Comparative cerebrospinal fluid metabolites profiling in glioma patients to predict malignant transformation and leptomeningeal metastasis with a potential for preventive personalized medicine

Abstract

Glioma shows progression presenting as malignant transformation or leptomeningeal metastasis (LM). However, longitudinal biopsy of brain parenchyma is difficult due to its critical location, whereas cerebrospinal fluid (CSF) can be obtained serially with a little invasiveness of puncture. Thus, if we could find a biomarker for glioma progression, we could predict such event and determine therapeutic interventions as early as possible. In this study, we examined whether cerebrospinal fluid (CSF) metabolome profiles can reflect glioma grade, difference with non-glial tumor, and LM status. We selected 32 CSF samples from glioma patients, and compared them with 10 non-tumor control and seven non-glial brain tumor (medulloblastoma) samples. A total of 10,408 low-mass ions (LMIs) were detected as a candidate of metabolites using mass spectrometry, and representative LMIs were identified via the Human Metabolome Database. Grade IV gliomas showed eight LMIs, including acetic acid, of higher levels (summed sensitivity and specificity > 180%) than grade III gliomas. Grade IV gliomas demonstrated more abundant 30 LMIs, including glycerophosphate, compared with medulloblastoma, but none was mutually exclusive. Phospholipid derivatives were significantly more abundant in LM (−) than LM (+) gliomas regardless of glioma grade. LMIs representative of LM (+) gliomas were derivatives of glycolysis. We also verified discriminative LMIs based on mean expression level of each LMI (Student t test, p < 0.05) and evaluated the differences of the above analyses. Over 90% of metabolite pathways indicated from two analytical models were common to each other. Non-targeted mass spectrometry of CSF metabolites revealed significantly different profiles across gliomas that possibly permitted differentiation between glioma grades, LM, and non-glial brain tumors.

Electronic supplementary material

The online version of this article (10.1007/s13167-020-00211-4) contains supplementary material, which is available to authorized users.

Introduction

Gliomas are incurable tumors that invade critical structures of the brain, preventing curative surgical resection. Although systemic chemotherapy has been attempted, it has met with limited success [1, 2]. Moreover, these gliomas underwent malignant transformation to higher tumor grades during or after the treatment, resulting in aggressive clinical behavior and changed genetic or epigenetic signatures [3]. For diagnosis and treatment decisions, tumor biopsy is required, but serial samples of gliomas are difficult to obtain because the brain is shielded by the cranium, and gliomas are often located in the eloquent area. Thus, diagnosis and treatment decisions largely depend upon neuroimaging such as magnetic resonance imaging (MRI) in case of difficult biopsy, but there are no standardized MRI criteria discriminating grade IV glioma from grade III [4]. Gliomas can also metastasize into the central nervous system (CNS), occasionally resulting in spread to the cerebrospinal fluid (CSF), but gadolinium-enhanced MRI sensitivity for leptomeningeal metastasis (LM) is around 60% [4].

Because CSF is in contact with the brain parenchyma and circulates throughout the neuroaxis, transporting neuro-metabolites, extracellular vesicles, and proteins, CSF has been studied for the evidence of CNS disease and to monitor treatment response as a so-called “liquid biopsy” [5, 6]. In particular, recent advances in metabolomics using high-resolution mass spectrometry (MS) have revealed features of the cancer environment that are expected to permit early diagnosis, monitoring of disease progression, and treatment response in cancer[7]. Metabolic biomarkers inferred from metabolomics could contribute to an advance of predictive, preventive, and personalized medicine (PPPM) in neuro-oncology [8–10]. Nevertheless, questions remain, such as whether the detection of a diluted metabolite, secreted from a relatively small tumor into a large amount of biofluids, reflect the unique characteristics of these tumors unless it is a pathognomonic byproduct of cancer cell metabolism. Previously, we described discriminative metabolomic profiles in 196 patients with different CNS tumors, including LM from solid primary cancers [11]. Using low-mass-ion discriminant equation (LOME) to select the best LMI profile, we could differentiate LM from the high-risk patients with brain metastasis. Other researchers have differentiated the CSF metabolomic profiles of malignant gliomas from those of controls including brain metastasis and non-cancerous patients using targeted MS [12, 13]. However, there is a need to explore CSF metabolomic profiles in gliomas to differentiate between WHO grade, leptomeningeal metastasis, and different invasiveness between glioma and non-glial primary brain tumors.

Here, we performed non-target MS of 32 glioma CSF samples and compared them to seven medulloblastomas and 10 non-tumorous controls (unruptured aneurysm). We compared the LMI profiles between (1) gliomas and controls, (2) gliomas of different WHO grades (III vs. IV), (3) primary brain tumors of the same WHO grade (IV) but different cellular origins and clinical behaviors (glioblastoma vs. medulloblastoma), (4) gliomas with or without LM, and (5) gliomas with different isocitrate dehydrogenase 1 (IDH1) mutation status. The discriminative LMIs were obtained experimentally using two different parameters of summed sensitivity/specificity and the difference of mean peak area of individual LMI between two groups of comparison. Furthermore, we examined the most prevalent metabolomic pathway designated by these discriminative LMIs using MetaboAnalyst (https://www.metaboanalyst.ca, ver. 4.0) to identify possible pathway dysregulation in particular glioma types and to compare the similarity between two analytical methods.

Materials and Methods

CSF collection

All CSF samples were obtained after Institutional Review Board (NCC-150002) approval in May 2015; written informed consent was obtained from each participant before the collection. The ethics approval was given; healthy controls submitted informed consent on their permission for operation of unruptured aneurysm, and in case of minors, their legal guardian submitted the informed consent. CSF samples were centrifuged (2000×g, 20 min) within 1 h of collection for cell down. A 50-μl sample of each supernatant was used for MS analysis.

CSF sample collection has been done under permission of NCC-150002 prospectively. However, this study design came out as a retrospective analysis of gliomas later using CSF archives and electronic medical record. CSF sampling was done from ventricle via Ommaya/shunt valve reservoir/extraventricular drainage catheter for controlling hydrocephalus or intraventricular/intrathecal chemotherapy or via lumbar puncture for the diagnosis of LM (Table 1). Ten samples of non-tumorous control (unruptured aneurysm) CSF were obtained during craniotomy from cisternal space, and seven samples of medulloblastoma were obtained via lumbar puncture for CSF cytology.

Table 1.

Clinical characteristics of 39 samples from 32 patients

Case	Gender/age	Diagnosis	WHO grade	LM status	LM examination (MRI/cytology)	IDH 1 status
1	M/56	Glioblastoma	IV	(−)	(−)/(−)	wt
2	F/33	DMG	IV	(−)	(−)/(−)	wt
3	F/47	Glioblastoma	IV	(−)	(−)/Atypical	wt
4	F/18	Glioblastoma	IV	(−)	(−)/ Atypical	n.d.
5	M/68	Glioblastoma	IV	(−)	(−)/(+)	wt
6	F/58	Glioblastoma	IV	(+)	(+)/(+)	n.d.
7	F/29	Glioblastoma	IV	(−)	(−)/(−)	n.d.
8	M/51	Gliosarcoma	IV	(+)	(+)/(−)	wt
			IV	(+)	(+)/(+)	wt
			IV	(+)	(+)/(−)	wt
			IV	(+)	(+)/(−)	wt
			IV	(+)	(+)/(−)	wt
9	F/48	Glioblastoma	IV	(−)	(−)/n.d.	wt
10	F/29	Glioblastoma	IV	(+)	(+)/(−)	wt
11	M/56	Glioblastoma	IV	(−)	(−)/(−)	wt
12	F/35	Glioblastoma	IV	(−)	(−)/n.d.	wt
13	F/57	Glioblastoma	IV	(−)	(−)/n.d.	wt
14	M/23	DMG	IV	(−)	(−)/n.d.	wt
15	M/74	Glioblastoma	IV	(−	(−)/(−)	wt
16	M/5	Brainstem glioma	III	(−)	(+)/Atypical	n.d.
		Brainstem glioma	III	(−)	(+)/(−)	n.d.
17	F/2.4	AA	III	(−)	(−)/(−)	n.d.
18	M/58	AA	III	(+)	(+)/Atypical	wt
19	F/71	AA	III	(+)	(+)/Atypical	n.d.
20	M/22	Brainstem glioma	III	(−)	(−)/(−)	n.d.
21	F/35	AA	III	(+)	(+)/Atypical	wt
22	F/47	Anaplastic ODG	III	(+)	(+)/Atypical	mt
		Anaplastic ODG	III	(+)	(+)/n.d.	mt
		Anaplastic ODG	III	(+)	(+)/Suggestive	mt
23	M/72	AA	III	(−)	(−)/n.d.	wt
24	F/41	ODG	II	(−)	(−)/n.d.	mt
25	M/32	Gliomatosis	II	(+)	(+)/Suggestive	n.d.
26	F/5	Medulloblastoma	IV	(−)	(−)/(−)	n.a.
27	F/2.8	Medulloblastoma	IV	(−)	(−)/(−)	n.a.
28	F/7	Medulloblastoma	IV	(+)	(+)/Atypical	n.a.
29	M/7	Medulloblastoma	IV	(−)	(−)/(−)	n.a.
30	F/10	Medulloblastoma	IV	(−)	(−)/Atypical	n.a.
31	M/11	Medulloblastoma	IV	(−)	(−)/Atypical	n.a.
32	M/7	Medulloblastoma	IV	(+)	(+)/(−)	n.a.

AA, anaplastic astrocytoma; DMG, diffuse midline glioma; IDH 1, isocitrate dehydrogenase 1; ODG, oligodendroglioma; LM, leptomeningeal metastasis; mt, mutated; n.a., not applicable; n.d., not done; wt, wild-type

Diagnosis of LM

CSF cytology findings positive or suggestive for metastasis were considered LM-positive. However, in the event of atypical cytology, MRI findings and previous CSF cytology were used as references for positive categorization as follows: (1) MRI was definite (leptomeningeal enhancement of cerebral hemisphere gyri, cerebellar folia, subependymal wall, spinal cord, and cauda equina) or suggestive (ventricular seeding or dural enhancement) of LM [14], and (2) previous CSF cytology was positive, and this sample was obtained after treatment for LM.

CSF extraction for metabolite profiling

Metabolites present in the CSF were extracted using a modified Bligh and Dyer method. In brief, 50 μl CSF were added to 1 ml deionized water. After vortexing, 2 ml methanol and 0.9 ml dichloromethane were added. After vortexing and incubating on ice for 30 min, 1 ml water and 0.9 ml dichloromethane were added, then the mixture was centrifuged (1000×g, 10 min, room temperature). Nitrogen gas was used to dry the supernatant for MS analysis. The extracted metabolites were analyzed by TripleTOF 5600+ system (Sciex, Framingham, MA) coupled with a Nexera X2 UHPLC system (Shimadzu, Kyoto, Japan). The metabolites were loaded onto an Atlantis T3 sentry guard cartridge (3 μm, 2.1 mm × 10 mm; Waters, Milford, MA), and separated by an Atlantis T3 column (3 μm, 2.1 mm × 100 mm; Waters) in a two-step linear gradient (solvent A, 0.1% FA in water; solvent B, 100% ACN; with 1% solvent B for 2 min, 1 to 30% solvent B for 6 min, 30 to 90% solvent B for 8 min, 90% solvent B for 4 min, 90 to 1% solvent B for 1 min, and 9 min in 1% solvent B). The MS system was set to perform one full scan (50 to 1200 m/z range) followed by tandem mass spectrometry (MS/MS) of the 10 most abundant parent ions (mass tolerance, 50 mDa; collision energy, 35%). Monoisotopic mass peak was used for the next analysis and the ratio of signal to noise (S/N) was set at 10:1.

Identification of discriminative LMIs

LMIs with outstanding discriminative ability (e.g., distinguishing the control from glioma groups) were identified using logarithmic peak areas [15]. Briefly, individual LMIs were assessed as follows. (1) For each LMI, a discrimination threshold value was determined where the sum of the sensitivity and specificity was highest, at increments of 0.01. When a range of adjacent threshold values showed the same highest discrimination performance, the highest and lowest threshold values (and their difference) were tabulated for later comparison. Each LMI was investigated in two modes. First, it was assumed that logarithmic peak areas of one group were mostly higher than those of the other group. Then a sample whose peak area was more/less than the discrimination threshold in the former group was assigned as a true-positive/false-negative case. A sample whose peak area was more/less than the discrimination threshold in the latter group was assigned as a false-positive/true-negative case. Next, it was assumed that logarithmic peak areas of the former group were mostly lower than those of the latter group. Then, a sample whose peak area was less/more than the discrimination threshold in the former group was assigned as a true-positive/false-negative case. A sample whose peak area was less/more than the discrimination threshold in the latter group was assigned as a false-positive/true-negative case. Finally, the sum of the sensitivity and specificity was calculated for the two modes. Then, the mode with the higher sum of the sensitivity and specificity was assigned as a discrimination mode of the LMI. The sum of the sensitivity and specificity at the discrimination mode of the LMI was considered in subsequence comparison. (2) A few discriminative LMIs were designated as possible candidates based on a “summed sensitivity/specificity” based on at least 80% sensitivity and specificity and a summed sensitivity/specificity greater than 180%.

Identification of metabolite ions

The MS and MS/MS spectra were submitted to the Formula Finder computational tool (SCIEX) that proposes probable elemental compositions within a specified mass tolerance of a given m/z ratio using the PeakView software (SCIEX). Using metabolite databases from the Human Metabolome Database (HMDB, www.hmdb.ca), putative specific compounds were found for the given m/z and listed in rank-order based on the MS and MS/MS data. The compounds list found from the HMDB was used for further analysis of metabolomic pathways in MetaboAnlayst (https://www.metaboanalyst.ca, ver. 4.0).

Statistical analyses

Categorical data were analyzed using chi square, Fisher’s exact, or Mann-Whitney U tests as appropriate. Quantitative data were analyzed using one-way analysis of variance with post hoc comparisons (Scheffe’s test). The Student t test and the false discovery rate (FDR) correction (Benjamini and Hochberg method) were examined to access the difference of mean peak area between groups. All statistical analyses were performed using STATA 10.0 (version 10.0, StataCorp LP, College Station, TX, USA) or R (version 3.6.0, The R Foundation).

Results

Clinical characteristics of patients

Clinical characteristics of 39 samples from 32 primary brain tumor patients are presented in Table 1. Three glioma patients provided nine samples (5 from case 8, 2 from case 16, and 3 from case 22), which were not replicates but different dates of sampling for pre-/post-procedure or treatment. Patients were 43.7% male and 56.3% female, and the median age was 34 years (range 2.4–74). Because the non-glioma group (n = 7) consisted of medulloblastoma only, the median age of these patients was 7.0 years, significantly younger than the median 47 years of the glioma group (n = 25, chi square test, p < 0.001). The control group (n = 10) showed the oldest median age, 61.5 years (range 50–72).

All gliomas and medulloblastomas were histologically confirmed except two cases of brainstem glioma (case 16 and 20). Among 25 patients with gliomas, glioblastoma (GBM) was the most common (n = 12, 48%). All GBM was primary except one case of gliosarcoma (case 8) being secondary GBM from pleomorphic xanthoastrocytoma. Two brainstem gliomas without biopsy-proven pathology (n = 2) were considered to be grade III. Of the 17 tumors of 23 CSF samples were examined by immunohistochemistry, four samples of oligodendroglioma (ODG) lineage showed an IDH1 mutation, and the remainder possessed wild-type IDH1. Eighteen samples were collected from 15 patients with LM. Among these, 13 samples were accompanied with CSF cytology, which was positive in 4 patients, atypical in 4 patients, and negative in 5 patients.

Among 32 glioma samples, recent therapeutic interventions were none in 12 samples (8 samples were obtained on operation day, 3 samples were postoperative period, and one received chemotherapy 5 years ago), systemic chemotherapy in 16 samples, intrathecal chemotherapy in 3 samples, and radiation in one sample.

Overview of CSF metabolomic profiling

A total of 10,193 LMIs in a peak table using Marker View (version 1.2.1, Sciex, Tokyo, Japan) were evaluated by MS of 49 CSF samples (32 gliomas, 7 medulloblastomas, and 10 controls, Supporting Information Table 1). Each LMI was identified by m/z and retention time (min), and relative expression level was denoted by differences in mean normalized logarithmic peak areas between groups of comparison. To evaluate the possibility of discriminative metabolomic profiles, supervised clustering heatmap was generated using MetaboAnlayst 4.0 (Fig. 1a). In the dendrogram, LMIs were classified into three clusters. The cluster 1 had higher expression in glioma group; the cluster 2, highest in the control group; the cluster 3, mixed profiles. In unsupervised clustering heatmap, all tumor samples separately clustered from non-tumor controls except for three tumor samples (WHO grades II, III, and IV, each, Fig. 1b). Furthermore, among 19 grade IV glioma samples, 11 out of 12 LM (−) samples clustered separately from the 7 LM (+) samples. These finding possibly suggests that CSF LMIs might enable differentiation between primary brain tumors and non-tumorous conditions as well as between gliomas with and without leptomeningeal metastasis.

Fig. 1.

Hierarchical clustering profiles of CSF LMIs. (a) Supervised clustering heatmap of CSF metabolomes was generated using the Pearson correlation–based distance measure and Ward’s linkage method provided from Metaboanalyst 4.0. (b) Unsupervised hierarchical clustering heatmap of metabolomes was generated using the Pearson correlation–based distance measure and Ward’s linkage method provided from Metaboanalyst 4.0. Color key indicates log₁₀ peak area value

Candidate LMIs discriminating gliomas from non-tumorous control

To select discriminative LMIs between groups of comparison, we performed two different methods; (1) the summed sensitivity/specificity (LMI-SSS) at a designated threshold level, and (2) the difference of mean peak areas (LMI-MPA) of each LMI. We identified 99 LMIs that were expressed significantly more abundant in glioma samples than in controls by LMI-SSS, and all of which showed complete discrimination (both 100% sensitivity and specificity). Additionally, 148 LMIs demonstrated abundance in gliomas at 97% sensitivity (one false-negative sample) (Supporting Information Table 1a). Based on the difference of LMI-MPA, we found 2674 discriminating LMIs (Student t test, p < 0.05 and |fold change| > 2). Among these, 2021 of which were more abundant in glioma samples than in controls, and 653 of which showed abundance in the non-tumorous control compared with glioma samples (Supporting Information Fig. 1a). However, the heatmap of 2021 LMIs abundant in gliomas failed to show distinctive profile according to glioma grade with this mean expression level selection (Supporting Information Fig. 1b).

Candidate LMIs discriminating gliomas of different WHO grade (III vs. IV)

Although there were hundreds of discriminative LMIs between grades II and III gliomas (Supporting Information Table 1b), we devoid of analyzing as the number of grade II gliomas was too small (only two cases). From LMI-SSS (> 180%), we found 61 LMIs exhibiting more abundance in grade III gliomas and eight LMIs with abundance in grade IV gliomas (Fig. 2a, scatter plot). Based on the LMI-MPA, 330 LMIs were significantly more abundant in grade III gliomas compared with that in grade IV, whereas only 21 LMIs showed significantly greater abundance in grade IV gliomas than that in grade III (Fig. 2a, volcano plot).

Fig. 2.

Discriminative LMI of CSF metabolites. Expression level and distribution of representative LMIs, which was overly expressed. (a) In grade III (scatter plot, left) and grade IV gliomas (scatter plot, right). (b) In grade IV non-gliomas (scatter plot, left) and glioma grade IV (scatter plot, right). (c) LM (−) (scatter plot, left) and LM (+) (scatter plot, right) grade III gliomas. (d) In LM (−) (scatter plot, left) and LM (+) grade IV gliomas (scatter plot, right). (Volcano plot of LMIs in each comparison) Metabolites which having more than 2 (a, b, and d) or 1.5 (c) fold change (FC) and Student t test p value < 0.05 (FDR-corrected) between two groups were colored in magenta

We identified the representative candidate metabolites from the LMI-SSS in the HMDB (Table 2). Sphingolipid metabolites, such as sphingosine and ceramide, which are involved in lipid signaling and apoptosis in cancer cells [16], were elevated in grade III relative to grade IV gliomas. LMIs more abundant in grade IV gliomas were glycol aldehyde, glyceric acid, and acetic acid, which are known to be related to dysregulated glycolysis and ketoacidosis.

Table 2.

Profiles of discriminative low-mass-ion candidates between G III and G IV

Status	Selected LMI (m/z*)	HMDB identifier	Candidate metabolite	Chemical formula	Class
G III > G IV	577.5154 577.5187 577.5221	HMDB0007300	Diacylglycerol	C36H64O5	Fatty acyl
	578.5191 578.5225	HMDB0010399	Lysophosphatidylcholine	C30H60NO7P	Lysophospholipid
	650.6416	HMDB0000831	N-Lignoceroylsphingosine	C42H83NO3	Sphingolipid
		HMDB0004956	Ceramide	C42H83NO3	Sphingolipid
	651.6494	HMDB0000918	1-Oleoyl-cholesterol	C45H78O2	Steroid derivative
		HMDB0056112	Diacylglycerol	C41H78O5	Glycerolipid
	684.1997	HMDB0059644	6β-Hydroxy-1,4,5,6-tetrahydronicotinamide-adenine dinucleotide	C21H31N7O15P2	Purine nucleotide
	859.5273	HMDB0009850	Phosphatidylinositol	C45H79O13P	Glyserophopholipid
		HMDB0116547	Phosphatidylglycerophosphate	C42H84O13P2	Glyserophopholipid
	859.5355	HMDB0009893	Phosphatidylinositol	C45H79O13P	Glyserophopholipid
		HMDB0116547	Phosphatidylglycerophosphate	C42H84O13P2	Glyserophopholipid
		HMDB0059616	All-trans-decaprenyl diphosphate	C50H84O7P2	Prenol lipid
G IV > G III	60.0458	HMDB0031645	Acetamide	C2H5NO	Carboximidic acid
		HMDB0001842	Guanidine	CH5N3	Organonitrogen
		HMDB0000906	Trimethylamine	C3H9N	Organonitrogen
	61.0303 61.0304	HMDB0000042	Acetic acid	C2H4O2	Carboxylic acid
		HMDB0003344	Glycolaldehyde	C2H4O2	Organooxygen
		HMDB0062508	N-Nitrosomethanamine	CH4N2O	Organonitrogen
		HMDB0000294	Urea	CH4N2O	Organic carbonic acid
		HMDB0000820	1-Propanol	C3H8O	Organooxygen
	73.9453	HMDB0034106	Methyl isothiocyanic acid	C2H3NS	Isothiocyanate
	107.0857	HMDB0006372	l-Glyceric acid	C3H8O4	Organic oxygen
	123.0913	HMDB0041570	Isopropylpyrazine	C7H10N2	Diazine
		HMDB0041571	Propylpyrazine	C7H10N2	Diazine
		HMDB0059873	3-Ethylphenol	C8H10O	Organic oxygen
		HMDB0029306	4-Ethylphenol	C8H10O	Organic oxygen

Candidate LMIs discriminating between WHO grade IV glioblastoma and medulloblastoma

Both glioblastoma and medulloblastoma are highly malignant primary brain tumors of WHO grade IV, but their cellular origins and clinical behaviors are different [17]. We found no mutually exclusive LMI by the LMI-SSS (Supporting Information Table 1c). However, at the summed sensitivity/specificity of > 180%, 30 LMIs showed higher expression levels in grade IV gliomas and 15 LMIs showed more abundance in the medulloblastoma group (Fig. 2b, scatter plot). Based on the LMI-MPA, 51 LMIs showed higher levels in medulloblastoma and eight LMIs were significantly more abundant in grade IV gliomas (Fig. 2b, volcano plot). Metabolites for representative LMIs discriminating grade IV gliomas and medulloblastomas by the LMI-SSS are identified and listed (Supporting Information Table 2). In glioblastoma, metabolites were carnitine and glycerophosphate (related to mitochondrial energy production) and phosphatidylinositol and phosphatidylserine (cell membrane phospholipid derivatives). Conversely, the representative LMIs of medulloblastoma were carboxylic acid or pteridine, which were not well-identified as products of biological pathway but seem to be intermediate organic compounds or chemical compounds.

Candidate LMIs discriminating gliomas with or without LM according to the same WHO grade

Among grade III gliomas, eight samples corresponded to LM (+) patients and three to LM (−). Based on the LMI-SSS, 94 LMIs were more abundant in LM (+) than LM (−) and 40 LMIs revealed greater peak areas in LM (−) than LM (+) at complete discrimination in grade III gliomas (Fig. 2c, scatter plot, and Supporting Information Table 1d). By the LMI-MPA, 48 LMIs were significantly more abundant in LM (+) relative to LM (−), and 97 LMIs showed significantly greater abundance in LM (−) than LM (+) grade III gliomas (Fig. 2c, volcano plot). In grade IV gliomas, 61 and 32 LMIs showed complete discrimination (Fig. 2d, scatter plot), with peak areas higher and lower, respectively, in LM (+) than in LM (−) by the LMI-SSS (Supporting Information Table 1e). Volcano plot from the LMI-MPA revealed 115 LMIs were significantly more abundant in G IV gliomas with LM (+) compared with LM (−); meanwhile, 548 LMIs showed significantly higher level in LM (−) than LM (+) grade IV gliomas (Fig. 2d, volcano plot).

Representative LMIs discriminating gliomas with or without LM by the LMI-SSS were identified at HMDB and listed in Table 3. In Table 3 comparing LM (+) and LM (-) gliomas, phospholipid derivatives, such as diacylglycerol and phosphatidylserine, major components of the cell membrane, were more abundant in LM (−) gliomas relative to LM (+) regardless of glioma grade. However, certain LMIs were more abundant in LM (+) than that in LM (−) gliomas according to glioma grade. In LM (+) grade IV gliomas, derivatives of abnormal glycolysis, including acetic acid, glycolaldehyde, and pyruvate aldehyde, predominated as representative LMIs. Meanwhile, the representative LMIs of LM (+) grade III gliomas were varied and included ketone derivatives (ethanone and lactone), catecholamines (norepinephrine and hydroxydopamine), pyruvaldehyde, and cysteine.

Table 3.

Profiles of discriminative low-mass-ion between leptomeningeal metastasis (+) and (−)

Status	Selected LMI (m/z*)	HMDB identifier	Candidate metabolite	Chemical formula	Class
G III LM(+) > LM(−)	60.0466 60.0459 60.0447	HMDB0001842 HMDB0031645 HMDB0003656	Guanidine Acetamide Acetaldehyde oxime	CH5N3 C2H5NO C2H5NO	Organooxygen Organooxygen Organooxygen
	61.0304	HMDB0003344 HMDB0000042	Glycolaldehyde Acetic acid	C2H4O2 C2H4O2	Organooxygen Carboxylic acids
	73.0662	HMDB0001167	Pyruvaldehyde	C3H4O2	Organic oxygen
	122.0955	HMDB0002017	1-Phenylethylamine	C8H11N	Organonitrogen
	165.0549	HMDB0060267	l-Fuculose	C6H12O5	Fatty acyls
	215.0158	HMDB0001213 HMDB0001031 HMDB0001351	1-Deoxy-d-xylulose 5-phosphate Deoxyribose 5-phosphate Deoxyribose 1-phosphate	C5H11O7P C5H11O7P C5H11O7P	Organooxygen Organooxygen Organooxygen
	319.2242	HMDB0013633	12-Ketoeicosatetraenoic acid	C20H30O3	Fatty acyls
	332.0947	HMDB0000905	Deoxyadenosine monophosphate	C10H14N5O6P	Purine nucleotide
	351.2016	HMDB0000268 HMDB0001425	Tetrahydrocorticosterone Estrone sulfate	C21H34O4 C18H22O5S	Steroids Steroids
	399.0329	HMDB0011606	4-alpha-Methyl-5-alpha-cholest-7-en-3-one	C28H46O	Steroids
	536.1656	HMDB0059635	Lipoyl-AMP	C18H26N5O8PS2	Purine nucleotide
	603.5273	HMDB0007503	Diacylglycerol (35:4)	C38H66O5	Glycerolipids
G III LM(−) > LM(+)	59.0516	HMDB0001659 HMDB0003366 HMDB0031652	Acetone Propanal Allyl alcohol	C3H6O C3H6O C3H6O	Organooxygen Organooxygen Alcohols
	111.1168	HMDB0002434	Hydroquinone	C6H6O2	Phenols
	125.0905	HMDB0004181	Methylimidazole acetaldehyde	C6H8N2O	azoles
	149.0232	HMDB0000227 HMDB0002050 HMDB0000606 HMDB0000567	Mevalonic acid Dihydroxyfumaric acid d-2-Hydroxyglutaric acid Cinnamic acid	C6H12O4 C4H4O6 C5H8O5 C9H8O2	Fatty acyls Keto acids Hydroxy acids Cinnamic acids
	629.5409	HMDB0007090	Diacylglycerol (37:5)	C40H68O5	Glycerolipids
	709.5068	HMDB0116692	Phosphatidylglycerol (31:0)	C37H73O10P	Glycerophospholipids
	887.5629 887.5587	HMDB0009900 HMDB0009888	Phosphatidylinositol (38:4)	C47H83O13P	Glycerophospholipids
	905.5981 905.5874	HMDB0116497	Phosphatidylglycerophosphate (40:5)	C46H82O13P2	Glycerophospholipids
	910.5443	HMDB0061595	Phosphatidylserine (9D5/13m5)	C49H84NO12P	Glycerophospholipids
G IV LM(+) > LM(−)	56.0509 56.0516	HMDB0061870	N-Methylene-ethenamine	C3H5N	Organonitrogen
	61.0293	HMDB0000042	Acetic acid	C2H4O2	Carboxylic acid
		HMDB0003344	Glycolaldehyde	C2H4O2	Organooxygen
	60.0460	HMDB0031645	Acetamide	C2H5NO	Carboximidic acid
		HMDB0001842	Guanidine	CH5N3	Organonitrogen
	61.0304	HMDB0003344	Glycolaldehyde	C2H4O2	Organooxygen
		HMDB0000042	Acetic acid	C2H4O2	Carboxylic acid
	73.0657 73.0668	HMDB0001167	Pyruvaldehyde	C3H4O2	Organooxygen
	98.9608	HMDB0000973	Hydrogen phosphate	H3O4P	Non-metal phosphate
	102.0914	HMDB0062600	N-Butylformamide	C5H11NO	Carboxylic acid
		HMDB0001252	Betaine aldehyde	C5H12NO	Organonitrogen
		HMDB0006509	Nervonyl carnitine	C6H16N	Organonitrogen
G IV LM(−) > LM(+)	377.1543	HMDB0000244	Riboflavin	C17H20N4O6	Pteridine
	384.2955	HMDB0013332	3-Hydroxy-5, 8-tetradecadiencarnitine	C21H37NO5	Carboxylic acid
	403.0912	HMDB0001117	4′-Phosphopantothenoyl cysteine	C12H23N2O9PS	Peptidomimetic
	433.2061	HMDB0114743	Lysophosphatidic acid	C21H37O7P	Glyserophopholipid
	722.5060	HMDB0009215	Phophatidylethanolamine	C41H72NO7P	Glyserophopholipid
	790.5873	HMDB0011321	Phosphatidylcholine	C46H80NO7P	Glyserophopholipid
	796.5393	HMDB0112816	Phosphatidylserine (37:5)	C43H74NO10P	Glyserophopholipid
		HMDB0113780	Phosphatidyl-N-methylethanolamine	C43H74NO10P	Glyserophopholipid
	810.5534	HMDB0116758	Phosphatidylserine (38:5)	C44H76NO10P	Glyserophopholipid
		HMDB0113782	Phosphatidyl-N-methylethanolamine	C44H76NO10P	Glyserophopholipid
	812.5632	HMDB0112575	Phosphatidylserine (38:4)	C44H78NO10P	Glyserophopholipid
	898.6161	HMDB0112923	Phosphatidylserine (44:3)	C50H92NO10P	Glyserophopholipid

Analysis for gliomas according to different IDH1 mutation status

Despite insufficient sample numbers (21 IDH1 wild-type samples from 15 patients vs. 4 IDH1 mutant samples from 2 patients), individual LMIs showed significantly different peak areas according to IDH1 mutational status (Supporting Information Fig. 2). Two LMIs showed complete discrimination as the peak areas of these LMIs were higher in IDH1 wild-type compared with IDH1 mutant, but no LMI showed exclusively higher peak areas in IDH1 mutants (Supporting Information Table 1f). However, a total of four and 32 LMIs showed discrimination at the LMI-SSS > 180% in IDH1 wild-type and mutant samples, respectively. Representative LMIs including 2-hydroxyglutarate (2-HG) discriminating between IDH1 mutation status are identified and listed (Supporting Information Table 3). The mean expression level of 2-HG, a well-known product of mutant IDH1 from alpha-ketoglutarate, was significantly higher in IDH1 mutant samples at a threshold level gap of 0.05 in log₁₀ scale compared with IDH1 wild-type samples (Fig. 3).

Fig. 3.

The mean normalized peak area value of a candidate metabolite of 2-hydroxyglutarate (2-HG). Low-mass-ions 149.0229 m/z at a retention time of 18.04 min. which were identified as a candidate metabolite of 2-HG. The mean expression level of 2-HG in IDH1 mutant gliomas is elevated at a threshold gap of 0.05 in log₁₀ scale compared with IDH1 wild-type gliomas

Metabolic pathway analysis using discriminative LMIs

In comparing grade III and grade IV gliomas (Fig. 4a), the LMI-SSS of grade III gliomas denoted two metabolic pathways, the most highly involved of which was sphingolipid metabolism, and the LMI-MPA were greatest the glycerophospholipid metabolism pathways, which were shared across these two different selection methods. The LMI-SSS of grade IV compared with grade III gliomas indicated 10 pathways, and the LMI-MPA indicated 11 pathways overrepresented in grade IV. Among these, nine pathways involving glycolysis or gluconeogenesis; interconversion of pentose and glucoronate; biosynthesis of folate biosynthesis; and metabolism of pyruvate, selenoamino acid, sulfur, taurine/hypotaurine, and vitamin B6 were common in both methods.

Fig. 4.

Metabolite enrichment analysis at KEGG pathway database. Representative LMIs selected based on sensitivity/specificity summation (left), and metabolites representing significantly different mean expression level (right), between (a) grade III and grade IV gliomas, (b) grade IV gliomas and medulloblastoma, (c) grade III gliomas with or without leptomeningeal metastasis (LM), and (d) grade IV gliomas with or without LM

The pathways identified by comparing grade IV gliomas and medulloblastoma were four in both methods (Fig. 4b). Among pathways overpresented in grade IV gliomas, sphingolipid metabolism was shared across these. In selection of pathways involved in medulloblastoma, folate biosynthesis is common among these.

In LM (+) or (−) grade III gliomas (Fig. 4c), the LMI-SSS and the LMI-MPA indicated three significantly involved metabolic pathways overrepresented in LM (−) tumors, respectively. Among those three pathways, glycero(phospho)lipid metabolism was common to both methods. The LMI-SSS and the LMI-MPA indicated 12 and 13 pathways, respectively, overrepresented in LM (+) tumors. Among those, 12 pathways were common to both methods, including glycolysis or gluconeogenesis; biosynthesis of fatty acids or folate; interconversion of pentose and glucoronate; and metabolism of glyoxylate/dicarboxylate, linoleic acid, pyruvate, selenoamino acid, sulfur, taurine/hypotaurine, and vitamin B6.

In LM (+) and (−) grade IV gliomas (Fig. 4d), the LMI-SSS and the LMI-MPA of LM (−) grade IV gliomas showed four and six overrepresented pathways, respectively. Among those, glycerophospholipid metabolism is common to both methods as it was in LM (−) grade III gliomas. In identifying pathways overrepresented in LM (+) grade IV gliomas, the LMI-SSS indicated nine pathways, and the LMI-MPA significantly involved in 10 pathways. Among these, all nine pathways indicated by the LMI-SSS were presented in the pathways of the LMI-MPA. Those were metabolism of pyruvate, “glycine, serine, and threonine,” sulfur, selenoamino acid, vitamin B6, “glyoxylate and diacarboxylate,” folate biosynthesis and “pentose and glucoronate interconversions” pathway.

Discussion

Our study is the first report of CSF metabolomic profiles that discriminate among glioma grade, WHO grade IV primary brain tumors (glioblastoma and medulloblastoma), and development of LM.

Advancement of MS to identify discriminating LMIs

High-performance liquid chromatography/MS (HPLC/MS) coupled with time-of-flight (TOF) MS is widely used to yield accurate qualitative and quantitative analysis in metabolomics. Previous work using matrix-assisted laser desorption ionization (MALDI)-TOF MS in 2005 was capable of identifying only 164 discernible peaks with 76% sensitivity for predicting LM from breast cancer [18]. The current MS technique provides much higher resolution when identifying LMIs in biofluids. Relative to targeted metabolomics, a non-targeted approach may offer global information on whole-individual LMIs in biofluids such as CSF; because the identification of LMI metabolites in non-targeted metabolomics is dependent on readily available HMDB information and tandem MS patterns of standard metabolites, the mass peak area of LMIs may provide information on the relative abundance of selected metabolites. In the present study, a total of 10,193 LMIs were evaluated by non-targeted metabolomics in 32 glioma, 7 medulloblastoma, and 10 non-tumor control CSF samples. This high-resolution triple-TOF MS identified metabolic profiles that can discriminate between detailed characteristics of glioma.

Clinical usefulness of CSF metabolomics profiles for glioma

Recently, studies searching for biomarkers of cancers in biofluids have proliferated to include methods such as circulating tumor cells, cell-free DNA, extracellular microRNA, and exosome sequencing. However, the minute concentration of biomarkers in serum and technical difficulties in identification and quantification of this mutation limits its clinical use. However, given the direct contact of CNS tumors with CSF, CSF should express such biomarkers more prolifically than serum. Although CSF sampling is somewhat invasive, it is less invasive than brain tumor biopsy, and using CSF as a liquid biopsy permits longitudinal sampling, which has special utility in gliomas because they are not a fixed form of disease but instead show progression to malignant transformation or LM. Although targeted MS-based metabolomic analysis of CSF has indeed identified signatures of malignant glioma[12, 19], the clinical usefulness of metabolomics analysis of CSF has not yet been verified through large sample sizes or prospective clinical trials.

However, the use of CSF as a liquid biopsy for glioma needs to be accompanied by the presence of a tumor. In a study of metabolite profiles of serum from patients with lung cancer [20], profiles differed between preoperative and postoperative samples in the same patient, indicating that serum metabolites depend on the existence of tumor tissue. Monitoring of glioma recurrence using CSF microRNA expression level was tried by Teplyuk et al. They tried to correlate the onco-miR expression level and tumor size on MRI [21]. But the evidence level was a case report and they could only suggest the relative change of microRNA expression level. Unfortunately, we could not find any study dealing with metabolite abundance and glioma recurrence yet. Nakamizo et al. performed targeted MS of CSF from 32 patients with histologically proven gliomas, which was collected at preoperative period [19]. They found that citric and isocitric acid levels discriminate glioma grades, and lactic acid levels were significantly higher in IDH1 mutant gliomas compared with IDH1 wild-type. However, tumor location and gadolinium enhancement did not show significant difference of their targeted LMIs. Recently, many clinical trials have been performed to inhibit IDH1 mutant gliomas with a concept of synthetic lethality using direct IDH1 mutant inhibitor, depletion of intermediate with provoking autophagy, and vaccination recognizing IDH1 R132H as a neo-antigen [22]. We expect an encouraging result from this trial of metabolic pathway target as a personalized medicine.

Relevance of identified discriminative LMIs for different gliomas

In malignant gliomas, the activation of EGFR signaling stimulates cellular membrane biogenesis and energy utilization, such as the Warburg effect, leading to increased fatty acid and glucose metabolism [23]. Furthermore, altered metabolic enzymes utilize the tricarboxylic acid cycle and oxidative phosphorylation differently than normal tissue [22]. Thus, the distinctive LMI profile in gliomas is likely due to an increase in the energy substrate or cell membrane components derived from these cells.

Indeed, we found LMIs related to sphingolipid metabolism abundant in grade III relative to grade IV gliomas. Inhibiting sphingosine kinase activity increases ceramide, and sphingosine levels inhibit cell proliferation and promote apoptosis [24]. Conversely, glycolaldehyde and l-glyceric acid were more abundant in grade IV compared with grade III gliomas, which is perhaps unsurprising given their role in energy metabolism [25, 26] and association with tumor aggressiveness [27], respectively. We identified several lipid derivatives differentially observed between glioblastoma and medulloblastoma. Due to the aberration of lipid metabolism not only in glioma but also in medulloblastoma, several studies have tried to identify a lipid marker to evaluate a disease status [27], whereas others describe elevation of choline-containing phospholipids in both glioblastoma and medulloblastoma tissues [28], the relative ratio of lipid components was different between the tumors, as measured by nuclear magnetic resonance, and these differences were manifest in the CSF samples. In our study, relatively higher m/z products (e.g., 887.5594, 888.5598, 905.5939) were found in glioblastoma than medulloblastoma CSF (Supporting Information Table 1). These metabolites were phosphatidylinositol, phosphatidylserine, or galabiosylceramide.

Differences among lipid metabolites in CSF metastasis have been previously reported in medulloblastomas, in which disease progression frequently includes LM. A mouse model of human medulloblastoma features a distinctive lipid signature of tumor tissue according to LM status, including reduced phosphatidic acids (PA), phosphatidylethanolamine (PE), and phosphatidylserine (PS) and increased phosphatidylinositol (PI) in metastasizing medulloblastoma tissues [29]. In our study, LM status was not balanced between grade IV gliomas and medulloblastomas. Thus, the difference of LMIs between these tumors should be cautioned to interpret. Pyruvaldehyde, a detrimental product of glucose, protein, and fatty acid metabolism, was increased in LM (+) relative to LM (−) grade III glioma. Because pyruvaldehyde is rapidly eliminated by the glyoxalase system [30], its accumulation could suggest a disorder in energy metabolism in LM (+) glioma.

In 2016, WHO classification of gliomas was revised, and IDH1 enzyme mutation status became a criterion for glioma classification [17]. The pathological role of IDH1 mutation in gliomas has not yet been described, but mutation occurs early in glioma development [31], and accumulation of its byproduct, 2-HG, may provoke malignant progression via hypermethylation [32]. In our study, LMI 149.0229, a candidate LMI of 2-HG, showed higher expression in IDH1 mutant than IDH1 wild-type samples, in keeping with others’ findings [13]. Among 21 IDH1 wild-type samples, four samples showed higher levels of 2-HG than IDH1 mutant samples (19% false negative at a sensitivity 81% and a specificity 100%); conversely, others have reported 15% false-positive rates [13]. All of our high-2-HG IDH1 wild-type samples were from fulminant, large glioblastomas, whereas three out of our four IDH1 mutant samples were from a patient with a small, anaplastic oligodendroglioma (0.5 × 1.5 cm at the largest diameter). Standardization of LMI expression based on tumor size should be considered in future studies.

Limitations and future directions

Limitations of this study were as follows: (1) LMIs are not specific molecules but instead suggest a candidate for targeted MS verification, (2) an absolute threshold level of LMIs in each glioma group was not analyzed, and (3) the relevance of metabolomic profiles to discriminate among types of primary brain tumors was not a proof-of-concept but, rather, suggested specific pathways based on metabolic pathway database analysis, and (4) as CSF sample collection in this study was not for study only but for instance of CSF cytology, or diversion, the number of samples is both limited and uneven to analyze each glioma grade and characteristics. Thus, we had to omit some meaningful analysis here.

In this study, we analyzed potential pathways involved in malignant transformation of gliomas (grade III vs. grade IV), leptomeningeal metastasis (LM (+) vs. LM (−)), and invasiveness (glioblastoma vs. medulloblastoma) using discriminative LMIs. The LMI-SSS might be a more appropriate approach for further investigation with targeted MS than the LMI-MPA in terms of exclusiveness. However, the exclusivity of discriminative LMIs should be re-evaluated with additional samples (or underwent external validation) unless the LMI was a metabolite unique to the glioma mutation, which never existed in the control group. Alternatively, profiling representative LMIs based on p value (the LMI-MPA) could be complemented using an algorithm such as LOME. Fortunately, differences among these two methods seemed not quite different in our study; as the pathways indicated by the LMI-SSS and the LMI-MPA were common to both methods, especially in comparing malignant characteristics (grade IV over grade IV, LM (+) over (−)), where more than 90% of designated pathways were common to each other.

Metabolomic profiles can reflect tumor behavior because they are the end product of tumor cells’ effort to survive and interact with the microenvironment. Relevant metabolomic targets downstream of predicted mutations will likely provide us with valid metabolomic targets as biomarkers [33–35]. As a chemical fingerprint combined with multi-omics data, metabolomic profiles of biofluids can help us understand the biology of glioma development and provide novel therapeutic targets [36–38].

Conclusions and expert recommendations

CSF metabolomic profiles from non-targeted MS analysis revealed significantly different profiles that varied according to glioma characteristics such as glioma grade, LM, and non-glial primary brain tumors of the same tumor grade. These results suggest that the important progression of glioma including malignant transformation and leptomeningeal metastasis could be predicted by CSF metabolomic analysis. Through a future validation trial of targeted MS based on candidate LMIs of this study, we expect that CSF analysis can support or replace the need for brain tumor biopsy and CSF cytology. Although we provide just discriminative metabolomic profiles, candidate molecules identified from HMDB indicate relevant metabolites such as lysophosphatidylcholine of membranous signaling control and acetic acid of ketone acidosis. As evidences of dysregulated metabolism and membrane moiety modulation in cancer cells have been built up, we expect to find a node to switch off glioma progression as a beginning of preventive medicine. CSF analysis as a liquid biopsy could verify genomics and transcriptomic/ proteomic profiles with recently advanced technology. Integrative analysis of CSF multi-omics results is expected to provide not only understanding of pathophysiology at molecular level but also finding druggable molecular targets. Regarding with an advancement of high-throughput drug screening system [39], we expect to find an inhibitor that achieve a personalized medicine in neuro-oncology as early as possible.

Abbreviations

CNS

central nervous system

CSF

cerebrospinal fluid

HMDB

Human Metabolome Database

HPLC/MS

high-performance liquid chromatography/MS

IDH1

isocitrate dehydrogenase 1

LM

leptomeningeal metastasis

LMIs

low-mass ions

LOME

low-mass-ion discriminant equation

MALDI

matrix-assisted laser desorption ionization

MS

mass spectroscopy

ODG

oligodendroglioma

PA

phosphatidic acids

PE

phosphatidylethanolamine

PPPM

predictive preventive personalized medicine

PS

phosphatidylserine

TOF

time-of-flight

Authors’ contributions

BC. Y, JH. L, and H-S. G designed and supervised the study, and analyzed data. JW. K, SH. S, H. Y, and K-Y. L contributed to acquisition of data. JH. I, TH. K, and KH. K contributed to the analysis and interpretation of data. JH. K and JB. P designed the study. All authors read and approved the final manuscript.

Funding information

This work was supported by grants from the National Cancer Center, Korea (NCC-1710871-3, 1910090-1 and 1910294-1), the Korea Health Industry Development Institute of Ministry of Health and Social Welfare, Republic of Korea (H1731340-2), and the Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and Future Planning (NRF-2017R1A2B4007859).

Compliance with ethical standards

Conflict of interest

The authors declare that they have no conflict of interest.

Ethics approval and consent to participate

The study procedure was approved by National Cancer Center and all patients signed consent forms. All CSF samples were obtained after Institutional Review Board (NCC-150002) approval in May 2015; written informed consent was obtained from each participant before the collection. The ethics approval was given in compliance with the Declaration of Helsinki.