www.impactjournals.com/oncotarget/ Oncotarget, Advance Publications 2017
A cerebrospinal uid microRNA signature as biomarker for
Johnny C. Akers1,*, Wei Hua2,*, Hongying Li3,*, Valya Ramakrishnan1, Zixiao Yang2,
Kai Quan2, Wei Zhu2, Jie Li1, Javier Figueroa1, Brian R. Hirshman1, Brittney Miller1,
David Piccioni4, Florian Ringel5, Ricardo Komotar6, Karen Messer3, Douglas R.
Galasko7, Fred Hochberg1, Ying Mao7,**, Bob S. Carter1,** and Clark C. Chen1,**
1 Center for Theoretical and Applied Neuro-Oncology, University of California, San Diego, CA, USA
2 Department of Neurosurgery, Huashan Hospital, Fudan University, Shanghai, China
3 Biostatistics Department, Moores Cancer Center, UC San Diego Health System, La Jolla, CA, USA
4 Department of Neurosurgery, Moores Cancer Center, UC San Diego Health System, La Jolla, CA, USA
5 Department of Neurosurgery, Klinikum rechts der Isar, Technische Universität München, Munich, Germany
6 Department of Neurological Surgery, Miller School of Medicine, University of Miami, Miami, FL, USA
7 Department of Neurosciences, University of California, San Diego, CA, USA
8 State Key Laboratory of Medical Neurobiology, Institutes of Brain Science, The Collaborative Innovation Center for Brain
Science, Fudan University, Shanghai, China
* These authors shared responsibility as rst authors
** These authors shared responsibility as senior authors
Correspondence to: Clark C. Chen, email: firstname.lastname@example.org
Keywords: extracellular vesicle, CSF, liquid biopsy
Received: April12, 2017 Accepted: May 19, 2017 Published: June 01, 2017
Copyright: Akers et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC-BY),
which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Purpose: To develop a cerebrospinal uid (CSF) miRNA diagnostic biomarker for
Experimental Design: Glioblastoma tissue and matched CSF from the same
patient (obtained prior to tumor manipulation) were proled by TaqMan OpenArray®
Human MicroRNA Panel. CSF miRNA proles from glioblastoma patients and controls
were created from three discovery cohorts and conrmed in two validation cohorts.
Results: miRNA proles from clinical CSF correlated with those found in
glioblastoma tissues. Comparison of CSF miRNA proles between glioblastoma
patients and non-brain tumor patients yielded a tumor “signature” consisting of
nine miRNAs. The “signature” correlated with glioblastoma tumor volume (p=0.008).
When prospectively applied to cisternal CSF, the sensitivity and specicity of the
‘signature’ for glioblastoma detection were 67% and 80%, respectively. For lumbar
CSF, the sensitivity and specicity of the signature were 28% and 95%, respectively.
Comparable results were obtained from analyses of CSF extracellular vesicles (EVs)
and crude CSF.
Conclusion: We report a CSF miRNA signature as a “liquid biopsy” diagnostic
platform for glioblastoma.
Glioblastoma, dened by the World Health
Organization (WHO) glioma classication as grade IV
astrocytoma, is the most common form of primary brain
cancer in adults [1, 2]. Diagnosis of the disease remains
a clinical challenge. First, error in diagnosis occurs in
up to 30% of the instances where clinical decisions are
based solely upon Magnetic Resonance Imaging (MRI)
. As such, diagnosis of the disease requires tissue
acquired through cranial surgery . However, morbidity
for biopsy surgical resection of glioblastoma involving
eloquent regions of the cerebrum can be as high as 10%
, with permanent neurologic injury for a subset of
these patients . The risk is higher for surgical resection
involving eloquent cerebrum . Second, a subset of brain
tumor patients present with co-morbidities that prohibit
consideration for surgery. Analysis of the Surveillance,
Epidemiology, and End Results (SEER) registry suggests
that ~20% of all aicted patients are medically too
ill to be considered for surgery . We propose that
these challenges can be addressed by development of a
minimally invasive “liquid biopsy” platform .
CSF is an appealing and accessible bio-uid for
glioblastoma “liquid biopsy”. The bio-uid lies in close
proximity to tumor tissue, often bathing tumor or its
associated microenvironment . The CSF can be
located in the brain or its ventricles, which we termed
“cisternal” CSF, or the lumbar region, which we termed
“lumbar” CSF. CSF in these compartments diers in
chemical compositions [11, 12], suggesting limited CSF
exchange between these two anatomic compartments.
Whether these dierences impact their diagnostic value
for glioblastoma remains an open question.
Extracellular Vesicles (EVs) are cell-secreted
vesicles that range 30-2000 nm in size that mediate
critical biologic functions, including cellular remodeling
and intracellular communication . Cancer cells exhibit
increased secretion of EVs, with secreted EVs containing
genetic contents reective of the cell of origin [9, 13]. In
this context, there is a growing interest in EVs derived
from bio-uids, including CSF , as a platform for
disease diagnosis .
Here, we examined miRNA proles of the CSF
EVs and “crude” CSF derived from glioblastoma patients.
miRNA is an attractive biomarker platform given its
stability in bio-uids , selective over-expression in
glioblastomas [13, 16, 17], and release by tumor cells into
the extracellular environment . Our results support
the utility of CSF miRNA proling as a “liquid biopsy”
platform for glioblastoma diagnosis.
miRNA proling of matched glioblastoma tumor
and CSF EVs in human subjects
We rst investigated whether the miRNA prole
from CSF mirrored that of the matched glioblastoma
specimen within the same subject, using the 15 subjects
with matched CSF and tumor tissue from Cohort 1.
Using a CT cut-o of 35, we found that 200-400 miRNAs
were detected in the glioblastoma specimens (median
313 species; range 238 to 351). Between 30-50% of
these miRNAs were detected in the matched EV CSF
(Figure 1A). However, the average CT value at which
these miRNAs were detected in CSF was increased by
~5 (Supplementary Figure 1), translating to a 30-fold
decrease in abundance. We plotted the level of each
detectable miRNA in CSF (Figure 1B, y-axis) against
its level in the glioblastoma sample (Figure 1B, x-axis)
and found correlation between CSF miRNAs and tumor
miRNAs for all 15 paired samples. These results suggest
that the miRNA content of CSF mirrors that of matched
Comparison of CSF fractions for number of
For select samples, miRNA proling was performed
for both CSF derived EV and crude CSF. In general, more
miRNA species were detected in the crude CSF relative
to EV. Nearly all miRNAs detected in the EVs were also
present in the crude CSF (Figure 1C).
Identication of a miRNA CSF signature which
can identify glioblastoma
Though all CSF samples were collected using the
same Standard Operating Procedure (SOP), signicant
variation in miRNA proles were found between CSF
derived from the rst three cohorts (cohorts 1, 2, and
3). To account for this variability, we used miRNA
proles derived from all three cohorts in our signature
development. Details of the analysis can be found in
Supplementary Figure 2. In brief, miRNAs with levels
that diered between glioblastoma and non-oncologic
CSF were identied using the criteria of FDR < 0.2
and log(fold-change) > 2 as described. From Cohort 1,
we identied 29 miRNAs. 3 miRNAs were identied
in Cohort 2. In Cohort 3, we identied 110 miRNAs
as dierentially expressed, with miR-21 having the
largest fold change as previously published . Based
on our cross-sample validation criteria, 24 miRNAs
were subsequently selected for signature development
(Supplementary Figure 3). In addition, three dierentially
regulated miRNAs which validated in one (but not two)
independent datasets were added to the candidate set.
miR-548a, stably expressed across the three data sets and
potentially useful as a reference miRNA, was also added
to the panel, yielding a total of 28 candidate miRNAs.
We then used LASSO  to develop a classier
from these 28 candidate miRNAs using cross-validated
minimum deviance as the model selection criterion
(Figure 2A). LASSO analysis indicated an optimal
classier consisting of 9 miRNAs, including 5 miRNAs
that were enriched (miR-21, -218, -193b, -331, and
Figure 1: miRNA analysis of matched glioblastoma tumor and CSF samples. miRNA prole of matched glioblastoma tumor
and CSF EV samples were analyzed using the TaqMan OpenArray platform. A. Venn diagrams indicating the unique and shared detectable
miRNAs between tumor tissue and CSF EVs. B. Correlation between miRNA proles of matched glioblastoma specimens and CSF. For
each patient, CT values of shared miRNAs in tumor specimen were plotted against CT values from CSF EVs. Pearson correlation coecient
was then calculated for each patient. The correlations were highly signicant for all matched pairs of tumor and CSF specimens. C. Venn
diagrams comparing the miRNA prole of crude CSF versus CSF EV. > 95% of miRNA found in CSF EVs were also represented in the
-374a) and 4 miRNAs that were depleted (miR-548c,
-520f, 27b, and 130b) in glioblastoma CSF (Figure 2B).
We then determined the optimal score cuto (0.4) below
which we classied a subject as non-glioblastoma and
above which we classied a subject as with a diagnosis
of glioblastoma. Both the signature coecients and the
cuto for classication as glioblastoma were documented
before proceeding to the validation step.
Correlation of the CSF miRNA signature score
with tumor volume
Pre-operative MRI was available for 11 of the
patients in Cohort 1. We created tumor volumes based on
the Agfa CD Viewer and related these to the CSF miRNA
gene signature scores. A positive correlation was observed
Figure 2: Identication of miRNA signature. Dierentially expressed miRNAs between glioblastoma and non-oncologic CSF
samples were selected from miRNA qPCR array based on FDR < 2 and log(fold-change) > 2 and cross-validated using multiple cohorts. A.
28 candidate miRNAs was used to train a classier with LASSO using a using cross-validated minimum deviance as the model selection
criterion, B. yielding a 9 miRNA signature.
Table 1: Patient demographics and samples
Discovery Discovery Discovery Validation Validation
UCSD, Munich, Miami
Cisternal and lumbar CSF
Age, Median (Range) 61
(22-84) 53.5 (29-74) 58
Female 17 32 13 5 23
Male 22 35 19 17 15
Glioblastoma 24 40 13 10 18
Normal/non-oncologic 15 27 19 12 20
Cisternal 26 0 32 22 0
Lumbar 13* 67 0 0 38
Tumor tissue yes no no no no
*All 13 lumber CSF samples from Cohort 1 were from the glioblastoma group
between miRNA signatures and tumor volumes (Figure
3A). Glioblastoma with volumes < 15 cc had lower
miRNA scores than those with > 15cc’s (P < 0.0001,
Validation of the CSF miRNA glioblastoma
We tested the performance of the 9-miRNA
signature in a prospective manner. Since most EV
miRNAs are also detected in crude CSF, we opted to
validate our signature using unfractionated crude CSF.
We prospectively collected and proled cisternal CSF
from an additional 22 patients (Cohort 4: 10 glioblastoma
and 12 non-oncologic patients). Using the cuto (0.4)
established during the discovery process, the signature
correctly identied 8/10 subjects with glioblastoma and
8/12 non-oncologic subjects, yielding a sensitivity of 80%
and specicity of 67%. The AUC was 0.75 (95% CI 0.53,
0.97) (Figure 4A).
We also prospectively collected and proled
lumbar CSF from 18 glioblastoma and 20 non-oncologic
patients (Cohort 5). Using the same coecients and
cuto score, the 9 miRNA signature correctly identied
5/18 glioblastoma subjects and 19/20 non-oncologic
subjects, yielding a sensitivity of 28% and specicity of
95%. The AUC was 83% (95% CI: 69%, 96%). (Figure
4B). Notably, few miRNA species were detected in the
lumbar CSF samples. These results suggest that cisternal
and lumbar CSF may dier in miRNA content. Notably,
these validation samples used whole CSF for the miRNA
assay, as described in methods.
13 lumbar glioblastoma CSF samples were collected
as a part of Cohort 1. We had compared the performance
of our miRNA signature in these samples in order to aord
direct comparison to that seen in the cisternal samples.
In the cohort 1 lumbar CSF samples, the 9 miRNA
signature correctly identied glioblastoma subjects in
3/13 glioblastoma samples yielding a sensitivity of 23%.
These results were comparable to those observed in the
validation cohorts, conrming our observation that the
diagnostic utility of the 9-miRNA signature is optimal
when applied to cisternal CSF (Supplementary Figure 4).
Validation of increased miR21 in a mouse
xenograft model of glioblastoma
miR-21  play a pivotal role in our signature. We
wished to determine whether glioblastoma growth induce
accumulation of miR-21 in the CSF and used a murine
xenograft model to achieve this end. We orthotopically
implanted the patient-derived glioblastoma neurosphere
line (JVJ), which expressed high levels of miR-21, into
nude mice. 4 weeks after injection, brain tissue and
murine CSF were collected from tumor bearing mice
and age-matched, mock injected nude mice (Figure 5A).
Both brain tissue and CSF were analyzed by qRT-PCR to
measure the level of miR-21. In all analyzed samples, we
found elevated miR-21 levels in the brain tissues (Figure
5B) and CSFs (Figure 5C) isolated from xenograft bearing
mice relative to control mice. This result suggest that
glioblastoma xenograft growth induce accumulation of
miR-21 in murine CSF.
Figure 3: Correlation of miRNA score with tumor volume. A. The tumor volume of 11 patients in Cohort 3 was plotted against
the CSF miRNA signature score, and the Pearson correlation coecient was calculated. B. Glioblastoma < 15 cc’s in volume showed a
lowered miRNA signature score relative to those with > 15cc’s.
In current clinical practice, CSF sampling is
not routinely performed in glioblastoma patients. The
sensitivity of CSF cytology as a diagnostic tool for
glioblastoma is ~10%  and below the threshold for
clinical utility. However, our study suggests potential
utility for CSF miRNA proling as a diagnostic platform
for glioblastoma. The miRNA detectable in human
and nude mouse glioblastoma specimens is detected
in matched CSF, though at a concentration that is
~30 fold lower. miRNA proles of CSF derived from
glioblastoma patients correlated well to the miRNA
proles of the matched tumor specimens. We developed
a nine miRNA CSF signature that discriminated CSF
of glioblastoma patients from those of patients without
history of brain cancer. We validated this signature using
prospectively collected CSF samples after development
and documentation of the original signature. For crude
CSF based assay, the sensitivity and specicity for
glioblastoma detection were 80% and 67%, respectively.
In contrast, for CSF derived from lumbar puncture, the
sensitivity and specicity for glioblastoma detection were
28% and 95%, respectively. It is important to note that
the miRNA reported here dier from those previously
reported to discriminate between types of brain cancer
, suggesting that our miRNA signature has limited
utility in discriminating between dierent forms of brain
cancers. These results suggest that distinct miRNA proles
may be required to address dierent clinical needs.
There has been signicant variability in the reported
miRNA proles in CSF derived from glioblastoma patients
[21-23]. We observed this variability in our own study,
where signicant variation in miRNA proles were found
between CSF derived from the three discovery cohorts
(Supplemental Figure 3). A major source of variability is
the CSF collection site (cisternal vs. lumbar). However,
even after correcting for site of collection, this variability
remained. It is worthwhile noting that the CSF samples
were collected in our study through a Standard Operating
Figure 4: Validation of miRNA signature. A. Performance of the 9-miRNA signature using crude cisternal CSF from an independent
collection of prospectively collected samples. B. Performance of the 9-miRNA signature using crude lumbar CSF from an independent
collection of prospectively collected samples.
Procedure (SOP) and processed identically post-collection.
The variability observed between these cohorts, in the
context of the published literature, suggests that CSF
miRNA proles are likely inuenced by physiologic
factors or perturbation that was not accounted for by the
SOP (e.g. circadian rhythm, fatigue, intake of medicine…
etc). As such, the robustness of the CSF miRNA signatures
are largely a function of the sample size, since larger
sample sizes aord a greater likelihood of minimizing the
undue inuence of any particular perturbation/physiology.
Our study is particularly important in this context, since
our study design is the only one in the literature that
derived the signature through three independent cohorts,
summing to 135 CSF samples. We subsequently validated
our results in another 60 prospectively collected CSF.
The scale of our study as well as the meticulous eort
devoted to validation is notable in the reported literature
of glioblastoma CSF biomarkers.
An important nding in this study is that the miRNA
contents of cisternal and lumbar CSF dier. We found that
less than half of the miRNAs detected in cisternal CSF
were detected in lumbar CSF (Supplementary Figure 5),
likely accounting for the fewer number of dierentially
expressed miRNAs found in cohort 2. This nding
suggests the two CSF compartments do not communicate
suciently for full equilibrium of miRNA contents.
Similar observations have been made for other proteins
and metabolites [11, 24]. For instance, IgG level decreases
progressive as the CSF moves from the site of intracranial
inammation to the lumbar sac . These dierences
bear relevance to CSF based diagnostics and warrant
consideration in future study design. For instance, separate
miRNA signatures may need to be developed for analysis
of clinical lumbar and cisternal CSF samples.
EVs have been touted as platforms for diagnostic
and prognostic biomarker interrogation [26-28]. The
isolation of these EVs from CSF introduces an additional
step during clinical sample processing , a step which
incurs increased cost and risk of contamination risk.
The step is necessary if 1) the biomarker of interest is
enriched in EVs or 2) if inhibitory factors prohibitive to
the analytical platform is present in the crude CSF. Our
analysis support neither hypothetical scenarios. When
we compared the miRNA proles of CSF EVs relative to
Figure 5: Direct release of miR-21 from glioblastoma xenograft in vivo. A. 20,000 human glioblastoma stem cells were
intracranially injected into nude mice. 4 weeks later, brain tissues and murine CSFs were collected from tumor bearing mice and age-
matched nude mice without the xenograft injection. B. Human miR-21 levels were elevated in the brain tissue of patient derived glioblastoma
xenograft bearing mice and undetectable in mice without xenograft implant. C. Human miR-21 levels were elevated in the CSF of patient
derived glioblastoma xenograft bearing mice and undetectable in mice without xenograft implant.
crude CSF, we found that > 95% of miRNAs found in EVs
(including miRNAs in our signature) were represented in
the crude CSF, suggesting that crude CSF may suce for
miRNA proling. Further supporting this hypothesis, the
performance of the 9 miRNA signature was comparable
when applied to CSF EV RNA (Supplementary Figure 6)
or crude CSF RNA (Figure 4).
The literature that examined altered miRNA
regulation in glioblastoma has expanded over the past
decade . It is notable that of the reported miRNA
that are signicantly over- or under-expressed in
clinical glioblastoma specimens , only miR-21 was
represented in our miRNA signature. As further validation
of our correlative clinical studies, we showed that murine
CSF miR-21 levels were elevated in murine CSF from
glioblastoma xenograft bearing mice (Figure 5). We did
not observe such increase for other miRNAs previously
reported to be over-expressed in glioblastoma, including
miR-16 [30-32] or miR-10b [17, 30, 33] (data not shown).
Our previous study suggested that > 90% extra-cellular
miR-21 were found in the EV fractions . Together,
these results suggest that glioblastoma harbor biologic
mechanisms that facilitate the exportation of miR-21
through EV secretion. This interesting hypothesis awaits
While our miRNA signature performed well as a
diagnostic tool in cisternal CSF, opportunities for obtaining
these samples are admittedly limited. Such samples can be
obtained only from patients with an Ommaya reservoir or
a ventriculo-peritoneal shunt. Because these procedures
involve placement of an indwelling catheter that is in
direct communication to cisternal CSF, serial samples
can be safely acquired in this patient population. As such,
clinical testing of the cisternal CSF signature is feasible
in the subpopulation of glioblastoma patients with an
indwelling shunt system. Moreover, serial sampling of
cisternal CSF from this patient population may aord a
minimally invasive platform for tracking glioblastoma
disease burden. We are in the process of collecting and
testing CSF from recurrent glioblastoma patients to further
test the utility of our miRNA signature.
In sum, our study provides a proof-of-principle
study demonstrating the plausibility of CSF miRNA
proling as a “liquid biopsy” platform for glioblastoma
diagnosis and provides the basis of future validation of
MATERIALS AND METHODS
Clinical specimen collection and image analysis
Five cohorts of patients totaling 195 subjects
provided CSF for these studies (Table 1). The CSF studies
were approved by IRB boards at University of California
San Diego (UCSD) (Cohorts 1, 3, 4, and 5), Technische
Universität München (TUM)(Cohorts 1), University
of Miami Hospital (UMH)(Cohorts 1), and Huashan
Hospital(Cohorts 2,5). All studies were in conducted in
accordance with the principles expressed at the declaration
at Helsinki. Each patient was consented in writing by a
research coordinator prior to CSF collection. Median
age ranged from 54 to 61 years across cohorts. Overall,
88 subjects were female and 107 were male, 111 had
diagnosis of glioblastoma and 84 had other non-oncologic
conditions. Cisternal and ventricular CSF (grouped
as “cisternal”) was collected on 80 subjects by drain
placement or cisternal aspiration at the time of craniotomy
prior to tumor manipulation. Lumbar CSF was collected
on 115 subjects, through lumbar puncture or lumbar
drain. Collected CSF specimens were ltered (0.8µm
lter), immediately frozen and stored at -80°C. 1 cc of
CSF was utilized as the input for all miRNA analysis. The
UCSD cohort was additionally consented for analysis of
MR images. Volumetric measurements of available pre-
operative MR images were carried out with Agfa CD
Viewer 4.5.1 using the formula Volume = (L × W × H)/2,
where L is the greatest length, W is the greatest width, and
H is the greatest depth or height of the tumor . Patients
that received bevacizumab were excluded from MR image
Extracellular vesicle (EV) isolation
The EV fraction was isolated by dierential
centrifugation as previously described . CSFs
were diluted 1:1 with 1x PBS (Mediatech) prior to
centrifugation. Samples were centrifuged at 2,000×g for
20 min to remove cellular debris. The supernatant was
further centrifuged at 120,000×g for 2 h in a Type 70 Ti
rotor (Beckman) to pellet the EVs. All centrifugation steps
were performed at 4°C. EV pellets were resuspended in
PBS and stored at -80°C.
RNA was extracted from each sample using the
miRCURY™ RNA Isolation Kit (Exiqon). Samples
assayed were EV, supernatant and tissue from cohort 1;
EV and supernatant from cohort 2; EV from cohort 3;
and whole CSF from cohorts 4 and 5. Four microliters
of RNA extract (4-20ng/µl) was used as input for
microRNA proling on the TaqMan® OpenArray® Real-
Time PCR System using the manufacturer’s instructions
(Life Technologies). Manufacturer’s cartridges consisted
of 818 TaqMan qPCR assays arranged on 384 well
plates, with primers targeting 754 miRNA species, and
16 replicate wells of one negative and 3 positive RNA
controls. Megaplex™ RT Primers, Human Pool A v2.1 and
Megaplex™ RT Primers, Human Pool B v3.0 were used
for the reverse transcription step. Megaplex™ PreAmp
Primers, Human Pool A v.2.1 and Megaplex™ PreAmp
Primers, Human Pool B v3.0 were used for the PreAmp
step. The samples within each of the 3 discovery cohorts
were assayed on the same date using the same reagents.
The validation samples (cohorts 4 and 5) were assayed in
two dierent batches on two dierent dates and data was
combined for analysis.
Data normalization, QC, and preprocessing
miRNA species with C
value ≥35 were considered
below the detection threshold. In tumor tissue samples,
CT values for the query miRNAs were normalized using
the mean of the positive controls (RNU44, RNU48,
U6-rRNA). For the CSF samples, the positive control
miRNAs were not uniformly expressed at high levels
across samples. For the discovery cohorts global mean
normalization was performed in which normalized
CT values were calculated as the raw CT value minus
the arithmetic mean of all expressed miRNAs in the
sample . For the validation cohorts, the data was
rst normalized within each sample as before using the
global mean normalization. Then the batch eect from
the two assay dates was removed using an empirical
Bayes approach (ComBat)  with assay date and two
confounding variables (pathology and CSF collection site)
included in the adjustment model. The batch-corrected
data were then combined for analysis.
Statistical approach to training and validation of
The classier was trained using the three discovery
cohorts (Cohorts 1, 2, and 3). When both supernatant
and EV miRNA was available within a cohort, we used
the fraction with the higher median number of detected
miRNA species for analysis within that cohort. For
cohorts with low detection rates, we used both fractions
with a Bonferroni correction for the two comparisons.
Dierentially expressed CSF miRNAs between
glioblastoma and non-oncologic subjects were identied
using the limma Bioconductor package , with FDR
< 0.2 and log (fold-change) > 2 as criteria. We then
required candidate miRNAs from a given cohort to
replicate as dierentially expressed in at least 2 additional
discovery data sets, including Cohort 1 EV, Cohort 2 EV
+ supernatant, Cohort 3 EV and also including tissue
miRNA data from TCGA . The replication criterion
was a two-sided p-value < 0.05 (from limma, or a t-test for
TCGA) and the same direction of dierential expression;
this test of replication has overall Type I error rate ~1%.
This candidate selection plan was pre-specied and
Candidate miRNAs were carried forward to a
multivariate model to discriminate glioblastoma from non-
oncologic controls using L1-penalized logistic regression
. The model was trained with Cohort 3 using glmnet
package in R with lambda chosen by cross validation .
The signature and optimal cuto score to discriminate
cases from controls in Cohort 3 were documented. The
prediction error of the classier with pre-determined
cuto was then evaluated data from whole CSF, using the
prospectively collected independent validation Cohorts 4
and 5. For correlation analyses, the Pearson correlation
coecient was calculated using Graphpad Prism 6.
Orthotopic xenograft model
Dissociated glioblastoma stem cells JVJ (2x104
cells in 4 μl HBSS) were stereotactically injected into the
brains of nude mice at age 6 weeks old. The coordinates
were: 1.8 mm to the right of bregma and 3 mm deep from
the dura. Aged-matched nude mice were used as controls.
Four weeks after injection, CSF samples were collected
from the cisterna magna as previously described .
Mock injection with vehicle control was carried out for
Quantitative reverse transcriptase-polymerase
chain reaction (qRT-PCR)
For the detection of tissue miRNA, RNA was
extracted from homogenized mouse brains using Qiagen
miRNeasy Mini Kit. cDNA was synthesized using
TaqMan miRNA Reverse Transcription Kit and miRNA-
specic stem-loop primers (Applied Biosystems),
followed by qPCR using SsoAdvanced™ Universal
Probes Supermix (Bio-Rad) and miRNA specic Taqman
assay on a Bio-Rad CFX96 instrument. For the detection
of miRNA from murine CSF, collected CSF was lysed
directly in buer containing 50mM Tris pH 8, 140mM
NaCl, 1.5mM MgCl2, 0.5% NP40, and 0.1% BSA, then
reverse transcribed using SuperScript® VILO™ cDNA
synthesis kit. The cDNA was pre-amplied for 15 cycles
using Taqman PreAmp Mastermix prior to PCR detection
with miR-21 Taqman assay.
Taqman miRNA assay for miR-10b, miR-16, and
miR-21 were purchased from ThermoFisher Scientics.
We thank Shirley Phillips
and Steven Lockton from
Regulus Therapeutics for performing the miRNA proling
CONFLICT OF INTERESTS
The authors declare that there is no conict of
Grant acknowledgement: 4UH3TR000931-03
awarded to BSC. 1RO1NS097649-01 and BWF
1006774.01 awarded to CCC. International S&T
Cooperation Program of China, 2014DFA31470 awarded
to YM and CCC.
This paper has been accepted based in part on peer-
review conducted by another journal and the authors’
response and revisions as well as expedited peer-review
1. Bartek J, Jr., Ng K, Bartek J, Fischer W, Carter B, Chen CC.
Key concepts in glioblastoma therapy. J Neurol Neurosurg
Psychiatry. 2012; 83: 753-60. doi: jnnp-2011-300709 [pii]
2. Ng K, Kim R, Kesari S, Carter B, Chen CC. Genomic
proling of glioblastoma: convergence of fundamental
biologic tenets and novel insights. J Neurooncol. 2012; 107:
1-12. doi: 10.1007/s11060-011-0714-2.
3. Vuorinen V, Hinkka S, Farkkila M, Jaaskelainen J.
Debulking or biopsy of malignant glioma in elderly people -
a randomised study. Acta neurochirurgica. 2003; 145: 5-10.
4. Gonda DD, Warnke P, Sanai N, Taich Z, Kasper EM, Chen
CC. The value of extended glioblastoma resection: Insights
from randomized controlled trials. Surg Neurol Int. 2013; 4:
110. doi: 10.4103/2152-7806.117173.
5. Air EL, Leach JL, Warnick RE, McPherson CM. Comparing
the risks of frameless stereotactic biopsy in eloquent
and noneloquent regions of the brain: a retrospective
review of 284 cases. J Neurosurg. 2009; 111: 820-4. doi:
6. Bernays RL, Kollias SS, Khan N, Brandner S, Meier
S, Yonekawa Y. Histological yield, complications, and
technological considerations in 114 consecutive frameless
stereotactic biopsy procedures aided by open intraoperative
magnetic resonance imaging. J Neurosurg. 2002; 97: 354-
62. doi: 10.3171/jns.2002.97.2.0354.
7. Skrap M, Mondani M, Tomasino B, Weis L, Budai R,
Pauletto G, Eleopra R, Fadiga L, Ius T. Surgery of insular
nonenhancing gliomas: volumetric analysis of tumoral
resection, clinical outcome, and survival in a consecutive
series of 66 cases. Neurosurgery. 2012; 70: 1081-93;
discussion 93-4. doi: 10.1227/NEU.0b013e31823f5be5.
8. Noorbakhsh A, Tang JA, Marcus LP, McCutcheon B,
Gonda DD, Schallhorn CS, Talamini MA, Chang DC,
Carter BS, Chen CC. Gross-total resection outcomes
in an elderly population with glioblastoma: a SEER-
based analysis. J Neurosurg. 2014; 120: 31-9. doi:
9. Santiago-Dieppa DR, Steinberg J, Gonda D, Cheung
VJ, Carter BS, Chen CC. Extracellular vesicles as a
platform for ‘liquid biopsy’ in glioblastoma patients.
Expert Rev Mol Diagn. 2014; 14: 819-25. doi:
10. Wang Y, Springer S, Zhang M, McMahon KW, Kinde
I, Dobbyn L, Ptak J, Brem H, Chaichana K, Gallia GL,
Gokaslan ZL, Groves ML, Jallo GI, et al. Detection of
tumor-derived DNA in cerebrospinal uid of patients with
primary tumors of the brain and spinal cord. Proceedings of
the National Academy of Sciences. 2015; 112: 9704-9. doi:
11. Rubalcava MA, Sotelo J. Dierences between Ventricular
and Lumbar Cerebrospinal Fluid in Hydrocephalus
Secondary to Cysticercosis. Neurosurgery. 1995; 37: 668-
12. Weisner B, Bernhardt W. Protein fractions of lumbar,
cisternal, and ventricular cerebrospinal uid. Journal
of the Neurological Sciences. 1978; 37: 205-14. doi:
13. Akers JC, Ramakrishnan V, Kim R, Skog J, Nakano I,
Pingle S, Kalinina J, Hua W, Kesari S, Mao Y, Breakeeld
XO, Hochberg FH, Van Meir EG, et al. MiR-21 in the
extracellular vesicles (EVs) of cerebrospinal uid (CSF):
a platform for glioblastoma biomarker development. PLoS
One. 2013; 8: e78115. doi: 10.1371/journal.pone.0078115
14. Akers JC, Ramakrishnan V, Nolan JP, Duggan E, Fu CC,
Hochberg FH, Chen CC, Carter BS. Comparative Analysis
of Technologies for Quantifying Extracellular Vesicles
(EVs) in Clinical Cerebrospinal Fluids (CSF). PLoS One.
2016; 11: e0149866. doi: 10.1371/journal.pone.0149866.
15. Akers JC, Ramakrishnan V, Yang I, Hua W, Mao Y, Carter
BS, Chen CC. Optimizing preservation of extracellular
vesicular miRNAs derived from clinical cerebrospinal uid.
Cancer Biomark. 2016; 17: 125-32. doi: 10.3233/CBM-
16. Chan JA, Krichevsky AM, Kosik KS. MicroRNA-21 Is an
Antiapoptotic Factor in Human Glioblastoma Cells. Cancer
Research. 2005; 65: 6029-33. doi: 10.1158/0008-5472.can-
17. Ciafrè SA, Galardi S, Mangiola A, Ferracin M, Liu CG,
Sabatino G, Negrini M, Maira G, Croce CM, Farace MG.
Extensive modulation of a set of microRNAs in primary
glioblastoma. Biochemical and Biophysical Research
Communications. 2005; 334: 1351-8. doi: http://dx.doi.
18. Suzuki HI, Katsura A, Matsuyama H, Miyazono K.
MicroRNA regulons in tumor microenvironment.
Oncogene. 2015; 34: 3085-94. doi: 10.1038/onc.2014.254.
19. Friedman J, Hastie T, Tibshirani R. Regularization Paths for
Generalized Linear Models via Coordinate Descent. Journal
of statistical software. 2010; 33: 1-22. doi:
20. J. Chris Balhuizen, Gerard Th. A. M. Bots, Aart Schaberg,
Fré T. Bosman. Value of cerebrospinal uid cytology for
the diagnosis of malignancies in the central nervous system.
Journal of Neurosurgery. 1978; 48: 747-53. doi: 10.3171/
21. Drusco A, Bottoni A, Lagana A, Acunzo M, Fassan M,
Cascione L, Antenucci A, Kumchala P, Vicentini C,
Gardiman MP, Alder H, Carosi MA, Ammirati M, et al. A
dierentially expressed set of microRNAs in cerebro-spinal
uid (CSF) can diagnose CNS malignancies. Oncotarget.
2015; 6: 20829-39. doi: 10.18632/oncotarget.4096.
22. Teplyuk NM, Mollenhauer B, Gabriely G, Giese A, Kim
E, Smolsky M, Kim RY, Saria MG, Pastorino S, Kesari S,
Krichevsky AM. MicroRNAs in cerebrospinal uid identify
glioblastoma and metastatic brain cancers and reect
disease activity. Neuro-Oncology. 2012; 14: 689-700. doi:
23. Baraniskin A, Kuhnhenn J, Schlegel U, Maghnouj
A, Z√∂llner H, Schmiegel W, Hahn S, Schroers R.
Identication of microRNAs in the cerebrospinal uid as
biomarker for the diagnosis of glioma. Neuro-Oncology.
2011. doi: 10.1093/neuonc/nor169.
24. Merritt HHHH, Fremont-Smith F. (1937). The cerebrospinal
uid: W. B. Saunders).
25. Ramos-Kuri M, Montoya R, Padilla A, et al.
Immunodiagnosis of neurocysticercosis: Disappointing
performance of serology (enzyme-linked immunosorbent
assay) in an unbiased sample of neurological patients.
Archives of Neurology. 1992; 49: 633-6. doi: 10.1001/
26. Gonda DD, Akers JC, Kim R, Kalkanis SN, Hochberg
FH, Chen CC, Carter BS. Neuro-oncologic Applications
of Exosomes, Microvesicles, and Other Nano-Sized
Extracellular Particles. Neurosurgery. 2013; 72: 501-10.
27. Taylor DD, Gercel-Taylor C. MicroRNA signatures of
tumor-derived exosomes as diagnostic biomarkers of
ovarian cancer. Gynecol Oncol. 2008; 110: 13-21. doi:
S0090-8258(08)00343-0 [pii] 10.1016/j.ygyno.2008.04.033.
28. Nilsson J, Skog J, Nordstrand A, Baranov V, Mincheva-
Nilsson L, Breakeeld XO, Widmark A. Prostate cancer-
derived urine exosomes: a novel approach to biomarkers
for prostate cancer. Br J Cancer. 2009; 100: 1603-7. doi:
6605058 [pii] 10.1038/sj.bjc.6605058.
29. Thery C, Amigorena S, Raposo G, Clayton A. Isolation and
characterization of exosomes from cell culture supernatants
and biological uids. Curr Protoc Cell Biol. 2006; Chapter
3: Unit 3 22. doi: 10.1002/0471143030.cb0322s30 [doi].
30. Moller HG, Rasmussen AP, Andersen HH, Johnsen
KB, Henriksen M, Duroux M. A systematic review of
microRNA in glioblastoma multiforme: micro-modulators
in the mesenchymal mode of migration and invasion. Mol
Neurobiol. 2013; 47: 131-44. doi: 10.1007/s12035-012-
31. Dong H, Luo L, Hong S, Siu H, Xiao Y, Jin L, Chen R,
Xiong M. Integrated analysis of mutations, miRNA and
mRNA expression in glioblastoma. BMC Syst Biol. 2010;
4: 163. doi: 10.1186/1752-0509-4-163.
32. Rao SA, Santosh V, Somasundaram K. Genome-wide
expression proling identies deregulated miRNAs in
malignant astrocytoma. Mod Pathol. 2010; 23: 1404-17.
33. Lavon I, Zrihan D, Granit A, Einstein O, Fainstein N,
Cohen MA, Cohen MA, Zelikovitch B, Shoshan Y,
Spektor S, Reubino BE, Felig Y, Gerlitz O, et al. Gliomas
display a microRNA expression prole reminiscent of
neural precursor cells. Neuro Oncol. 2010; 12: 422-33. doi:
34. Monga SPS, Wadleigh R, Sharma A, Adib H, Strader D,
Singh G, Harmon JW, Berlin M, Monga DK, Mishra L.
Intratumoral Therapy of Cisplatin/Epinephrine Injectable
Gel for Palliation in Patients With Obstructive Esophageal
Cancer. American Journal of Clinical Oncology. 2000; 23:
35. Hygino da Cruz LC, Rodriguez I, Domingues RC,
Gasparetto EL, Sorensen AG. Pseudoprogression and
Pseudoresponse: Imaging Challenges in the Assessment
of Posttreatment Glioma. American Journal of
Neuroradiology. 2011; 32: 1978-85. doi: 10.3174/ajnr.
36. Mestdagh P, Van Vlierberghe P, De Weer A, Muth D,
Westermann F, Speleman F, Vandesompele J. A novel
and universal method for microRNA RT-qPCR data
normalization. Genome Biology. 2009; 10: 1-10. doi:
37. Johnson WE, Li C, Rabinovic A. Adjusting batch eects in
microarray expression data using empirical Bayes methods.
Biostatistics. 2007; 8: 118-27. doi: 10.1093/biostatistics/
38. Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W,
Smyth GK. limma powers dierential expression analyses
for RNA-sequencing and microarray studies. Nucleic Acids
Research. 2015; 43: e47. doi: 10.1093/nar/gkv007.
39. TCGA. Comprehensive genomic characterization denes
human glioblastoma genes and core pathways. Nature.
2008; 455: 1061-8. doi: nature07385 [pii]10.1038/
40. Liu L, Du K. A Technique for Serial Collection of
Cerebrospinal Fluid from the Cisterna Magna in Mouse.
Journal of Visualized Experiments : JoVE. 2008: 960. doi: