Corticosteroid-dependent association between prognostic peripheral blood cell-free DNA levels and neutrophil-mediated NETosis in patients with glioblastoma

Abstract

Purpose:

Non-invasive prognostic biomarkers to inform clinical decision-making are an urgent unmet need for the management of patients with glioblastoma (GBM). We previously showed that higher circulating cell-free DNA concentration [ccfDNA] is associated with worse survival in GBM. However, the biology underlying this is unknown.

Experimental Design:

We prospectively enrolled 129 patients with treatment-naïve GBM with blood drawn prior to initial resection (baseline) and at time of first post-radiotherapy MRI. We performed ccfDNA methylation deconvolution to determine cellular sources of ccfDNA. ELISA was performed to detect citrullinated H3 (citH3), a marker of neutrophil extracellular traps (NETs). Multiplex proteomic analysis was used to measure soluble inflammatory proteins.

Results:

We found that neutrophils contributed the highest proportion of prognostic ccfDNA. The percentage of ccfDNA derived from neutrophils was correlated with total [ccfDNA], but only in patients receiving pre-operative corticosteroids. At baseline and on-therapy, [citH3] was significantly higher in the plasma of patients with GBM receiving corticosteroids compared to corticosteroid-naïve GBMs or no-cancer controls. Unsupervised hierarchical clustering of ccfDNA methylation patterns yielded two clusters, with one enriched for patients with the NETosis phenotype and who received corticosteroids. Unsupervised clustering of circulating inflammatory proteins yielded similar results.

Conclusions:

These data suggest neutrophil-mediated NETosis is the dominant source of prognostic ccfDNA in patients with GBM and may be associated with glucocorticoid exposure. If further studies show that pharmacological inhibition of NETosis can mitigate the deleterious effects of corticosteroids, these plasma markers will have important clinical utility as non-invasive correlative biomarkers.

Introduction

We previously demonstrated in two independent GBM cohorts that patients with high baseline levels of plasma circulating cell-free DNA (ccfDNA) have poor survival, even after adjusting for other prognostic factors (1,2). The biological basis of this association, however, is not clear, as patients with GBM generally have low or undetectable levels of circulating tumor-derived cell-free DNA (ctDNA) in plasma (1–4). Using deconvolution of plasma methylation data across several healthy and cancer cell types, the predominant tissue of origin for ccfDNA in patients with non-CNS tumors has been identified as white blood cells (WBCs), with neutrophils being the largest WBC source (5,6). However, patients with glioma were not included in these studies. In one prior report, glioma patient serum contained high and variable amounts of neutrophil ccfDNA. However, IDH mutation status was mutated or unknown in ~85% of patients, making it unknown whether patients fulfilled the WHO criteria for GBM, and neither prognostic associations nor sources of neutrophil ccfDNA were explored (7). A subsequent glioma study also identified neutrophils as a dominant source of DNA, although this group utilized post-operative whole blood, suggesting the majority of extracted DNA was cellular, not ccfDNA (8). Overall, the origins of prognostic ccfDNA in patients with GBM remain unknown.

The mechanisms by which neutrophils influence GBM patient outcomes are not well understood. Increased neutrophil infiltration of the GBM tumor microenvironment was shown to promote tumor progression and necrosis and reduce overall survival (OS) in genetically altered mice (9). A study of patients with primary and metastatic brain tumors demonstrated neutrophilic infiltrate in GBM tumors, and tumor-associated neutrophils had a pro-tumor and immunosuppressive phenotype (10). One potential explanation for neutrophil-mediated pro-tumor effects is the release of neutrophil extracellular traps (NETs), webs of decondensed chromatin decorated with citrullinated histones and cytoplasmic proteins that are extruded from activated neutrophils to form web-like extracellular structures. While NETs provide a first line of defense against bacteria and viruses (11,12), NETosis has also been implicated in illnesses including SARS-CoV-2, thrombosis, autoimmune disorders, and some cancers (13–16). Immunofluorescence tissue staining detected varying levels of the NETosis marker citrullinated H3 (citH3) (17) in a cohort of 10 high-grade gliomas of unknown IDH mutation status (16). However, to our knowledge, evidence of systemic NETosis in the setting of GBM has not been demonstrated for a large, rigorously enrolled patient cohort.

Interplay between neutrophils and corticosteroids, which are widely used to manage symptoms of patients with GBM, has recently been explored. Franco et al. performed RNA sequencing of various glucocorticoid-treated cell types from healthy donors and found neutrophils to have a higher number of glucocorticoid-responsive genes than other WBC subtypes (18). Consistent with this, steroid dose-dependent increases in circulating neutrophils were measured in a cohort of healthy individuals (19). A blood-based methylation approach for predicting epigenetic response to dexamethasone treatment in patients with glioma found neutrophils to be a better indicator than other leukocytes (20). In murine breast cancer models, glucocorticoids have been shown to induce neutrophils to form NETs and establish a metastasis-promoting microenvironment (21).

Given this background, we sought to identify the sources of prognostic ccfDNA for patients with GBM and determine whether neutrophil-derived ccfDNA is associated with NETosis markers. Accounting for the known effects of corticosteroid exposure on neutrophil transcriptional programs, we also explored whether corticosteroid dosing impacts the relationship between ccfDNA levels and NETosis.

Methods

Patient Population

A study of adult patients with newly diagnosed, histopathologically confirmed glioblastoma was performed at the University of Pennsylvania between February 2018 and November 2022 under IRB protocol #828164. Potential subjects with radiographic suspicion for high-grade gliomas were prospectively approached for written informed consent prior to biopsy or initial surgical resection. Preoperative whole blood samples were collected from patients in the operating room immediately prior to resection, from those who provided written informed consent. Healthy patients who were undergoing routine colonoscopy and endoscopy procedures were consented from 2015 to present under IRB protocol #822028, and blood was collected prior to their procedure. Both studies were performed in accordance with the principles of the Declaration of Helsinki. Baseline demographic and clinical variables were abstracted from the medical record, including age, sex, O6-methylguanine-methylatransferase (MGMT) promoter methylation status, exposure to pre-operative corticosteroids, and the extent of surgical resection (biopsy or partial resection vs near or gross total resection). Representativeness of study participants was collected (Supplemental Table 1). Patients were prospectively followed for overall survival (OS), with OS calculated as the number of months (30.4 days) from blood draw to death or censor. Date of death was abstracted from medical records or publicly available death data. Median follow-up was 58.39 months for the full 129 patients with GBM in our cohort. Following surgery, patients remained on the study if 1) histopathology confirmed a diagnosis of GBM, 2) the tumor was determined to be IDH1/2 wildtype by targeted NGS, and 3) the patient received standard of care adjuvant therapy including radiation with or without concomitant temozolomide chemotherapy following surgery. Exclusion criteria included patients with a concurrent cancer, those currently being treated for a previously diagnosed tumor, and any patient enrolled onto a clinical trial or receiving experimental therapy. Corticosteroid dosing for the 48 hours prior to resection (“pre-operative”) and prior to first post-radiation MRI (“post-radiation”) were obtained through abstraction of the electronic medical record. Oral prednisone dosage was converted to dexamethasone equivalents by dividing the prednisone dose by 6.66.

To obtain an age- and sex-matched no-cancer control cohort, whole blood samples were collected from 32 healthy donor patients undergoing a routine upper or lower endoscopy at the University. Neither age (median = 66, range = 39–85 years old, N=32) nor sex (40.3% female) was significantly different from patients in the GBM cohort (Table 1, N=129) with P=0.7669 (Mann-Whitney test) and P>0.9999 (Fisher’s Exact test), respectively, for the archival samples selected for this control cohort. Healthy patients were 18 years of age or older and non-pregnant with no history of cancer or inflammatory diseases known to be associated with NETosis, organ transplant, viral hepatitis, HIV, or pancreatic disease and no recent surgical procedures or traumas. These patients were consented and enrolled under IRB protocol #822028 and similarly provided written informed consent. Chart review was performed to ensure the no-cancer control patients had not been exposed to systemic steroids prior to blood draw.

Table 1.

Baseline characteristics of study participants for all patients included in the study (N=129), as well as for patients with ccfDNA concentration less than or equal to (N=65) or greater than (N=64) the median value and the statistics comparing these two groups (Mann-Whitney test (^*) and Fisher’s exact test (^†)).

	All Patients	ccfDNA ≤ Median	ccfDNA > Median	P-value
N	129	65	64
Age
Median (Min-Max)	66 (39–85)	65 (39–85)	67.5 (49–81)	0.1376*
<65, n (%)	59 (46%)	34 (52%)	25 (39%)	0.1585†
>65, n (%)	70 (54%)	31 (48%)	39 (61%)
Sex, n (%)
Female	52 (40%)	28 (43%)	24 (38%)	0.5914†
Male	77 (60%)	37 (57%)	40 (63%)
Race, n (%)
African American	5 (4%)	2 (3%)	3 (5%)	0.5455†
Asian	2 (2%)	0 (0%)	2 (3%)
Other	2 (2%)	1 (2%)	1 (2%)
Unknown	9 (7%)	3 (5%)	6 (9%)
White	111 (86%)	59 (91%)	52 (81%)
Ethnicity, n (%)
Not Hispanic or Latino	122 (95%)	63 (97%)	59 (92%)	0.4249†
Hispanic	3 (2%)	1 (2%)	2 (3%)
Not Reported	4 (3%)	1 (2%)	3 (5%)
Pre-operative corticosteroids, n (%)
Unexposed	23 (18%)	15 (23%)	8 (13%)	0.1670†
Exposed	106 (82%)	50 (77%)	56 (88%)
Surgical resection, n (%)
Near Total/Gross Total	90 (70%)	50 (77%)	40 (63%)	0.1014†
Biopsy Only/Partial	38 (29%)	15 (23%)	23 (36%)
Unavailable	1 (1%)	0 (0%)	1 (2%)
MGMT promoter methylation status, n (%)
Methylated	53 (41%)	28 (43%)	25 (39%)	0.7213†
Unmethylated	76 (59%)	37 (57%)	39 (61%)
Imaging Volume
N	122	63	59
Total (contrast-enhancing+ T2/FLAIR), Median (Min-Max)	62178 (4749–250603)	47672 (6302–250603)	73833 (4749–187373)	0.0427*
Karnofsky Performance Status (KPS), n (%)
≤60	5 (4%)	0 (0%)	5 (8%)	NA
>60	26 (20%)	14 (22%)	12 (19%)
Unknown	98 (76%)	51 (78%)	47 (73%)

^*Mann-Whitney Test,

^†Fisher’s Exact Test

Specimen Collection and Plasma Isolation

Whole blood samples were collected in either K₂EDTA (Becton Dickinson) or Streck^® Cell-Free DNA (Streck) blood collection tubes. K₂EDTA samples were stored and processed at 4C; samples were banked within 2 hours of collection. Whole blood was centrifuged at 1,900 × g for 10 minutes; the plasma supernatant was isolated and centrifuged at 3,000 × g for 15 minutes (swinging bucket rotor, brake-on). Streck samples were stored and processed at room temperature; samples were banked within 5 days of collection. Streck whole blood was centrifuged at 1,600 × g for 10 minutes; the plasma supernatant was then collected and centrifuged one or two additional times at 4,122 × g for 15 minutes (swinging bucket rotor, brake-off). All plasma samples were aliquoted at 1mL and stored at −80C for future use. As shown in Supplemental Figure 1, and given limitations in the volume of blood collected, not all assays could be completed for all patients. Given this, results for the maximum possible sample size are shown for each assay with the N’s listed in each figure legend.

ccfDNA Extraction, Quantification, and Methylation Profiling

Extraction of ccfDNA from plasma was performed using the QIAamp MinElute ccfDNA Mini Kit (Qiagen) according to the manufacturer’s instructions. Quantification was performed using a SYBR Green-based qPCR assay for a 115 bp amplicon of the human ALU repeat (22). Amplification was performed using forward primer 5’-CCT GAG GTC AGG AGT TCG AG-3’ and reverse primer 5’-CCC GAG TAG CTG GGA TTA CA-3’ (Integrated DNA Technologies). DNA standard (Promega) curve dilutions and cfDNA samples were diluted 1:10 in nuclease-free water. Power SYBR Green PCR Master Mix (Applied Biosystems) was prepared according to the manufacturer’s instructions; qPCR was performed on 1uL of sample and carried out in quadruplicate. Enzymatic conversion and PCR amplification of 10ng of ccfDNA was performed using the NEBNext^® Enzymatic Methyl-seq Kit (New England BioLabs). Samples were not sheared prior to enzymatic conversion. The PCR-amplified product was shipped to the University of Minnesota Genomics Center where it underwent a slightly modified processing protocol that excluded bisulfite conversion prior to hybridization to Infinium Methylation EPIC v2 BeadChip Arrays (Illumina).

Tissue DNA Extraction, Quantification, and Methylation Profiling

Tissue tumor specimens obtained at time of resection were fixed in formalin and embedded in paraffin. In brief, sections (10 μm) of formalin-fixed and paraffin-embedded (FFPE) tissue were cut, DNA extracted, and processed for methylation analysis on EPIC arrays (Illumina). For 25 samples, DNA was extracted using clinically validated protocols using Maxwell Promega, and DNA methylation profiling was performed at NYU Pathology using Illumina EPIC v1 array. DNA extraction for 54 samples was performed at University of Pennsylvania from 5 tissue sections per sample using the QIAamp DSP DNA FFPE Tissue Kit (Qiagen) according to manufacturer’s instruction. The only modification was use of 180uL of Deparaffinization Solution (Qiagen) instead of xylene. DNA was quantified using Qubit dsDNA BR Assay Kit (Invitrogen, Thermo Fisher Scientific) and 500 ng was shipped to the University of Minnesota Genomics Center (UMGC) for DNA methylation profiling using Illumina EPIC v2 array. Both laboratories followed manufacturer’s instructions and previously published protocol (23). All array output iDAT files from EPIC arrays were analyzed at NYU using clinically validated CNS tumor DNA methylation classifier v12 (24).

Citrullinated Histone 3 ELISA

Plasma citH3 concentration (ng/mL) was measured by sandwich ELISA (Cayman Chemical Cat No. 501620). Plasma samples banked at −80C were thawed at room temperature and diluted 2-fold in provided Assay Buffer. Samples were run on a DS2 Automated ELISA instrument (Dynex Technologies) according to assay instructions, with TMB color developed for 1 hour at room temperature.

Radiographic Measurements

Pre-operative brain MRI was performed using the brain tumor imaging protocol of the University of Pennsylvania on a 3-T magnet (Trio, Skyra or Prisma; Siemens), which included an axial T1-weighted 3-dimensional magnetization-prepared rapid gradient echo sequence before and after contrast administration, a postcontrast axial fluid-attenuated inversion recovery sequence. Structural MRI data were analyzed for every subject by a board-certified neuroradiologist (A.N.) and tumor segmentation was performed in a semi-automated fashion using ITK-SNAP (RRID:SCR_002010) and subsequently tumor volumes were measured.

Plasma ccfDNA and Tissue DNA Methylation Array Data Processing

Methylation signals were extracted from iDAT files and preprocessed using the R package SeSAMe. The signal set was background-corrected with noob, and probe detection P-values were calculated using the pOOBAH algorithm implemented in SeSAMe (25,26). To account for the highly fragmented nature of ccfDNA and to recover as much signal as possible, the detection cut-off was chosen to be 0.2, higher than the conventionally used value of 0.05. Probes with detection p-value less than this cut-off were masked and removed from subsequent analysis, as well as probes known to underperform (27). Beta values (Methylated probe signal/(Methylated probe signal + Unmethylated probe signal)) were calculated fArom the signal set using unmasked probes. Duplicate probe values were averaged using SeSAMe’s collapseToPfx function. Non-CpG and sex chromosome probes were excluded.

Reference Methylomes Used in Deconvolution

Published reference methylomes were obtained from the Gene Expression Omnibus (GEO, RRID SCR_005012) and are provided in Supplemental Table 2.

Deconvolution

Sample data was obtained from both EPIC v1 and EPIC v2 arrays (25 tissue DNA samples were obtained from EPIC v1 and all remaining were EPIC v2). Additionally, reference data were obtained from EPIC v1 or the HM450K arrays. Therefore, only CpGs overlapping between the EPIC v1, EPIC v2, and HM450K arrays were utilized in deconvolution. Further, CpGs reported to have poor translatability between EPIC v1 and EPIC v2 were excluded (28) (list available as part of EPIC v2 documentation at https://support.illumina.com/content/dam/illumina-support/documents/downloads/productfiles/methylationepic/InfiniumMethylationEPICv2.0ProductFiles(ZIPFormat).zip), resulting in a set of 267,126 CpGs that could be analyzed across the 3 arrays (Supplemental Table 3). Finally, CpGs with missing data for samples or references were excluded resulting in a final set of 154,990 CpGs. The FeatureSelectV4 algorithm from the MethylCIBERSORT package was used to generate a 2,031 CpG signature from the reference data for this set.

Mean and standard deviation values for each cell-type were calculated from the complete reference betas matrix to yield an average reference matrix and standard deviation reference matrix. The estimateCellComposition function from the SeSAMe R package was used for deconvolution of sample betas utilizing these reference matrices for the signatures vector described above. This function uses an iterative approach to obtain a set of coefficients that minimizes the absolute deviation between the linear combination of the average reference matrix and the matrix of sample betas. To account for biological noise, the average reference matrix was perturbed according to the standard deviation reference matrix for 20 iterations of the deconvolution, and the averaged results were reported as final cell fraction estimates. These cell fraction estimates are reported as % cell type ccfDNA or tissue DNA (i.e. % neutrophil ccfDNA).

Differential Methylation and Pathway Analysis

All ccfDNA data was obtained utilizing the EPIC v2 array, therefore the set of all autosomal CpGs on this array was considered in subsequent analyses (902,485 CpGs). As tissue DNA data were obtained from both EPIC v1 and EPIC v2 arrays, similar exclusion to that listed above for deconvolution was performed, however overlap with 450K was unnecessary resulting in a larger set of 521,016 CpGs. Unsupervised hierarchical Euclidean clustering analyses were performed with R’s hclust function. After limiting to CpGs listed above, any CpGs with missing sample data were excluded, and clustering was based on the top 2,000 most variable probes by standard deviation. Continuous biomarker values were n5 normalized using clusterSim (RRID:SCR_023743) function data. Normalization and a 1 to 9 scale were applied for display on the heatmap.

Differential methylation analyses for CpG sets described above were performed with limma (RRID:SCR_010943). Significant differential methylation was defined as adjusted P-value < 0.05 and |Δbeta|>0.1. Methylation gene set analysis was performed with methylGSA, using methylRRA with the method option set to “GSEA” (29). Gene ontology (GO) gene sets with sizes between 25 and 500 were included in the analysis (RRID:SCR_002811, SCR_020938). Ranking of gene sets (used to determine top 10 gene sets in bubble plots) was based on lowest adjusted p-value then highest normalized enrichment score. A custom NETosis gene set was curated by combining RNA sequencing-based NETosis signatures across five publications, with considerable overlap (30–34). This custom gene set was tested using the methylRRA-GSEA method described above. Heatmaps were generated using wheatmap (35) and scales (RRID:SCR_019295). Volcano and bubble plots were generated with ggplot2 (RRID:SCR_014601) and forcats from tidyverse (RRID:SCR_019186). The BioVenn R package was used to calculate overlap of differentially methylated CpG lists and produce Venn diagram visualizations (36).

Olink Multiplex Proteomic Analysis

EDTA plasma was shipped to Olink Proteomics (Watertown, MA) overnight on dry ice. Samples were run on the Olink Inflammation Panel which tested for the presence of 92 human inflammatory proteins. Normalized protein expression (NPX) values were calculated and reported by Olink. Of 98 samples sent to Olink, results for 5 were excluded, including 3 that did not meet Olink QC parameters and 2 that had profiles dramatically different than the remaining 93. Hierarchical Euclidean clustering was performed as described for methylation data using n5 renormalized Olink NPX values. Differential expression analysis was performed on NPX values with limma and significantly differentially expressed proteins were defined as those with adjusted P-value < 0.05. Venn diagrams were produced as described for methylation data. Correlation analysis of NPX values was performed in GraphPad Prism (RRID:SCR_002798).

Statistical Analysis

Kaplan-Meier, log-rank, and Cox regression analyses were performed in Stata/IC 16.1 (Stata Corp., RRID:SCR_012763). Spearman correlation, multiple linear regressions, Fisher’s exact test, Mann-Whitney test, Dunn’s Multiple Comparison test, and ROC analyses (all two-sided) were performed in Graphpad Prism 9.5.1 (RRID:SCR_002798). Mann-Whitney test and/or Spearman correlation were used for comparison of two continuous variables. ROC analysis was used for comparison of a continuous variable between two groups. Multiple linear regression was used to adjust for the contribution of two continuous variables on a third; ANOVA parameter P-values for variables of interest are reported. Fisher’s exact test was used for comparison of categorical variables between two groups. Dunn’s multiple comparisons test was used for comparison of continuous variables between three groups with adjustment for multiple testing. All median dichotomizations were performed within the cohort being tested at less than or equal to the median vs greater than the median. Survival statistics were generated using Kaplan-Meier curves to estimate survival function, log-rank test for comparing survival functions between two groups, and Cox model to estimate hazard ratios.

Data Availability

The de-identified participant data underlying the findings in this manuscript are available from the corresponding author upon reasonable request. They were not made publicly available because of patient privacy concerns.

Results

Neutrophils are the dominant source of prognostic ccfDNA for patients with GBM

Whole blood was obtained prior to primary resection from 129 patients with newly diagnosed GBM. A post-treatment draw was obtained at the time of first post-radiation MRI in 39 of the patients. For some patients, there was insufficient blood to perform all assays (sample sizes for each assay shown in Supplemental Figure 1, Supplemental Table 4); therefore, sample sizes for each assay are reported below and/or shown in each figure legend. Baseline (pre-operative) patient characteristics are displayed in Table 1. Consistent with our earlier work, total ccfDNA concentration above the median was associated with worse OS at baseline (HR=1.50 [1.04–2.17], log-rank P=0.0295) and post-radiotherapy (HR=4.32 [1.96–9.51], P=0.0001) (Supplemental Figure 2). In our cohort, 106 patients (82%) were treated with corticosteroids in the 48 hours prior to initial resection and blood draw (total dose received = 1 to 54 mg; median = 20 mg). When pre-operative steroid dose, age, MGMT promoter methylation status, resection status (biopsy only/partial vs total/near total), and total tumor volume were added in a multivariate analysis, baseline [ccfDNA] remained significantly associated with OS (HR=1.53 [1.02–2.29], P=0.040; N=121). A significant association was also observed for multivariable analysis at time of first post-radiation MRI (Supplemental Figure 2).

To determine ccfDNA tissue of origin, we used published reference methylomes (5,37–41) (see detailed list in methods) to perform deconvolution of plasma ccfDNA methylation profiles. The percentage of ccfDNA from neutrophils was higher for patients with GBM than for any other cellular source analyzed. Percent neutrophil ccfDNA was significantly higher for patients with GBM (median 46.58%, range 17.10–69.88%) than no-cancer controls (20.39% (11.56–44.00%), P <0.0001) (Figure 1A). No-cancer controls did not receive systemic corticosteroids prior to blood draw. There was no other tissue of origin for which the % ccfDNA was significantly higher for patients with GBM relative to no-cancer controls. Total ccfDNA concentration was positively correlated with % neutrophil ccfDNA for patients with GBM (P < 0.0001, r=0.4669), but not controls (P=0.3221, r=-0.1839) (Figure 1B). No other source identified by our deconvolution algorithm was significantly positively associated with [ccfDNA] for patients with GBM (Supplemental Figure 3). Methylation deconvolution analysis was next performed for tissue samples obtained at the time of first resection. Median neutrophil % tissue DNA was 3.36% (1.47%-51.84%) of total tissue DNA (Figure 1C). There was no significant association of % neutrophil DNA for ccfDNA vs. tissue (P=0.2294, r=0.1488) (Figure 1D). The % neutrophil tissue DNA was not significantly associated with steroid dose, either as a categorical or continuous variable (Supplemental Figure 4). Taken together, these results provide systemic evidence suggesting neutrophils as the dominant source of prognostic ccfDNA for patients with GBM.

Figure 1. Sources of DNA in peripheral blood and tissue.

Shown in (A) is deconvolution of plasma ccfDNA methylation results for 110 patients with glioblastoma (GBM) and 31 age- and gender-matched no-cancer controls with percentage of ccfDNA from WBC and tissue types as indicated. Mann-Whitney test with Bonferroni correction was utilized to compare groups with p-value indicated as *P<0.05, **P<0.01 and ****P<0.0001. Black horizontal bars indicate median values. In (B), the correlation of % neutrophil ccfDNA with total ccfDNA concentration is shown for 110 patients with GBM (left) and 31 no-cancer controls (right). In addition, tissue was obtained, and methylation analysis of extracted DNA performed for 79 patients with GBM. Shown in (C) is the deconvolution of cellular sources of tissue DNA. Shown in (D) is the correlation of % neutrophil DNA from tissue DNA vs plasma ccfDNA for the 67 patients with both tissue and ccfDNA methylation analysis. All statistical tests are two-sided with no adjustment for multiple testing (unless indicated otherwise).

Neutrophil ccfDNA is strongly associated with NETosis marker citrullinated histone 3

Given the high proportion of plasma ccfDNA from neutrophils, we reasoned this might be associated with NETs (42). To assess this, we performed ELISA to quantify levels of the NETosis marker citH3 (17). We found that plasma [citH3] was significantly elevated in patients with GBM (1.006 ng/mL, 0.063–5.938) compared to no-cancer controls (0.348 ng/mL, 0.085–2.440; P<0.0001) (Figure 2A). Plasma [citH3] was significantly positively associated with % neutrophil ccfDNA for patients with GBM (P<0.0001, r=0.6485) but not for no-cancer controls (P = 0.8527, r=-0.0400) (Figure 2B), suggesting that homeostatic proliferation and apoptosis could be a more dominant source of neutrophil ccfDNA for no-cancer controls. Plasma [citH3] was not significantly positively associated with any other WBC subset (Supplemental Figure 5), and there was no correlation between [citH3] and age (P=0.8880, r=0.0134). When compared to methylation deconvolution analysis of tissue, plasma [citH3] was not significantly associated with % neutrophil tissue DNA (Supplemental Figure 6).

Figure 2. Association of % neutrophil ccfDNA with the NETosis marker citrullinated histone 3.

ELISA was performed on baseline plasma to determine the concentration of citrullinated histone 3 (citH3). Shown in (A) are results for 113 patients with glioblastoma (GBM) and 25 no-cancer controls with associated Mann-Whitney P-value. Shown in (B) are correlations of plasma [citH3] with % neutrophil ccfDNA as determined by methylation deconvolution for the 107 patients with glioblastoma (GBM, left) and 24 no-cancer controls (right). Of the 113 patients with citH3 results, only 107 also had % neutrophil ccfDNA data. In (C), ELISA-determined plasma citH3 values are again shown but the GBM patient cohort is split between those who received (“exposed”) or didn’t receive (“unexposed”) pre-operative steroids. Dunn’s multiple comparison P-values are listed. In (D) the associations of plasma [citH3] and % neutrophil ccfDNA are shown for patients with GBM who were exposed to steroids (left, N=86) and those who weren’t (right, N=21). All statistical tests are two-sided with no adjustment for multiple testing (unless indicated otherwise).

Corticosteroids are routinely given to patients with GBM and have also been shown to induce NET formation in experimental models of breast cancer (21). Consistent with this, we observed a significant correlation for [citH3] and pre-operative steroid dose (P=0.0006, r=0.3279). Given that baseline tumor volume was also significantly correlated with [citH3] levels (P=0.0007, r=0.3225), we performed a multiple linear regression and found that [citH3] remained significant after adjusting for tumor volume (P=0.0156; N=106) (Supplemental Figure 7). At baseline, [citH3] was higher at a median of 1.159 ng/mL (0.063–5.938) for patients who received pre-operative steroids compared with those who did not (0.564 ng/mL (0.1420–2.310); P=0.0143). CitH3 levels for the patients unexposed to steroids were not significantly different from those of no-cancer controls (P=0.3297; Figure 2C). Similar results were obtained at baseline for % neutrophil ccfDNA (Supplemental Figure 8). Consistent with these baseline findings, [citH3] and % neutrophil ccfDNA were both significantly higher at time of post-radiotherapy MRI for patients who received steroids vs those who did not (Supplemental Figure 9). A strong correlation of [citH3] with % neutrophil ccfDNA was observed in patients receiving steroids (P<0.0001, r=0.6225) but slightly less strong in patients who did not receive steroids (P=0.0337, r=0.4649; Figure 2D). Similar results were obtained for the correlation of total [ccfDNA] with % neutrophil ccfDNA (Supplemental Figure 10). Taken together, this analysis suggests NETosis as a major source of plasma ccfDNA in patients with GBM, although NET-derived DNA is only elevated in the setting of corticosteroid exposure.

Unsupervised hierarchical clustering of plasma ccfDNA methylation results

To further explore the plasma ccfDNA methylation patterns for evidence of NETosis, we performed unsupervised hierarchical clustering for 110 patients based on the 2,000 most variably methylated CpGs, resulting in two main clusters (Figure 3A). Compared to Cluster 2, Cluster 1 was enriched for patients with significantly higher % neutrophil ccfDNA, higher plasma [citH3], and patients who had received pre-operative steroids (Figures 3B–3D); there was no difference in tumor molecular subtypes between Cluster 1 vs 2 (P=0.6391) (Supplemental Figure 11) (24). More CpGs were hypomethylated (N=15,133) than hypermethylated (N=1,037) in Cluster 1 vs Cluster 2 (Figure 3E). To identify gene sets associated with the methylation profiles of patients in Cluster 1 vs 2, methylation gene set enrichment analysis was used (full results in Supplemental Table 5). Among the top ten pathways (Figure 3F), four are related to specific neutrophil granules (red arrows) which are functional elements of NETosis. To specifically assess NETosis-associated genes, we next compiled and applied a bespoke set of 69 NETosis-associated genes (Supplemental Table 6) (30–34) yielding a highly significant (P=0.0004, normalized enrichment score (NES)=1.40) association between patients in Cluster 1 vs Cluster 2. Unsupervised clustering of tissue DNA methylation results yielded two major clusters, however, neither was enriched for patients with high % neutrophil ccfDNA, high plasma [citH3], or patients who received steroids (Supplemental Figure 12A–D). Although a large number of CpGs were differentially methylated for patients in cluster 1 vs 2, no neutrophil-associated pathways were identified in the top 10 gene ontology (GO) gene sets (Supplemental Figure 12E and 12F; Supplemental Table 7). Differential methylation between the two clusters did not yield significant enrichment of our bespoke NETosis gene set (P=0.0509, NES=1.21). As expected, there was a significant difference in the molecular tumor subtypes identified for Cluster 1 vs 2 (P<0.0001) (Supplemental Figure 13A). Volcano plots dichotomized at median % neutrophil plasma ccfDNA, median plasma [citH3], and steroid exposure were unremarkable (Supplemental Figure 13B–D).

Figure 3. Unsupervised clustering of plasma ccfDNA methylation data, differential methylation, and pathway analysis.

Shown in (A) is an unsupervised clustering analysis of the 2000 most variably methylated CpGs in the plasma ccfDNA of 110 patients with glioblastoma (GBM). The heatmap depicts the normalized beta values for these CpGs, as well as normalized values for % neutrophil ccfDNA, citH3 concentration (N=107), and pre-operative steroid dose (rows under heat map) for each patient (columns). For citH3, grey indicates sample unavailable for analysis. For pre-operative steroids, white indicates a zero value, i.e., patients unexposed to pre-operative steroids. In (B, C, and D), the associations of % neutrophil ccfDNA, plasma [citH3], and pre-operative steroid dose (respectively) with Cluster 1 and Cluster 2 from (A) are shown as values with Mann-Whitney P-value (top) and non-parametric ROC analysis (bottom). In (D, right), this association is shown for pre-operative steroid exposure (exposed vs. unexposed; Fisher’s exact test). The volcano plot in (E) depicts the association of all CpGs measured with these two clusters; red dots indicate significantly hypomethylated or hypermethylated with an FDR-adjusted P-value > 0.05 and |Δβ| > 0.1, with the number of CpGs meeting these criteria listed below the plot. The top 10 Gene Ontology (GO) gene sets (selected first by ordering by p-value, and when two or more pathways had the same p-value, then ordered by NES) associated with this differential methylation profile are shown in (F) with red arrows identifying pathways associated with the NETosis process. To generate the bubble plot shown, pathways were re-ordered based on gene set size. The Venn diagram shown in (G) demonstrates the overlap of significantly differentially methylated CpGs for patients dichotomized at median [citH3] or median % neutrophil ccfDNA and for patients dichotomized by pre-operative steroid exposure. All statistical tests are two-sided with no adjustment for multiple testing (unless indicated otherwise).

Given that Cluster 1 was enriched for patients with high % neutrophil ccfDNA, the NETosis marker citH3, and for patients who received steroids, we reasoned there might be overlap in the differentially methylated CpGs for patients who were high vs low for these three characteristics. To assess this, we first identified differentially methylated CpGs for patients with % neutrophil ccfDNA or [citH3] levels above vs below the median, and for patients who received any vs no steroids. For patients with high vs low % neutrophil ccfDNA or [citH3], or with any vs no pre-operative steroids, the number of hypomethylated CpGs far exceeded those that were hypermethylated (Supplemental Figure 14, and full gene set lists in Supplemental Tables 8, 9 and 10). When we compared the three sets of differentially methylated CpGs, we found that, of the 4,929 CpGs differentially methylated for patients with high vs low [citH3], 4,902 (99.5%) were also differentially methylated for patients with high vs low % neutrophil ccfDNA and 462 (9.4%) were also differentially methylated for patients who were steroid exposed vs unexposed (Figure 3G). Taken together, these results suggest a strong association between % neutrophil ccfDNA and plasma-based measures of NETosis in patients with GBM.

Plasma inflammatory markers

We next performed targeted proteomics of 92 circulating inflammatory biomarkers to identify further evidence of the NETosis phenotype in plasma. Unsupervised hierarchical clustering yielded the heatmap shown in Figure 4A. Similar to what was shown above for the plasma methylation analysis in Figure 3, one of two main clusters was enriched for patients with higher % neutrophil ccfDNA, higher [citH3], and exposure to pre-operative corticosteroids (Figure 4B). Fourteen differentially expressed proteins were identified when comparing patients with high vs low % neutrophil ccfDNA, with 13 underexpressed proteins and one overexpressed protein, oncostatin M (OSM) (Figure 4C, Supplemental Table 11). When patients with high vs low plasma [citH3] were compared, 12 differentially expressed proteins were identified, including 8 underexpressed. Four proteins were significantly overexpressed: OSM, extracellular newly identified receptor for advanced glycation and end-products binding protein (S100A12, alias EN-RAGE), interleukin-10 (IL10), and hepatocyte growth factor (HGF). When patients who received any vs no pre-operative corticosteroids were compared, 37 proteins were differentially expressed, including 34 underexpressed. Among the 3 overexpressed proteins (OSM, LIFR, and IL18R1), OSM was most highly and significantly overexpressed. The intersections of the significantly differentially expressed proteins from the three volcano plots are depicted as a Venn diagram in Figure 4D. Among the 12 proteins differentially expressed for high vs low [citH3], 8 (66.7%) were also differentially expressed for patients with high % neutrophil ccfDNA. All 14 markers that were differentially expressed for patients with high vs low % neutrophil ccfDNA were also differentially expressed for patients with any vs no pre-operative steroid exposure. Eight common proteins were differentially expressed in all 3 comparisons, with OSM upregulated and the other 7 downregulated. This targeted proteomic analysis is consistent with the ccfDNA methylation analysis above and further suggests an association between pre-operative steroid exposure and NETosis.

Figure 4. Analysis of plasma inflammatory biomarker levels.

Shown in (A) is an unsupervised clustering analysis of the plasma levels of 92 inflammatory biomarkers for N=93 patients with glioblastoma (GBM). The heatmap depicts the n5 re-normalized values for these biomarkers (red to blue), normalized values for % neutrophil ccfDNA, citH3 concentration, and pre-operative steroid dose (orange rows under heat map), as well as the unsupervised hierarchical cluster indicated by the ccfDNA methylation profile from Figure 3A (green row under heat map), for each patient (columns). For % neutrophil ccfDNA, grey indicates sample unavailable for analysis. For pre-operative steroids, white indicates a zero value, i.e., patients unexposed to pre-operative steroids. In (B, dot plots from left to right), the associations of % neutrophil ccfDNA (N=87), plasma [citH3], and pre-operative steroid dose with cluster 1 vs cluster 2 from (A) are shown as values and Mann-Whitney P-value (top) and non-parametric ROC analysis (bottom). Additionally, in B (bar charts far right), this association is shown for pre-operative steroid exposure (exposed (black) vs. unexposed (white); Fisher’s exact test), and also for the ccfDNA clusters from Figure 3A (ccfDNA Cluster 1 (light green) vs ccfDNA Cluster 2 (dark green); Fisher’s exact test). The volcano plots in C (left to right) depict the association of the 92 inflammatory biomarkers with % neutrophil ccfDNA dichotomized at the median value, [citH3] dichotomized at the median value, and pre-operative steroid exposure (exposed vs. unexposed); red dots indicate P-adjusted >0.05. The Venn diagram shown in (D) demonstrates the overlap of significantly differentially expressed proteins from C. All statistical tests are two-sided with no adjustment for multiple testing (unless indicated otherwise).

Discussion

We and others have previously demonstrated that total plasma ccfDNA concentration is an independent prognostic biomarker in patients with GBM (1,2,43). Recognizing that peripheral blood levels of tumor-derived ccfDNA are scant in patients with GBM (3,4), we sought to determine the cellular origins of plasma ccfDNA and to better understand the underlying biology of its prognostic impact with an eye toward future therapeutic implications. Utilizing deconvolution analysis of ccfDNA methylation patterns and ELISA-based detection of citH3, we identified neutrophils and NETosis to be the dominant cellular source of prognostic ccfDNA in a prospectively enrolled cohort of patients with GBM. ELISA-based detection of NETosis marker citH3, gene set analysis of ccfDNA methylation pathways, and targeted analysis of inflammatory circulating markers were consistent with this NETosis phenotype.

Few studies have examined ccfDNA sources in the setting of GBM. In a small cohort of 48 glioma patients with mostly IDH-mutant or IDH mutation-unknown tumors, Sabedot and colleagues found neutrophils to be the dominant source of ccfDNA in patient serum, with a median % neutrophil ccfDNA > 80% (7). This is in contrast to a median of 46.58% for our cohort, perhaps due to the plasma used in our study being a cleaner input than serum (44,45). Molinaro, et al. also found neutrophils to be the largest contributor of DNA in the blood of patients with GBM vs controls (8). However, whole blood was used instead of plasma, suggesting that cellular DNA, rather than ccfDNA, was the dominant input. More recently, Mattox, et al. identified neutrophils as the predominant source of ccfDNA among patients with various solid tumors, although their cohort did not include primary brain tumors (6). In contrast, we enrolled a large cohort of 129 patients with GBM, all uniformly fulfilling the current 2021 WHO criteria of having IDH-WT mutational status. In addition, all baseline plasma was banked from blood samples collected immediately prior to initial resection and before any systemic therapy. Thus, we present the only study, to our knowledge, that examines the sources of prognostic ccfDNA in a large, consistently defined GBM patient population, and with optimal blood collection and processing techniques.

The role of circulating neutrophils in the peripheral blood in GBM is not well understood, although several studies have demonstrated that high absolute neutrophil counts and/or high neutrophil-to-lymphocyte ratios are associated with worse outcomes (46–48). NETosis has been associated in pancreatic cancer with resistance to immunotherapy (49). It has recently been established that NETs produced by neutrophils in the peripheral blood have important roles in promoting neoplastic progression and metastatic seeding in cancer (50), although until our study there has not been an evaluation of NETosis in the blood of patients with GBM. The majority of our patients had elevated % neutrophil ccfDNA and [citH3], a specific extracellular trap marker, relative to no-cancer controls, suggesting a much higher level of NET formation in patients with GBM than no-cancer controls. It is also most likely that, for patients with GBM, neutrophils are the dominant source of extracellular traps. In our cohort, neutrophils are a much higher source of plasma ccfDNA than monocytes and eosinophils. Moreover, only neutrophil % ccfDNA strongly positively correlated with [citH3], whereas there was no significant association for monocyte or eosinophil % ccfDNA. Taken together, our data suggest that NETosis in the peripheral blood of patients with GBM is likely responsible for much of the detectable plasma ccfDNA, which may explain the prognostic value of ccfDNA levels given the known pro-tumor effects of NETosis (10,12,42,50).

Depending on the context, neutrophils in the tumor microenvironment can either promote tumor progression by enhancing angiogenesis, tissue remodeling, and immunosuppression (9,10), or can facilitate tumor cell killing and activate anti-tumor adaptive immune responses (51). While our study was not designed to resolve the role of tumor-associated neutrophils, we do observe interesting, hypothesis-generating differences between matched tissue and plasma DNA methylation patterns. While neutrophils were identified by our deconvolution analysis as a tissue of origin for both plasma and tumor DNA, most patients had low amounts of tissue DNA derived from neutrophils. Whether the neutrophil ccfDNA identified by our methylation analysis was shed by cells within the tumor microenvironment or in adjacent vasculature is impossible to determine with our assay. We found no significant association between % neutrophil DNA for tissue vs plasma. These findings are consistent with that of Maas et al., who found phenotypic and functional differences between peripheral blood and tumor-associated neutrophils in a recent study of patients with metastatic and primary brain tumors, including IDH-WT gliomas (10). Nevertheless, as further inroads are made to therapeutically target pro-tumor neutrophil activity, peripheral blood may serve as a source of non-invasive biomarkers of NETosis.

Importantly, we observed that NET-derived DNA in the plasma, as measured by [citH3], was only elevated in GBM patients in the setting of corticosteroid exposure. This was observed for both the pre-operative blood draw (baseline) and also at the time of first post-radiotherapy MRI. In addition, [citH3] was correlated with corticosteroid dosage, and we found considerable overlap in the peripheral blood ccfDNA methylation and proteomic patterns of our patients with % neutrophil ccfDNA or [citH3] above the median value and those who were exposed to pre-operative corticosteroids. Notably, the positive association of [citH3] with steroid dose for our cohort remained significant after adjusting for tumor volume. Taken together, these data suggest that the NETosis we observed is unlikely to purely be a function of tumor size or disease severity, but rather may, in part, be driven by corticosteroid use. Given the observational nature of our study, however, evaluation of whether steroids can cause NETosis in animal models of GBM will be essential to prove this hypothesis. Such an outcome would be consistent with the findings of He et al. who showed in mouse models of breast cancer that glucocorticoid treatment could induce NET formation in vivo (21).

Our study also implies that pharmacologic interventions may be needed to mitigate or block the pro-tumor effects of NETosis in GBM. For instance, Deng et al. showed in mouse models of colorectal and lung cancer that the pharmacologic inhibition of PADI4, an enzyme strongly expressed in neutrophils and known to affect histone H3 and H4 citrullination, could reduce primary and metastatic tumor growth (52). He et al. noted for their in vivo breast cancer models that glucocorticoid-induced NETs were independent of PADI4 enzyme activity but were inhibited by CDK4/6 inhibitors (21). However, in our cohort, we found marked citH3 elevation, suggesting the NETosis observed in our patients was in part PADI4-dependent. Recent studies have explored NETosis targeting with DNAses to reduce circulating NET levels, and other targeted agents (21). In light of these studies and our observational clinical data in patients with GBM, we have commenced preclinical evaluation of therapies aimed to block or reduce the effects of NETosis in this disease.

Our study has some limitations. Our results would be strengthened by a cohort of patients from an outside institution to independently validate our findings, and such efforts are underway. In addition, patients were included in our study only if they were newly diagnosed, IDH-wildtype GBM and treated with standard of care chemoradiation, suggesting that our findings need to be assessed separately before being considered generalizable to other settings such as recurrent GBM, IDH-mutant gliomas, and/or clinical trials of experimental therapies. While our proteomic analysis included a large panel of 92 inflammatory markers, future studies might generate additional hypotheses as to the biological underpinnings of prognostic ccfDNA by performing mass spectrometry or other more comprehensive and unbiased proteomic approaches. Furthermore, given we analyzed stored rather than fresh plasma, we were unable to perform neutrophil functional assays. For epigenetic analysis, we utilized an enzyme-, rather than bisulfite conversion-based approach to DNA preparation to minimize destruction of input material and maximize signal detection. Nevertheless, our targeted approach to methylation analysis of plasma and tissue DNA could be made more comprehensive by performing whole genome methylation analysis.

In summary, NETosis is a major source of prognostic ccfDNA in the peripheral blood of patients with GBM, and this finding may lead to broader, potentially therapeutic implications for patients with GBM. Our results suggest that focused experiments using in vivo models will be essential to establish the role of the tumor microenvironment, corticosteroid dosing, and perhaps other factors in driving NETosis. Our work also suggests that such animal models should be used to evaluate and provide pre-clinical evidence for the efficacy of therapeutic targeting of neutrophils and NETosis in GBM. If we are able to establish that corticosteroid treatment drives NETosis in animal models of GBM, additional studies can be done to determine if co-administration of NET inhibitors can mitigate the harmful effect of corticosteroids. Further work is ongoing to elucidate the underlying mechanism of our observations here and the implications for patients with GBM, Covid-19, or other conditions with evidence of NETosis-associated sequelae.

Statement of Translational Relevance.

Glioblastoma is the most prevalent primary brain tumor in adults and yet plasma-based biomarkers of survival are lacking. While we had previously shown that total ccfDNA is associated with worse overall survival, the source of the ccfDNA had not been identified. Here we show that neutrophil-mediated NETosis is the dominant source of plasma ccfDNA in patients with GBM, suggesting a potential explanation for the association between ccfDNA levels and survival. This work lays the foundation for exploration of NETosis as a therapeutic target, especially for patients receiving corticosteroids.

Abbreviations List

CAR T-cell

Chimeric Antigen Receptor T-cell

ccfDNA

Circulating cell-free DNA

CDK4/6

Cyclin-dependent kinase4/6

citH3

Citrullinated Histone 3

CNS

Central Nervous System

CpGs

Cytosine-guanine dinucleotides

ctDNA

Circulating tumor DNA

FFPE

Formalin-fixed and paraffin-embedded

GBM

Glioblastoma

GO

Gene ontology

HGF

Hepatocyte growth factor

HIV

Human Immunodeficiency Virus

IDH

Isocitrate Dehydrogenase

IDH-WT

Isocitrate Dehydrogenase Wildtype

IL10

Interleukin-10 protein

IL18R1

Interleukin-18 receptor 1

K₂EDTA

Dipotassium ethylenediaminetetraacetic acid

LIFR

Leukemia Inhibitory Factor Receptor

MGMT

O6-methylguanine methyltransferase

NES

Normalized Enrichment Score

NET/NETosis

Neutrophil Extracellular Traps(osis)

NGS

Next generation sequencing

NPX

Normalized protein expression

OS

Overall Survival

OSM

Oncostatin M

PADI4

Peptidyl arginine deiminase 4

qPCR

Quantitative Polymerase chain reaction

ROC

Receiver operating characteristic

SARS-CoV-2

Severe acute respiratory syndrome-coronavirus 2

WBCs

White blood cells

WHO

World Health Organization