We prospectively collected the clinical data of 27 glioma patients admitted to Nanjing Jinling Hospital (Nanjing, China) from December 2020 to July 2021. All patients were confirmed with glioma by magnetic resonance imaging (MRI). CSF samples (4–8 mL) were collected from the lumbar puncture before surgery. Tumor tissue samples were collected from surgical specimens, and part of removed tumor tissue was used for histopathologic diagnosis. 4 CSF samples were collected from patients with encephalitis as normal controls. In accordance with the Declaration of Helsinki, written informed consent was obtained from all individuals or guardians, and the study was approved by the Institutional Review Board of Nanjing Jinling Hospital (2022DZKY-093-01) and local regulations.

Cell-free DNA (cfDNA) was extracted using an Apostle MiniMax High Efficiency cfDNA Isolation Kit (APOSTLE) according to the manufacturer’s instructions. Fresh tissue DNA was extracted using a QIAamp DNA Tissue & Blood Kit (Qiagen) according to the manufacture’s recommendations. Extracted cfDNA and DNA were quantified by Qubit4.0 (Life Technologies) according to the manufacturer’s protocols. Quantitative levels of CSF cfDNA were measured in haploid genome equivalents per milliliter (hGE/mL), which was calculated as the result of total cfDNA concentration and the mean allele fraction of somatic mutations.

NGS libraries of tumor tissue DNA were prepared with KAPA HyperPlus Kit (Roche). cfDNA libraries were established with the xGenTM Prism DNA Library Prep Kit (IDT). After hybridization of DNA and cfDNA libraries, the captured libraries were amplified by polymerase chain reaction (PCR) according to the manufacturer’s protocols. As described previously, we performed the targeted sequencing by NGS with the panel containing 131 genes and 4 chromosomes (131 + 4 panel) related to brain tumors (see Supplementary Table S1 for gene list) [12]. Sequencing was carried out using NovaSeq 6000 system in this study.

All CSF and tumor samples were measured for single nucleotide variants (SNVs), copy number variants (CNVs), and chromosomal variations. Initial sequencing reads were trimmed and filtered with fastp (v.2.20.0) and then aligned against the human reference genome (hg19) using BWA-mem (v.0.7.17). The realignment of local regions was performed to improve the accuracy of variant detection around Indels using ABRA2.

We used VarDict (v.1.5.7) for variant detection. The results of variant detection were annotated to multiple databases using SnpEff and ANNOVAR software, including dbNSFP (v.4.2a), COSMIC (v.9.6), gnomAD (v.3.11), ClinVar and more. According to annotation results, the filtration was further conducted, and the background baseline was used to filter systematic false positive results. In combination of Intervar results and pathogenic interpretation from databases including ClinVar, somatic and germline mutations were finally identified. Threshold for mutation calling is ≥0.5% variant allele frequency (VAF) in CSF and ≥1.0% VAF in tumor tissue.

The original sequencing depth was counted, which was standardized to exclude the influence of the amount of sequencing data. The CNVkit algorithm was used to construct the background distribution model of the copy number with normal genomic sequencing. CNVs were determined based on the difference between the copy number in each region of the sample and the background model. The structural variations of chromosomes were determined according to the copy number distribution of the whole chromosome arm. CNVs were analyzed by CNVkit (dx1.1). We used CSF from the patients with non-tumor diseases as the baseline, and mutations that were identified as false positive were excluded.

Hematoxylin and eosin staining and immunohistochemistry (IHC) were performed on 4 µm-thick sections using standard protocols. IHC stains contains Ki-67 (Roche), and IDH1 R132H (Agilent).

All cases were comprehensively diagnosed and reviewed by the two pathologists (NYL and NW) using histology, as well as molecular features of tumor DNA according to the 2021 WHO CNS classification criteria and the Consortium to Inform Molecular and Practical Approaches to CNS Tumour Taxonomy group in 2018–2020 [2, 13–16].

Statistical analysis and plots were conducted by Graphpad Prism 5 and IBM SPSS statistics software version 19. Quantitative ctDNA differences across different characteristics of patients were assessed using Mann-Whitney U test. Spearman’s rank correlation coefficient test was used to evaluate the association between the tissue DNA and CSF ctDNA with SNVs and CNVs. p-values lower than 0.05 were considered as statistically significant. To assess the possibility of CSF ctDNA test for glioma diagnosis, we performed a receiver operating characteristic (ROC) curve analysis. Consistency of diagnosis was examined by Kappa measure of agreement. The Kappa values were categorized as “almost perfect” (0.8–1.0), “fair to good” (0.4–0.8), or “poor” (below 0.2) [17].

Twenty-seven gliomas were recruited for the CSF and tumor biopsy tissue collection. Histopathological subtypes included glioblastoma (GBM, n = 13), astrocytoma (n = 6), pilocytic astrocytoma (n = 2), pleomorphic xanthoastrocytoma (n = 2), oligodendroglioma (n = 1), ganglioglioma (n = 1), glioneuronal tumor (n = 1), and gliosarcoma (n = 1). The median age of the participants was 54 years (range 15–78 years) and 66.7% (n = 18) were male. These patients were confirmed pathologically and classified into histological grade I (n = 3), grade II (n = 6), grade III (n = 4), grade IV (n = 14). After surgery, 63.0% (n = 19) patients received standard Stupp treatments, including adjuvant radiotherapy and temozolomide (TMZ) chemotherapy, 11.1% (n = 3) patients received adjuvant radiotherapy. The characteristics of these patients were presented in Table 1.

We measured the ctDNA level among 27 CSF samples and found the ctDNA level showed significant differences across age (p = 0.025, Figure 1A), Ki-67 index (p = 0.013, Figure 1B), MRI signal (p = 0.004, Figure 1C), and tumor subtypes (p = 0.001, Figure 1D). There was no significant association of ctDNA levels with other clinical characteristics, including tumor size (p = 0.064), ventricular involvement (p = 0.303), tumor number (p = 0.644) and location (p = 0.762), as well as involvement of cerebral cortex (p = 0.769) (Figures 1E–I).

Twenty-seven patients were all detected using 131 + 4 panel, and 3 of them were excluded on the grounds of their undetectable ctDNA concentration. We identified at least 1 tumor-related gene mutation or chromosome variant from CSF ctDNA of 22 patients, consistent with the number of tumor tissue DNA samples (Figure 2). In contrast, we did not detect any oncogenic alterations in four individuals with encephalitis. Based on the cancer targeted panel, 32 mutant genes were detected in 24 gliomas. Of them, TERT was the most prevalent gene. TERT promoter mutations were detected in 12 (50%) of patients via CSF, while were identified in 16 (66.7%) of patients though tumor tissue. The concordance with the TERT mutation status between CSF and tumor tissue yielded 83.3%. The mutations were detected both in tumor tissue and CSF samples, apart from FLT4 and CIC mutations. Abnormal gene copy number was found in 10 patients for CSF ctDNA. The coincidence between CSF and tumor tissue for CDKN2A/B deletion was 87.5%. However, 8 patients presented with PTEN deletion in tumor DNA but not detected in CSF ctDNA. Alterations of chromosome chr 1p, chr 19q, chr 7, or chr 10 were detected in 11 CSF samples.

We examined the concordance between SNVs and CNVs detected in tumor/CSF pairs. Comparison of CSF ctDNA and tumor DNA showed a moderately preferential correlation in the variant allele frequency (r = 0.64, p < 0.0001; Figure 3A) and a high correlation in the copy number (r = 0.86; p < 0.0001; Figure 3B). The most mutated genes in tumor samples and CSF ctDNA were TERT, TP53, EGFR, PTEN, PIK3CA, and PDGFRA (Figure 3C), which were completely shared in 18 patients, yielding an overall concordance rate of 81.8% between tumor and CSF ctDNA samples (Figure 3D). In addition, 5 out of 24 patients showed a gene concordance of less than 50%. Interestingly, the sensitivity of CSF in glioma was less than the tumor tissue in the detection of +7/−10 and TERT mutations, but presented a superior detection ability for TP53 mutations and EGFR amplification (Figure 3E).

To further explore the performance of CSF ctDNA test, ROC curves for integrated diagnosis of all gliomas were generated (Figure 4). CSF ctDNA performed well in diagnosis of GBMs with the area under the curve (AUC) of 1 (95% CI: 1–1; Figure 4A). We also tested the possibility of CSF ctDNA as a strong classifier of non-GBM diagnosis (AUC = 0.82, 95% CI: 0.65–0.99, Figure 4B). When 3 cases with low ctDNA concentration was excluded, an AUC of 0.91 (95% CI: 0.77–1, Supplementary Figure S1A) was reached. Additionally, for patients with enhanced MRI signal, the AUC value was 1 (95% CI: 1–1; Supplementary Figure S1B).

We established a new integrated diagnostic method by combining histology with tumor DNA and CFS ctDNA as the final diagnosis, which was used as the positive control. The diagnostic diagram of histology, and histology combined with CSF ctDNA and tumor DNA molecular features was revealed in Figure 5. The diagnosis of 5 patients were amended based on the integrated histological and molecular criteria, including 1 GBM and 4 astrocytomas. Among these, the diagnosis of 4 patients was rectified based upon histology and tumor DNA. In addition, there were 3 patients whose diagnosis was corrected depending on the histological and CSF ctDNA data. Kappa measure of agreement was used to determine whether the diagnosis was consistent and found the agreement between CSF and positive control was better. The agreement between CSF ctDNA and positive control was 91.6% (Kappa, 0.800), indicating a high concordance. Meanwhile, the conformity between histology and positive control was 79.1% (Kappa, 0.565), and the consistency in diagnosis between tumor DNA and positive control was 95.8% (Kappa, 0.895). The genotyping consistency of GBM was completely matched for CSF.

We found that CSF ctDNA sequencing could identify the glioma subtyping and determine the patient risk level. IDH1/2 mutation was highly associated with the prolonged overall survival (OS).17 Case 11, a 59 year-old woman, was diagnosed as anaplastic pleomorphic xanthoastrocytoma, WHO Ⅲ, in September 2019, and recurred in March 2021 with histologically confirmed as GBM, WHO IV (Figure 6A). IDH1 was positive at the time of initial diagnosis but negative at the time of recurrence according to immunohistochemical results. Next, tumor DNA sequencing failed to capture the mutation in IDH1, however, CSF ctDNA sequencing showed a rare mutation of R132C in IDH1, and corrected the tumor diagnosis into IDH1-mutated astrocytoma, WHO IV, according to CSF test. This patient experienced recurrence in the left temporal-occipital lobe, with greater possibilities of secondary GBM with IDH1 mutation. The remaining 4 genes were found as the shared alterations.

We also found different mutations between tumor DNA and CSF ctDNA sequencing. Case 13, a 78 year-old man with GBM, +7/−10 and ATM mutation were found in the CSF ctDNA but not in the tumor DNA, at the same time, loss of CDKN2B was only found in tumor DNA (Figure 6B). Three shared somatic mutation genes were identified between the tumor and the CSF, namely TERT, PI3KCB, and CDKN2A (Figure 6C). Importantly, these results suggested that the CSF ctDNA analysis can be invoked as an appropriate diagnostic tool to classify and stratify for some cases.