MRD4U: A path to development for personalized liquid biopsy for children with central nervous system tumors

Abstract

Background

Liquid biopsy assays using cerebrospinal fluid (CSF) can revolutionize care for children with central nervous system (CNS) tumors by enabling precise monitoring of therapeutic responses and detecting recurrence or measurable residual disease (MRD). These assays can detect cell-free, circulating tumor DNA (ctDNA) via somatic alterations, though accurately measuring low-abundance ctDNA in CSF is challenging.

Methods

Our research focused on the optimization of next-generation sequencing library preparation from cell-free DNA (cfDNA), evaluating four commercial kits to address the low nucleic acid yield in CSF-derived cfDNA. The selected kit minimized false positives and detected somatic variants at 5% variant allele frequency using 0.1 ng input of synthetic cfDNA, suitable for low-volume CSF samples.

Results

We then applied our optimized workflow to six children with CNS tumors using a personalized hybrid-capture sequencing strategy (“MRD4U”), in which individualized panels were designed based on each patient’s tumor sequencing. Using MRD4U, we identified ctDNA in two samples, even though neither patient had radiographic or clinical evidence of disease at the time of liquid biopsy. Notably, one ctDNA-positive patient developed radiographic recurrence four months later, demonstrating the assay’s potential to detect molecular relapse ahead of conventional clinical measures.

Conclusions

These findings demonstrate applicability of our personalized MRD4U assay in early detection of disease recurrence. Unlike non-targeted or tumor-agnostic CSF liquid biopsy approaches, MRD4U leverages patient-specific genomic information to enable sensitive, tumor-informed monitoring that can be deployed across a wide range of pediatric CNS tumors. Our approach is broadly applicable to any tumor type with existing genomic data, enabling ctDNA detection across diverse diagnoses. Ultimately, this strategy may inform clinical decision-making and enable earlier therapeutic intervention.

Background

In recent years liquid biopsy-based testing, particularly utilizing next-generation sequencing (NGS) of cell-free DNA (cfDNA) from plasma to identify circulating tumor DNA (ctDNA), has gained traction in research studies and clinical trials [1, 2]. cfDNA consists of DNA fragments circulating freely in bodily fluids like blood, urine, or cerebrospinal fluid (CSF), released through processes such as cell death [3]. A specific subset of cfDNA originating from tumor cells is termed ctDNA, the detection of which allows for non-invasive monitoring of disease status. The appeal of liquid biopsy lies in its advantages over traditional tissue biopsies: it is less invasive, more cost-effective, and more sensitive than traditional imaging techniques [4, 5]. Initially, the use of liquid biopsy focused on treatment decisions for various cancers including non-small cell lung, prostate, ovarian, and breast cancer, enabling tailored treatments based on specific genetic alterations detected via ctDNA analysis [6,7,8]. Historically, most liquid biopsy studies have focused on common diagnoses in adults and have used blood as the sole source of biofluid [9,10,11,12,13,14,15]. Furthermore, in adult studies, liquid biopsy assays are designed to identify ctDNA by leveraging the fact that many types of adult tumors share commonly mutated genes. For example, in colorectal cancer ~ 80% of tumors have a mutation in the gene APC and in acute myeloid leukemia, > 75% of tumors have mutations in one of three genes: FLT3, DNMT3A, or NPM1 [16, 17]. Therefore, assay designs for detecting mutations in ctDNA in common adult cancers are usually straightforward, employing targeted sequencing panels capable of identifying mutations across specific recurring gene subsets.

While liquid biopsy has revolutionized cancer screening, diagnosis, therapy selection, and monitoring response to treatment via detection of minimal or measurable residual disease (MRD) in certain adult tumors, its application specifically in pediatric central nervous system (CNS) tumors presents intriguing challenges. For example, children with CNS tumors typically present with an intact blood brain barrier which prevents tumor shedding into the periphery, precluding the use of blood-based liquid biopsies [18]. Rather, CSF liquid biopsies are a more appealing route since many tumors anatomically abut ventricles which are reservoirs to produce CSF, a biofluid which can contain cfDNA depending on disease status [19,20,21,22]. CSF presents additional challenges for this type of testing; most notably, the yield of extracted cfDNA is relatively low, making NGS library preparation challenging [19, 23]. In addition, pediatric CNS tumors exhibit a diverse genomic landscape, with few common gene mutations shared across the many disease entities [24, 25]. While some liquid biopsy studies in pediatric brain tumors have explored general approaches such as genome sequencing of CSF cfDNA to detect copy number alterations, these methods have limited sensitivity in detecting low levels of MRD [23, 26,27,28]. Other brain tumor liquid biopsy studies have used CSF cfDNA to perform targeted sequencing with either a static panel or exome approach and have successfully demonstrated feasibility and clinical utility [27, 29,30,31]. However, static panels often lack generalizability due to the diverse mutational landscapes across different brain tumor types, while exome sequencing, although broadly applicable, demands extremely high sequencing depth and is thus costly. Our personalized approach, as described below, balances these factors by utilizing smaller, patient-specific panels, enabling sensitive ctDNA detection from CSF in a more targeted, cost-effective manner.

Here, we describe the development of a liquid biopsy assay for pediatric CNS tumors by first identifying the most suitable methodology for maximizing assay sensitivity with low input cfDNA. In light of the recent recommendations set forth by the Association for Molecular Pathology and College of American Pathologists Liquid Biopsy Working Group for cfDNA assay validations, we report on several key pre-analytical considerations and assay performance metrics including a thorough evaluation of various NGS library preparation kits, the use of appropriate controls, assessment of decreasing levels of input cfDNA, and determination of sensitivity and limit of detection for our assay [32]. Our progression towards implementing this type of workflow in the pediatric CNS tumor patient population began with foundational laboratory tasks aimed at investigating cfDNA extraction from 44 CSF samples and assessing feasibility of detecting ctDNA in the cell-free extracts using a simple targeted amplicon sequencing approach. These preliminary investigations laid the groundwork for our subsequent endeavors. Focusing then on a unified library preparation protocol, we adopted a two-pronged approach to detect ctDNA in CSF in real-time from pediatric patients with CNS tumors as a surrogate for measuring MRD. First, we used a tumor-informed approach to design personalized hybridization capture panels (MRD4U; “MRD for you”) with known sequence variants identified previously through tumor exome sequencing. Second, low-pass genome sequencing (lpGS) allowed us to identify specific copy number variations (CNVs) known from prior tumor sequencing. Applying our approach to a cohort of six children with embryonal central nervous system tumors, we detected tumor-derived DNA in the CSF (“ctDNA+”) from two patients. Both ctDNA + children had negative findings on magnetic resonance imaging (MRI), a reported absence of malignant cells based on clinical cytological analysis at the time of our liquid biopsy assay, and notably– one of them developed radiographic disease recurrence months later, thus highlighting the increased sensitivity and early detection capabilities of our assay.

Methods

Specimen collection and processing

Written informed consent was obtained from the patient and/or legal guardians under a research study approved by the Institutional Review Board (IRB) at Nationwide Children’s Hospital (STUDY00003080). All research presented in this study was performed in accordance with relevant guidelines and regulations as set forth by the IRB at Nationwide Children’s Hospital. The volume of cerebrospinal fluid (CSF) collected was determined by the physician based on factors such as patient age, weight, and the quantity required for standard-of-care clinical laboratory analyses. CSF was collected into sterile vials or syringes via one of the following procedures: lumbar puncture, external ventricular drain (EVD), or ventriculoperitoneal (VP) shunt placement (Supplementary Table S2). The CSF samples were first transported to a clinical laboratory for standard-of-care assessment including cytological analysis. Residual CSF samples were then transferred to appropriately sized conical tubes compatible with an Eppendorf 5810R centrifuge and centrifuged at 2000x g for 5 min at 4 °C. Following centrifugation, the CSF supernatant (CSF-SN) was carefully aspirated using a sterile serological pipette without disturbing any cell pellet present, transferred to 2.0 mL cryovials, and stored at −80 °C for varying lengths of time until needed for extraction (Table 1). Prior to cell-free DNA (cfDNA) extraction from CSF-SN, the samples were thawed and transferred to a 1.5 mL Eppendorf tube and centrifuged at > 12,000 g for 15 min. The CSF-SN was aspirated for cfDNA extraction. In addition, we also requested a blood sample to be taken at the same time as CSF collection, but only when there was a clinical need for a blood draw. For these samples, blood was collected in a purple top EDTA tube. Blood was combined into an appropriate-sized conical tube compatible with an Eppendorf 5810R and centrifuged at 1,200 x g for 10 min at 15–20℃. Plasma was separated from red cells using a sterile serological pipette and transferred to a labeled 15 mL or 50 mL conical tube. The conical tube containing plasma was re-centrifuged at maximum speed for 15 min at 4℃. After centrifugation, any visible fat layer on the surface of the plasma was removed using a sterile wooden stick or serological pipette. The plasma was aliquoted into a 2.0 mL cryovial, with 0.5-1 mL of plasma, and stored at −80℃. Both CSF-SN and plasma were used for cfDNA extraction using Quick-cfDNA/cfRNA Serum and Plasma kit (Zymo Research; catalog No. R1072) following the manufacturer’s protocol with the following modification: for maximum recovery, the initial cfDNA eluate was dispensed back onto the same spin column filter followed by incubation for 2 min and a second centrifugation at 12,000 x g for 30 s. For the extraction process, 1 mL of AcroMetrix TM Multi-Analyte ctDNA Plasma Control (AcroMetrix, catalog No. 957563) was utilized as a positive control to ensure the effectiveness of the DNA extraction. cfDNA quantity and quality were evaluated with the Qubit 1X dsDNA HS assay (Invitrogen; catalog No. Q33231) and Bioanalyzer High Sensitivity DNA kit (Agilent; catalog No. 5067 − 4626). Based on the Qubit dsDNA HS assay specifications, the lower limit of detection is 0.1 ng. Samples with DNA concentrations below this threshold were recorded as 0.0 ng when calculating yield per mL of extracted CSF.

Table 1 The status of treatment during biofluid collection, the collection methodology employed, the volume of fluid subjected to extraction, and the resulting yield

Targeted DNA sequencing

Targeted amplicon deep sequencing (“Libricon”) followed our previously described protocol optimized for genomic DNA [33]. To evaluate the effectiveness of this amplicon approach using cfDNA as input material, we conducted serial dilutions of the Pan-Cancer cfDNA reference standards v1 (Twist BioScience; catalog No. 104549). These dilutions varied the input amounts from 10 ng, 5 ng, 1 ng, 0.5 ng, and 0.25 ng for each of the reference standard variant allele frequencies (VAFs) controls: 5%, 2%, 1%, 0.5%, 0.25%, 0.1%, and 0%. Our initial assessment used primers designed to target the BRAF p.V600E variant (Supplementary Table S5) known to be present in the Pan-Cancer cfDNA reference standards at the aforementioned VAFs. For deep targeted sequencing of patient-specific variants in CSF cfDNA we also, when available, used the patient’s tumor DNA as a positive control for detecting the targeted tumor variant. Additionally, we included human reference DNA GM24143 (Horizon Discovery; catalog No. GM24143) as a negative control. All controls used 10 ng of input, if available from extractions, into the targeted amplicon workflow. To identify patient-specific somatic and germline variants, we leveraged the patient’s prior tumor and comparator blood exome sequencing data. We selected a patient germline target that was a known pathogenic variant and/or absent from the gnomAD database [34]; for tumor variants, we chose the highest VAF coding or splicing variant based on the prior exome sequencing data and designed primers to amplify this region (Supplementary Table S5). The entire CSF cfDNA extract was used as input into the Libricon workflow, splitting into two separate PCR reactions, one targeting the somatic tumor variant and the other targeting the germline variant. The final libraries were sequenced on an Illumina iSeq100 platform using paired-end 150 bp chemistry aiming for > 10,000x coverage. The resulting data were analyzed following the procedure described in our ‘Variant Calling’ method section. Outliers for the targeted BRAF assessment were identified using the Grubbs’ test for outliers and subsequently removed. For the patient Libricon data displayed in Fig. 2, only samples in which we successfully identified the patient germline variant in CSF cfDNA at > 25% VAF were included for plotting and analysis, and all other results were considered “inconclusive.”

Comparative NGS assay development: workflow and custom hybridization probe panel design

Custom probe panel design for assay development using reference standards

A custom hybridization probe panel was designed using the Integrated DNA Technologies (IDT) xGen™ MRD Hybridization Panel portal. The panel was designed to target known variants (25 SNV and 24 indel variants representing 39 unique genes; (Supplementary Table S6) present in the cfDNA Pan-Cancer Reference Standards. The panel design encompassed a per-probe length of 120 bp and employed a 3x tiling approach across the target regions. The resulting probe panel was used to evaluate the sensitivity of detecting known variants in the Twist cfDNA control using probe hybridization capture enrichment in four distinct library preparation workflows as described below.

New England biolabs NEBNext^® ultra™ II library Prep kit for Illumina^®

The NEBNext^® Ultra™ II Library Prep Kit for Illumina^® (catalog No. E7645L) was used to create pre-capture libraries using 10 ng of Twist cfDNA Pan-Cancer Reference Standards: 5% (Twist-5), 1% (Twist-1), and 0% (Twist-0) VAF. The pre-capture library construction process adhered to the recommended protocol, except for the following modifications: during ligation, the IDT xGen™ UDI-UMI adapter (catalog No. 10005903) was utilized at a concentration of 1.5 µM. Subsequently, the ligation products underwent purification using a 1.2X ratio of SPRI Select beads (Beckman Coulter; catalog No. B23319) and were eluted in 20 µl of nuclease-free water. Nine cycles of PCR were performed, using 4 µl of ligation product in duplicate 50 µl total volume reactions for each sample. The PCR reactions used the NEBNext^® Ultra™ II Q5 Master Mix and universal primers specific to the Illumina adapter sequences P5 and P7, following the recommended cycling conditions. After cycling, the duplicate PCR reactions were combined, purified with SPRI Select beads at a 1.2X ratio, and eluted in 22 µl of nuclease-free water. DNA quantity and quality were assessed using the Qubit 1X dsDNA HS assay (Invitrogen; catalog No. Q33230) and the Bioanalyzer High Sensitivity DNA kit (Agilent; catalog No. 5067 − 4626). Subsequently, approximately 300 ng of each library were pooled together for capture enrichment using the custom IDT xGen™ MRD panel (described in methods ‘custom probe panel design’) and the IDT xGen™ Hybridization and Wash Kit (catalog No. 1080577) with overnight probe hybridization. Post-capture PCR involved 14 cycles of amplification, utilizing 5 µl of post-capture product in duplicate 50 µl total volume reactions for each sample. The PCR reactions used the NEBNext^® Ultra™ II Q5 Master Mix, following the recommended cycling conditions. After cycling, the duplicate PCR reactions were combined, purified with SPRI Select beads at a 1.2X ratio, and eluted in 22 µl of nuclease-free water. DNA quantity and quality were assessed using the Qubit 1X dsDNA HS assay and the Bioanalyzer High Sensitivity DNA kit. The resulting libraries were sequenced on an Illumina NovaSeq6000, generating paired-end 100-bp reads. The resulting fastq files were randomly downsampled to 45 M paired-end reads for each sample, with subsequent read alignment to human reference build GRCh38 (https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.26/), and marked duplicate reads computed from BAM files using Picard.

TwinStrand biosciences duplex library Prep kit

Next, the TwinStrand Biosciences Duplex Library Prep Kit (manual version 1.1, November 2021) was used to generate pre-captured libraries from 10 ng of Twist-0, Twist-1, and Twist-5 controls. The library construction process followed the manufacturer’s guidelines, with a few specific modifications. First, half of the pre-capture libraries, ranging from 400 to 500 ng each, were utilized as input for hybridization and capture. These captures were performed as single-reaction processes, employing the panel described in methods ‘custom probe panel design,’ using the TwinStrand hybridization and wash buffers. Following the recommended cycling conditions, the first post-capture PCR used 16 cycles of amplification with all the resulting PCR product then used as input for the second round of hybridization and capture. In this second step, the post-capture PCR was conducted with 6 cycles, again adhering to the kit’s recommended cycling parameters. To assess the quality and quantity of the final libraries, we used the Qubit 1X dsDNA HS assay (Invitrogen; catalog No. Q33230) and the Bioanalyzer High Sensitivity DNA kit (Agilent; catalog No. 5067 − 4626). The final libraries were sequenced on an Illumina NovaSeq6000, generating paired-end 100-bp reads. Analysis involved the use of the DNANexus platform, which required fastq files and a target bed file to perform the TwinStrand DuplexSeq FASTQ to VCF Parallel App (4.3.0) application. This application performs error corrected alignment and variant calls across multiple samples. The fastq files for each sample were randomly downsampled to 45 M paired-end reads prior to uploading to DNANexus. The bed file provided included the coordinates for the targeted Twist variants used in the design of the ‘custom probe panel.’ The output files include aligned BAM files to the human reference build GRCh38 (https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.26/).

Claret bioscience SRSLY^® PicoPlus NGS library Prep kit

The single-stranded DNA library prep kit from Claret Bioscience, SRSLY^® PicoPlus Kit (catalog No. CBS-K250B-96) with UDI-UMI primers (catalog No. CBS-UMI-96) was used to create pre-captured libraries from varying input amounts (10 ng, 5 ng, 1 ng, 0.5 ng, and 0.1 ng) of Twist-0, Twist-1, and Twist-5 controls. The library construction process adhered to the manufacturer’s guidelines, with a few specific modifications: during ligation, the 10 ng input samples used the Claret Biosciences NanoPlus SRSLY NGS Adapters (catalog No. CBS-K155B-96), whereas all other inputs were prepared using the standard PicoPlus SRSLY NGS Adapters. The number of PCR cycles performed pre- and post-capture varied based on the initial sample input amounts. Specifically, pre-capture libraries generated from 10 ng, 5 ng, and 1 ng of starting input material underwent 10 cycles of PCR. In contrast, pre-capture libraries generated from 0.5 ng starting input material underwent 12 cycles of amplification, and the 0.1 ng starting material underwent 14 cycles of amplification. PCR conditions followed the recommended cycling parameters. After amplification, the PCR reactions were purified using the specific single-sided SPRI Select workflow and eluted in 22 µl of nuclease-free water. The DNA quantity and quality were assessed using the Qubit 1X dsDNA HS assay (Invitrogen; catalog No. Q33230) and the Bioanalyzer High Sensitivity DNA kit (Agilent; catalog No. 5067 − 4626). For samples that had 10 ng of starting input material, approximately 500 ng of each pre-capture library were combined into a single pool for capture enrichment. In a separate reaction, 185 ng of each pre-captured library derived from samples starting with 5 ng and 1 ng of Twist cfDNA were pooled together. In a third reaction, 140 ng of each pre-captured library constructed from samples starting with 0.5 ng and 0.1 ng of Twist cfDNA were combined for capture enrichment. The capture enrichment process used the panel described in methods ‘custom probe panel design’ with the IDT xGen™ Hybridization and Wash Kit (catalog No. 1080577). Probe hybridization was performed overnight. The post-capture PCR was conducted with 14 cycles of amplification for the Twist cfDNA samples with starting inputs of 10 ng, 5 ng, and 1 ng of cfDNA and 13 cycles of amplification for the 0.5 ng and 0.1 ng inputs. Each sample underwent duplicate PCR reactions, with 5 µl of post-capture product in a total volume of 50 µl per reaction. The PCR reactions used the NEBNext^® Ultra™ II Q5 Master Mix, following the recommended cycling conditions. For purification, the duplicate PCR reactions were combined and treated with SPRI Select at a 1.2X ratio. The elution was performed using 22 µl of nuclease-free water. To assess DNA quantity and quality, we employed the Qubit 1X dsDNA HS assay (Invitrogen) and the Bioanalyzer High Sensitivity DNA kit (Agilent). The resulting libraries were sequenced on an Illumina NovaSeq6000, generating paired-end 100-bp reads. The resulting fastq files were randomly downsampled to 45 M paired-end reads for each sample, with subsequent read alignment to human reference build GRCh38 (https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.26/), and marked duplicate reads computed from BAM files using Picard.

Integrated DNA technologies (IDT) xGen™ CfDNA & FFPE DNA library Prep kit

The xGen™ cfDNA & FFPE DNA Library Prep Kit (catalog No. 10010207) library prep kit from Integrated DNA Technologies (IDT) was used to create pre-capture libraries from varying input amounts (10 ng, 5 ng, 1 ng, 0.5 ng, and 0.1 ng) of the Twist-0, Twist-1, and Twist-5 controls. The library prep process adhered to the manufacturer’s guidelines, with a few specific modifications: the number of PCR cycles performed pre- and post-capture varied based on the initial sample input amounts. The number of pre-capture PCR cycles performed varied from 10 cycles (Twist cfDNA sample inputs of 5 ng and 1 ng), 11 cycles (Twist cfDNA sample input of 10 ng), and 14 cycles (Twist cfDNA sample input of 0.5 and 0.1 ng). The resulting purified PCR products were assessed using the Qubit 1X dsDNA HS assay (Invitrogen; catalog No. Q33230) and the Bioanalyzer High Sensitivity DNA kit (Agilent; catalog No. 5067 − 4626). For samples that had 10 ng of starting input material, approximately 500 ng of each pre-capture library were combined into a single pool for capture enrichment. In a separate reaction, 185 ng of each pre-captured library derived from samples starting with 5 ng and 1 ng of Twist cfDNA were pooled together. In a third reaction, 140 ng of each pre-captured library constructed from samples starting with 0.5 ng and 0.1 ng of Twist cfDNA were combined for capture enrichment. The capture enrichment process used the panel described in methods ‘custom probe panel design’ with the IDT xGen™ Hybridization and Wash Kit (catalog No. 1080577). Probe hybridization was performed overnight. The post-capture PCR was conducted with 14 cycles of amplification for the Twist cfDNA samples with starting inputs of 10 ng, 5 ng, and 1 ng of cfDNA and 11 cycles of amplification for the 0.5 ng and 0.1 ng inputs. Each sample underwent duplicate PCR reactions, with 5 µl of post-capture product in a total volume of 50 µl per reaction. The PCR reactions used the NEBNext^® Ultra™ II Q5 Master Mix, following the recommended cycling conditions. For purification, the duplicate PCR reactions were combined and treated with SPRI Select at a 1.2X ratio. The elution was performed using 22 µl of nuclease-free water. To assess DNA quantity and quality, we employed the Qubit 1X dsDNA HS assay (Invitrogen) and the Bioanalyzer High Sensitivity DNA kit (Agilent). The resulting libraries were sequenced on an Illumina NovaSeq6000, generating paired-end 100-bp reads. All samples were downsampled to 45 M paired-end reads. Strand consensus reads using fixed in-line UMIs were produced and aligned to the human reference genome GRCh38(https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.26/) using the fgbio best practices fastq to consensus pipeline (https://github.com/fulcrumgenomics/fgbio/blob/main/docs/best-practice-consensus-pipeline.md). Briefly, this used bwa mem with alt-aware alignment, then grouped reads by UMI, followed by consensus read extraction using fgbio (2.1.0, using --error-rate-post-umi = 30 --min-reads = 3). Consensus reads were then aligned again to the reference with alt-aware bwa mem, filtered, and sorted.

Pilot study in pediatric brain tumor cohort

We enrolled six pediatric patients for MRD4U, disclosed here as; LB002 was an adult and therefore excluded from this manuscript. Our pilot study utilized a dual-pronged approach from patient CSF-cfDNA: (1) patient-specific hybridization capture panel approach (MRD4U; “MRD for you”) targeting multiple somatic and germline variants identified via prior tumor-normal testing and (2) low-pass genome sequencing (lpGS) to detect tumor-specific copy number variation (CNV). Patients enrolled on our study all had either clinical or research exome sequencing performed on their tumor previously at our institution, and those data was available to us as part of our liquid biopsy IRB-approved study. We have previously described our exome sequencing method along with variant calling and CNV detection workflows [35]. CSF-cfDNA libraries were prepared (using the entirety of the extracted cfDNA) following the IDT workflow described above. The resulting library was then split, and > = 125 ng was used for MRD4U capture and > 10 ng for lpGS. MRD4U: Per patient, we created xGen™ MRD hybridization panels designed to include up to 500 probes per panel at low cost and a rapid turnaround time of five business days from when the order is placed (Integrated DNA Technologies). Each patient had a custom panel that included 10 germline targets (private to the patient and absent from gnomAD) and 20–100 personalized somatic variants, which were chosen based on the highest VAF if more than 100 variants were present in the tumor exome data. For all panels, we also included a static list of 26 targets either present in the Twist Pan-Cancer cfDNA reference standard controls and/or genes or variants reported to be commonly altered in pediatric brain tumors (Supplementary Table S8) [25, 36, 37]. A tiling approach was used to design probes for all somatic and germline targets, with each target having three different probes covering the area of interest, to yield up to 500 total probes per panel (Supplementary Fig. S1). A minimum of 125 ng of the amplified CSF-cfDNA IDT library was used for MRD4U capture along with (when available) 1–10 ng of patient plasma cfDNA and the following controls: Twist-5, Structural Multiplex cfDNA Reference Standard HD786 (Horizon Discovery; catalog No. HD786), and human reference DNA GM24143 (ref-gDNA). Pooling and hybridization capture was performed according to the manufacturer’s protocol for xGen Hybridization and Wash (Integrated DNA Technologies; catalog No. 1080584) (Supplementary Table S9). Final libraries were sequenced on Illumina NovaSeq 6000 using a run recipe of 2 × 100 (Read1 = 109; i7 = 8; i5 = 8; Read2 = 109). Variant calling was performed as described below. lpGS: At least 10 ng of the CSF-cfDNA final IDT library (pre-capture) was used for lpGS and sequenced on NovaSeq 6000 using a run recipe of 2 × 100 (Read1 = 109; i7 = 8; i5 = 8; Read2 = 109) aiming for 5x coverage. CNVs were called and plotted by comparing to a panel of normals (PON) using our previously described approach [35]. We generated the PON using plasma cfDNA from ten anonymous blood donors, using 1 ng of cfDNA prepared with the IDT xGen cfDNA & FFPE DNA Library Prep Kit as described above. Tumor-specific CNVs, known from the patient’s prior exome sequencing, were then searched for in the CSF-cfDNA lpGS data.

Variant calling and copy number variation detection

Variant calling and VAF calculations for both the Libricon workflow and the hybridization capture workflows were generated utilizing SAMtools mpileup and mpileup2cns VarScan (version 2.3.4) commands to generate read counts for the variants of interest from raw, UMI-deduplicated BAM files. VarScan2 and GATK were used to assess CNVs and loss of heterozygosity across all chromosomes [38]. In the Libricon study, only samples in which we successfully identified the patient germline variant in CSF cfDNA at > 25% VAF were considered for positive or negative ctDNA detection; all others were considered to be “inconclusive” based on our assay. In the MRD4U pilot study, variant calling was refined using processed controls concurrently run with patient samples. These controls included the Twist Pan-Cancer cfDNA at a 5% VAF (Twist-5), the Horizon Discovery Structural Multiplex Reference Standard (HD786), and the commercial genomic DNA reference standard GM24143 (ref-gDNA). These controls served to identify background noise and potential sequencing or alignment artifacts. Variants identified in any of the control samples were excluded from the patient data to ensure that only non-artifactual patient-specific alterations were considered. To classify a sample as ctDNA+, it must have exhibited alternative read support of at least three consensus reads for at least one variant, a cutoff based on the low input testing with 0.1 ng of the Twist-0 standard in the IDT workflow described above.

Results

Pediatric and AYA patient sample collection

Our research focused on analyzing 45 biofluid samples (44 CSF and 1 plasma) from 25 pediatric, adolescent and young adult (AYA) patients, collected during distinct phases of their clinical treatment for CNS tumors (Supplementary Table S3). The samples served diverse purposes within our study: some were for refining extraction methods, while others were instrumental in optimizing molecular workflows to detect ctDNA through targeted amplicon sequencing (referred to as “Libricon”) and deep sequencing techniques; referred to as “Vignette_#” throughout (Supplementary Table S4). Additionally, a select group of samples (referred to as “LB00#”) were incorporated into a pilot study to detect ctDNA as a surrogate for measuring the presence of MRD, called MRD4U (“MRD for you”) (Table 2). This innovative approach combined tumor-informed targeted capture enrichment with lpGS to enhance the measurement of ctDNA, offering a comprehensive, multimodal assessment of MRD (Fig. 1A-B).

Table 2 Matrix plots for library kit bakeoff

Fig. 1

Illustration of MRD4U workflow for detecting circulating tumor DNA and liquid biopsy sample utilization. (A) This diagram illustrates the analysis of CSF for ctDNA in pediatric and AYA patients as part of the MRD4U pilot study. Initially, the patient’s tumor is profiled using exome sequencing to identify tumor-specific variants. These variants guide the creation of a personalized probe-based capture panel, named MRD4U, used to analyze cfDNA extracted from the patient’s CSF-SN. Concurrently, part of the CSF-derived cfDNA library is subjected to lpGS to detect CNVs. (B) Sankey plot displaying how biofluid samples from pediatric and AYA patients were allocated and analyzed across different workflows. Relationships between the blocks are represented by the flowing lines, with their width corresponding to the number of patients. Abbreviations: CSF - cerebrospinal fluid, ctDNA - circulating tumor DNA, AYA - adolescent and young adult, cfDNA - cell-free DNA, CSF-SN - CSF supernatant, lpGS– low-pass genome sequencing, CNVs - copy number variations

The diagnoses at the time of sample collection included medulloblastoma (n = 34 biofluid samples obtained), ependymoma (n = 4), atypical teratoid rhabdoid tumor (ATRT; n = 2), atypical choroid plexus papilloma (n = 1), brainstem lesion (n = 1), mature teratoma (n = 1), and pineoblastoma (n = 1) (Supplementary Table S1). Samples were obtained at diagnosis– 10 to 14 days post initial surgery– (n = 8), at relapse (n = 8), during treatment (n = 8), or post radiation (CSI; n = 8), with additional samples collected at end of therapy (n = 5) or post-chemotherapy (n = 3). A limited number of biofluid samples were without disclosed treatment status during time of collection (n = 5). Most patients had only a single biofluid sample collected (n = 17), with three patients providing two samples and five patients having three or more samples for assessment. Out of the 45 liquid biopsy samples, 44 were CSF, primarily collected through lumbar puncture (n = 38) (Table 1). The remaining CSF samples were obtained through an external ventricular drain (n = 3), during ventriculoperitoneal shunt replacement (n = 1), or through an unspecified collection method (n = 2).

In most cases, CSF was first sent clinically for cytology, cell count, and/or culture analyses; remaining additional CSF was used for ctDNA analyses. The volume of CSF extracted for ctDNA analyses varied between 0.3 and 3.4 ml (average = 1.2 ml), with most samples (n = 29) being ≤ 1.0 ml (Table 1 and Supplementary Table S2). The CSF collected during ventriculoperitoneal shunt replacement had the largest volume at 2 mL, followed by CSF obtained via external ventricular drain (average = 1.5 mL), lumbar puncture (average = 1.2 mL), and CSF collected through unspecified means (average = 1.0 mL). The collected CSF was stored at −80 °C until extraction, with the duration between collection and extraction averaging 107 days and varying from a minimum of 6 days to a maximum of 806 days. In contrast, the single plasma sample isolated from blood exhibited a shorter duration of 16 days between plasma separation and extraction.

The extracted DNA yield varied, with 20 of 44 CSF samples falling below our limit of detection (see methods). Among quantifiable samples, the highest DNA concentration was observed in CSF collected via external ventricular drain (average 55.9 ng/mL), followed by shunt replacement (2.7 ng/mL), lumbar puncture (2.0 ng/mL), and unspecified methods (0.9 ng/mL) (Table 1 and Supplementary Table S2). We examined whether DNA yield was associated with treatment status at the time of CSF collection. We specifically focused on CSF collected through lumbar puncture, as this method of collection was consistent across all defined treatment statuses, and lumbar punctures are routinely performed for standard of care disease evaluations for many CNS tumors, especially CNS embryonal tumors. CSF collected at diagnosis yielded the highest average DNA concentration (6.8 ng/mL), followed by samples at relapse (3.5 ng/mL), post-radiation (1.3 ng/mL), mid-treatment (1.2 ng/mL), end of therapy (0.4 ng/mL), and mid-chemotherapy (0.0 ng/mL) (Supplementary Fig. S2 and Table 1).

Targeted amplicon sequencing to detect ctdna in CSF

There are inherent challenges associated with evaluating MRD using circulating cfDNA, including degradation and low yield of the cfDNA extract. Furthermore, ctDNA is often at low levels warranting a deep and/or targeted sequencing approach to provide higher coverage at the regions of interest. As such, our initial strategy involved generating targeted amplicons (“Libricons”) followed by deep sequencing to assess for feasibility of detecting ctDNA from CSF. We have used this method previously to successfully confirm low-level somatic mosaic variants (as low as 0.1% variant allele frequencies (VAF)) in resected brain tissue from children with epilepsy [39]. The method’s prior utility, however, relied on 10 ng of genomic DNA as input. To assess the Libricon approach effectiveness with cfDNA, we conducted an evaluation using serial dilutions of the Twist Pan-Cancer cfDNA reference standards, a control containing known sequence variants at specific VAFs. Specifically, we designed primers to target the BRAF p. V600E single nucleotide variant (SNV) within the Twist control, varying input amounts from 0.25 to 10 ng and exploring a range of VAFs from 0 to 5%. This approach allowed us to define the lower limit of detection (LOD) and sensitivity for our Libricon approach. Average coverage for the BRAF Libricons was ~ 27,500x (Range: 282x − 141,600x). Our findings indicate a strong concordance between the expected and observed BRAF p.V600E VAFs for inputs of 10 and 5 ng, demonstrated by high R [2] values of 0.96 and 0.93, respectively (Fig. 2A). When the input was reduced to 1 and 0.5 ng, we noted a decrease in concordance, marked by increased variation between expected and actual VAFs, with an R [2] value of 0.65. At the lowest input level of 0.25 ng, the discrepancy between expected and observed VAFs further increased, especially at the lower expected VAFs, evidenced by an R [2] value of 0.59. When using the Twist Pan Cancer cfDNA control with a 5% VAF, despite the observed VAF ranging between 2 and 6%, the BRAF p.V600E was detected across all input levels (0.25 ng − 10 ng), including replicates. Below 5%, the LOD for 1 ng input was 2% VAF and for inputs of 5 and 10 ng, the LOD was 0.25% VAF.

Next, we applied our optimized cfDNA Libricon method to a subset of our patient cohort (n = 24 samples), for cases from whom residual tumor genomic DNA (gDNA) was also available to generate Libricons in parallel (Fig. 1B and Supplementary Table S4). First, we demonstrated that Libricon sequencing is concordant with exome sequencing in detecting similar VAFs, by initially identifying the somatic variant with the highest VAF from prior exome sequencing of tumor DNA (refer to Methods section) and subsequently generating Libricons, targeting this specific variant from tumor gDNA. We achieved sequencing coverage exceeding 10,000x for these Libricons, and our analysis demonstrated that the VAFs detected by the Libricon method closely matched those identified by exome sequencing (Fig. 2B-C). Using this same subset of patients, we then targeted the same somatic variant in the CSF cfDNA extract. The range of tumor gDNA VAFs for targeted variants was 35–90% and of the 24 CSF cfDNA samples evaluated, 5 were found to be positive for ctDNA, with a lower LOD of 0.37% VAF detected in cfDNA (Figs. 1B, 2C and Supplementary Table S4). Our initial patient cohort included longitudinal collection of CSF for three children, all of whom were diagnosed with medulloblastoma. We performed Libricon sequencing and subsequently plotted longitudinal analyses for these patients to track any changes in ctDNA/MRD status throughout their course of treatment (Fig. 2D). For Vignette_14, ctDNA was not detected at any of the sampled timepoints (negative circulating tumor DNA; ctDNA-). In contrast, Vignette_17 was positive for circulating tumor DNA (ctDNA+) at the time of diagnosis but subsequently ctDNA- following radiation treatment, suggesting a potentially favorable response to therapy. Vignette_18 presented a more complex pattern, initially testing ctDNA + post-radiation, then negative at the later time of clinical relapse, indicating fluctuating levels of ctDNA that may either reflect dynamic changes in tumor behavior or treatment response or, rather, a limited sensitivity of the Libricon approach given that only one somatic variant is targeted. While our Libricon approach has proven effective for detecting circulating tumor DNA from low yields of CSF cfDNA, its scalability to target multiple loci is inherently limited. This limitation arises from the need to optimize multiplexed primer pairs, a process further complicated by the necessity to tailor primer pairs on a per-patient basis based on unique genomic profiles. To enhance the comprehensiveness of our approach, we next sought to optimize a sample preparation and hybridization capture panel strategy to profile a broader array of known tumor variants in our cohort.

Fig. 2

Comprehensive analysis of ctDNA detection using Libricon sequencing. (A) Sensitivity and precision of the Libricon approach in detecting the BRAF p.V600E variant using Twist PanCancer cfDNA controls at VAFs of 0%, 0.10%, 0.25%, 0.50%, 1.00%, 2.00%, and 5.00%. Libricons were created with DNA inputs ranging from 0.25 ng to 10 ng, illustrating the detection method’s performance at various levels, and visually representing the observed VAFs. (B) Comparison of the VAF of the targeted variant from exome sequencing (displayed as a blue bar) with the VAF derived from tumor genomic DNA analyzed using the Libricon approach (purple bar). (C) Targeted variant VAF as determined by Libricon on tumor DNA and in CSF cfDNA, with ctDNA + in CSF cfDNA marked by an orange circle. (D) Displays longitudinal Libricon results for three patients, charting disease state at CSF collection against the observed VAF, with ctDNA + indicated by orange bars. Abbreviations: VAF - variant allele frequency, ES - exome sequencing, CSF - cerebrospinal fluid, cfDNA - cell-free DNA, ctDNA+ - positive circulating tumor DNA

Comparative analysis of Commercially-Available NGS sample Preparation kits for CfDNA

This study aimed to compare the effectiveness of four commercial NGS library preparation kits designed for either low input or highly fragmented DNA—key traits of cfDNA—focusing on their capabilities for single-stranded DNA processing and error correction via unique molecular identifiers (UMIs) (Table 3). Using Twist Pan-Cancer cfDNA reference standards, we evaluated kits from New England Biolabs (NEBNext Ultra II, abbreviated as NEB), Claret Bioscience (SRSLY PicoPlus, abbreviated as Claret), TwinStrand Biosciences (Duplex Library Prep, abbreviated as TwinStrand), and Integrated DNA Technologies (cfDNA & FFPE DNA Library Prep, abbreviated as IDT) based on the quality of sequencing data. These quality metrics included the precision of identifying targeted variants and the accuracy of VAFs from libraries prepared with 10 ng of Twist cfDNA control samples at VAFs of 5% (Twist-5), 1% (Twist-1), and 0% (Twist-0), targeting 25 single nucleotide variants (SNVs) and 24 insertion/deletion (INDEL) events across 39 genes using a probe-based capture enrichment strategy.

Table 3 MRD4U summary of findings

Bioinformatically, we randomly downsampled the total number of sequenced paired-end reads to be 45 million for each sample and observed variation in both coverage and the number of variants detected across the different kits. Ideally, at 0% VAF, representing the control scenario with no expected variants, the detection of any variants indicates false positives. Here, a variant is considered detected if it is supported by more than one read or more than one consensus read. Using the Twist-0 control and preparing it with four different NGS kits, we observed 39 of the 49 variants detected with the Claret workflow, 22 in the NEB-prepared sample, 12 from the IDT preparation, and 6 in the TwinStrand sample (Fig. 3A; Supplementary Tables S6 and S7). Notably, both the IDT and TwinStrand kits incorporate fixed inline UMIs that can be used to form a consensus read during data analysis. This strand consensus process could potentially account for the fewer variants detected in Twist-0, as it more stringently filters out sequencing errors and potential artifacts, thus reducing the likelihood of false positives. Consequently, the observed low VAFs in Twist-0 for TwinStrand (average 0.13%, range of 0–2.48%) and IDT (average 0.17%, range 0–4.68%) are consistent with the expected outcomes of the thorough consensus filtering. In comparison, the variants detected in Twist-0 processed with NEB and Claret exhibited higher average VAFs of 0.35% (range 0–8.58%) and 0.37% (range 0–8.42%), respectively. Notably, performing strand consensus impacts overall coverage at the variant of interest loci, leading to TwinStrand having the lowest average duplex consensus coverage for the Twist cfDNA controls (Twist-0: 375x; Twist-1: 309x; Twist-5: 342x), while Claret reported the highest average coverage (Twist-0: 1,851x; Twist-1: 1,851x; Twist-5: 1,508x) in the normalized downsampled data. While there were distinct differences in coverage between the kits, several variants detected within the Twist-0 control were supported by only a few reads. The Claret sample had 24 detected variants supported by one or two reads, the NEB sample had 19, the TwinStrand sample had three variants with read support of one or two consensus reads, and IDT had eight variants with read support of one or two consensus reads. These results suggest the need for custom filtered analysis to mitigate workflow-specific artifacts while maintaining a balance between coverage and read support thresholds, highlighting the challenges in specificity and revealing the diverse background noise levels across the kits.

Fig. 3

Performance evaluation of NGS kits for cfDNA variant detection by input and VAF. (A) The plot displays the variant detection capabilities of four NGS kits—NEBNext Ultra II (NEB, green), SRSLY PicoPlus (Claret, blue), Duplex Library Prep (TwinStrand, orange), and cfDNA & FFPE DNA Library Prep (IDT, purple)—using cfDNA samples. Variants are shown as triangles for SNVs and circles for INDELs, mapped against observed VAFs. Colors represent different kits, demonstrating their precision and accuracy in identifying variants at controlled VAFs (5%, 1%, 0%) with 10 ng of Twist Pan-Cancer cfDNA. (B) This plot evaluates the performance of Claret (blue) and IDT (purple) kits under varying inputs from 0.1 to 10 ng, illustrating their efficacy in variant detection across different Twist controls (−0, −1, −5), plotted against observed VAFs. Abbreviations: NGS - next-generation sequencing, NEB - New England Biolabs, IDT - Integrated DNA Technologies, cfDNA - cell-free DNA, SNVs - single nucleotide variants, INDELs - insertions/deletions, VAF - variant allele frequencies

Next, we evaluated variant detection within the Twist-1 and Twist-5 controls which demonstrated differences in sensitivity and precision among the kits at these VAF thresholds. For Twist-1 and Twist-5, NEB identified 43 variants in each sample, achieving average VAFs of 0.83% for Twist-1 and 2.57% for Twist-5. Claret detected 48 variants in both controls, with average VAFs reaching 0.86% in Twist-1 and 3.09% in Twist-5. TwinStrand showed a lower detection rate, with 34 variants in Twist-1 (average VAF 0.52%) and 42 in Twist-5 (average VAF 2.02%). Lastly, the IDT workflow resulted in 40 variants detected in Twist-1 at an average VAF of 0.67% and 45 variants detected in Twist-5, with an average VAF of 2.83%.

Given these collective results, along with Claret’s ability to construct sequenceable libraries from single-stranded DNA, thus improving coverage, and IDT’s use of strand consensus to minimize background noise in variant detection, both kits were selected for further evaluation. This assessment used the Twist − 0, −1, and − 5 controls with inputs varying from 0.1 to 5 ng, aiming to evaluate kit efficacy under a wide spectrum of challenging input conditions. The same probe panel targeting 25 SNVs and 24 INDELs, as previously described, was used to determine kit efficacy in variant detection at these varying input amounts (Fig. 3B; Supplementary Tables S6 and S7). For both Claret and IDT kits at the minimal 0.1ng input, the number of variants detected in the Twist-0 control (Claret: 12 variants detected; IDT: 4) closely matched those found in the Twist-1 control (Claret: 9; IDT: 3), and for Claret, the Twist-5 control (13 variants detected). However, more than twice as many variants were found in the IDT prepared Twist-5 control (10 variants detected) than the IDT prepared Twist-0 control (4 variants detected). This deviation highlights that, despite the overall reduced sensitivity at such low DNA input, IDT’s strand consensus methodology may still discern true variants, at a VAF greater than 5%, with effectiveness at a 0.1 ng input. At the 0.5 ng input level, the Claret workflow resulted in eight detected variants in Twist-0 with an average VAF of 0.51% (range: 0–7.35%), 16 variants in Twist-1 with an average VAF of 0.93% (range: 0–9.78%), and 28 variants in Twist-5, showcasing an average VAF of 2.07% (range: 0–13.33%). Conversely, IDT’s performance at the same input level revealed four variants detected in Twist-0 with an average VAF of 0.19% (range: 0–4.23%), 11 variants in Twist-1 with an average VAF of 0.44% (range: 0–3.31%), and an increase to 34 variants in Twist-5, where the average VAF was 2.85% (range: 0–13.51%). At 1 ng and 5 ng inputs, Claret and IDT kits demonstrated improved variant detection and observed VAF, with Claret detecting up to 43 variants and IDT up to 45 variants in the Twist-5 control. Considering the IDT kit’s effective variant detection with the Twist Pan-Cancer cfDNA controls at minimal inputs, this kit was subsequently selected for evaluating patient samples for ctDNA.

Assessment of liquid biopsy for detecting MRD in children with brain tumors

In this pilot study, liquid biopsy (LB) was employed as a novel approach to detect ctDNA as a surrogate for MRD in patients with central nervous system tumors after various stages of treatment. We collected CSF samples from six patients (Supplementary Table S2) to assess MRD at various timepoints including mid-chemotherapy (LB001 and LB004), post-radiation therapy (LB007), and completion of their treatment regimen (LB003, LB005 and LB006). This study employed a multimodal approach combining our optimized hybridization capture of cfDNA plus lpGS (Fig. 1A). A distinct feature of our study was the development of personalized “MRD4U” panels for each patient. All patients undergoing MRD4U had prior tumor sequencing in either the research or clinical setting and from those data, we chose up to 100 patient-specific known somatic variants to be targeted in our assay (Supplementary Figure S1). In addition, we included private, heterozygous germline variants and a “static backbone” comprising probes against common pediatric CNS tumor alterations and against the Twist Pan-Cancer cfDNA variants for positive processing controls (Supplementary Table S8).

In this LB cohort, the extracted cfDNA yields from CSF were below the limit of detection for our quantification method, except for a single sample (LB007) which yielded 1.1 ng of cfDNA (Supplementary Table S2). Despite most yields being classified as “undetectable” with our current method, we successfully targeted and detected patient-specific germline variants in all cfDNA samples, indicating that patient-specific cfDNA was present in each extract (Table 2 Supplementary Table S10). The entirety of each CSF cfDNA extract was used for IDT library preparation, and the resulting amplified product was then split: >= 125 ng was used for MRD4U capture and > 10 ng for lpGS Supplementary Table S9). Leveraging the strand consensus capability of the highlighted IDT library kit enabled us to achieve an average targeted variant loci consensus coverage of 46.0x (range = 4.9x-173.1x) across the six CSF-derived cfDNA samples (Supplementary Table S10).

We identified two ctDNA + in our LB cohort: LB005 (end of therapy) and LB007 (post-radiation) (Table 2). For LB005, the personalized MRD4U panel targeted 50 tumor-informed somatic variant targets, of which 11 were detected in the patient’s CSF-derived cfDNA with a mean VAF of 37.2% (range: from 17.4 to 70.8%) (Table 2 and Fig. 4A-B). The MRD4U panel also targeted 10 patient-specific germline variants, which in LB005 CSF cfDNA achieved an average consensus coverage of 63.7x and showed an average VAF of 51.2%. In patient LB007, the personalized MRD4U panel, designed with 100 tumor-informed variant targets, detected 52 of those variants in the CSF-derived cfDNA with an average VAF of 10.9%, ranging from 1.3 to 35.6%. For the patient’s targeted 10 germline variants, we observed an average consensus coverage of 122.7x with an average VAF of 43.9% (Supplementary Table S10). In LB007, we observed a significant correlation between tumor VAF and CSF cfDNA VAF among variants detected in both samples, with a wide range of allele frequencies in the CSF; across both LB005 and LB007, ctDNA was consistently detectable in the CSF for variants with tumor VAFs greater than 5%, with detection rates increasing in the 10–20% and 20–50% bins (Supplementary Fig S3). Neither LB005 nor LB007 had evidence of disease based on radiographic or clinical findings at the time of MRD4U; clinical cytological analysis was negative for malignant cells, and both brain and spine imaging (MRI) was negative for evidence of recurrent or metastatic disease, respectively. Subsequent disease imaging, however, in LB005 confirmed disease recurrence and is described in greater detail in the Case Study section below. LB007 unfortunately passed away due to disease; longitudinal follow-up of patients characterized as ctDNA- (LB001, LB003, LB004, and LB006) confirmed that these patients have remained in remission without any reoccurrence or relapse of cancer 10–13 months following CSF sample collection.

Fig. 4

Comprehensive analysis of genetic variants and copy number variations in patient LB005 from the MRD4U study. Panels A and B illustrate the VAF of variants detected in patient LB005 from the MRD4U study. Panel A presents SNVs, with somatic variants depicted as closed circles and germline variants as closed triangles, plotted against VAFs on the y-axis and various samples on the x-axis: ref-gDNA (green), Twist-5 control (blue), HD786 control (orange), and CSF-SN (purple), alongside original tumor genomic DNA VAFs (pink) for context. Panel B mirrors this for INDELs. Panels C and D detail CNVs in LB005, with Panel C showing CNV changes in tumor genomic DNA from exome profiling, and Panel D displaying CNVs in CSF cfDNA identified via lpGS. Abbreviations: VAF - variant allele frequencies, SNV - single nucleotide variant, ref-gDNA - reference genomic DNA, CSF - cerebrospinal fluid, CSF-SN - CSF supernatant, INDEL - insertion/deletion, CNV - copy number variation, lpGS– low-pass genome sequencing, cfDNA - cell-free DNA

Prior to MRD4U hybridization capture, a portion of each patient’s CSF cfDNA library was saved for lpGS to assess for known tumor copy number variations (CNVs). First, as a positive control for our lpGS assay we sequenced the Horizon Structural Multiplex cfDNA reference standard HD786, which contains known focal amplifications. To evaluate efficiency of CNV detection using low coverage, the control was prepared with the same IDT kit with varying inputs (1 ng and 5 ng) and sequenced to different depths. The 5 ng input reached an average coverage of 22.9x, while the 1 ng inputs were sequenced to 18.7x and a lower depth of 4.0x. Known focal amplifications were successfully detected in the 1 ng input with low coverage (Supplementary Fig S4A). To establish a comparator dataset for somatic CNV calling in our CSF cfDNA samples, we generated a “panel of normals” using cfDNA isolated from ten anonymous blood donors, prepared using the same IDT kit with 1 ng input (refer to Methods section). Average coverage for lpGS in all LB samples was 2.8x (range = 1.0x-5.5x) (Supplementary Table S11). We found that except for LB004 (1.0x coverage), all lpGS results were interpretable at coverages > 1.5x.

Known tumor CNVs were detected only in LB005 and LB007, supporting the ctDNA + finding noted in our MRD4U hybridization capture assay (Supplementary Table S11 and Supplementary Fig. S4B). All others were reported to have no CNVs is the CSF cfDNA, except LB004, whose sample had too low coverage to make a conclusive finding. Notably for LB005, prior tumor exome sequencing indicated a gain of chromosome arm 7q, losses of chromosome arms 11p and 20p, and an isochromosome 17q [(i)17q] (Fig. 4C), all of which were positively detected and visualized in the CSF cfDNA lpGS at 2.6x coverage (Fig. 4D).

A case study highlighting early detection capabilities

LB005 is a female diagnosed at 10 years of age with localized (M0) non-WNT/non-SHH (Group 4) medulloblastoma. She underwent gross total resection of posterior fossa disease, followed by adjuvant proton craniospinal irradiation and maintenance chemotherapy as per the Children’s Oncology Group (COG) ACNS0331 regimen, though was only able to complete six of nine planned chemotherapy cycles due to significant infectious and gastrointestinal toxicity. Approximately two years following completion of chemotherapy, imaging demonstrated relapsed disease with distant spinal metastases (two bulky masses in cervical and thoracic spine, without intracranial disease), for which she underwent resection of both masses (confirming group 4 medulloblastoma), followed by craniospinal proton re-irradiation. She was then treated with maintenance chemotherapy as per COG ACNS0821 regimen with temozolomide, irinotecan, and bevacizumab, though was only able to complete three cycles due to substantial hematologic toxicity leading to prolonged delays, despite several dose reductions. The patient subsequently underwent close surveillance with regular MRI, at which time she was enrolled on our research liquid biopsy study as “LB005.” A lumbar puncture performed approximately 6 months following chemotherapy completion was negative for malignant cells by cytology, though results of our MRD4U assay were positive for presence of ctDNA in CSF, indicative of residual or recurrent disease, as described above. Brain MRI at this time remained stable from recent prior scans, without radiographic evidence of recurrent disease (Fig. 5A-B). Repeat MRI approximately three months after the CSF MRD4U ctDNA + research finding demonstrated new/increased small enhancing, diffusion restricting nodules around the posterior fossa surgical cavity, worrisome for relapsed intracranial disease (Fig. 5C-D). Short interval follow-up MRI then one month later showed further increase in size and number of these enhancing, diffusion-restricting nodules, now convincing for disease recurrence (Fig. 5E-F).

Fig. 5

LB005 MRI from three longitudinal timepoints. Imaging at the time of MRD4U liquid biopsy in (A&B): Axial T1-weighted (T1-W) post contrast image in (A) shows no definite enhancing foci along the fourth ventricular margin while axial diffusion weighted imaging (DWI) in (B) shows no evidence of foci or restricted diffusion along the fourth ventricular margin. Imaging three months post MRD4U liquid biopsy in (C&D): Axial T1-W post contrast image in (C) shows an enhancing nodule along the fourth ventricular margin (arrow) while Axial DWI in (D) shows restricted diffusion of this nodule (arrow). Imaging four months post MRD4U liquid biopsy in (E&F): Axial T1-W post contrast image in (E) shows enlarging nodule along the fourth ventricular margin (arrow) while axial DWI in (F) shows restricted diffusion of this nodule (arrow)

Discussion

Despite significant advancements in liquid biopsy research for pediatric brain tumors, challenges in standardizing methodologies persist, particularly in adapting to low input, highly fragmented cfDNA and establishing robust analytical pipelines for potential transition into clinical application [32]. Our study addresses the challenge of working with limited CSF volumes from children with CNS tumors, in which we obtained an average of 1.2 mL, yielding 2.02 ng/mL of cfDNA on average (range: 0-16.6 ng/mL) in our study from lumbar puncture-derived CSF samples; other similar studies have shown wide ranges of extraction yields and similar average as ours [27, 30]. Our comparative analysis of four commercially available NGS library preparation kits evaluated their performance with these low input amounts, their ability to process the single-stranded DNA commonly found in cfDNA, and their error correction capabilities using unique molecular identifiers (UMIs) and strand consensus variant calling. The most effective kit demonstrated the best capacity to minimize false positives, a crucial attribute when analyzing low variant allele frequencies (VAFs) in circulating cfDNA. This was particularly evident when we sequenced a negative control, Twist-0, where no somatic variants were expected. Additionally, the selected preparation method excelled in detecting somatic variants at extremely low DNA inputs, down to 0.1 ng inputs efficiently in our study using synthetic controls. Similarly, samples in our LB pilot study, with inputs at or below the 0.1 ng threshold, still yielded positive findings. This demonstrates the flexibility of our assay, which can work with significantly lower DNA inputs compared to the 2–5 ng lower limit commonly reported in other studies [23, 26, 27]. Assays amenable to low input are especially important for pediatric CSF applications.

Upon optimizing this workflow, we implemented a multi-modal liquid biopsy approach that combined lpGS for detecting CNVs with our personalized, tumor-informed targeted sequencing strategy MRD4U for somatic variants. The MRD4U workflow targets mutations identified from each patient’s prior tumor sequencing data. However, a limitation of MRD4U is its reliance on diagnostic tumor profiling, which may miss spatial or temporal heterogeneity including relapse-specific alterations or secondary malignancies [40]. Integrating lpGS as a complementary assay helps address this gap by enabling detection of evolving genomic changes without prior knowledge. Our decision to adopt a personalized panel design, instead of a broader static panel, is based on several factors. Primarily, pediatric CNS entities represent individualized diseases and while recurrent hotspot variants are observed in some cases, the mutational profile is diverse even amid genes within recurrently altered signaling pathways. We opted for a method (hybridization capture) to target a diverse array of mutations. This approach contrasts with our demonstrated Libricon technique or digital droplet PCR methods, both of which require extensive optimization to scale the number of targets they can assess. Additionally, unlike gene-specific panels that apply only to certain tumor types - such as targeting BRAF in low-grade gliomas or H3K27M in diffuse midline gliomas - our chosen method offers broader applicability [27, 41,42,43,44]. Moreover, while many “pan-cancer” panels exist, these are generally designed around adult cancers and are often too extensive to achieve deep coverage practically, particularly when using strand consensus for variant calling [45]. The development of personalized panels for targeted sequencing in our study relies on prior tumor sequencing, which is becoming more routine in pediatrics, as seen in the Children’s Oncology Group Molecular Characterization Initiative [46]. Additionally, all tumors, including fusion-driven cancers, harbor some somatic variants at relatively high VAFs; even in low-mutation burden tumors, passenger mutations are typically present and detectable using high-sensitivity sequencing approaches such as deep exome [25]. In our study, tumor variants present at > 5% VAF were detectable in matched CSF using our current MRD4U workflow, supporting its applicability even in tumors with low mutational burden or without recurrent driver mutations. In our LB pilot study, the minimum number of somatic variants to be targeted was 20 (LB004), with tumor VAFs ranging 2–70%. Our workflow evaluation with synthetic cfDNA controls successfully demonstrated circulating VAF detection as low as 5% even with low input starting material.

Tailoring patient-specific panels ensures that sequencing coverage is directed toward known tumor variants for a given individual, which is crucial when the available coverage is expected to be relatively low. In our study, the average somatic target consensus depth in MRD4U ranged from 4.9x to 173.1x and for lpGS, it ranged from 1.0x to 5.5x. The highest targeted coverage, observed in LB ctDNA + patients (43.3x and 173.1x), highlights the potential limitations in detecting ctDNA/MRD in patients with lower coverage, likely due to minimal amounts of cfDNA initially available for library preparation, although further studies with a larger number of patients are needed to definitively determine the optimal coverage depth. Our selected method for library preparation used known fixed UMI sequences, added to initial DNA molecules, to allow for strand consensus calling. This strategy decreases base calling errors which is especially important in liquid biopsy samples which often have either low input or low VAF. In our pipeline, we require a minimum of three consensus reads to positively identify a somatic variant, reducing the chance of false positive calls.

In our study we applied our optimized MRD4U workflow to a small, exploratory cohort of six children and successfully identified two as ctDNA + using CSF. Interestingly, in one of these cases, the presence of ctDNA was detected three months prior to radiographic evidence of recurrence. Due to the experimental nature of our research assay and in adherence with consent and protocol guidelines, these liquid biopsy results were not shared with either the clinicians or the patients. While our workflow was designed with clinical implementation in mind and emphasizes speed, cost-efficiency, and feasibility with panel design and probe synthesis achievable within five days, it is important to note limitations. Our cohort is heavily weighted towards medulloblastoma cases, reflecting real-world accessibility as CSF is more readily collected in this patient population due to routine lumbar punctures for disease staging and surveillance. This imbalance, however, limits ability to draw broader conclusions about associations between ctDNA detectability and tumor types. Although our MRD4U approach is theoretically applicable to any patient with prior tumor sequencing, broader implementation may be constrained by institutional infrastructure for custom assay development. Larger, diverse cohorts and further real-world testing are needed to fully evaluate the clinical utility and scalability of this personalized strategy across pediatric CNS tumors.

Conclusions

The capacity to detect ctDNA in CSF holds great promise in monitoring pediatric brain tumors, particularly in scenarios where traditional imaging and cytological methods may fail to identify residual or recurrent disease, as demonstrated in our exploratory pilot study and many other previous studies [23, 28, 29]. This capability could support earlier prediction of relapse and more precise measurement of responses to therapeutic interventions, thereby informing treatment decisions and monitoring strategies. Additionally, since lumbar punctures are frequently part of standard care for some patients, CSF remains readily accessible for these assays. Our personalized “MRD4U” approach, which tailors capture panels to each patient’s tumor profile, offers a scalable and cost-conscious alternative to static or genome-wide methods, and may improve sensitivity across diverse pediatric CNS tumor types.