Limit of detection of ddPCR system

To develop and test our LB monitoring method, we initially conducted a limit of detection experiment for H3F3A K27M and BRAF V600E mutation. We spiked the DNA of a medulloblastoma cell line (D425Med) with decreasing concentration of either H3F3A K27M or BRAF V600E mutated DNA and reached a limit of detection of 0.0025% for H3F3A K27M and 0.01% for BRAF V600E, respectively (Supplementary Fig. 2 and 3). For future analyses, we kept the same threshold settings for the channels.

Patient characteristics and treatment

We analyzed 35 patients with a median age of 9.3 years (range 0–33, Fig. 1a–b, Supplementary Table 1,) 27 cases were classified as grade 3/4 glioma (HGG) tumors and 8 as grade 1/2 tumors (LGG). In 22 patients, H3F3A K27M was diagnosed, 12 had BRAF V600E mutation, and one was identified with both mutations. 23 patients had a biopsy, in eight patients, a partial resection or near total resection was performed, and in four patients, a complete resection was achieved. First-line chemotherapy was temozolomide in 23 cases, while seven received targeted therapy with trametinib/dabrafenib, vemurafenib, nivolumab/nimotuzumab or dendritic cell vaccination. Five patients were treated with carboplatin/vincristine. Two patients did not receive systemic therapy. All except eight received at least one course of radiation therapy and in twelve cases, re-irradiation therapy was conducted. The median overall survival of DMG H3K27M diagnosed patients was 22.18 months (Supplementary Fig. 4a). For BRAF V600E, we divided the cohort into LGG (WHO grade 1 and 2) and HGG (WHO grade 3 and 4) and reached a 5-year overall survival of 100% in the LGG group (n = 3), whereas the HGGs (n = 4) displayed a median survival of 93 months (Supplementary Fig. 4b). The BRAF V600E MAF detection in plasma samples showed no significant differences between WHO grades 1–4 (shown in Supplementary Fig. 4c). In total, 23 patients of the cohort had succumbed to their disease, while the remaining 12 patients were alive at the time of analysis.

Validation of H3F3A K27M and BRAF V600E mutation in tumor tissue vs CSF and plasma

First, we analyzed the tumor tissue of 20 patients in our cohort and isolated the genomic DNA (11 H3F3A K27M, 8 BRAF V600E and 1 H3F3A K27M/BRAF V600E double mutation) to validate our ddPCR method. All 20 tumor tissues displayed the known mutations with mean MAF of 34% in H3F3A K27M (ranging from 8 to 80%) and 25% in BRAF V600E (ranging from 4 to 40%), respectively (Fig. 1a). Subsequently, we screened our LB samples for the nearest time point to the surgery date where tumor was measured in MRI (CSF and plasma, if available, showed in Supplementary Table 3) and analyzed the matched LB MAF profile. In all CSF samples, we were able to detect H3F3A K27M with a mean value of 9% (ranging from 1.8 to 60%). In plasma samples, four of five patients showed the H3F3A K27M mutation with a mean value of 0.05% (range 0.02–0.3%) and all four patients had the BRAF V600E mutation with a mean of 0.1% (ranging 0.02–0.2%, Fig. 2b–d). For BRAF V600E, we had one matched CSF sample at diagnosis and detected a mean value of 3.12% (Fig. 2e). However, no other matched samples at diagnosis of BRAF V600E-positive cases were available. To validate the sensitivity of the detection system, we included CSF samples of other pediatric brain tumor entities (medulloblastoma n = 5, DMG/pons glioma H3F3A WT n = 3, pilocytic astrocytoma n = 2, HGNET-BCOR n = 1, ependymoma n = 1, LGG n = 1, GBM IDH1 MT n = 1,) and a serum pool of 20 non-tumor bearing pediatric patients.

Fig. 2

Quality assessement overview of liquid biomarkers: a mutation allele frequency (MAF) detection in different analytes, H3F3A K27M detection in b CSF vs tumor tissue, and c) plasma. BRAF V600E detection in d plasma and e CSF vs tumor tissue. Asterisks indicate significance (one-way Anova; ***P < 0.001 analyzed with GraphPad Prism), error bars indicate mean ± S.D.; n.d. not detected in the sample. f Specificity and sensitivity of liquid biomarkers in different analytes compared to non-tumor controls. Sensitivity, specificity and predictive values were calculated according to Trevethan R, 2017

Moreover, we performed a quality check of our detection platform and screened the available CSF and plasma samples of our cohort (Fig. 2f). Including the previously mentioned matched biopsy samples, an additional five CSF (total n = 13) and 13 plasma samples (total n = 19) were available for H3F3A K27M (Supplementary Fig. 5a–b), and four more CSF samples (total n = 5) as well as 7 plasma samples for BRAF V600E detection (total n = 12) (Supplementary Fig. 5c–d). Based on these data, we calculated the predictive power as described by Trevethan in 2017 [37]. In the H3F3A K27M cohort, we reached a sensitivity of 84.61% for CSF and 73.68% for plasma and 100% specificity in both the sources. For the BRAF V600E cohort, we achieved a sensitivity of 83.33% and 100% specificity in the plasma group. For CSF, we found a sensitivity of 80% with 100% specificity (Fig. 2f). Furthermore, we performed receiver operation characteristics (ROC) curves to check the quality of the liquid biomarker of all LB samples and determined an area under the curve (AUC) of 0.92 in CSF and 0.70 in plasma for H3F3A K27M and 0.87 in CSF and 0.91 in plasma for BRAF V600E (Fig. 3a–d).

Fig. 3

Receiver operation characteristics (ROC) curve of H3F3A K27M marker in a CSF (n = 13) and b plasma (n = 19). ROC curve of BRAF V600E marker in c CSF (n = 5) and d) plasma (n = 12). ROC curves were analyzed in IBM SPSS statistics 27 and mapped with GraphPad Prism

LB detection rates differ between tumor and sampling site/time

Next, we correlated the primary tumor location (thalamic, pontine, hemispheric, optical pathway and spinal) to the successful detection of the biomarkers. Except for a higher H3F3A K27M mutation detection in CSF of patients with a thalamic location (MAF range from 1.8 to 60%, *p < 0.05), we did not observe a significant difference among tumor locations as shown in Fig. 4a–b. There was a tendency though for a better detection result in thalamic tumors followed by pontine and spinal gliomas. However, this tendency might have been confounded by the fact that most of the available CSF samples were from cases with tumors located in the thalamic area. Hence, we can not exclude a selection bias. Interestingly, the same trend in thalamic tumors was observed for plasma in the H3F3A K27M group (Fig. 4c). In the BRAF V600E sample collection, the optical pathways glioma (OPG) displayed the highest detection rate (Fig. 4d). However, there was no significant difference between the three brain locations.

Fig. 4

Tumor localization and LB detection. a–d Detection of H3F3A K27M and BRAF V600E classified to tumor location in CSF and plasma. e Pie chart of successful liquid biopsy according to puncture localization. f Detailed bar chart—tumors were grouped according to their localization and linked to successful LB detection. Asterisks indicate significance (students t-test; *P < 0.05, **P < 0.01 analyzed with GraphPad Prism), n.s. no significance, error bars indicate mean ± S.D. Parts of the figure were created with Biorender. OPG, optical pathway glioma. Δ indicated no tumor visible in MRI at LB sampling and ddPCR was negative for respective biomarker

Consequently, we analyzed the potential impact of the sampling time and site and grouped our CSF samples into intra-operatively collected CSF, and post-operatively collected CSF from ventricular or lumbar puncture sites. The most sensitive sampling method regardless of the tumor location and sampling site was the intra-operatively collected CSF (100%, cisternal/subarachnoidal n = 3, ventricular n = 1, biopsy canal n = 1) followed by ventricular CSF (93%, n = 30) and the lumbar puncture site (66%, n = 3) as shown in Fig. 4e. Subsequently, we analyzed the correlation between tumor location and puncture site/time (Fig. 4f). In the pontine group (n = 11), we had three intra-operative CSF samples with 100% (3/3) detection, the ventricular site showed a positive signal in six of seven samples (85%) and in the single case obtained by lumbar puncture (100%). One CSF sample of an OPG case was obtained intra-operatively and showed a positive signal (n = 1). In the thalamic group (n = 21), we could detect the underlying alteration in 100% of ventricular CSF. In this group, no other detection sites were available. With respect to hemispheric tumors, mutation detection by LB was positive in one (n = 1) intra-operative sample but not in one lumbar puncture sample, notably, at time of LB sampling, this case had no visible tumor in the MRI (n = 1). The spinal cohort included one patient with CNS metastasis at several sites. In this case, analysis was successful (100%), whereas the ventricular puncture did not show mutation detection. In summary, we identify a potential impact of the CSF sampling site in relation to tumor localization which is of potential relevance for LB detection in brain tumors.

Radiological features

Next, we correlated LB parameters to radiological findings. We compared the contrast-enhancing and non-contrast-enhancing tumor volume to matched LB measurements in plasma (n = 19) and CSF (n = 9). Within our limited cohort, we did not detect an overall correlation between tumor volume and MAF in either plasma or CSF (Supplementary Fig. 6). Importantly, all patients with leptomeningeal tumor dissemination showed detectable ctDNA in CSF pointing towards a particular potential for detection and tumor surveillance in this patient group (Supplementary Fig. 7a). In contrast, positive LB detection in plasma was not associated with leptomeningeal metastasis, and in both liquids, the presence of tumor necrosis had no impact on LB (Supplementary Fig. 7b–d).

Longitudinal follow-up in plasma and CSF

Finally, we analyzed the longitudinal monitoring opportunities of our biomarkers in CSF and/or plasma of seven patients. Four patients harbored a H3F3A K27M mutation (LB_MUV_01, LB_MUV_08, LB_MUV_09 and LB_MUV_11), two a BRAF V600E mutation (LB_MUV_03 and LB_MUV_27) and one case was characterized by both mutations (LB_MUV_02). Figure 5 depicts a swimmer plot, indicating the longitudinal monitoring time points, and showing the increasing or decreasing levels of detected biomarkers in CSF or plasma and the concurrent MRI results. We could demonstrate the increase of mutational DNA fraction of H3F3A K27M in the CSF and plasma in patients with a tumor progression in the MR images (Fig. 6a–d, Supplementary Fig. 8a–b, Supplementary Fig. 9) and partially decreased levels for BRAF V600E mutation during targeted drug treatment with trametinib and dabrafenib (Fig. 6b, Supplementary Fig. 10). However, in LB_MUV_19, a slight increase of H3F3A K27M was detected in the last obtained plasma, while the tumor mass in the last MRI was shrinking (Supplementary Fig. 9c). It is worth noting that in three patients (LB_MUV_01, LB_MUV_02 and LB_MUV_08), we could compare tumor tissue and liquids obtained on the day of surgery and detected positive results in our detection platform in all available liquids as shown in Fig. 6e–f and Supplementary Fig. 8c. These results support the strong significance of ddPCR-based LB detection and its clinical relevance for diagnosis and longitudinal treatment monitoring.

Fig. 5

Swimmer plot of longitudinal LB monitoring of 7 cases including all MRI and LB time points. The first squares in the timeline represent the initial MRI and LB time point of each patient; more squares in the timeline depict a stable disease; triangles demonstrate an increase or decrease in in tumor growth and LB mutation detection. The BRAF V600E patients represent one LGG WHO grade 1 (LB_MUV_27 diagnosed as OPG) and one HGG WHO grade 4 (LB_MUV_03 diagnosed as gliosarcoma)

Fig. 6

Longitudinal monitoring of 2 cases. Treatment history including MAF of LB and tumor volumes obtained from MRI (cm2) of a LB_MUV_01 and b LB_MUV_02. Matched MRI images (coronal T2-weighted MRI) to LB MAF detection for c LB_MUV_01 and d LB_MUV_02. Red arrow marks the tumor spread. Initial MAF of H3F3A K27M in tissue vs initial pre-surgery LB samples: e CSF pre-surgery and f plasma