Molecular features unique to circulating tumor DNA enable the tumor-naïve liquid biopsy of glioblastoma

Abstract

Glioblastoma is the most common and deadly brain cancer in adults. Although the disease disrupts the blood-brain barrier and exposes the tumor to the systemic circulation, using circulating cell-free DNA as a non-invasive biomarker for diagnosing glioblastoma remains elusive. The main obstacles are the confounding effects from artifactual variants in cell-free DNA caused by errors during next-generation sequencing and real variants in cell-free DNA originating from white blood cells that acquired somatic mutations during normal aging. Here, we developed and validated TrueR sequencing to overcome next-generation sequencing-related false positives in cell-free DNA. Subsequently, TrueR identified molecular features unique to tumor-derived cell-free DNA (i.e., circulating tumor DNA) that are detectable in blood during a tumor-naïve search. Specifically, variants associated with glioblastoma circulating tumor DNA were exclusively present in cell-free DNA (i.e., absent in white blood cell DNA) and showed gene-specific subclones. These features were uncommon in low-grade glioma, stroke, and age-related somatic mosaicism of white blood cells. We also detail a variant-agnostic method, demonstrating that copy number gains and losses in short cell-free DNA fragments (<90 bp) support the detection of glioblastoma. Finally, we present evidence that our findings extend beyond glioblastoma, supporting the broader advancement of the liquid biopsy.

Introduction

Glioblastoma (IDH-wildtype) is the most common and deadly brain cancer in adults^1,2. Because other neurologic diseases resemble glioblastoma on clinical imaging, the histopathologic evaluation of surgically acquired tumor tissue remains the diagnostic standard^3,4. Invasive diagnostic procedures, however, risk inducing a new neurologic deficit, which reduces survival time^5–8. Additionally, a second invasive procedure for gross total or subtotal tumor resection to meet the standard of care for glioblastoma management is often required after the diagnosis is established⁸. Thus, non-invasive diagnostic methods are highly sought to initiate optimal care before or instead of risk-associated procedures^9,10.

Next-generation sequencing (NGS) of cell-free DNA in blood (i.e., circulating cell-free DNA, ccfDNA) is an emerging molecular tool for the non-invasive diagnosis of human cancers (i.e., the liquid biopsy)¹¹. Although the blood-brain barrier theoretically reduces tumor-derived ccfDNA (i.e., circulating tumor DNA, ctDNA) from central nervous system (CNS) tumors, glioblastoma is supported by a network of malformed capillaries that cause hemorrhage and blood-brain barrier disruption^12–14. Studies have consistently demonstrated that contrast-enhancing tissues on imaging, present in >95% of glioblastomas because of the disrupted barrier¹⁵, contain the highest density of glioblastoma tumor cells¹⁶. Therefore, glioblastoma-derived cell-free DNA comes into direct contact with the circulation following tumor cell death. Yet, detecting glioblastoma ctDNA remains challenging. A targeted search based on a single tumor variant identified ctDNA in <16% of patients¹⁷. A blind search identified ccfDNA somatic mutations in 50% of patients, but corresponding mutations in biopsies of the primary tumors were absent¹⁸. While the latter study may have been secondary to glioblastoma’s mutational heterogeneity, the lack of ccfDNA mutations in tumor DNA also highlights the uncertainty of discerning a true glioblastoma ctDNA signal from confounding factors such as clonal hematopoiesis of indeterminate potential (CHIP). CHIP artifacts, the age-related acquisition of low-frequency somatic mutations in white blood cells (WBCs), occur in cancer-driver genes and enter the ccfDNA pool via apoptosis with a similar allele frequency as low-frequency ctDNA^19–21. Therefore, identifying molecular features that distinguish glioblastoma ctDNA from CHIP is critical to advance non-invasive diagnostics, especially since both glioblastoma and CHIP occur more commonly in older individuals.

Beyond CHIP, several characteristics of glioblastoma complicate ctDNA detection. First, a tumor-informed search based on a focal tissue biopsy does not encompass all mutations due to spatial mutational heterogeneity^22–24. Second, glioblastoma rarely metastasizes beyond the CNS²⁵, causing low variant allele frequency (VAF) ctDNA that may be indistinguishable from NGS-associated noise during a tumor-naïve search (i.e., without a priori knowledge of specific tumor mutations to target)²⁶. Errors accrue throughout an NGS workflow (i.e., PCR errors, sequencing artifacts, alignment errors, etc.), giving rise to nonreference alleles (NRAs) in ccfDNA sequence reads with a similar allele frequency as ctDNA, causing false positives. Finally, necrotic debris is adjacent to abnormal vessels in >85% of tumors^27–29 and may be an enriched source of cell-free DNA. However, cell-free DNA from necrosis is >10,000 bp in length³⁰, and traditional library preparations targeting the sequence-ready lengths of apoptotic cell-free DNA ( < 700 bp) do not access the necrosis-derived ccfDNA component. Given the challenge of detecting glioblastoma ctDNA, methods that enhance the signal-to-noise ratio, include necrosis-derived cell-free DNA, or both, merit consideration.

Here, we systematically establish glioblastoma ctDNA detection by studying glioblastoma in an orthotopic xenograft model and in humans. For variant detection, we build upon our True2 platform, a validated, duplex-based sequencing approach previously used to discover gene-specific subclones in the tumor and peritumoral edema of adult diffuse gliomas³¹. Specifically, True2 is adapted to the inherently short fragment lengths of ccfDNA to yield TrueR, which has a reduced false positive rate and improved sensitivity for detecting ultralow frequency somatic mutations. Using TrueR, we consider short ccfDNA fragments and ccfDNA derived from necrosis to improve ctDNA detection. Subsequently, we uncover unique molecular attributes associated with glioblastoma ctDNA, which can be leveraged to differentiate glioblastoma from low-grade glioma, stroke, and age-related somatic mosaicism. We also provide evidence that our findings extend to cancers beyond glioblastoma and may prove valuable to the general advancement of non-invasive ccfDNA diagnostics.

Results

Detection of glioblastoma ctDNA in an orthotopic mouse xenograft model

To elucidate factors affecting glioblastoma ctDNA detection, the human MGG8 cell line was implanted into the brain striatum of NSG mice (Fig. 1A). Using more than a third of each animal’s total plasma volume for ccfDNA extraction (Supplemental Table S1), the mean percent ctDNA by ddPCR was 2.5%, but the range varied from 0% to 5.9% (Fig. 1B). Notably, animals with the highest and lowest ctDNA fractions exhibited similarly large tumors. Next, we used a custom-designed NGS panel (115 genes, 114 kb; Supplemental Table S2) to target 38 single nucleotide variants (SNVs) in ccfDNA present in the cell line and all xenograft tumors but absent in mouse wildtype DNA (Supplemental Table S3). The increased number of targets using NGS enabled detection of the MGG8 variants in 5 of 5 animals (100%), but the percentage of detected MGG8 variants ranged between 10.5% and 100% (Fig. 1C). Percent plasma volume and ccfDNA concentration were not correlated with variant detection (Supplementary Fig. S1).

Fig. 1. Detection of glioblastoma ctDNA in a xenograft model.

Green fluorescence identifies MGG8 tumors and cells in two mice (A; 3 F, solid arrow; and 2 M, dashed arrow). The proportion of ctDNA in ccfDNA as measured by droplet digital PCR (ddPCR) varied between animals (B). The percentage of the 38 MGG8 variants detected was similarly variable between animals (C). A PDGFRA variant, which was amplified in MGG8, was the most abundant variant detected (D). In B–D, data points share a common legend to identify individual animals, and arrows identify data points corresponding to images in (A). The PDGFRA variant was biased towards shorter fragment lengths compared to mouse wild-type ccfDNA (E). The greatest ratio between ctDNA and ccfDNA occurred in fragments <90 bp (E, purple shaded box). Limiting fragment lengths to <90 bp significantly reduced the detection rate of variants (F). For variants in the euploid state, variant allele frequency (VAF) in MGG8 was correlated with an increased variant count in ctDNA (G).

The most abundant variant detected in ccfDNA was in PDGFRA (Fig. 1D), attributable to a large PDGFRA amplification in the MGG8 cell line (Supplemental Fig. S2). Human ccfDNA (i.e., ctDNA) exhibited a shorter fragment length compared to mouse wildtype ccfDNA, and the ratio of ctDNA to ccfDNA was greatest for inserts <90 bp (Fig. 1E). However, an in silico selection of inserts <90 bp to enrich for ctDNA decreased variant detection (Fig. 1F), indicating that a pre-specified fragment length threshold may miss ctDNA with a low VAF. For variants in the euploid state, a higher VAF improved detection (Fig. 1G). For variants in the heterozygous state, monosomy and trisomy did not uniformly affect detection because variants did not consistently occur on the lost or gained chromosome (Supplementary Fig. S3). Overall, our xenograft experiment demonstrated that even under ideal conditions (i.e., large tumors, a high percentage of total plasma volume, identical molecular characteristics, and a tumor-informed search using many targets), glioblastoma ctDNA detection varies. Regardless, targeting molecular events that increase variant abundance (e.g., amplification), exploiting the short fragment length bias of ctDNA, or both will increase glioblastoma ctDNA detection in humans. *P < 0.05.

Validation of TrueR sequencing

Because necrosis is common in glioblastoma, we reasoned that large (>10,000 bp), necrosis-derived cell-free DNA may be present in plasma, adherent to buffy coat cells, or both. Therefore, we searched for glioblastoma-derived somatic mutations in buffy coat DNA (i.e., buffy DNA), unsheared ccfDNA, and sheared ccfDNA. Although shearing ccfDNA extracted from plasma would theoretically enable access to necrosis-derived ccfDNA, apoptotic ccfDNA that already has a size profile compatible with short-read technologies (<700 bp)³² is also sheared, causing very short fragments (Fig. 2A). Using DNA from 14 healthy controls (Supplemental Table S4), True2 sequencing revealed more false positives in sheared ccfDNA than buffy DNA and unsheared ccfDNA (Fig. 2B), with more variability in the number of false positives between sheared ccfDNA samples (Fig. 2C). Hypothesizing that NRAs associated with true positives occur randomly in sequence read positions, while false positives are caused by error accumulation at the ends of reads due to incomplete adapter trimming, we integrated a position filter into True2 to produce TrueR, an improved NGS workflow (Supplementary Fig. S4). TrueR requires a true positive to have a portion of NRAs (i.e., count fraction) in the center portion of a read (i.e., window; Fig. 2D). Because sheared ccfDNA exhibited the greatest error, sheared ccfDNA from healthy controls was used to determine the optimal window size and count fraction to minimize false positives, which were determined to be 40% and 0.3, respectively (Fig. 2E). Using the same window size (40%) for buffy DNA and unsheared ccfDNA, we found that a count fraction of 0.1 yielded a similar false positive rate, consistent with their lower error profile (Fig. 2E). Using these optimal settings for each sample type, false positives were reduced on average to <1 across the entire 114 kb panel (Fig. 2F). Using glioblastoma tumor DNA, we confirmed that NRAs associated with TERT promoter mutations occurred throughout the read sequence, including the central 40% window, regardless of VAF (Fig. 2G). Moreover, the improved error suppression lowered the threshold for discerning true positives, thereby increasing sensitivity based on the original validation set for True2 (Fig. 2H)³¹. TrueR sensitivity for SNVs with a VAF ≥ 0.1% was 98.0%, and sensitivity for variants with a VAF < 0.1% was 47.4%.

Fig. 2. Validation of the position filter in TrueR sequencing.

Shearing ccfDNA generates shorter fragment lengths compared to unsheared ccfDNA as determined by paired-end NGS (A). Total false positives for the 114 kb panel were greatest in sheared ccfDNA (N = 7) compared to buffy DNA (N = 14) and unsheared ccfDNA (N = 7), regardless of the nonreference allele (NRA) count threshold used to define a true variant (B, points represent mean values). Using the original True2 workflow, a subset of sheared ccfDNA samples demonstrated ≥2 false positives, which did not occur in buffy DNA and unsheared ccfDNA (C). A cartoon (D) depicts the position filter that seeks to further suppress NGS-associated noise by requiring a portion of NRAs (i.e., count fraction) be present in the center portion of a read (i.e., window) to reduce false positives caused by the accrual of errors at the ends of reads. In (E), an optimal window size and nonreference allele fraction within the window were selected to minimize false positives for sheared ccfDNA (N = 7; window size: 40%, NRA fraction: 0.3) and buffy DNA (N = 14) and unsheared ccfDNA (N = 7; window size: 40%, NRA fraction: 0.1), where points correspond to mean values. Compared to True2, TrueR improved error suppression, allowing for a lower NRA threshold (≥4) to identify variants (F). In tumor DNA (G), NRAs associated with TERT promoter mutations were present across the entire read length regardless of variant allele frequency. The reduced threshold of TrueR improved variant detection at lower variant allele frequencies (H, lines are the mean from two replicates).

Tumor-informed detection of glioblastoma ctDNA

Using our 115 gene panel, TrueR profiled tumor somatic mutations from 49 glioblastoma patient tumors (8 patients had ≥2 samples from different tumor regions; Supplemental Table S5). Using a single sample for each patient, mutation VAF ranged from 0.05% to 94.7% (median, 0.26%), and 62.1% of mutations had a VAF < 1% (Supplementary Fig. S5A, Supplementary Data 1). Despite variant discovery across diverse VAFs, a median of only 5 mutations per patient were detected in tumor DNA, and 81.6% of tumor samples had <10 mutations, providing fewer targets per patient than the xenograft model (Supplementary Fig. S5B). TERTp (c.228 C > T, c.250 C > T) mutations were present in 89.8% of patients, and the most commonly mutated genes were PTEN, TP53, EGFR, and NF1 (Supplementary Fig. S5C). In patients with multiple biopsies, the unique mutation count increased significantly relative to a single tissue sample (Supplementary Fig. S5D). Similarly, gene-specific subclones were present in 36.7% of patients, which increased to 75.0% in patients with multiple tissue biopsies (Supplementary Fig. S5E). Using all available samples, gene-specific subclones were identified in 23 genes (20.2%), most commonly occurring in EGFR (22.4%) and NF1 (16.3%; Supplementary Fig. S5F). However, few genes on our panel included full exon coverage, preventing a comprehensive assessment of gene-specific subclones (Supplementary Fig. S6). Copy number variations (CNVs) in tumor DNA were also mapped, demonstrating characteristic glioblastoma alterations such as EGFR and PDGFRA amplification, trisomy 7/monosomy 10, and CDKN2A codeletion, among others (Supplementary Fig. S7).

The NRA counts corresponding to the 341 unique somatic mutations detected in tumor DNA (Supplementary Data 1) were ≤1 in buffy DNA and unsheared and sheared ccfDNA from 14 healthy controls. Therefore, the TrueR threshold for variant detection during the tumor-informed search was reduced from ≥4 to ≥2 to maximize sensitivity. Glioblastoma-derived somatic mutations were uncovered in buffy DNA, unsheared ccfDNA, and sheared ccfDNA (Fig. 3A), most commonly in TERTp and EGFR (Fig. 3B). Although PTEN was the second most commonly mutated gene in tumor DNA, PTEN mutations were not detected in blood. NRAs associated with TERTp mutations occurred in the center of reads regardless of VAF, supporting the use of the sequence position filter for true variant detection (Fig. 3C). VAFs of somatic mutations in blood ranged from 0.033% to 3.01%, with no difference in VAF between buffy DNA, unsheared ccfDNA, and sheared ccfDNA (Fig. 3D). All tumor somatic mutations found in buffy DNA were also detected (NRA count ≥1) in ccfDNA (i.e., shared), but a subset of ccfDNA mutations occurred uniquely (Fig. 3E), where unique was defined as a buffy DNA NRA count equal to 0 or a buffy DNA to ccfDNA count ratio <0.01. The latter criterion prevented an NGS error with an NRA count of one in buffy DNA from eliminating a corresponding variant in ccfDNA with a high NRA count (i.e., >100). Overall, our investigation of sheared ccfDNA did not improve the detection of glioblastoma ctDNA. Subsequent analyses were restricted to unsheared ccfDNA, enabling the inspection of insert length profiles associated with somatic mutations detected in ccfDNA. Regardless, our pursuit of necrotic ctDNA and the associated deep sequencing of buffy DNA revealed that unique ccfDNA mutations may support glioblastoma ctDNA detection during a tumor-naïve search.

Fig. 3. A tumor-informed search identifies glioblastoma-derived somatic mutations that occur uniquely in ccfDNA.

Using a tumor-informed search, glioblastoma-derived somatic mutations were identified in buffy DNA, unsheared and sheared ccfDNA (A). For glioblastoma patients with a somatic mutation detected in blood during a tumor-informed search (N), the most common genes with a variant are shown in (B). Nonreference alleles associated with the TERT promoter (TERTp) mutation occur within the center 40% of a read (C, pink box) in buffy DNA, sheared and unsheared ccfDNA, regardless of variant allele frequency (VAF). The corresponding VAF in tumor DNA is shown above each distribution. The VAF of glioblastoma ctDNA was broad (D). In (E), all glioblastoma-derived somatic mutations in buffy DNA co-occurred in ccfDNA (i.e., shared, blue) while a subset of variants in unsheared and sheared ccfDNA were unique (i.e., absent in buffy DNA, purple). The VAFs of variants co-occurring in buffy DNA and unsheared ccfDNA were strongly correlated (F, gray dashed unity line is shown). The distribution of VAFs between shared and unique unsheared ccfDNA variants was similar (G). However, the fragments associated with unique ccfDNA variants were biased towards shorter lengths compared to shared ccfDNA variants and ccfDNA from healthy controls (H). A variant associated with an amplification in tumor DNA was more likely to yield a mutation unique to ccfDNA (I). **P < 0.01, ns = not significant.

The VAF of somatic mutations in buffy DNA and unsheared ccfDNA were strongly correlated (Pearson’s r = 0.81, P < 0.001; Fig. 3F). The VAF of shared and unique ccfDNA mutations was similar (Fig. 3G). However, unique ccfDNA mutations exhibited a shorter overall fragment length compared to shared ccfDNA mutations and ccfDNA from healthy controls (Fig. 3H). Unique ccfDNA mutations were more commonly associated with a corresponding gene amplification in tumor DNA compared to shared ccfDNA mutations (Fig. 3I). Consistent with the xenograft data, mutations associated with an amplification improved ctDNA detection.

Detection of glioblastoma-derived somatic mutations using a tumor-naïve search

Next, TrueR was used for somatic mutation discovery during an unbiased search in buffy DNA and unsheared ccfDNA from healthy controls (N = 33, age 19–85 yrs) and participants with stroke (N = 10), low-grade glioma (N = 10 IDH-mutant astrocytoma, WHO grade 2–3; N = 10 IDH-mutant and 1p/19q-codeleted oligodendroglioma, WHO grade 2–3), and glioblastoma (N = 37) (Supplemental Table S4, S5). Somatic mutations detected in buffy DNA increased with age in all cohorts (Supplementary Fig. S8A, B), a finding echoed in ccfDNA mutation counts (Fig. 4A, B). For participants >60 years, buffy DNA mutation counts were similar between glioblastoma and healthy controls (Supplementary Fig. S8C), but the ccfDNA mutation counts were higher in glioblastoma compared to healthy controls (Fig. 4C). Affected genes in glioblastoma and healthy controls were similar, demonstrating the confounding effects of age-related somatic mosaicism on ctDNA detection (Fig. 4D).

Fig. 4. A tumor-naïve search for somatic mutations in buffy DNA and ccfDNA uncovers unique mutations and subclones specific to glioblastoma.

The number of somatic mutations discovered in ccfDNA increases with age in both non-tumor (A) and brain tumor (B, GBM = glioblastoma, LGG = low-grade glioma) participants. In older participants (>60 yrs), more ccfDNA mutations were present in glioblastoma patients than healthy controls (C). Although EGFR mutations were in a greater proportion of glioblastoma patients, the four most commonly mutated genes in ccfDNA from glioblastoma patients were also mutated in healthy controls (D). Unique ccfDNA mutations were significantly more common in glioblastoma (E). Increasing the nonreference allele (NRA) count threshold from four to five to define a ctDNA variant, reduced the detection of glioblastoma patients but yielded 100% specificity (E). Gene-specific subclones were uniquely present in glioblastoma (F). In (G), the genes associated with a unique ccfDNA mutation and subclone (S) from all participant cohorts are shown. A glioblastoma patient with co-occurring stage IV melanoma exhibited both unique ccfDNA mutations and a gene-specific subclone at high variant allele frequencies (VAFs, H). ***P < 0.001, *P < 0.05, ns = not significant; numbers in parenthesis identify sample size.

We observed two key differences between somatic mutations in ccfDNA associated with glioblastoma compared to the other cohorts. First, mutations unique to ccfDNA (i.e., absent in buffy DNA) occurred in 13 of 38 glioblastoma patients (34.2%) and 2 of 63 controls (3.2%; Fig. 4E). Increasing the NRA count threshold from ≥4 to ≥5 for TrueR variant detection reduced unique ccfDNA mutations in glioblastoma to 8 patients (21.1%) but eliminated such events in all 63 controls (0%, Fig. 4E). Second, gene-specific subclones occurred in 3 of 38 glioblastoma patients (7.9%) and none of the 63 controls (Fig. 4F). Although subclones occur regularly in glioblastoma³¹, ccfDNA subclones may be a combination of CHIP artifacts and glioblastoma ctDNA, and discerning the origin of each may not be feasible. Regardless, subclones were not observed in controls, thereby providing an additional molecular event to distinguish glioblastoma patients. Notably, the genes corresponding to glioblastoma ctDNA detection were distinct from those with mutations in the control cohort (Fig. 4G).

Because unique ccfDNA and subclonal ccfDNA events occurred with a low VAF (Supplementary Fig. S9), we investigated whether ccfDNA uniqueness and subclones were random observations rather than truly tumor-derived. We analyzed the ccfDNA from a glioblastoma patient (IDH-wildtype, TERT c.228 C > T positive, EGFR amplified) initially excluded from the analysis because a diagnosis of stage IV melanoma with unknown primary (BRAF wildtype, TERT c.250 C > T positive) was made 39 days after the histologic diagnosis of glioblastoma (Supplement 1). A glioblastoma tumor-informed search for three somatic mutations in ccfDNA identified TERT c.228 C > T with a VAF of 0.070% as the lone variant present. A blind search uncovered eight somatic mutations unique to ccfDNA (Fig. 4H). Six of the unique ccfDNA mutations exhibited high VAFs (>10%), inconsistent with glioblastoma ctDNA (Fig. 4H), such as the melanoma’s TERT c.250 C > T mutation, uniquely present in ccfDNA at a VAF of 11.8%. Importantly, two subclones of SPTA1 unique to ccfDNA were also identified, both with a VAF > 19% (Fig. 4H). Therefore, somatic mutations unique to ccfDNA and subclones in ccfDNA are true molecular events that can be leveraged to detect ctDNA from glioblastoma and other cancers.

Glioblastoma-derived CNVs in ccfDNA

Identification of CNVs, which are common in glioblastoma (Supplementary Fig. S7) and occur with or without a somatic mutation, offers an alternative approach to variant-based ctDNA detection. Because our xenograft data and tumor-informed search demonstrated that gene amplifications in tumor DNA improved ctDNA detection and were associated with a shorter ccfDNA fragment length, we hypothesized that glioblastoma-derived CNVs occur in shorter ccfDNA inserts. We used an in silico selection of incremental ccfDNA lengths to profile the CNV of EGFR in a patient with a large amplification in tumor DNA and co-occurring somatic mutation detected in ccfDNA. We observed that fragments <90 bp demonstrated the strongest evidence of the EGFR amplification (Fig. 5A). A significant correlation in copy number between tumor DNA and unsheared ccfDNA <90 bp from glioblastoma patients provided further support (Fig. 5B). Subsequently, we compared all participant cohorts using EGFR copy number from ccfDNA <90 bp, revealing 5 of 38 (10.5%) glioblastoma patients exceeded the upper limit of copy number from all comparison cohorts (Supplementary Fig. S10). All five patients had an EGFR amplification in tumor DNA, establishing ccfDNA <90 bp as an effective strategy for detecting glioblastoma ctDNA.

Fig. 5. Copy number variations (CNVs) in short (<90 bp) ccfDNA fragments distinguish glioblastoma patients from controls.

In (A), measuring copy number across ccfDNA fragment lengths using an EGFR somatic mutation amplified in glioblastoma tumor DNA found the greatest evidence of CNV occurred with inserts <90 bp. The uptick for all sizes (gray arrow) arises from the inclusion of ctDNA associated with the dinucleosome (i.e., >200 bp). The dashed gray line is euploidy. A significant correlation was present between tumor EGFR copy number and copy number of EGFR in ccfDNA <90 bp (B). A heatmap (C) of the 17 genes studied for CNV in ccfDNA <90 bp shows CNVs that exceed the extremes of the control cohorts occurring in isolation or in combination in more than half of glioblastoma patients. The gray box identifies genes predominantly associated with copy number loss. For glioblastoma patients with (+) or without (-) an EGFR or PDGFRA amplification in tumor DNA, the detection rate in ccfDNA using fragments <90 bp is shown in (D). A subset of glioblastoma patients had copy number loss of TP53 in ccfDNA (E). TP53 copy loss commonly occurred in conjunction with other copy number alterations (F), where font size indicates the number of glioblastoma patients and shaded circles indicate genes with CNVs that also occurred independently of TP53. GBM = glioblastoma. *P < 0.05, ns = not significant.

We also sought to identify CNVs that distinguished glioblastoma patients from the other cohorts. This was complicated by inter-subject variation in genes with few hotspot panel positions, particularly since the selection of ccfDNA fragments <90 bp reduced read-depth (Supplementary Fig. S11). Thus, we studied the 17 genes with the lowest quartile of variance from the 35 healthy controls and the highest quartile of gene positions (Supplementary Fig. S11). CNVs were evident in a subset of genes but absent in others (Supplementary Fig. S12). Overall, 20 of 37 (54.1%) glioblastoma patients demonstrated a CNV gain/loss that exceeded the upper/lower bounds of the controls in at least one gene, and eight patients (21.6%) exhibited alterations in multiple genes (Fig. 5C). For EGFR and PDGFRA, ≥20% of patients with a tumor DNA amplification were detected (Fig. 5D), and a corresponding somatic mutation was absent in 4 of 6 patients (66.7%). Unexpectedly, 9 of 37 (24.3%) patients demonstrated copy number loss of TP53 (Fig. 5E), which co-occurred with a second gene gain/loss in 7 of 9 (77.8%) patients (Fig. 5F). None of the 9 patients demonstrated copy number loss of TP53 in tumor DNA (Supplementary Fig. S7), which supports spatial heterogeneity of copy number in glioblastoma, as previously described^22–24. Collectively, our data support CNVs in ccfDNA <90 bp as an alternative pathway to ctDNA detection.

The enhanced signal from an amplified somatic mutation improves ctDNA detection

Next, we sought to identify patient and tumor features affecting ctDNA detection. A neuroradiologist blinded to non-imaging data (e.g., ccfDNA concentration, somatic mutations, etc.) quantified tumor volume, estimated percent necrosis, and determined presence/absence of hemorrhage and contrast-enhancement using pre-operative clinical magnetic resonance images (Fig. 6A, Supplemental Table S5). Necrotic content (Fig. 6B) and hemorrhage (Fig. 6C), but not tumor size, increased ccfDNA concentration (Supplementary Fig. S13). Necrotic content was associated with age, but tumor size, hemorrhage, and ccfDNA concentration were not (Supplementary Fig. S14). Notably, ccfDNA concentration did not discern glioblastoma patients from other cohorts (Fig. 6D). In fact, stroke patients demonstrated significantly greater ccfDNA concentrations than all others, while no other group differed significantly.

Fig. 6. Somatic mutations on an amplified allele improves ctDNA detection.

In (A), magnetic resonance images of glioblastoma (arrowheads) in three patients showing necrosis (asterisk, left panel), contrast enhancement (CE; green arrow, left panel), and minimal contrast enhancement (yellow arrow, middle panel) on CE-T1w images, and hemorrhage (blue arrow, right panel) on gradient echo (GE). A mild association was present between necrotic content and concentration of circulating cell-free DNA ([ccfDNA], B). Intratumoral hemorrhage increased the concentration of ccfDNA (C). Stroke patients had the highest concentration of ccfDNA among the study cohorts (D). Patients with (+) and without (-) detection of circulating tumor DNA (ctDNA) were not distinguishable using image-based or other metrics (E). Tumor DNA with a gene amplification (e.g., EGFR and PDGFRA) and a co-occurring single nucleotide variant (SNV; i.e., gene (+), snv (+)) was more likely to yield ctDNA detection compared to non-amplified SNVs (F). GBM glioblastoma, LGG low-grade glioma. ***P < 0.001, **P < 0.01, ns = not significant.

Results from the tumor-informed search of sheared and unsheared ccfDNA were used to maximize sample size and sensitivity to identify factors that affect ctDNA detection. The negative ctDNA control cohort consisted of unsheared ccfDNA with no molecular evidence of glioblastoma ctDNA (i.e., absent somatic mutation detection during tumor-informed and tumor-naïve search, no copy number variation in ccfDNA). Two patient samples (one ctDNA-positive, one ctDNA-negative) were removed because of low percent ccfDNA, yielding 23 ctDNA-positive and seven ctDNA-negative patients (Supplementary Data 1). Although larger tumor size and presence of hemorrhage were associated with ctDNA detection (Fig. 6E), a multiple logistic regression analysis did not identify a metric differentiating ctDNA-positive and ctDNA-negative patients (Supplemental Table S6). All variables showed a substantial overlap in values between positive and negative ctDNA patients, and a clear patient or tumor feature governing ctDNA detection was absent. Notably, a difference in overall survival between ctDNA-positive and ctDNA-negative patients was not observed (Supplementary Fig. S15).

Because our xenograft model and tumor-informed search demonstrated amplified somatic mutations were readily detectable in ccfDNA, we hypothesized that the amplification of a mutation increases ctDNA detection. We studied somatic mutations in EGFR and PDGFRA because of the large amplifications identified in tumor DNA from a subset of patients. Fifty mutations from 29 patients were classified into three categories (Supplemental Table S7): (1) gene amplification with a co-occurring amplified mutation (N = 12; e.g., 6-fold amplification, VAF of somatic mutation >10%), (2) gene amplification with a co-occurring non-amplified mutation (N = 17; e.g., 6-fold amplification, VAF of somatic mutation <2%), and (3) somatic mutation without gene amplification (N = 23). A somatic mutation occurring on an amplified allele was significantly more detectable compared to the other categories (Fig. 6F). However, in one patient with a large EGFR amplification (G46, log₂ = 7.25) and three amplified EGFR subclones, only one of the subclones was detected in ccfDNA, consistent with tumor heterogeneity and detection variability. Notably, no somatic mutations occurring in a minor subclone or on the opposite allele of an amplification were detected. Overall, data from our xenograft model, tumor-informed search, and CNV analysis demonstrate that amplifications with or without a co-occurring somatic mutation improve glioblastoma ctDNA detection.

Discussion

We have identified molecular features that enable glioblastoma ctDNA detection without a tumor-informed search. Using TrueR in combination with deep sequencing of buffy DNA and ccfDNA uncovered two variant-based ccfDNA characteristics that define ctDNA: (1) uniqueness (i.e., absent in buffy DNA), and (2) gene-specific subclones. We demonstrated that both features can be effectively leveraged to support differentiating glioblastoma from low-grade glioma, stroke, and natural aging. Additionally, both events extended beyond glioblastoma to a metastatic melanoma patient, providing compelling evidence for a broader use in non-invasive cancer diagnostics. We also established a panel-based, variant-agnostic method for glioblastoma ctDNA detection via the analysis of CNVs from short ccfDNA fragments (<90 bp). Though we were unable to distinguish tumor and tumor-related features (except for amplified somatic mutations) that govern glioblastoma ctDNA detection, our approach demonstrates the presence of glioblastoma ctDNA in plasma, and the detection rate will strengthen with methods that further improve the signal-to-noise ratio.

The source of somatic mutations in the buffy DNA of glioblastoma patients during a tumor-informed search, which included detection of classic TERT promoter mutations, was most likely circulating tumor cells (CTCs). The combination of several observations supports this conjecture. Glioblastoma CTCs have been previously reported in humans³³. CTCs, like WBCs, undergo apoptosis and generate ccfDNA³⁴. Thus, the strong correlation in VAFs between concurrently detected (i.e., shared) somatic mutations in buffy DNA and ctDNA was consistent with the presence of CTCs in buffy DNA diluted by WBCs, and CTC-derived ccfDNA proportionately diluted by WBC-derived ccfDNA. Additionally, the fragment length profile of ctDNA with a shared somatic mutation was concordant with apoptosis, not necrosis, and the size distribution was similar to WBC-derived ccfDNA. In contrast, the length profile of ctDNA associated with unique somatic mutations (i.e., absent in buffy DNA) was shorter, echoing prior studies that have documented the shorter fragment length of ctDNA compared to ccfDNA from non-cancerous cells (mononucleosome peak of ~140 bp vs. ~167 bp, respectively)^35,36. Since unique ctDNA and shared ctDNA both arise from tumor cells, the modest difference in fragment length may be a temporal artifact of extracellular exposure to deoxyribonucleases (DNases). Specifically, the additional exposure of ctDNA to DNases in the tumor microenvironment before entering the circulation would reduce fragment size compared to ctDNA that enters the circulation directly via CTC apoptosis³⁷. To investigate the discrepancy in ctDNA fragment length, future study designs would benefit from an integrated evaluation of glioblastoma-derived CTCs and ctDNA, as previously accomplished in metastatic breast cancer, where a strong agreement between mutational profiles of CTCs and ctDNA has been shown³⁸.

The clinical necessity of distinguishing between ctDNA and ccfDNA variants derived from somatic mosaicism of WBCs was established in a gastric cancer study that showed unique ccfDNA mutation detection was associated with a higher recurrence risk and shorter survival³⁹. Our study demonstrates the importance of separating ctDNA from CHIP artifacts in ccfDNA diagnostics by showing the commonness of hematopoietic-related variants in non-cancer patients. Additionally, we extend the molecular criteria for ctDNA detection to include gene-specific subclones, which are common in diffuse gliomas^31,40. Their absence in low-grade glioma ccfDNA suggests the more robust blood-brain barrier disruption by glioblastoma generates a greater signal⁴¹, distinguishing glioblastoma from low-grade glioma, as well as stroke and somatic mosaicism. Future studies could improve the detection of gene-specific subclones with full exon coverage of genes rather than the hotspot approach used herein.

Although we illustrated the limitations of using ccfDNA fragment lengths to detect somatic mutations in low abundance, we successfully demonstrated the use of short ccfDNA fragments to isolate glioblastoma-derived CNVs. Our panel-based approach found that fragments <90 bp were most strongly enriched for ctDNA, allowing us to observe copy number gains and losses in glioblastoma that exceeded comparison cohort limits. The reduction of TP53 copy number in ccfDNA <90 bp was unexpected, particularly since TP53 loss was not commonly observed in corresponding glioblastoma tumor DNA. However, prior studies have demonstrated TP53 loss in 19-23% of glioblastoma patients^42,43, which is similar to the ccfDNA detection rate reported here. Moreover, TP53 loss varies spatially in glioblastoma^24,42, and tumor phylogeny indicates TP53 loss is a late-phase event²⁴. In contrast, EGFR amplification has been described as an early-phase event with intratumoral homogeneity²⁴, further supporting our observed correlation between EGFR copy number in ccfDNA and tumor DNA. Thus, our detection of TP53 loss in ccfDNA may reflect the broader sampling of tumor DNA afforded by ccfDNA, previously described for somatic mutations in heterogeneously diverse cancers, such as multinodular hepatocellular carcinoma⁴⁴.

Our data indicate gene amplifications in glioblastoma improve ctDNA detection. EGFR and PDGFRA are the most commonly amplified genes in glioblastoma, occurring in approximately 50% and 20% of patients, respectively^45–47. Thus, ccfDNA as a non-invasive means to diagnose glioblastoma may be biased towards the subset of tumors harboring an amplification in one of these genes. However, glioblastoma gene amplifications have been shown to occur in a variety of other genes (e.g., GLI, MDM2, PIK3CA, etc.)^48,49, and detection in ccfDNA may be improved by full exon coverage rather than a hotspot approach. Beyond gene-specific amplifications, glioblastoma is also characterized by chromosomal instability. Trisomy 7, monosomy 9p, monosomy 10, and trisomy 19 are the most commonly altered chromosomes in glioblastoma, with at least one event occurring in >75% of tumors^50,51. Inclusion of genes from these chromosomes for the detection of copy number changes, singularly or in combination, may improve the panel-based detection of glioblastomas with or without gene amplification. Additional markers of the genomic instability that characterizes glioblastoma, such as segmental aneuploidies and structural variants^52,53, are non-uniform, heterogeneous, and generally tumor-specific, presenting a challenge to panel-based technologies in the absence of a bespoke, tumor-informed search. However, low-pass whole genome sequencing (LP-WGS), the use of ultra-low read depths (e.g., <1X) to survey the entire genome for aneuploidy, may be a complementary tumor-naïve approach to improve detection of ctDNA from non-amplified glioblastomas. LP-WGS has been shown to detect glioblastoma-derived cell-free DNA in cerebrospinal fluid⁵⁴. Although adaptation of LP-WGS to the relatively low ctDNA content of glioblastoma may prove challenging, in silico selection of cell-free DNA fragments in plasma between 90 and 150 bp improves LP-WGS detection of cancers with a high ctDNA content³⁶. Thus, the integration of in vitro strategies into workflows that enrich for short ccfDNA fragments, such as single-stranded DNA library preparation methods^55,56, may be necessary to support LP-WGS detection of glioblastoma ctDNA.

Detection of glioblastoma ctDNA remains challenging. Because we did not identify a clear imaging-based, demographic, or ccfDNA-based governor of ctDNA detection, factors affecting ctDNA shedding and dilution merit consideration. For example, glioblastoma has a lower apoptotic rate compared to other primary brain tumors, and expression of survivin, a member of the inhibitor of apoptosis family, has been associated with a worse prognosis⁵⁷. PTEN, a regulator of apoptosis⁵⁸, was the second most commonly mutated gene in our cohort, but PTEN variants were not detected in ccfDNA. Thus, analysis of apoptotic activity in tumor tissue to correlate with the presence/absence of ctDNA may provide a stronger understanding of biological determinants that affect ctDNA shedding. The inclusion of tumor cell mitotic activity may also prove beneficial because non-glioblastoma studies, such as colorectal and non-small cell lung cancers, have demonstrated a correlation between more proliferative tumors and ctDNA content^59,60. Systemic inflammation, which varies between glioblastoma patients⁶¹, increases ccfDNA from normal cells⁶², diluting the signal from ctDNA. Therefore, the collection of systemic inflammatory markers concomitant with ccfDNA acquisition may support a patient-specific understanding of the presence/absence of glioblastoma ctDNA detection. To overcome biologic mechanisms that complicate glioblastoma ctDNA detection, we note the emergence of technologies, such as focused ultrasound⁶³, that increase the release of glioblastoma ctDNA. Exogenously stimulated ctDNA release, combined with the molecular features specific to ctDNA described herein, may be the necessary next step to broadly diagnose glioblastoma non-invasively through enhancement of the ctDNA signal-to-noise ratio.

There are several limitations to our study to be considered. Shearing the ccfDNA from a subset of glioblastoma patients further reduced the sample size for downstream analyses. For example, our study of copy number variation was limited to the unsheared ccfDNA cohort. Sheared ccfDNA samples could not be included since enzymatic shearing altered the fragment length of apoptotic and necrotic ccfDNA. However, the design of future studies seeking to detect glioblastoma ctDNA may benefit from our observation that shearing ccfDNA did not improve ctDNA detection, despite the rich necrotic component of glioblastoma. The absence of an association between ctDNA detection and overall survival should be interpreted with caution. Glioblastoma survival is affected by a variety of factors (e.g., age, Karnofsky performance status score, extent of resection, etc.)⁶⁴ that could not be accounted for given the sample size. The role, if any, of pre-operative ctDNA detection on prognosis or treatment response may be better assessed in a more homogenous and/or larger glioblastoma population. Each control cohort was also relatively small. Thus, the study design did not include partitioning each group into training and testing cohorts to evaluate each molecular feature’s performance characteristics in distinguishing glioblastoma. Nevertheless, our identification of new molecular features that distinguish ctDNA continues the advancement of glioblastoma ctDNA detection.

Methods

GFP-ffluc MGG8 cells and orthotopic mouse xenograft model

MGG8 stem-like cells were a gift from Dr. Hiroaki Wakimoto at Massachusetts General Hospital (Boston, MA)⁶⁵, and cultured as previously described³¹. Cells were transduced with a pLL-EF1a-rFLuc-T2A-GFP-mPGK Puro Lentivirus (SBI, Inc.) supplemented with 5 µg/mL polybrene (EMD Millipore). Twenty-four hours post-transduction, cells were treated with 5 ng/mL puromycin prior to isolation of GFP+ cells by cell sorting on an Aria BD flow cytometer (BD Biosciences, Inc.). Afterwards, sorted cells were grown in the presence of antibiotics and antimycotics (Gibco) for 2-3 passages, before returning to normal growth media. Cell cultures were also tested for mycoplasma (Lonza).

All animal procedures were performed in compliance with the University of Washington IACUC guidelines (IACUC protocol 4370-01). For xenograft implantation, NSG mice (8–12 weeks old) were anesthetized with Avertin (0.3-0.8 mg/kg). While anesthetized, the scalp was shaved and sterilized with successive betadine and 70% isopropyl alcohol swabs. Animals were placed in a stereotactic frame to implant 1.5–2 × 10⁵ GFP⁺-MGG8 cells intracranially into the striatum (Bregma +0.5 mm, lateral 2.0 mm, depth 2.5–3 mm). After surgery, animals were given buprenorphine (1.2 mg/kg) as an analgesic. Animals were monitored and euthanized by an AVMA-approved method (Avertin-overdose, 2 mg/kg, and clipping of the diaphragm) for blood collection and brain harvest when body weight was reduced by 20% (5–6 weeks post-implantation). Blood was collected via the vena cava, then animals were transcardially perfused with saline before fixation with 4% paraformaldehyde (ThermoFisher, Inc.). Blood was collected in EDTA microtainers (BD Biosciences, Inc.) and centrifuged at 2500 × g for 10 min at 4˚ C to isolate plasma. Total volume of plasma removed was estimated using a hematocrit of 50% and 78 mL/kg body weight at the terminal procedure for total blood volume. Plasma samples were stored at −80˚ C and coded to ensure subsequent analysis was blinded.

A total of eight NSG mice were used as controls (one male, two female) and in the xenograft model (three male, two female). After euthanasia and fixation, brain tissue was harvested and post-fixed in 4% paraformaldehyde overnight. Each brain was equilibrated to 10%, 20%, and then 30% sucrose with 24 h incubations at each step, sectioned (60 µm sections) on a sliding microtome, a 1-in-6 series of the tumor (though the anterior-posterior axis) was stained with Tomato lectin-594 (Vector Labs) and DAPI (ThermoFisher, Inc.), and imaged on a Leica SP8 confocal microscope. Confocal micrographs were compiled and visualized with ImarisViewer (Oxford Instruments). A separate 1-in-6 series was mounted onto glass slides for tumor DNA collection in MGG8 xenografts and normal brain tissue in age-matched controls (NSG mice).

A neuropathologist outlined regions of dense tumor on glass slides of xenograft tissues, followed by macrodissection. Corresponding anatomic regions of brain tissue from control mice were similarly identified and macrodissected. Tissue DNA was extracted using the QIAmp DNA Blood Mini Kit (Qiagen). ccfDNA was extracted from mouse plasma using the QIAmp Circulating Nucleic Acid Kit (Qiagen) following the manufacturer’s protocol with minor adjustments for the low volume: (1) PBS was used to increase plasma volume to 500 µL, and (2) ccfDNA was eluted from the column using 30 µL of IDTE applied twice.

Droplet digital PCR (ddPCR)

To detect human ccfDNA, we used an IDH1 R132H assay from Bio-Rad (dHsaMDS651418010; Bio-Rad Laboratories, Hercules, CA). This is a FAM/HEX MUT/WT assay, but only the wildtype allele was used. To detect wildtype mouse ccfDNA we used an in-house design detecting mouse IDH1:

MmIDH1_FWD: TAGTGACAGGCTGGGTAAA

MmIDH1_REV: ATGCTCACTCAAATGACTTACT

Probe: CATCATTGGCCGACATGCATATG (HEX-ZEN-IABlq)

The ddPCR tests were performed in a single-channel amplitude multiplex design in ddPCR Multiplex Supermix (Bio-Rad) on a QX200 ddPCR system (Bio-Rad). Mouse primers were added at 1000 nM final concentration, and mouse HEX-mIDH1 probe at 400 nM final concentration. The Bio-Rad human IDH1 assay was added at 0.8X final input (720 nM primers, 200 nM probes). HindIII restriction enzyme (NEB) was added at 5U per reaction. PCR cycling was performed with 45 cycles of 15 s at 94 °C and 60 s at 60 °C. QX Manager software was used for data analysis. Data were gated using a single channel (HEX) multi-amplitude approach. Copy numbers of human and mouse IDH1 and % human ctDNA in mouse ccfDNA were calculated from absolute observed positive droplet counts in each relevant analysis sector. Before testing on animal samples, performance of primer/probe sets was evaluated using (1) 100% sheared human genomic DNA, (2) 100% sheared mouse genomic DNA, and (3) a serial dilution of sheared human genomic DNA into sheared mouse genomic DNA (50, 25, 10, 5, 2.5, 1%) test at 1 ng input per well (Supplementary Fig. S16). Subsequently, human and mouse IDH1 R132 wildtype was measured using 3 µL of the ccfDNA elute, representing 10% of the total ccfDNA extracted from each animal.

Patient recruitment, sample collection, and DNA extraction

All study participants were adults (≥18 years) recruited from Huntsman Cancer Institute and the University of Utah. Patients with no prior history of cancer with (1) imaging evidence of a mass concerning for a primary brain tumor, (2) an ischemic stroke within the past 30 days, or (3) healthy individuals at the outpatient lab for routine blood work were recruited for enrollment. All study procedures were approved by the Institutional Review Board at the University of Utah (IRB #00089989), and the research was conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent.

Brain tissue collection from tumor patients was consistent with our prior description³¹, and all tissue samples were snap-frozen in liquid nitrogen. Whole blood was collected in Streck BCT tubes via venipuncture. For brain tumor patients, blood was collected within 24 h before surgery. Whole blood was centrifuged at 1900 × g for 20 min followed by isolation of the buffy coat and plasma. The isolated plasma was centrifuged at 12,100 × g for 10 min at 4˚C to remove residual cellular debris. The plasma supernatant was removed and stored at −80˚C along with the buffy coat within 24 hours of collection.

Buffy coat DNA and frozen tumor tissue DNA were extracted, consistent with our prior procedures³¹. The QIAmp Circulating Nucleic Acid Kit was used to extract ccfDNA from plasma following the manufacturer’s protocol except the ccfDNA was eluted in IDTE applied twice to the column. Sample quality was assessed and percent ccfDNA documented using the 4150 TapeStation (Agilent) and Cell-free DNA Screen Tape (Agilent) system. Low-quality ccfDNA was defined as percent ccfDNA <20% by TapeStation analysis. A Qubit 2.0 fluorometer (Thermo Fisher Scientific) was used to measure the concentration of all extracted DNA and ccfDNA.

Library preparation and NGS

DNA and ccfDNA libraries were generated using duplex adapters (IDT; Coralville, IA) with dual barcodes and dual indexes consistent with our prior procedures³¹. Library input for buffy/tumor DNA was 100 ng. For ccfDNA, the maximum amount of ccfDNA available up to 100 ng was used. Mouse ccfDNA and human ccfDNA libraries were generated without enzymatic shearing using the NEBNext Ultra II DNA kit following the manufacturer’s instructions. A subset of human ccfDNA libraries underwent enzymatic shearing using the NEBNext Ultra II FS DNA kit following the same protocol as buffy/tumor DNA.

A single hybridization capture of library DNA (500 ng) was accomplished in a single-plex reaction using a custom-designed capture-enrichment panel (115 genes, 114 kb, Supplemental Table S2, Supplement 2, IDT) as previously described³¹. All mouse-related samples and human participant samples underwent paired-end (151 × 151) NGS using the Illumina NovaSeq 6000 (v1.5) and Illumina NovaSeq X (10B reagent kit), respectively. For healthy control samples used in TrueR development, FASTQ files were trimmed to 160 million total paired reads after merging data from two sequencing runs. For all other samples, a target of >100 million total paired reads was used during sample pooling.

Targeted search for MGG8 variants

Duplex collapsed consensus sequences (GRCh37) were generated from FASTQ files using FGBIO as previously described³¹. Potential SNVs (somatic mutations and single nucleotide polymorphisms) were identified from NRAs with a VAF > 10% and a minimum 200X duplex consensus read depth in cells. Variants subsequently found in all five tumor DNA samples with a VAF > 10% and ≤1 NRA in four control mouse samples (three brain tissues, one plasma sample) were considered potential targets. The final list of targets included only a single variant from each gene with the greatest mean VAF from cells and tumors (Supplemental Table S3). For comparison of NRA counts and VAF, genes in the euploid state on copy number analysis (log₂ value between 0.485 and 1.485) were used to mitigate the effects of amplification, monosomy, and trisomy. For comparison of NRA counts and copy number variation, variants with a VAF between 20% and 70% were included to isolate the heterozygous state. In mouse control samples, variants with a VAF > 40% and 200X minimum read depth absent in cell DNA were considered unique to mouse DNA. For comparison of insert lengths, PDGFRA chr4:55141055 A > G (human) and RPL5 chr1:93301869 C > T (mouse) variants were used because of their relative abundance in ccfDNA.

Bioinformatics

Duplex collapsed consensus sequences (GRCh37) using the True2 workflow were generated using a position mask to exclude error-prone locations (Supplement 3) as previously described³¹. The output from the True2 workflow was passed through a position filter to remove potential variants without sufficient NRA support from the central portion of reads (Supplement 4). Variants that passed the combined True2 and position filter workflow (i.e., TrueR) were annotated using Variant Effect Predictor⁶⁶ as previously described³¹.

Copy number analysis

Singleton collapsed consensus sequences (GRCh37) using FGBIO were generated from FASTQ files trimmed to 50 and 100 million total paired reads for tumor DNA and ccfDNA, respectively, followed by copy number analysis as previously described³¹ except for minor modifications which included: (1) normalization to median read depth for each sample rather than the median value across all samples, (2) for EGFR, only exons 8–28 were used to mitigate EGFRvIII mutation presence, absence, and heterogeneity on copy number, and (3) gene-specific adjustment based on the mean value from young healthy controls. The latter adjustment accounts for differences in read-depth variability between genes.

Brain tumor image analysis

Pre-operative brain magnetic resonance images (MRIs) were assessed on each patient by a board-certified neuroradiologist with a certificate of added qualification in neuroradiology with over 10 years of experience, blinded to molecular and pathologic results. There was some heterogeneity in imaging acquisition, with many patients having their baseline, pre-operative MRI examinations performed at varying institutions with inconsistent imaging techniques and different magnet strengths (1.5 T and 3 T). Therefore, imaging assessments were performed on sequences available on every patient’s pre-operative MRI examination following minimum standards for brain tumor imaging, including T2 fluid-attenuating inversion recovery (FLAIR), T2, diffusion-weighted imaging, T2*-weighted imaging, and post-contrast T1 imaging⁶⁷. For each patient, the mass was measured in perpendicular diameters on a single slice on the axial T1-post contrast sequence for enhancing tumors and axial T2/FLAIR for non-enhancing tumors, according to Response Assessment in Neuro-Oncology 2.0 criteria⁶⁸. The longest craniocaudal dimension was also measured to calculate volume. In those with multifocal or multicentric neoplasms, the location of the largest mass was recorded and measured. The presence of enhancement and hemorrhage was recorded as a dichotomous variable (present/absent). The percentage of the total mass that was necrotic was estimated from the necrotic portion of the total mass.

Statistical analysis

All statistical tests were performed with GraphPad Prism 10 (version 10.1.0). The independent and paired t-test were used for comparisons between groups and paired data, respectively. Pearson’s correlation was used to evaluate associations between groups. Fisher’s exact test was used for categorical data. A multiple logistic regression analysis was used to evaluate metrics associated with the presence vs. absence of ctDNA detection. Differences in overall survival were evaluated with Kaplan-Meier curves and the log-rank (Mantel-Cox) test. Results were considered statistically significant for P < 0.05.

Supplementary materials

A single merged PDF contains supplementary information in the following order: (1) supplemental tables, (2) supplemental figures, (3) a clinical vignette of the glioblastoma patient concurrently diagnosed with metastatic melanoma has been included as ‘Supplement 1’, (4) the BED file used during analysis has been included as ‘Supplement 2’, (5) the blacklist file used to exclude specific genomic coordinates based on error has been included as ‘Supplement 3’, (6) commented Python scripts for the TrueR position filter have been included as ‘Supplement 4’, and (7) the title and legend for Supplementary Data 1.

Author contributions

H.U., D.S., S.H., H.B., M.P., and R.J. conceptualized the study and developed the methodology. H.U., D.S., and R.J. supervised the study. H.U., D.S., S.H., H.B., M.P., H.C., and R.J. provided resources to conduct the study. H.U., S.M., R.M., M.K., J.M., L.B., D.C., C.S.D., K.A., S.P., H.C., and R.J. recruited participants and performed study enrollment. H.U., A.F., D.N., and YZ provided bioinformatic software used in the study. C.J.D. was the study neuropathologist, and H.B. was the study neuroradiologist. H.U., D.S., H.B., C.J.D., S.H., L.B., B.H., and S.S. performed the investigation. H.U. and A.T. conducted the statistical analyses. H.U. wrote the main manuscript text with specific components written by D.S., S.H., and H.B. All authors reviewed the manuscript.

Data availability

Data used to compile results and figures are provided as an Excel spreadsheet in the Supplemental Data Files as ‘Supplementary Data 1’. All FASTQ files are available from the Sequence Read Archive (SRA) database (https://www.ncbi.nlm.nih.gov/sra/PRJNA1327960).

Competing interests

Author S.H. is an employee of ARUP Laboratories, a commercial company, but declares no non-financial competing interests. All other authors declare no financial or non-financial competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01496-w.