabolic map of breast cancer, based on their tissue metabolomics data to reveal alterations in energy, amino acid, and nucleotide metabolism (33).

Some of the included studies have used cross-omic approaches involving both metabolomics and proteomics (64, 80, 108). Two studies focused on changes in the metabolome based on altered protein levels (109, 110). For example, Brockmöller and colleagues correlated the expression levels of the enzyme glycerol-3-phophate acyltransferase with increased levels of phospholipids and phosphatidylcholines (109). This emphasizes that increased membrane synthesis accompanied by high cellular turnover is a hallmark of cancer and can be mirrored both on the metabolic as well as on the protein level. In the latter study is an example for (i). In contrast, changes in the metabolome can also influence other cellular layers (ii), for example the epigenome: pathways corresponding to epigenetic mechanisms, such as S-adenosylmethionine (SAM) and methylated nucleosides indicate connections between epigenomic and metabolic alterations in cancer. Henneges and colleagues investigated urinary nucleoside levels of patients with breast cancer in a targeted fashion. They reported that methyltransferase activity was altered by monitoring pathophysiologic patterns of methylated nucleosides and an increase of S-adenosylhomocysteine (SAH), the product of methylation reactions (75). In addition, SAH is a potent inhibitor of methyltransferase activity. Its increased excretion suggests the ongoing methylation capacity of cancer cells. Woo and colleagues, also found altered methylated nucleosides in the urine of patients with breast, ovary, and cervical cancer (111). Chen and colleagues reported only 1-methyladenosine together with two unknown components to be upregulated in hepatocellular carcinoma (112). A comprehensive study by Putluri and colleagues investigated tissue and urine from bladder cancer patients and revealed cross connections between elevated SAM and DNA methylation patterns (80). Furthermore, the authors measured altered promotor methylation of genes involved in xenobiotic metabolism (i.e., CYP1A1 and CYP1B1) and their consequent silencing by epigenetic mechanisms. A very comprehensive study led by Sreekumar, analyzed urine, tissue, and plasma samples of patients with prostate cancer (73). They could show that a single metabolite, namely sarcosine, was increased in prostate cancer progression and metastasis. The authors were further able to show that sarcosine is involved in the mechanism of a tumor's invasiveness, by performing cell culture and knockout experiments of the enzymes involved in sarcosine metabolism. In contrast, Struys and colleagues showed that serum sarcosine levels alone, were not enough to distinguish controls from patients with elevated prostate-specific antigen (PSA) or prostate cancer (113). Moreover, Jentzmik and colleagues reported that urinary sarcosine did not improve the discrimination of patients with prostate cancer from patients with no evidence of malignancy (114). They also reported no association with tumor stage or Gleason score. The relevance of sarcosine as a marker for prostate cancer is therefore controversial.

Thirty-three studies with more than 85% correct classification rate of patients with cancer indicate that metabolomics has the potential of identifying novel diagnostic markers. It is furthermore a high-throughput technique, which is fast and cost efficient. However as yet, metabolomics cannot be considered as a standard tool in clinical oncology. Once consistent marker sets have been discovered and replicated in independent populations, it will be critical to promote them to prospective settings for validation.

Secondary applications include tumor classification or grading, which can aid pathologic assessment or patient prognosis. A few studies with low patient numbers showed promising results by being able to differentiate recurrence or metastasis from primary disease (19, 47, 53, 55). Following patients over a longer period of time gives the additional opportunity to monitor changes in the metabolome during treatment. While screening procedures are limited to minimal invasive sampling techniques (i.e., blood, urine) additional samples such as tumor tissue can be investigated from patients who undergo surgery. This will help to better understand systematically the overall picture of metabolic changes in tumors. Together with studies targeting therapy control, metabolomics has the ability to become a valuable tool for a future personalized medicine approach.

Method standardization is an important factor in metabolomics and has been extensively discussed by other authors (115, 116). The analytic platforms themselves (LC-MS or GC-MS) are robust. Problems are reported by several authors due to differences in sample pretreatment in different hospitals (50). If sample processing is not identical, systematically biased results are to be expected. For tissue metabolomics, contamination of tumor cells with surrounding normal tissue cells can furthermore confound the analysis. Therefore, techniques such as microdissection should be used before sample homogenization (117). Fasting status of patients as a possible confounder is controversially discussed by different authors. Some do not note any problems for urinary metabolomics (50), whereas in serum, differences due to fasting status have been reported (39, 51). Ma and colleagues discussed decreases in valine and arachidonic acid as a result of prolonged fasting during surgery, by investigating serum levels pre- and post-surgery (118). Krug and colleagues studied the absolute metabolic changes due to various types of challenges such as fasting, meal frequency, and physical exercise which showed large intraindividual variation (119). It is important to develop biomarkers that are independent of fasting status, to make them truly cancer specific and suitable for the clinical situation, when fasting status cannot always be monitored.

Another important issue is the assessment of unidentified metabolites. Within the present work, unknown metabolites are frequently not included in the data analysis, which may be considered as a conservative approach. However, this could lead to the omission of important information, because unknowns are frequently found to be altered (37, 38, 120). Silva and colleagues were able to discover a new and unidentified metabolite with important discriminating abilities for ovarian cancer, using accurate MS (52). Even if metabolites are not initially identified by library-based searches, they should still be reported as unknowns because they may contain valuable information and should be followed up.

Multivariate statistical approaches such as PCA or PLS-DA/OPLS-DA are very frequently used to differentiate between groups, i.e., cancer versus control. Because of the generally low sample numbers compared with the numbers of metabolites, overfitting is a common problem for supervised statistical methods. PLS-DA models were reported to be especially susceptible (38). It is therefore important to keep in mind that independent validation of putative markers is essential. This can be achieved by a replication study: within the present work, only a minority of 18 of 106 studies (17%) further validated their set of discovered biomarkers in a separate study population.

Many studies claim to aim for the early detection of cancer, but it is important to note that no study used a prospective study design, following a cohort of healthy individuals over time until the onset of cancer. This type of validation within a prospective screening is essential for early detection, as outlined in the five phases of screening for biomarkers approach by the Early Detection Research Network (ERDN) of the National Cancer Institute (121). A prospective design for patient prognosis was also implemented in only a few studies (19, 47, 54). Validation in prospectively collected samples will be essential for clinical translation.

Most of the altered metabolites are either not specific for one particular cancer or can be influenced by other metabolic conditions, such as comorbidities, for example, lactate and gentisic acid are altered in alkaptonuria (122), various metabolites of the energy metabolism in diabetes, or urea in liver cirrhosis (123). Confounding due to comorbidities was investigated by Yu and colleagues. The group used five different forms of gastric disease as controls for gastric cancer and was not able to differentiate between them (124). To achieve a higher specificity the implementation of diseased controls within the biomarker development is essential. For example, most authors investigating HCC included diseased controls in their datasets (23, 100–103). This has the advantage of reducing confounding factors due to inflammatory processes, frequently accompanied by HCC, such as hepatitis or cirrhosis. Some groups also included patients with benign tumors in their datasets (65, 79, 96, 105). This also enhances the possible specificity of the final validation metabolite set.

Another solution to increase specificity is to use multiple “omics” strategies in combination. Few authors compared their developed metabolomic markers with conventional biomarkers (101, 112, 125). Those who did reported an improved sensitivity of metabolomic markers over conventional tumor markers, whereas Ikeda and colleagues reported higher specificity for proteomic tumor markers (125). Wang and colleagues showed that if metabolic biomarkers and the traditional tumor marker AFP are used together, a sensitivity of 96.4% and a specificity of 100% was reached (101). Chen and colleagues also reported that combined datasets (1-methyladenosine and AFP) perform better than one marker alone (112).

This review was limited to MS-based metabolomics. Therefore, NMR analysis and non-MS methods were not included. Furthermore, targeted studies measuring less than 10 metabolites were considered as being non-omics approaches and therefore excluded. In contrast, the strength of this review is its comprehensive systematic investigation of 106 original research papers investigating 21 different types of cancers in 7 matrices. Therefore, a representative overview of the MS-based metabolomics in cancer research is presented.

On the basis of this review, metabolomics should still be considered as a discipline at an early stage of development, but has the ability to become a valuable tool in epidemiology and clinical oncology. The results show that its role in cancer research may be primarily early detection and screening applications. In addition, initial studies in translational areas such as investigations of patient prognosis and therapy control indicate that metabolomics may evolve into a flexible tool in translational medicine. Moreover, additional tumor tissue profiling may be useful for precise tumor classification in a pathologic setting. Nevertheless, standardization is needed regarding sample preparation, normalization, and data analysis. To enhance specificity and to further validate potential metabolic biomarkers as markers for early detection, large validation studies with a prospective study design are needed. Finally, cross-omic approaches, merging different disciplines in systems biology can help overcome drawbacks of a single discipline.

No potential conflicts of interest were disclosed.

Conception and design: D.B. Liesenfeld, N. Habermann, C.M. Ulrich

Development of methodology: D.B. Liesenfeld, N. Habermann, C.M. Ulrich

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): D.B. Liesenfeld, N. Habermann, R.W. Owen

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): D.B. Liesenfeld, C.M. Ulrich

Writing, review, and/or revision of the manuscript: D.B. Liesenfeld, N. Habermann, R.W. Owen, A. Scalbert, C.M. Ulrich

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): D.B. Liesenfeld, C.M. Ulrich

Study supervision: N. Habermann, C.M. Ulrich