Transcriptomic profiles of CSF from patients with and without medulloblastoma
Most studies attempting to profile CSF have focused on circulating tumor DNA (ctDNA) due to the relative ease of analysis of stable DNA fragments, including in MB [12, 32, 33]. Despite CSF also containing RNAs, due to their low abundance and lability, most studies have used targeted approaches to profile miRNAs and mRNAs in CSF from patients with various CNS tumors. Recognizing the need to systematically profile RNAs in biofluids due to their biomarker potential, Hulstaert et al. recently published a comprehensive atlas of the extracellular transcriptomes of human biofluids, including CSF, but their analysis was limited to a comparison of profiles of patients with hydrocephalus and glioblastoma and no MB patient was profiled . There has yet to be a comprehensive and systematic analysis of RNA species in the CSF of MB patients.
We therefore established global transcriptomic differences in the CSF of patients with (n = 40) and without (n = 11) MB representing different molecular subtypes (Additional file 4: Table S1). Each CSF samples showed varied read counts and mapping rate (Fig. 1a). By both principal component analysis (PCA) and unsupervised clustering, CSF samples separated into two distinct groups according to the presence or absence of MB (Fig. 1b–d). Although, there was no clear separation into molecular subtypes, one hundred and ten genes were differentially expressed in CSF samples from patients with and without MB (Fig. 1c, Additional file 4: Table S2, and Additional file 4: Fig. S1; log2 fold-change (FC) < -2 or > 2; adjusted p-value < 0.05) that were enriched for several pathways by geneset enrichment analysis (GSEA) : TGF-β signaling (SKI, FKBP1A, ID2, RHOA, BMPR1A; false discovery rate (FDR) 2.59E-04), TNF-α signaling via NF-kB (TSC22D1, DUSP1, ID2, KLF9, FOS, IL6ST, SAT1; FDR 1.19E-03), and adipogenesis (ALDH2, CMPK1, APOE, UQCR10, TOB1, YWHAG; FDR 4.51E-03). TGF-β has previously been implicated in the progression of MB , perhaps by suppressing the anti-tumor effects of cytotoxic T cells , and the other identified pathways warrant further exploration.
We next examined expression of circular RNAs (circRNAs), a novel class of non-coding (nc)RNAs with a covalently closed loop structure derived from the host gene’s RNA splicing by back splicing. Although generally present at low abundance , since circRNAs do not have exposed ends, they are more resistant to degradation and more stable than linear RNAs , making them ideal biomarkers for detection in human biofluids including blood , saliva , semen , urine , and CSF . CircRNA expression levels in CSF were low, ranging from mean read counts 203 to 1850 in samples from MB patients and only 8.57 ± 5.09 in normal samples. Nevertheless, 10 circRNAs were differentially expressed between MB and non-MB groups (log2 FC < -1 or > 1; adjusted p-value < 0.1) (Fig. 1e, Additional file 4: Table S3, Additional file 1). Of these, circ_463 was the most upregulated and abundant circRNA in MB CSF, as confirmed by qRT-PCR (Fig. 1f).
Circ_463, also known as ciRS-7 or CDR1as, was originally identified as a highly expressed circRNA in human and mouse brains . It contains 73 miR-7 seed targets and functions as a miR-7 sponge with an unknown role in the brain . In cancers, ciRS-7 promotes growth and metastasis of esophageal squamous cell carcinoma , and its silencing in melanoma drives IGF2BP3-mediated invasion and metastasis . In multiple myeloma, its expression is downregulated in immunomodulatory drug resistant cell lines, and depletion of ciRS-7 increased the CpG methylation of its host gene LINC00632 . While there have been a few very recent reports of circRNA expression in MB tissues and cells demonstrating potential oncogenic function for overexpressed transcripts [49,50,51], this is the first circRNA analysis of CSF in MB patients. Therefore, circ_463 appears to be pleiotropic, with overexpression in CSF samples of MB patients suggesting a novel oncogenic role in this context.
The metabolic differences in CSF from patients with and without medulloblastoma
Global metabolomics has become an important unbiased approach to identify diagnostic, prognostic, and predictive biomarkers in human disease [17, 52], and altered metabolism is a hallmark of cancer cells, which need to adapt to their nutrient-poor microenvironment to sustain their viability . Although it is clear that cancer cells have altered metabolism, it is less clear to what extent this influences the CNS microenvironment and the CSF. Like other tumors, several studies have established that metabolism is altered in primary and recurrent MB, including decreased fatty acid oxidation, increased lipogenesis, and a glycolytic phenotype reflected in the detection of MB by 18FDG-PET . However, there have been fewer comprehensive studies of the CSF metabolome in CNS tumors and in MB specifically. Metabolite analysis of the CSF in glioma patients identified differences in the abundance of 43 metabolites compared with controls , while in MB, Reichl et al. detected upregulation of hypoxia-induced proteins and metabolites (up-regulation of tryptophan, methionine, serine and lysine) in MB CSF . However, the full metabolomic landscape of CSF in MB has not been accurately or fully quantified.
Therefore, we performed comprehensive untargeted metabolic profiling of the brain CSF samples using ultra high-pressure liquid chromatography and high-resolution mass spectrometry (UHPLC-HRMS). Metabolite data were collected in a randomized manner to avoid bias. Using flank feature filtering (BFF) to eliminate false peaks, 3995 true metabolic features were identified, of which 352 metabolites were identified as level 1 (highest level of confidence in the annotation). Similar to the transcriptomic profiles, PCA and unsupervised clustering of differentially expressed metabolites revealed clear separation of metabolic profiles between normal and MB CSF (Fig. 2a and c) but not between different molecular subtypes. The majority of differentially regulated metabolites (FC > 1.5; FDR p < 0.05) were upregulated in MB samples (Fig. 2b and Additional file 4: Table S4). Exploratory pair-wise metabolite profile discrimination between normal and different MB sub-groups confirmed that differentially expressed metabolites clearly distinguished different molecular subgroups of MB (Additional file 4: Fig. S2). Uniquely elevated (Additional file 4: Fig. S3) and downregulated (Additional file 4: Fig. S4) metabolites in the different MB subtypes were analyzed using volcano plot-based differential statistical analysis (p-value < 0.05, fold change ≥ 1.5).
We next performed KEGG metabolic pathway analysis of significantly differentially expressed metabolites (hypergeometric test, relative betweenness centrality, p-value < 0.05) (Fig. 3a). The TCA cycle, alanine, aspartate, and glutamate metabolism, and arginine biosynthesis pathways were all upregulated in MB, particularly in SHH, group 3/4, and group 4 tumors. Given that CSF metabolic profiles did not discriminate between molecular subgroups, we established which metabolites were uniformly expressed in all MB subtypes and might therefore be candidate diagnostic biomarkers for MB. α-ketoglutarate (Fig. 3b), fumarate (Fig. 3c), hydroxypyruvate (Fig. 3d), malate (Fig. 3e), and succinate (Fig. 3f) from the TCA cycle and N-acetyl-aspartate (Fig. 3g) from the alanine, aspartate, and glutamate metabolism pathway were all significantly elevated in all different sub-groups of MB; citrate, isocitrate, and trans-aconitate (Additional file 4: Fig. S5A-C; TCA cycle) and GABA (Additional file 4: Fig. S5D; alanine, aspartate, and glutamate metabolism) showed minor but significant downregulation in MB. For validation, α-ketoglutarate, fumarate, malate, and succinate (Fig. 3h) from the TCA cycle and N-acetyl-aspartate were all significantly upregulated by targeted quantification (Fig. 3i). Finally, anserine (Additional file 4: Fig. S5E; histidine and beta-alanine metabolism) and S-(5′-adenosyl)-L-methionine (arginine biosynthesis; Additional file 4: Fig. S5F) were significantly upregulated and 5-oxo-L-proline (glutamine and glutamate metabolism; Additional file 4: Fig. S5G) significantly downregulated in MB compared with normal. Collectively, these data suggest that a broad range of metabolites in the CSF, particularly those involved in the TCA cycle, distinguish MB from normal. This is consistent with a more general model of proliferating MB cells not only using the TCA cycle to fuel the need for reducing equivalents in the form of NADPH  but to provide metabolic precursors for the biosynthesis on non-essential amino acids, since upregulated α-ketoglutarate indicates (i) a continuous supply of glutamine maintaining the integrity of the cell cycle ; (ii) maintaining the cell’s ability to synthesize citrate for energy production and de novo lipogenesis, since α-ketoglutarate is oxidized to oxaloacetate to maintain citrate production and oxaloacetate can be converted to malate and then pyruvate to produce NADPH in a glucose-independent manner .
Lipidomic alterations in medulloblastoma CSF
Lipids are fundamental and abundant biomolecules in cells that have structural, transport, energy storage, and cellular signaling roles. Unsurprisingly, therefore, they all play critical roles in many diseases including cancer ; however, there is little available information on the lipid profiles of human MB. Tissue analysis suggests that human MBs may have high lipid levels, at least in contrast to other pediatric brain tumors , and a lipidomic analysis of a mouse model of SHH MB determined 34 upregulated lipids associated with metastasis . Given that biofluid lipidomes might provide a rich source of biomarkers and provide insights into the underlying biology of MB, we proceeded to examine CSF lipid profiles.
Using LipidMatch, 727 lipids were identified in all samples including predicted lipids (Additional file 2), and 14 of these were differentially expressed based on fold change threshold 1.5 and p-value < 0.05 (11 lipid species elevated and 3 downregulated) in the CSF of MB patients compared with normal (Fig. 4a). To understand the role of specific lipids in MB, we conducted lipid class analysis between MB and normal. Total triacylglycerols (TGs; n = 171) were significantly upregulated in MB (Fig. 4b) and diacylglycerols (DGs; n = 17) (Fig. 4c), monogalactosyldiacylglycerol (MGDG; n = 19) (Fig. 4d), cholesterol ester (CE; n = 14) (Fig. 4e), phosphatidylcholine (PC; n = 85) (Fig. 4f), N-hexadecanoyl hexosylceramide (HexCer; n = 6) (Fig. 4g), sphingomyelin (SM; n = 51) (Fig. 4h), and oxidized lipids including oxLPC (lysophosphatidylcholine; n = 2), oxLPE (lysophosphatidylethanolamine n = 1), oxPC (n = 20), oxPE (n = 21), and oxTG (n = 13) (Fig. 4i) were significantly downregulated in MB compared with normal. Together with the increase in α-ketoglutarate noted above, these lipid profiles might reflect a state of hypoxia in MB CSF because (i) cancer cells accumulate TGs due to hypoxia ; and (ii) hypoxia can create a deficit of glucose-derived acetyl-CoA, requiring the conversion of α-ketoglutarate into citrate so that it can be then used to generate acetyl-CoA . From the practical perspective, CE, HexCer (CerG1), and SM may be promising CSF biomarkers for MB.
Integrative analysis of transcriptome, lipidome, and metabolome
Given that the transcriptome, lipidome, and metabolome are integrated and interrelated biological systems that modulate phenotype, we next performed a multivariate analysis to integrate the molecular changes characterizing the CSF of MB patients using the data integration analysis for biomarker discovery DIABLO method in the mixOmics R package . The DIABLO method identified several important features discriminating cancer from normal through interrogation of correlations between the three omics datasets.
The first component of sparse partial least-squares discriminant analysis (sPLS-DA)  of the combined transcriptomic, metabolomic, and lipidomic datasets clearly discriminated normal from MB CSF samples (Fig. 5a), with the transcriptomic and metabolomic data showing the highest discriminatory capacity and correlations (Fig. 5b and Additional file 4: Fig. S6). To obtain the best discriminative features, the minimum loading coefficient for the first component of sPLS-DA was set at ± 0.15 for each data block. This filtering (Fig. 5c and d) identified n = 19 transcripts, n = 28 metabolites, and n = 16 lipids that best distinguished MB from normal samples (Fig. 5e). Among 19 RNA transcripts, ten were validated by qRT-PCR (Additional file 4: Fig. S7). The integration of data using multi-omics tools is indispensable for cancer metabolism studies . Finally, to visualize the between-omics correlations in the DIABLO analysis, a Circos plot (Fig. 5f) revealed a number of strong positive and negative correlations; for example, UFM1 was positively correlated with S-adenosyl-L-methionine (Pearson's r = 0.76) and LPC 17:0 (Pearson's r = 0.6) and LPC 17:0 was positively correlated with S-adenosyl-L-methionine (Pearson's r = 0.66). UFM1 (ubiquitin-fold modifier 1) has been identified as an important factor associated with microcephaly by affecting cell cycle regulation and cancer development  while, in a preliminary study, S-adenosyl-L-methionine found to modulate cell cycle progression in cancer . We further analyzed the MAGIC (Medulloblastoma Advanced Genomics International Consortium (https://plone.bcgsc.ca/project/magic) ; Additional file 4: Fig. S8 and Additional file 4: Fig. S9) datasets and found 17 out of the 19 differentially expressed RNAs in different MB subtypes (Fig. 5f).