Early experience with cerebrospinal fluid cell-free DNA molecular profiling in a neuro-oncology practice

Abstract

Purpose

Cerebrospinal fluid (CSF) cell-free DNA (cfDNA) is increasingly used as a liquid biopsy in neuro-oncology for diagnosing lesions of uncertain etiology, distinguishing progression from pseudoprogression, and ruling out leptomeningeal disease (LMD). We assessed our institution’s early experience with a new CSF cfDNA molecular profiling assay for detecting variant alleles (VA) and copy-number alterations (CNA) across these indications in our neuro-oncology practice.

Methods

Beginning in August 2021, providers could request research CSF cfDNA sequencing from clinically indicated lumbar punctures. Following extraction, targeted next-generation sequencing (NGS) and/or low-pass whole-genome sequencing (LPWGS) were performed. VAs and CNAs were summarized by indication and integrated with clinical features, cfDNA yield, matched tissue data, and overall survival.

Results

CSF samples from seventy-six consecutive patients were analyzed and retrospectively classified into three categories: lesions of uncertain etiology (46%), question of CNS progression (21%), or ruling out LMD (26%). Median cfDNA yield was 0.27 ng/mL (IQR = 0.05–0.61 ng/mL). VA and/or CNA detection, hereafter designated “tumor calls,” were identified in 20 patients (26%), with 29 patients having negative tumor calls (38%) and 27 patients (36%) having insufficient cfDNA yield/quality. A copy number burden threshold of 7.42 accurately identified tumor-specific CNAs (AUC 0.998) and was associated with poorer survival (HR 3.4, p = 0.006). Overall survival for patients with positive tumor calls was significantly lower than for patients with negative tumor calls or low yield (p = 0.003).

Conclusion

Despite low cfDNA abundance, CSF NGS and LPWGS provide informative data across neuro-oncologic indications. Further work is underway to optimize the assays for low input CSF cfDNA.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11060-026-05799-7.

Introduction

Central nervous system (CNS) malignancies, including gliomas and CNS metastases, are typically only accessible intra-operatively at the time of biopsy or resection. As such, magnetic resonance imaging (MRI) is utilized for disease diagnosis and monitoring. However, both surgery and imaging have limitations across key neuro-oncologic diagnostic questions. For newly discovered CNS lesions, similar radiographic findings between distinct disease entities, such as high-grade glioma versus primary CNS lymphoma, hampers definitive diagnosis via MRI [1–3]. Tissue sampling is required for definitive diagnosis but is associated with risks including intracranial hemorrhage, stroke, or injury to eloquent brain function [4, 5]. Moreover, tissue biopsies can be non-diagnostic in up to 10–15% of cases [6, 7]. Another area of unmet need in neuro-oncology practice is the delineation of true tumor progression from treatment-related changes, such as radiation necrosis or pseudoprogression [8, 9]. However, the sensitivity and specificity of conventional MRI for detecting disease progression are poor, ranging from 30–70% [10, 11], with perfusion sequences slightly improving accuracy [12]. Finally, diagnosis of leptomeningeal disease (LMD) is similarly challenging due to the low frequency of positive cerebrospinal fluid (CSF) cytology even in patients with significant clinical decline, and the limited role of surgical biopsy in these cases [13, 14].

CSF cell-free DNA (cfDNA) sequencing has been proposed to help address these areas of unmet need across lesions of unknown etiology and question of LMD or CNS tumor progression [15, 16]. Multiple studies have demonstrated 50–60% sensitivity of CSF cfDNA for diagnosis of CNS tumors, including at the time of initial diagnosis and suspected disease progression, most frequently in cohorts consisting largely of patients with gliomas [17–19]. Sensitivity of CSF cfDNA for detection of LMD exceeds 90% in multiple series [14, 20], in part due to enrichment for patients with previously established cytology-positive LMD. While multiple studies have demonstrated the performance of CSF cfDNA for specific clinical indications [14, 17], fewer groups have reported on the real-world experience of deploying cfDNA sequencing in neuro-oncology practices where multiple clinical indications are encountered regularly [18].

To that end, starting in 2021, a new CSF cfDNA sequencing assay was made available at our institution on a research-basis to our neurology, oncology, and neurosurgery departments at provider request for any patient undergoing CSF sampling, most commonly via lumbar punctures. We herein report our real-world, institutional experience with CSF cfDNA sequencing spanning diverse indications, including diagnostic yields and clinical context, for low-pass whole genome sequencing (LP-WGS) and next-generation sequencing (NGS).

Methods

Patient referrals, informed consent, and CSF processing

Beginning in August 2021, providers could request evaluation of cell-free DNA from clinically indicated lumbar punctures under an IRB-approved research protocol at the Mayo Clinic. Patients with a clinical question for whom CSF cfDNA analysis was desired and who were undergoing a lumbar puncture for CSF acquisition were referred to the study by their provider(s). In patients for whom CSF was tested to investigate an unknown diagnosis or to rule out LMD, CSF was obtained prior to any treatment. In those where CSF was sampled for question of progression versus treatment-related changes, CSF was obtained during or after treatment, consistent with the clinical context in which the question arises.

Patients were approached to provide their informed consent to the brain tumor biomarkers study (NCT04692324) which allows for collection of CSF samples via research lumbar punctures or collection of residual CSF from a clinical LP. Patients were also co-enrolled in the Mayo Clinic Neuro-Oncology biorepository (IRB#12-003458) for ongoing sample storage. Per the IRB approval for NCT04692324, patients could provide consent for return of research results from cfDNA testing to them and their requesting providers for interpretation and use at the provider’s clinical discretion. Although testing was performed on a research basis and was not intended or required to guide clinical decision-making, we did not seek to restrict access to potentially informative results when requested by a treating provider, particularly if other diagnostic information was limited and if the clinician interpreted the data in the context of the totality of available clinical evidence and their confidence in the result. The cut-off for inclusion in this report was August 2021 to December 2024. Inclusion criteria for this report included CSF from an LP and, if multiple samples were collected for that patient, evaluation of the first LP result.

After the LP, CSF was maintained on ice or at 4 °C, typically for less than 1–2 h. Once in the laboratory, CSF was then centrifuged at 400xG for 10 min at 4 °C, aliquoted, and frozen on dry ice for shipping to Predicine, Inc. Samples were stored at -80 °C until cfDNA extraction. CSF sample annotations included collection date, sample color, collection time, processing, and freezing times. In patients with known cancers for whom CSF was requested for question of leptomeningeal disease or progression versus treatment-related changes, pathology and treatment details were retrospectively annotated by chart review.

Cell-free DNA extraction

CSF samples were collected from seventy-six patients, with a median volume of 4.6 mL (range: 1.2–7.5 mL, IQR: 3.8–4.7 mL). All CSF cfDNA extraction and sequencing was performed at Predicine, Inc. cfDNA was isolated from CSF using the QIAamp Circulating Nucleic Acid Kit (Qiagen, catalog 55114). Purified cfDNA was assessed for quantity and quality on a Qubit fluorometer (Thermo Fisher Scientific) and a Fragment Analyzer (Agilent Technologies). Both Qubit and Fragment Analyzer measurements were used together to cross-validate cfDNA yield (Supplementary Table 1) as well as quality. Total DNA yield was determined by Qubit, while the cell-free DNA fraction (60–600 bp) was calculated based on size distribution and fragment integrity profiles obtained from the Fragment Analyzer.

Following CSF cfDNA extraction, samples were analyzed via either NGS and/or LPWGS (Fig. 1A) based on cfDNA concentration and quality (LPWGS: 1–5 ng, NGS > 5 ng), with the exception of when rescue NGS or LPWGS was performed at the provider’s request (n = 10 rescue CNV at < 1 ng yield; n = 2 and 3 rescue CNV + NGS at yields < 1 ng and 1–5 ng, respectively). Given that this assay was performed on a research basis, decisions to rescue some samples did not have prespecified criteria and were evaluated on a case-by-case basis when samples were near the cut-off for analysis and may be informative based on their lesion/clinical course. The median CSF volume undergoing cfDNA extraction was 4.6 mL (range: 1.2-7.5mL, IQR 3.8–4.7). The median cfDNA concentration was 0.27 ng/mL (range: 0.0-14.3, IQR 0.04–0.61). Across the seventy-six CSF samples, thirty-seven samples underwent CNV only (49%) and twelve underwent CNV and NGS testing (16%). Of the 76 samples, only two NGS QC failures occurred (3%).

Fig. 1.

Overview of the workflow and cohort. A. Workflow of the deployment of a research CSF cfDNA assay on clinically indicated lumbar punctures in our neuro-oncology practice, beginning from clinician request through cfDNA extraction and subsequent NGS and/or LP-WGS. B. Summary of the complete clinical cohort, including age, sex, survival status as of the end of the study, tissue diagnosis (when performed), and summary of LP-WGS/NGS results (positive = copy number alteration (CNA) and/or variant allele (VA) detected, negative = no CNA or VA detected, or low yield = unable to proceed with NGS/LP-WGS). C (i) Spearman correlations of cell-free DNA mass yield (ng/mL) to CSF volume (mL) for the three tumor call groups. (ii) A Kruskal-Wallis test was performed to evaluate the differences in CSF volume across the three tumor call groups (positive, negative, and low yield). (iii) ANOVA was used to compare the cfDNA concentration (ng/mL) between positive and negative tumor call and low yield groups

Evaluating each clinical indication separately (Fig. 1B), of the 21 patients with query of CNS progression versus treatment-related changes, 2 samples underwent NGS and LPWGS testing (9.5%), 12 underwent LPWGS only (57.1%), and 7 did not proceed to further testing (33.3%). In the question of LMD cohort (n = 20 patients), 2 samples underwent NGS + LPWGS (10.0%), 10 underwent LPWGS only (50.0%), and 8 did not undergo either (40.0%). Finally, in the 35 patients who underwent an LP for a CNS lesion of unknown etiology, 8 patients underwent NGS and LPWGS (22.9%), 15 underwent LPWGS only (42.9%), and 12 did not proceed with testing (34.3%).

Library preparation, next-generation sequencing and low-pass whole genome sequencing

After extraction, up to 30 ng of cfDNA was used for library preparation, beginning with end-repair, A-tailing, and adapter ligation. Adapter-ligated fragments were then PCR-amplified. Amplified libraries were evaluated on a Tape Station (Agilent Technologies), and only those with adequate yield (library DNA ≥ 0.03 µg) proceeded to hybrid capture. Hybrid capture was performed with biotin-labeled DNA probes. Each library was incubated overnight with the Predicine 152-gene targeted panel and Dynabeads M-270 Streptavidin. Unbound DNA was removed by washing, and the captured fragments were re-amplified by PCR. Final enriched libraries were assessed again on a Tape Station and then sequenced on an Illumina NovaSeq 6000 using paired-end 2 × 150-bp kits.

The PredicineCARE hybrid-capture NGS 152-gene panel was used to detect genomic alterations, including single-nucleotide variants, small insertions/deletions, copy-number changes, and gene fusions [21]. Sequencing reads were aligned to hg19, followed by somatic variant calling and filtering according to the company’s established pipeline [21, 22]. Briefly, as previously described [23], adaptor trimming is first performed followed by barcode checking and correction. FASTQ files are aligned to hg19 via the Burrows-Wheeler Aligner alignment tool. Paired-end reads from the same molecules are merged as single-strand fragments to create consensus bam files; single-strand fragments from the same double-stranded DNA are also merged, allowing for suppression of sequencing and PCR errors. Then, to identify variants, the threshold for positive tumor call/VAF was if the VAF was supported by at least three independent DNA fragments, including at least one duplex fragment, and if they met a minimum log-odds (LOD) threshold of ≥ 13 [24]. The LOD metric reflects the log-likelihood ratio comparing observed variant evidence against the reference model, incorporating expected mutation frequencies and false-positive rates.

The PredicineSCORE LP-WGS assay (“CNV” in figures) was used to derive genome-wide copy-number profiles and CNB (copy number burden) scores. The average sequencing depth of was 3X across all samples. Individual 1-Mb segments were first corrected for GC content and mappability, followed by normalization to the mean segment profile of normal reference samples using the ichorCNA algorithm [25]. Briefly, as previously described [26], arm-level CNV deviation was then calculated as the mean normalized CNV value across each chromosome arm. The CNB score was generated by summing the absolute z-scores of these arm-level deviations and applying a log2 transformation. Higher CNB scores reflect greater genome-wide CNV aberration relative to the normal background. CNA plots were manually reviewed by the two co-first authors to evaluate positive or negative tumor calls, blinded to the patient’s clinical course and imaging.

For further details on the analytical validation of the PredicineCARE and PredicineSCORE (CNB) assays, please see the supplementary materials of our previously published manuscript [26], including limits of detection, reportable ranges, associated allele frequencies or copy numbers, accuracy, precision, and analytical specificity.

Data availability

Sample volumes and cfDNA concentration metadata are included in Supplementary Table 1. Raw genomic data.

Results

Overview of patient cohort

Seventy-six consecutive lumbar CSF samples met inclusion criteria between August 2021 and December 2024. At the time of summative analysis, requests for lumbar CSF cfDNA evaluation were retrospectively divided into three different clinical categories, including a neurological lesion of unknown etiology from magnetic resonance imaging (MRI) that had not been biopsied (n = 35 patients, 46% of requests), question of progression versus treatment-related changes on imaging (n = 21 patients, 28% of cases), and ruling-out of LMD based on clinical changes and/or new contrast-enhancement on MRI along the meninges (n = 20 patients, 26% of cases) (Fig. 1A). Cohort demographics are summarized in Table 1 (or Supplementary Table 2 based on tumor call). The CNS progression cohort consisted largely of patients with gliomas (13/21 patients, 61.9%), wherein serial MRIs were insufficient for confirming true disease progression from treatment-related changes. In contrast, most patients in the LMD cohort had known systemic cancer and had developed MRI and/or clinical findings for which LMD was on the differential diagnosis (14/20 patients, 70.0%).

Table 1.

Clinical and radiographic characteristics of the cohort

	[ALL]	CNS progression?	LMD?	Unknown diagnosis	P overall
	N = 76	N = 21	N = 20	N = 35
Age at collection (years; mean, SD)	52.2 (16.4)	49.9 (14.3)	53.8 (15.5)	52.6 (18.3)	0.695
Sex:					0.358
F	39 (51.3%)	10 (47.6%)	13 (65.0%)	16 (45.7%)
M	37 (48.7%)	11 (52.4%)	7 (35.0%)	19 (54.3%)
Survival Status					0.078
Alive	50 (65.8%)	10 (47.6%)	13 (65.0%)	27 (77.1%)
Deceased	26 (34.2%)	11 (52.4%)	7 (35.0%)	8 (22.9%)
Tissue Diagnosis					.
Glioma	23 (30.3%)	13 (61.9%)	2 (10.0%)	8 (22.9%)
Inflammatory	3 (3.95%)	0 (0.00%)	0 (0.00%)	3 (8.57%)
Lymphoma	4 (5.26%)	1 (4.76%)	0 (0.00%)	3 (8.57%)
Metastasis	20 (26.3%)	4 (19.0%)	14 (70.0%)	2 (5.71%)
Non-diagnostic	4 (5.26%)	0 (0.00%)	0 (0.00%)	4 (11.4%)
Other CNS neoplasm	4 (5.26%)	1 (4.76%)	3 (15.0%)	0 (0.00%)
Unknown	18 (23.7%)	2 (9.52%)	1 (5.00%)	15 (42.9%)
PredicineCARE completed? (NGS)					0.365
No	64 (84.2%)	19 (90.5%)	18 (90.0%)	27 (77.1%)
Yes	12 (15.8%)	2 (9.52%)	2 (10.0%)	8 (22.9%)
PredicineSCORE completed? (CNV)					0.886
No	27 (35.5%)	7 (33.3%)	8 (40.0%)	12 (34.3%)
Yes	49 (64.5%)	14 (66.7%)	12 (60.0%)	23 (65.7%)
CNV call					0.957
Low Yield	27 (35.5%)	7 (33.3%)	8 (40.0%)	12 (34.3%)
Negative	34 (44.7%)	9 (42.9%)	8 (40.0%)	17 (48.6%)
Positive	15 (19.7%)	5 (23.8%)	4 (20.0%)	6 (17.1%)
NGS call					0.701
Low yield	64 (84.2%)	19 (90.5%)	18 (90.0%)	27 (77.1%)
Negative	2 (2.63%)	0 (0.00%)	0 (0.00%)	2 (5.71%)
Positive	10 (13.2%)	2 (9.52%)	2 (10.0%)	6 (17.1%)
Tumor call (CNV and/or NGS call)					0.976
Low yield	27 (35.5%)	7 (33.3%)	8 (40.0%)	12 (34.3%)
Negative	29 (38.2%)	9 (42.9%)	7 (35.0%)	13 (37.1%)
Positive	20 (26.3%)	5 (23.8%)	5 (25.0%)	10 (28.6%)
Cytology results					1.000
Negative	48 (87.3%)	7 (87.5%)	16 (88.9%)	25 (86.2%)
Positive	7 (12.7%)	1 (12.5%)	2 (11.1%)	4 (13.8%)
Enhancement on MRI					0.010
No	14 (18.4%)	4 (19.0%)	0 (0.00%)	10 (28.6%)
Yes	62 (81.6%)	17 (81.0%)	20 (100.0%)	25 (71.4%)
LMD suspected on MRI					<0.001
No	50 (65.8%)	16 (76.2%)	4 (20.0%)	30 (85.7%)
Yes	26 (34.2%)	5 (23.8%)	16 (80.0%)	5 (14.3%)
Ventricular contact on MRI					<0.001
No	39 (51.3%)	2 (9.52%)	18 (90.0%)	19 (48.6%)
Yes	37 (48.7%)	19 (90.5%)	2 (10.0%)	16 (45.7%)
Contact with CSF space other than ventricle on MRI					0.054
No	30 (39.5%)	4 (19.0%)	9 (45.0%)	17 (48.6%)
Yes	46 (60.5%)	17 (81.0%)	11 (45.0%)	18 (51.4%)
Maximum tumor dimension (mm; mean, SD)	25.4 (20.1)	35.7 (28.1)	16.8 (9.93)	23.6 (15.4)	0.012
Prior systemic therapy?					0.001
No	28 (36.8%)	2 (9.52%)	6 (30.0%)	20 (57.1%)
Yes	48 (63.2%)	19 (90.5%)	14 (70.0%)	15 (42.9%)
Prior radiation therapy?					<0.001
None	41 (53.9%)	3 (14.3%)	6 (30.0%)	32 (91.4%)
Yes	35 (46.1%)	18 (85.7%)	14 (70.0%)	3 (8.57%)

Twenty of the thirty-five patients with unknown diagnosis at the time of CSF collection subsequently had tissue results (57%). In the LMD cohort, cytology had been obtained for 18/20 patients (90%) and was negative in 16 of the 18 patients (89%), consistent with the known poor sensitivity of cytology for LMD detection [13]. Forty eight of the 76 patients (63.2%) previously received systemic therapy, such as but not limited to chemotherapies or immunotherapies, and 35 (46.1%) had received radiation therapy, such as but not limited to intensity modulated radiation therapy or stereotactic radiosurgery. Radiographically, 62 of the 76 patients had contrast-enhancement on MRI (81.6%). 48.7% of patients had a lesion in contact with ventricular CSF, and 60.5% patients had a lesion(s) with contact to a CSF space other than a ventricle.

Detecting tumor calls across clinical indications

Tumor calls were defined as a detectable variant allele and/or copy number variant (“positive” tumor call); samples that could not proceed with further testing were designated as “low yield.” Across all three indications, 20 had a positive tumor call (26.3%), 29 had a negative tumor call (38.2%), and 27 were low yield (35.5%). Evaluating based on each indication, in the query of CNS progression versus treatment-related changes group, 5 samples had a positive tumor call (23.8%), 9 had a negative tumor call (42.9%), and 7 were low-yield (33.3%). In the query LMD group, 5 samples had positive tumor calls (25.0%), 7 had negative tumor calls (35.0%), and 8 were low yield (40.0%). Finally, in the unknown diagnosis group, 10 samples had a positive tumor call (27.0%), 13 had negative tumor calls (37.1%), and 12 were low yield (34.3%). In summary, the distribution of positive versus negative tumor calls and low yield samples were relatively similar across indications.

Of note, it is known that cfDNA yield is CNS tumors is highly heterogeneous, in part but not fully explained by tumor-intrinsic factors like tumor-CSF contact [27]. Consistent with this heterogeneity, increasing CSF volume did not correlate with cfDNA mass yield (Spearman R = 0.12, p = 0.32, Fig. 1Ci), nor was there a statistically significant difference in the volumes across CSFs with tumor call (Kruskal-Wallis p = 0.80; Fig. 1Cii). CSF samples with positive tumor calls had higher cfDNA concentrations relative to those with negative tumor calls or low yields (median yield = 0.91 ng/mL, vs. 0.30 ng/mL and 0.00 ng/mL respectively; Fig. 1Ciii) with no significant effect noted for CSF sample volume on ANOVA (p = 0.253). Thus, VA or CNV detection was associated with cfDNA concentration rather than CSF volumes.

Evaluating chromosomal alterations and variant alleles

Having evaluated overall positive versus negative tumor calls across samples, we then turned to determining the specific CNAs and VAs detected via LPWGS or NGS in the positive tumor call samples. Via LPWGS, detected chromosomal alterations were diverse across samples (Fig. 2A), but included 1p/19q co-deletion in a patient with a grade 3 oligodendroglioma (e.g., patient 35) as well as gain of chromosome 7 and loss of chromosome 10 in patients with known glioblastoma (e.g., patient 285), consistent with molecular diagnostic criteria for those tumors [28]. Overall, higher chromosomal complexity was identified in the LMD and CNS progression cohorts than in the unknown diagnosis cohort. We then sought to identify a CNB threshold indicative of positive CNA detection. A CNB score of 7.42 was determined post-hoc to optimally separate CSF samples within this cohort based on manually calling positive CNA calls from those with negative CNA calls across testing indication subcohorts (Fig. 2Bi), with an AUC of 0.998 (Fig. 2Bii). Indeed, patients with a CNB score ≥ 7.42 had significantly worse overall survival compared to those with CNB < 7.42 (log-rank test, p = 0.003), even after controlling for query indication (HR: 3.37, 95% confidence interval: 1.41–8.01, p = 0.006; Fig. 2Biii-iv).

Fig. 2.

Copy number variant landscape in CSF cfDNA. (A) Oncoprint of copy number variants in samples that underwent LP-WGS, with correlative clinical annotations per Fig. 1B. (B)(i) Density plots showing the distribution of CNB scores for samples with negative (pink) and positive (blue) CNA calls, stratified by clinical query. The dashed vertical line indicates the CNB score threshold used for classification. Separation of the density curves illustrates the relative distribution of CNB scores between negative and positive tumor calls within each clinical context. (ii) ROC analysis demonstrating discrimination of tumor-positive versus tumor-negative samples (AUC = 0.998). (iii) Kaplan–Meier curves for overall survival stratified by CNB score (< 7.42 vs. ≥ 7.42). (iv) Multivariable Cox proportional hazards model evaluating the association of CNB score ≥ 7.42 with survival, as compared to comparing the associated of query LMD or unknown diagnosis against CNS progression. Hazard ratios with 95% confidence intervals are shown

Next, we evaluated the somatic mutation landscape across the 12 CSF samples that underwent NGS, 10 of which had detectable variant alleles (Fig. 3). Similar to CNVs, the landscape across this diverse patient cohort was highly heterogeneous, including detection of relevant tumor-associated mutations such as TP53, APC, TERT, BRAF, IDH1, and MYD88, among others. Most patients had multiple mutations detected in their CSF via NGS, with the distribution of VAs varying substantially across patients and mutations (Supplementary Fig. 1). Three of the patients had clinical NGS panels performed on surgically acquired tissue. After filtering to identify genes present in common across all NGS panels, each of the patients had mutations private to CSF (Supplementary Fig. 2). In two of the three patients, a total of 5/11 (45.4%) and 10/11 (90.9%) were shared between tissue and cfDNA. The third patient did not have tissue mutations common across NGS panels. In summary, CSF cfDNA NGS and LPWGS revealed a chromosomally and mutationally diverse genomic landscape, consistent with the diversity of indications and pathologies present in our cohort.

Fig. 3.

Variant allele landscape in CSF cfDNA. Oncoprint of variant alleles in samples that underwent NGS with variant alleles detected, with correlative clinical annotations per Fig. 1B

Tumor calls and overall survival

Prior studies in medulloblastoma and LMD have shown that increased CSF cfDNA burden is a negative prognostic indicator [29, 30], perhaps due to increased cell turnover into CSF when there is sufficient tumor-CSF contact, suggesting that a negative cfDNA result could still have clinical utility. To that end, we sought to test the hypothesis that patients with positive tumor calls (CNA and/or VA detection) in CSF cfDNA have a decreased overall survival. Across all reasons for CSF analysis, patients with positive tumor calls had significantly poorer survival than patients with negative tumor calls or low yield CSF (log-rank p = 0.003) (Fig. 4A). When evaluating at the cohort-level, this finding was identified in the unknown diagnosis group (log rank p = 0.024) (Fig. 4B), with a similar but non-significant trend in patients for whom CSF was obtained to rule out LMD (p = 0.083) (Fig. 4C). In the CNS progression versus treatment-related changes group, there was no difference in overall survival between patients with positive tumor calls versus those with negative tumor calls or low cfDNA yield in any of the patient cohorts evaluated (Fig. 4D). This finding is likely indicative of diverse amounts of recurrent tumor burden in relation to treatment-related changes within this patient cohort.

Fig. 4.

Survival and tumor call. Kaplan-Meier curve estimates of overall survival probability in the positive tumor call, negative tumor call, and low cfDNA yield groups in the (A) entire cohort, as well as the three separate indications for CSF cfDNA testing, (B) lesion of unknown etiology, (C) ruling out LMD, and (D) querying progression versus treatment-related changes. Hatch marks indicate data censoring

Representative patients

To illustrate case examples of CSF cfDNA testing across various indications, we highlight three representative cases where positive CSF cfDNA would have been diagnostically useful in each of the three clinical indications. First, for the clinical query of lesions of unknown etiology, patient 315 was a 72 year old male with a recent past medical history of untreated chronic lymphocytic leukemia who presented to the emergency department with subacute onset of confusion and generalized weakness. Magnetic resonance imaging revealed multiple diffuse areas of T2 hyperintensity involving the bilateral cerebral hemispheres, cerebellum, and brainstem (Fig. 5A). First, a lumbar puncture was obtained that was positive for the patient’s known CD5 + B cell lymphoproliferative disorder including monotypic kappa B-cells expressing CD19, CD20, CD22, CD5 (dim), CD23 (partial), CD38, and CD200 that matched their CLL profile from peripheral blood obtained the year prior. CSF was also positive for HHV6 and EBV. Given the rarity of CNS involvement with CLL and concern for viral encephalitis, the patient subsequently underwent left parietal needle biopsy that demonstrates CD3 + T cells, scattered CD20 dimly positive B cells, and co-expression of PAX5, CD5, and weak CD23 on B-cells, demonstrating CLL involvement. At the time of the LP, CSF cfDNA had also been requested which demonstrated amplification of chromosome 12, consistent with known Trisomy 12 CLL, in addition to multiple KRAS and BRAF mutations that are known to be enriched in trisomy 12 CLL [31]. After the biopsy and multi-disciplinary discussions, the patient was initiated on Zanubrutinib, but his hospital course was complicated by GI bleeding and respiratory infections. He subsequently passed away approximately two months after the LP.

Fig. 5.

Patient examples across clinical indications. Within each clinical indication of (A) lesion of unknown etiology, (B) ruling out LMD, and (C) querying progression versus treatment-related changes, the following are shown for each representative patient: imaging at the time of CSF sampling, copy number plot, and detected variant alleles and their frequency. * indicates a variant of unknown significance. Imaging includes (A) axial T2-FLAIR Brain MRI for patient 315, (B) axial T1-post gadolinium MRI with yellow arrow pointing to disseminated disease for patient 440, and (C) axial T1-post gadolinium MRI in patient 35

As an example of LMD detection via CSF cfDNA, patient 440 was a 38 year old male with known metastatic adenocarcinoma of the gastroesophageal junction, treated in the past with FOLFOX, trastuzumab, pembrolizumab, FOLFIRI, nivolumab, and trastuzumab deruxtecan, as well as stereotactic radiosurgery to a right cerebellar metastasis (Fig. 5B). In oncology follow-up, the patient presented with continued general deconditioning and new multifocal sites of suspected leptomeningeal metastases from the cervical to lumbar spine. Decision at the time was to evaluate CNS response to fam-trastuzumab deruxtecan. A month later, following mild interval response to therapy, a lumbar puncture was performed that revealed an extensive copy number burden as well as mutations in TP53, NTRK1, and EGFR, consistent with suspected LMD. Due to continued mixed response to therapy, the patient subsequently underwent craniospinal irradiation and passed away two months thereafter.

Finally, patient 35 was a 51 year old female who presented for surveillance of a known grade 3 oligodendroglioma, initially diagnosed and resected fourteen years prior, after which she underwent treatment with multiple rounds of temozolomide, radiation, and two repeat resections (Fig. 5C). At the time of presentation, she had completed her sixth cycle of TMZ. While imaging was temporarily stable, enhancing ependymal nodules were under observation and CSF evaluation was requested out of caution due to multiple prior recurrences. CSF testing demonstrated co-deletion of chromosomes 1p and 19q, as well as the IDH1 R132H mutation, consistent with her known grade 3 oligodendroglioma. VAs were also detected in TERT, NOTCH1, and SDHB, among others, that were consistent with her tissue genomics. Two months later, MRI demonstrated significant enlargement of multiple non-enhancing ependymal nodules in the left and right lateral ventricles as well as the fourth ventricle, in addition to a new enhancing right pericaudate nodule. The patient was subsequently enrolled on an IDH inhibitor trial at an outside institution. In sum, CSF cfDNA testing demonstrated positive concordance with clinical status in these patients within the three testing indications.

Discussion

As brain tumors are frequently only accessible through invasive neurosurgical procedures, cfDNA-based liquid biopsy assays are of interest for disease diagnosis and monitoring. In contrast to non-CNS cancers, the utility of plasma cfDNA detection and sequencing for VAs or CNAs in patients with brain tumors is limited by the blood-brain barrier (BBB) [32], which limits analytes from diffusing out of the central nervous system (CNS) into systemic circulation. Moreover, in patients with known systemic disease, positive plasma cfDNA is unhelpful to assess for CNS dissemination. As such, cerebrospinal fluid (CSF) cfDNA sequencing has gained increasing traction due to its improved proximity to the tumor relative to plasma [33, 34]. Within the literature and clinical practice, there now exists a broadening set of indications for CSF cfDNA sequencing¹ [14, 15, 17], including diagnosis of lesions of unknown etiology, delineation of progression from treatment-related changes, and evaluation of leptomeningeal disease. While multiple studies have reported on the performance of CSF cfDNA for each specific indication [14, 15, 17, 18], few have evaluated the implementation of CSF analyses in real-world neuro-oncology settings with diverse patient populations [18].

In this context, our experience reported herein demonstrates the ability to deploy CSF cfDNA sequencing within our diverse neuro-oncology practice with low rates of NGS QC failure despite the low cfDNA abundance of the lumbar CSF compartment. Within our series, sufficient CSF was available for analysis in 64.4%, with positive tumor calls, including VA and/or CNAs, were made in 40.8% of these analyzable patients, comprising 26.3% of the total cohort. Recent series from Hickman et al. [18] and Iser et al. [35] reported mutational detection rates of 53% and 67% respectively from lumbar CSF in patients with gliomas and other brain tumors. Lower detection rates in our study as compared to that of prior reports likely reflect the broader referral base in this study, wherein providers frequently requested cfDNA sequencing for patients with substantial diagnostic uncertainty, including lesions of unknown etiology and ruling out of LMD. Moreover, further technical improvements are needed and are underway to optimize this assay to the lower cfDNA yield in these CSF samples. CSF yielded insufficient DNA for analysis in 36% of patients and revealed no evidence of tumor despite adequate cfDNA abundance in 38% of cases. Importantly, however, low abundance of cfDNA and negative tumor calls was a significantly positive prognostic indicator based on overall survival, namely in the unknown diagnosis or ruling out of LMD cohorts, consistent with non-CNS tumor diagnoses. Prior studies in neuro-oncology and systemic cancers have demonstrated that cfDNA burden is a negative prognostic indicator [17, 36, 37]. However, one must also consider the location of the lesion in question relative to the CSF space sampled, wherein fully parenchymal disease with minimal CSF contact is highly unlikely to be positive via cfDNA testing [27]. In contrast, in patients with parenchymal brain metastases with tumor-CSF contact, positive cfDNA yield may not indicate ongoing LMD.

Across all studies, including our own, CSF samples with detectable VAs or CNAs consistently exhibit higher cfDNA concentrations than samples with no calls or low yields. Factors such as tumor–CSF contact have been shown to influence overall cfDNA yield and may help identify patients in whom cfDNA testing is more likely to succeed [15, 17, 35]. However, even after accounting for such variables, substantial inter-patient variability in cfDNA release persists and cannot be fully mitigated by simply increasing input volume. Although total cfDNA yield in CSF is generally lower than in plasma, multiple studies have shown that it is markedly more enriched for tumor-derived material in patients with CNS malignancies [33, 38]. Accordingly, continued development of low-input sequencing platforms is poised to enhance the diagnostic utility of CSF in neuro-oncology [39].

In patients with positive tumor calls, VAs and CNAs revealed a diverse genomic landscape that was captured via CSF cfDNA. Such VAs included disease-defining mutations like MYD88 in a patient with PCNSL as well as targetable mutations like IDH and EGFR. Clinical sequencing of tumor tissue was unfortunately infrequently performed during the duration of this study for correlation of CSF to tissue genomics and was thus only available for comparison in three patients, demonstrating shared mutations between tissue and CSF. However, prior studies have identified instances of private mutations in CSF that were not detectable in CSF and vice versa, including our own longitudinal series [26, 40, 41]. Given the spatial heterogeneity of intracranial tumors [42, 43], such private alterations in CSF may occur due to a different region of a heterogeneous tumor having contact with the sampled CSF space as compared to the region sampled by tissue biopsy. Alternatively, in patients for whom CSF sampling occurs at a different timepoint than tissue sequencing, such as when querying LMD or CNS progression, these differences could reflect ongoing tumor evolution, particularly if patients underwent radiation or chemotherapeutic treatment between sampling timepoints. In patients with systemic disease that has disseminated to the CNS, including LMD, such alterations can be indicative of genomic features facilitating survival within the CNS microenvironment [44, 45]. Within neuro-oncology practice, CSF cfDNA may thus serve as a complementary source of genomic insights to help assess disease evolution and therapeutic response.

While we evaluated single timepoint samples for queries of potential CNS progression and LMD, longitudinal samples are more optimal to evaluate disease trajectory, wherein increasing VAF or CNB could indicate increasing disease burden. Indeed, our prior study in cranial CSF has demonstrated the initial feasibility of correlating longitudinal VAs and CNAs with changes in disease burden on a single-patient basis [26]. Whether lumbar CSF could have sufficient sensitivity for disease monitoring in real-world neuro-oncology practices remains to be evaluated, and would likely require high sensitivity platforms suitable for low cfDNA yields. Moreover, repeated lumbar punctures are less convenient for patients than CSF access devices and may have lower sensitivity than intracranial CSF sampling depending on the location of the lesion relative to CSF, although some studies have reported similar genomic profiles in lumbar and cranial CSF [17]. However, these pitfalls are countered by improved availability when needed, since few patients have a pre-existing cranial CSF access devices-placement of which carries inherent risks, including infection [46]. Importantly, cfDNA posterior fossa and spinal disease may be better reflected in CSF from the lumbar cistern than lateral ventricle due to directionality of CSF flow [27]. Indeed, prior work in children with medulloblastoma demonstrated the ability to monitor MRD and identify disease progression and longitudinal tumor evolution using lumbar CSF [30]. Future studies will be needed to evaluate the monitoring performance of lumbar CSF in patients with detectable VAs and/or CNAs across other diagnoses. An additional remaining question is in what circumstances results derived from lumbar CSF could reduce the need for tissue biopsies in patients with lesions of unknown etiology or in suspected disease progression. Histopathologic confirmation remains the gold standard across most settings, though CSF cfDNA may enhance management in cases where tissue diagnosis is not safely achievable, spanning lymphoma after steroids, LMD, lesions not safely reachable by surgery, complementing monitoring opportunities. As CSF cfDNA testing becomes more available across institutions, additional experience will be gained in real-world neuro-oncology practices to better define if and when CSF results could obviate the need for more invasive tissue sampling.

This study had several limitations. Across all three clinical indications, identifying a robust standard against which to compare CSF cfDNA sequencing results is challenging. For patients who do not undergo a biopsy for a lesion of unknown etiology, which represented a large portion of the cohort, we are limited to the subsequent survival of those patients. In the case of unknown diagnosis, these patients often had prolonged survivals (Fig. 4), suggesting that low yield or negative tumor calls were truly likely negative in this cohort. Nevertheless, tissue biopsies to delineate progression from pseudoprogression are increasingly endorsed in neuro-oncology [47], but were not yet frequently performed at our institution when these CSF samples were acquired. As such, tissue was often not available for correlation to CSF tumor calls, including for correlation of tissue versus CSF CNAs to support the use of the CNB = 7.42 cut-off. Additionally, this cut-off was determined in a post-hoc unblinded fashion without an independent cohort. Moreover, for patients with LMD, cytology is often viewed as the gold standard for diagnosis, but has poor sensitivity [13, 14]. Accordingly, cytology was infrequently positive when it was obtained, limiting correlations to positive tumor calls. Given such limitations, we utilized survival outcomes as indirect corroboration for cfDNA findings, based on the well-established prognostic implications of LMD, in similar manner to diagnosis of CNS malignancy in the case of lesions of unknown etiology. Cases evaluated for possible progression versus pseudoprogression changes likely represent a heterogeneous mixture in which varying proportions of recurrent tumor coexist with treatment-related changes, likely explaining minimal differences in survival between patients with versus without positive tumor calls. Finally, tissue sequencing was infrequently performed clinically at our institution at the time of this study for this cohort, including for both our patients with gliomas and brain metastases, limiting comparison of CSF VAFs to matched tissue. This is especially relevant for low-frequency or missense variants without biological associations to CNS malignancies, as these could represent sequencing artifacts. Such variants should thus be interpreted cautiously and in the context of the patient’s overall clinical course. These limitations underscore the need for continued comparison of CSF-based assays against both retrospective clinical trajectory review, and, when feasible, gold standards such as tissue analysis to more definitively resolve diagnostic ambiguities and support the integration CSF cfDNA assays into routine neuro-oncology practice.

In conclusion, CSF cfDNA analysis can be deployed across multiple clinical indications in real-world neuro-oncology practice, including evaluation of lesions of uncertain etiology, ruling out of leptomeningeal disease, and querying true disease progression versus pseudoprogression. Since the assay was first implemented for CSF testing at our institution, requests for CSF analysis have increased substantially over time (Supplementary Fig. 3), reflecting growing clinical interest in this approach. Efforts are currently underway to further develop and refine a low-input CSF cfDNA pipeline. Although positive tumor calls, including both variant alterations and copy number alterations, were identified in approximately one quarter of patients in this cohort, CSF cfDNA testing may also provide clinical value in certain negative cases when interpreted within the broader clinical context and disease trajectory, particularly when evaluating an unknown diagnosis or ruling out leptomeningeal disease.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

CRC, TCB, SJ, and PD conceptualized the project. CRC, AEW, JR, DOY, TCB, JL, CD, and YH performed sample/data acquisition and processing. RK, CRC, CD, and YH performed data analysis. CRC, JL, TCB performed protocol and project administration. CRC, RK, and TCB wrote the original draft. All authors reviewed the manuscript.

Funding

CRC was supported by the National Institute of Health T32GM145408. TCB and AEW were supported by NINDS R61 NS122096 and NCI R37CA276851. TCB was also supported by Mayo Clinic Center for Individualized Medicine and CCaTS award UL1TR002377, Humor to Fight the Tumor, the Brain Cancer Research Alliance, and Lucius & Terrie McKelvey.

Data availability

Sample volumes and cfDNA concentration metadata are included in Supplementary Table 1. Raw genomic data are under deposition at GEO.

Declarations

Ethical approval

The Mayo Clinic Institutional Review Board approved all studies included in this manuscript. All portions of the studies were performed according to the Declaration of Helsinki. All patients provided written informed consent for use of their CSF and electronic medical record data, in addition to providing consent for publishing.

Competing interests

CD, JL, YH, XW, and PD report employment with Predicine, Inc. SJ reports employment with Predicine, Inc. and ownership of Predicine, Inc. stock.