Tumor and Cerebrospinal Fluid microRNAs in Primary Central Nervous System Lymphomas

Primary central nervous system lymphoma (PCNSL) is a rare, highly aggressive, extranodal form of non-Hodgkin lymphoma, predominantly diagnosed as primary diffuse large B-cell lymphoma of the central nervous system (CNS DLBCL). Fast and precise diagnosis of PCNSL is critical yet challenging. microRNAs, important regulators in physiology and pathology are potential biomarkers. In 131 patients with CNS DLBCL and with non-malignant brain lesions (n-ML), miR-21, miR-19b and miR-92a, miR-155, miR-196b, miR-let-7b, miR-125b, and miR-9 were examined by RT-qPCR in brain biopsy samples (formalin-fixed paraffin-embedded tissues, FFPET; CNS DLBCL, n = 52; n-ML, n = 42) and cerebrospinal fluid samples (CSF; CNS DLBCL, n = 30; n-ML, n = 23) taken for routine diagnosis. FFPET samples were split into study and validation sets. Significantly higher CSF levels of miR-21, miR-19b, and miR-92a were identified in PCNSL but not in n-ML, and differentiated PCNSL from n-ML with 63.33% sensitivity and 80.77% specificity. In FFPETs, miR-155 and miR-196b were significantly overexpressed and miR-let-7b, miR-125b, and miR-9 were downregulated in PCNSL as compared to n-ML. Combined miR-155 and miR-let-7b expression levels in FFPETs discriminated PCNSL and n-ML with a 97% accuracy. In conclusion, tissue miR-155, miR-196b, miR-9, miR-125b, and miR-let-7b expression profiles differentiate PCNSL from n-ML. PCNSL CSFs and the relevant biopsy samples are characterized by specific, different microRNA profiles. A logistic regression model is proposed to discriminate between PCNSL and non-malignant brain lesions. None of the examined microRNAs influenced overall survival of PCNSL patients. Further ongoing developments involve next generation sequencing-based profiling of biopsy and CSF samples.

. MicroRNA expression in formalin-fixed paraffin-embedded tissue samples of cerebral lesions from patients with CNS DLBCL (n = 52 for the top and middle rows, and n = 11 for the bottom row) and with non-malignant brain lesions (n = 42 for the top and middle rows, and n = 10 for the bottom row).
As shown by the ROC analysis ( Figure 2), combined miR-155 and miR-let-7b expression levels in brain FFPET samples presented the best discrimination power between CNS DLBCL and n-ML (98% specificity and 96% sensitivity, AUC = 0.988). Supplementing the combination of miR-155 and miR-let-7b with miR-196b or miR-125b did not improve the discrimination power. The following logistic regression model was built based on a combination of miR-let-7b and miR-155 expression levels to predict CNS DLBCL diagnosis: alpha = −2.664 − 0.1225[miR-let-7b] + 9.32[miR-155]. To validate the model, i.e., the discrimination power of miR-155 and miR-let-7b expression measurements, an independent, blinded set of brain biopsies was examined. Out of 17 samples of CNS DLBCL, and 17 samples of n-ML, all but one were correctly classified (Table S1). The falsely classified one was a CNS DLBCL sample predicted as a n-ML. However, further detailed analysis of this FFPET sample revealed it contained almost exclusively normal cells of adjacent tissue, thus the sample was not representative for the tumor. After the validation step that showed highly consistent results between the study and validation groups (Figures S1 and S2), data were pooled according to the final diagnosis (CNS DLBCL vs. n-ML) and the pooled series were statistically re-analyzed. Figure  1 shows the results of the pooled data.

miR-21, miR-19b, and miR-92a Levels in CSFs and Brain Biopsies from Patients with CNS DLBCL vs. n-ML
The levels of CSF miR-21, miR-19b, and miR-92a ( Figure 3) were significantly higher in patients with CNS DLBCL than in patients with n-ML (miR-21: median 12.628 and 6.804, respectively, p = 0.016738; miR-19b: median 1.759 and 1.139, respectively, p = 0.006055; and miR-92a: median 2.820 and 1.628, respectively, p = 0.001135). To validate the model, i.e., the discrimination power of miR-155 and miR-let-7b expression measurements, an independent, blinded set of brain biopsies was examined. Out of 17 samples of CNS DLBCL, and 17 samples of n-ML, all but one were correctly classified (Table S1). The falsely classified one was a CNS DLBCL sample predicted as a n-ML. However, further detailed analysis of this FFPET sample revealed it contained almost exclusively normal cells of adjacent tissue, thus the sample was not representative for the tumor. After the validation step that showed highly consistent results between the study and validation groups (Figures S1 and S2), data were pooled according to the final diagnosis (CNS DLBCL vs. n-ML) and the pooled series were statistically re-analyzed. Figure 1 shows the results of the pooled data.

Survival Analyses
The overall survival of patients with CNS DLBCL is shown in Figure S3a. There were no significant differences in OS related to tumor microRNA expression levels ( Figure S3b).

CNS DLBCL Immunophenotypes and miR Expression
We found no miR expression to be related to the immunophenotype subtypes of the CNS DLBCL.

Discussion
We showed that FFPET samples presented significantly overexpressed miR-155 and miR-196b, while miR-9, miR-125b, and miR-let-7b were downregulated in CNS DLBCL, as compared to n-ML. We further validated these results in an independent set of samples and demonstrated that a combined assessment of miR-155 and miR-let-7b discriminated CNS DLBCL from n-ML with 97% accuracy. We are the first to show that microRNA expression in brain biopsy samples differentiates CNS DLBCL from n-ML. On top of that, a significance of cell content assessment in the samples was brought up. This could be useful for clinical practice when routine histopathology and immunochemistry is misleading [25,26], and in the biopsies of the so-called vanishing tumors, observed following steroid treatment (based on our own experience).
The increased expression of brain-specific miR-9 and miR-125b in n-ML may relate to their crucial role in a number of neurogenic processes, including microglial migration, regulation of the adult neural stem cells quiescence and activation, axonal branching and outgrowth, and astrogliosis during neurodegeneration [27][28][29][30]. The potent neurogenic role of miR-9 has been confirmed by the ectopic expression of miR-9/miR-124 in human adult fibroblasts. It triggered chromatin accessibility reconfiguration, DNA methylation, and mRNA expression changes, and induced a default neuronal state, and the conversion of fibroblasts to functional neurons [31]. Moreover, deregulated miR-9, miR-125b, and let-7b expression has been implicated in the pathogenesis of many neurologic diseases, including schizophrenia, Alzheimer disease, amyotrophic lateral sclerosis, Huntington's disease, ischemic stroke, and MS [32][33][34][35][36][37][38][39][40][41], what further pointed to the relevance of those miRNAs in brain tissues.
The increased expression of two inflammation-and lymphoma-associated miRNAs, miR-155 and miR-196b, in CNS DLBCL compared to n-ML, is in line with previously published data. PCNSL has been shown to express higher levels of miR-155 in than nodal DLBCL, NOS specimens [42,43]. miR-196b has been demonstrated to contribute to the pathogenesis of some types of leukemia [22], and, in acute myeloid leukemia, high miR-196b expression influenced prognosis [44].
Immunophenotypic subgroups of CNS DLBCL (CD5-positive, germinal-center B-cell (GCB)-type and non-germinal center B-cell (non-GCB)-type) presented no differences in the examined miRNA profiles of both CSF and FFPET specimens. Similarly, none of the microRNAs examined in the tumors related to OS in patients with CNS DLBCL. Takashima et al. [45] have recently suggested a miRNA expression signature of PCNSL tumors as a predictor of prognosis. However, the signature comprised other microRNAs that we examined here.
CSF microRNAs have been considered as novel potential biomarkers for patients with brain lesions [10]; still, the available data are scarce and further validation is necessary. While examining the CSFs, we found a significantly higher levels of miR-21, miR-19b, and miR-92a in patients with CNS DLBCL than in patients with n-ML, and no differences between the two series of patients in the CSF levels of miR-9, miR-125b, miR-155, and miR-196b. A set of CSF miR-21, miR-19b, and miR-92a differentiated CNS DLBCL from n-ML, with a specificity of 80.77% and a sensitivity of 63.33%. In PCNSL, Baraniskin et al. [24,46] were the first to demonstrate CSF miRNA assessment as a powerful tool for the diagnosis and follow-up of patients. In line with our results, they showed the increased levels of miR-21, miR-19b, and miR-92a [24]. However, the diagnostic accuracy they presented was much higher (95.7% sensitivity and 96.7% specificity). The reason of this discrepancy most probably lies in the reference groups. The control series we used comprised patients with benign brain neoplasms and diverse neurological disorders, while their series were dominated by MS cases. Notably, patients with MS have recently been shown to present diminished CSF miR-21 levels [47]. Considering the above, the diagnostic performance achieved by Baraniskin et al. [24] for CSF levels of miR-21, miR-19b, and miR-92a may have been exceedingly high. Consistently with our results on miR-125b, Drusco et al. [48] found no differences in the levels of CSF miR-125b between patients with PCNSL and n-ML, while there was a significant up-regulation of miR-125b in medulloblastoma and glioblastoma. It should be noted that although Drusco et al. [48] employed high throughput methods, in their series of patients with benign and malignant brain tumors of different origin, only 3 were diagnosed with PCNSL. CSF miR-155 levels previously studied in glioma patients were found not to differ from that of non-malignant controls, although a TCGA data analysis showed over 2-fold tumor tissue miR-155 overexpression [49].
Interestingly, the examined miRNA levels in CSFs and tissues from patients with CNS DLBCL did not match. Similar CSF patterns of miR-155, miR-196b, miR-9, and miR-125b, and different patterns of miR-21, miR-19b, and miR-92a between patients with n-ML and those with CNS DLBCL did not reflect the miR patterns in brain tumor samples. It is not clear why in the CNS DLBCL, CSFs, and the relevant biopsy samples are characterized by different microRNA profiles. We believe that inconsistencies between tumor and CSF miRNA expression were not linked to RNA degradation characterizing FFPET samples. A number of studies have proven FFPET-derived miRs to have enhanced stability [50]. Moreover, the inconsistencies concerned were also less prone to degradation, GC-rich miR-92a [51]. To explain the discrepancies between miR profiles of CSFs and tumors, one may speculate that while microRNAs detected in the brain biopsy specimens may originate mainly from lymphoma cells, CSF microRNAs may derive from other cell types besides lymphoma cells, such as those associated with the ventricular choroid plexus, ventricular system, the subarachnoid space, and spinal cord [52]. Moreover, since CNS DLBCL and neurodegenerative diseases are frequently accompanied by a blood-brain barrier dysfunction [53,54], circulating microRNAs may also contribute to disease-related CSF microRNA profiles. Interestingly, as recently reported, in healthy donors, the miRNA profiles of brain tissues and CSF exosomes were not identical, suggesting a selective secretion of miRNAs by brain tissues [55]. For example, brain-enriched miR-124a that has been identified in CSFs of patients with glioblastoma and CNS metastases, was not expressed in the tumor samples [49]. Also, proteome studies have shown discrepancies between CSF and the relevant PCNSL tissue expression [53]. Inconsistencies in microRNA profiles have also been observed between peripheral tumor specimens and the respective blood samples [56]. Interestingly, an inverse correlation between circulating and tumor-associated microRNA levels has also been observed [57]. Therefore, in line with previous suggestions by Witwer [58] that referred to circulating microRNAs in solid tumor patients, one cannot exclude that CNS DLBCL-or non-malignant disease-related changes in CSF microRNA levels may reflect a systemic response to the pathologies rather than deregulations in the brain lesion.
To further investigate microRNA profiles of biopsy tissues and CSFs of patients with CNS lesions, in order to potentially identify new microRNA molecules involved, to look into the biological contexts, and to verify the utility microRNA assessment of as an ancillary tool in the differential diagnosis and follow up of CNS DLBCL patients, we are currently using next generation sequencing approaches.

Patients
Brain FFPET (Tables S3 and S4) and CSF (Tables S2 and S5) Table S5) or with a final diagnosis of n-ML (n = 23, 13 women/10 men, median age 31, range 9-81, Table S2). The CSF samples were centrifuged at 170× g, to recover the cells for the routine cytological and FCM examinations. The leftover supernatants were centrifuged at 500× g for 10 min at 20 • C, aliquoted to 400 µL volumes, and stored at −70 • C.
Brain tumor biopsy samples were obtained by stereotactic biopsy or surgical resection of CNS tumors of patients with the initial clinical and/or MRI presentation suggesting PCNSL, and subsequently diagnosed with CNS DLBCL (n = 52, 32 women/20 men, median age 62, range 31-82, Table S3) or with n-ML (n = 42, 18 women/24 men, median age 47, range 18-78, Table S4). The first series of consecutive tumor samples from patients with CNS DLBCL (n = 35) and n-ML (n = 23) served as study groups, subsequent samples were blinded and served as a validation group for miRNA evaluation. The validation group included 17 CNS DLBCL, and 17 n-ML. Additionally, the study and validation series were pooled and re-analyzed.
For more details see Supplementary Methods.

Diagnostic Procedures: Immunophenotyping, Morphology and Proliferation Evaluation of CSF Cells
Immunophenotyping of CSF samples was performed by FCM. Concentrated cells isolated from CSFs by centrifugation were incubated with a panel of MoAbs (for staining procedure see [23], for a list of MoAbs see Table S7). For more details see Supplementary Methods.
Simultaneously, CYT were stained with a May-Grünwald-Giemsa for morphological evaluation.

Final Diagnosis
To set the final diagnosis considered in this study, all malignant and non-malignant HP/IHC biopsy specimens and FCM analyses were re-evaluated in 2018 by an experienced pathologist (GR) in the context of clinical characteristics, imaging results, and individual patient histories followed for several years. The final diagnosis of CNS DLBCL and n-ML considered histopatological (HP) criteria and IHC examination according to the 2017 WHO classification [2], and included immunohistochemical subgroups, CD5-positive, GCB-and non-GCB-types, distinguished by the Hans algorithm, i.e., based on CD10, BCL6, and multiple myeloma oncogene-1 (MUM1) expression, as proposed in the previous WHO 2008 classification [25]. For more details see Supplementary Methods.
In each case, systemic DLBCL involvement was ruled out. Only non-immunocompromised patients were included in this study. The clinical and pathomorphological characteristics of the patients are presented in Tables S2-S5.
The study was approved by the local Institutional Review Board (decision no. 4/2011/1/2012).

RNA Extraction and RT-qPCR
Total RNA was isolated from CSF samples with the Gene Matrix Universal RNA/miRNA Purification Kit (EURx, Gdansk, Poland), according to the manufacturer's instructions, with some minor modifications specified herein. Three hundred µL of the RL buffer (EURx) was added directly to the 400 µL of a frozen CSF sample. After complete thawing, the sample was vortexed, and 500 µL of the acid phenol:chloroform:IAA (Ambion, Foster City, CA, USA) was added. The sample was mixed briefly and centrifuged at 11.000× g for 3 min. Subsequently the upper aqueous phase was transferred to a homogenization column and centrifuged (11.000× g, 2 min). In addition, after spinning, following a suggestion of Alexander Baraniskin (personal communication), first, 60 µg of glycogen (Invitrogen, Carlsbad, CA, USA) was added to the flow-through; secondly, the Reverse Transcription (RT) reaction was performed using 10 µL, as the amounts of the isolated total RNA were below the sensitivity of the NanoDrop ND 1000 Spectrophotometer (NanoDrop Technologies, ThermoFisher Scientific, Madison, WI, USA).
Ten 20-µm thick sections of each FFPET sample were cut with a disposable blade. Total RNA was extracted using the RecoverAll™ Total Nucleic Acid Isolation Kit (Applied Biosystems, Foster City, CA, USA), according to the manufacturer's recommendations. RNA concentration and purity were measured with the NanoDrop ND 1000 Spectrophotometer (NanoDrop Technologies).

Statistical Analyses
The differences in miRNA expression levels between samples (both in CSF and FFPET) were assessed by the Mann-Whitney U test. Associations of miRNA expression and clinical variables were probed with the Kruskal-Wallis test.
The Cox proportional hazards model was used to analyze the correlation of clinical data and microRNA levels with overall survival (OS). A separate model was created for each miR variable. A multivariate model was applied to correlate OS with all three treatments. OS was calculated from the date of diagnosis to the date of death, or-for patients who were still alive-to the date of the last observation. A logistic regression model was built to predict diagnosis on the basis of miR-let-7b and miR-155 expression levels in FFPET samples of the study set and on the basis of miR-19b, miR-21, and miR-92a levels in the whole set of CSF samples.
The analyses were performed in R (version 3.4.1) with the aid of the survival package (version 2.42-6).

Conclusions
• Tumor miR-155, miR-196b, miR-9, miR-125b, and miR-let-7b expression levels are significantly different in PCNSL and in non-malignant brain lesions. • A logistic regression model is proposed to discriminate between PCNSL and non-malignant brain lesions.

•
We confirm the value of cerebrospinal fluid miR-21, miR-19b, and miR-92a profiles as potential CNS DLBCL markers. • PCNSL CSFs and the relevant biopsy samples are characterized by specific, different microRNA profiles.

•
The examined microRNA profiles do not influence overall survival of PCNSL patients.

Supplementary Materials:
The following are available online at http://www.mdpi.com/2072-6694/11/11/1647/s1, Figure S1: MicroRNA expression in formalin-fixed paraffin-embedded brain tumor biopsy samples of patients with CNS DLBCL and with non-malignant CNS diseases (Controls)., Figure S2: MicroRNA expression in formalin-fixed paraffin-embedded brain tumor biopsy samples of patients with CNS DLBCL and with non-malignant CNS diseases., Figure S3: Overall survival (OS) analyses of patients with CNS DLBCL., Table S1: Validation of the discrimination power of miR-155 and miR-let-7b expression measurements between CNS DLBCL and non-malignant tumors. Table S2: Patients with non-neoplastic cerebral tumors (n = 23): clinical diagnosis and miRNA expression level in the cerebrospinal fluid (CSF)., Table S3: Clinical and pathomorphological characteristics of CNS DLBCL patients (n = 52), along with miRNA expression levels in tumor biopsy samples. Table S4: Non-neoplastic cerebral or cerebellum tumor patients (n = 42): pathomorphological diagnosis and tumor miRNA expression levels. Table S5: Patients with CNS DLBCL (n = 30): clinical diagnosis and miRNA expression level in the cerebrospinal fluid (CSF). Table S6: The clones of Dako antibodies used for immunohistochemistry analyses. Table S7: Antibodies used for flow cytometry. Table S8: MicroRNA sequences with GC content.