Cerebrospinal fluid circulating tumor DNA profiling in central nervous system metastases from solid tumors: A prospective multicentre translational analysis of the BrainStorm (OncoDistinct-006) study

Abstract

Background

Central nervous system (CNS) metastases represent a complication in solid tumors and often harbour unique genomic alterations diverging from those in primary tumor tissue or plasma; however, direct sampling of CNS lesions carries procedural risks, limiting its molecular profiling. The main aim of the present analysis was to assess the feasibility of Cerebrospinal Fluid (CSF) ctDNA profiling and to descriptively compare genomic alterations detected in matched tumor tissue, plasma, and CSF samples, including in patients with leptomeningeal metastases (LM).

Patients and methods

BrainStorm (NCT04109131) is a prospective, international, multicenter study enrolling patients at high risk of developing CNS metastases from different solid tumors. In the present analysis, we evaluated matched samples of tumor tissue, plasma, and CSF ctDNA from patients who developed CNS metastases. Tumor tissue was profiled using the OncoDEEP® 638-gene NGS panel, while matched plasma and CSF ctDNA were sequenced with OncoFOLLOW® tumor-informed panel (plus 50 hotspot genes). Descriptive statistics and Jaccard index-based concordance analyses were performed to compare mutational profiles across compartments.

Results

CSF ctDNA was detectable in 21 out of the 25 evaluable CSF samples (84%) and in 10 out of 11 patients with LM (91%). Overall concordance among genomic alterations in samples was 52% between tissue and plasma, 45% between tissue and CSF, and 75% between plasma and CSF. In the subset of 11 patients with LM, concordance was 39% (tissue-plasma), 42% (tissue-CSF) and 91% (plasma-CSF). Overall, seven alterations were identified exclusively in CSF or CSF and tissue but not in plasma, including variants in TP53, PIK3CA, and ERBB2.

Conclusions

These preliminary findings indicate that CSF ctDNA can provide a sensitive, compartment-specific molecular portrait of CNS disease, complementing plasma-based analyses, and supporting its integration as a minimally invasive tool to capture the evolving genomic landscape of CNS metastases.

Clinical trial registration number

NCT04109131.

Highlights

• •Cerebrospinal fluid (CSF) ctDNA was detectable in 84% of patients with central nervous system (CNS) metastases.
• •Concordance rates were 75% between plasma and CSF overall and 91% in patients with leptomeningeal metastases.
• •Among all alterations, 13.7% were exclusive to CSF or shared with tissue but not with plasma, including TP53, PIK3CA, ERBB2.
• •CSF ctDNA offers sensitive compartment-specific profiling, complementing plasma for precision oncology in CNS metastases.

1. Introduction

Central nervous system (CNS) metastases represent a serious complication in patients with metastatic solid tumors, affecting up to 50% of patients during the course of their advanced disease [1,2], and their incidence is particularly high among patients affected by breast cancer (BC), melanoma, non-small-cell lung cancer (NSCLC) and small-cell lung cancer (SCLC) [[3], [4], [5], [6]]. The incidence of CNS metastases is constantly rising in recent decades, both due to improved diagnostic techniques such as magnetic resonance imaging (MRI) use, and because of prolonged survival in patients with metastatic disease treated with novel systemic therapies [7]. Nevertheless, the development of CNS metastases is still associated with important neurological decline, profound deterioration of prognosis and quality of life, and despite advances in local and systemic treatments, outcomes remain unsatisfactory, also for those patients with leptomeningeal metastases (LM) [[8], [9], [10]].

Studies have demonstrated that CNS metastases frequently harbour molecular alterations distinct from those observed in the primary tumor or in extracranial sites [11]. These differences may uncover novel therapeutic opportunities but also highlight the limitations of systemic therapies, many of which either fail to cross the blood-brain barrier or do not target CNS-specific oncogenic pathways [12]. In parallel, accumulating evidence shows that novel targeted agents and immunotherapies such as trastuzumab deruxtecan [13], tucatinib-based combinations [14], lorlatinib [15], and immune checkpoint inhibitors (particularly ipilimumab/nivolumab combination) [16] can achieve meaningful intracranial responses, further underscoring the clinical value of accurately characterizing the genomic landscape of CNS disease. Comprehensive molecular profiling is therefore critical to guide precision oncology strategies and to inform therapeutic decisions tailored to CNS involvement [17]. However, direct sampling of CNS lesions through resection or biopsy is invasive, rarely feasible in routine practice, and carries procedural risks [2,17].

Liquid biopsy approaches, particularly those analysing circulating tumor DNA (ctDNA), provide a minimally invasive alternative to capturing tumor heterogeneity and clonal evolution [18]. Plasma ctDNA is increasingly used in clinical practice to non-invasively characterize tumor mutational status and adapt treatment to detected targets [[19], [20], [21]]; however, it may not reliably reflect the genomic landscape of intracranial disease [22,23]. In contrast, cerebrospinal fluid (CSF) ctDNA offers a unique window into the CNS compartment and has demonstrated higher sensitivity in detecting CNS-specific alterations [24]. Previous studies comparing CSF with matched peripheral blood samples have shown that CSF ctDNA may better capture genomic alterations associated with intracranial disease, particularly when tumor DNA shedding into the systemic circulation is limited [25,26]. More recent prospective data have further supported the complementary molecular information provided by CSF and plasma profiling in patients with CNS metastases [27]. Nonetheless, systematic studies directly comparing matched tumor tissue, plasma ctDNA, and CSF ctDNA remain limited, and their clinical utility has yet to be fully defined [26,27].

The BrainStorm (OncoDistinct-006) study is an academic, international, multicenter, prospective study designed to create a clinical database and biospecimen collection to address unmet needs in patients with CNS metastases [28]. As part of this program, tumor tissue, plasma, and CSF samples have been prospectively collected from patients with metastatic solid tumors at the time of diagnosis of metastatic disease and at the time of development of CNS metastases.

In this context, we performed a translational analysis with the primary objective of characterizing the rate of detection of CSF ctDNA and the concordance of mutational profiles across the different biological compartments. The main aim was therefore to determine whether CSF ctDNA could serve as a reliable surrogate for tumor and plasma mutational profiling, while also identifying CNS-specific alterations with potential therapeutic and prognostic significance.

2. Methods

2.1. Study design and patient selection

This translational analysis was conducted within the framework of the prospective, international, multicenter BrainStorm (OncoDistinct-006) trial (NCT04109131) [28]. The BrainStorm trial is structured into three sequential parts, each addressing a specific phase of disease evolution: I) Part A (i.e., before CNS metastases diagnosis): this cohort includes patients with newly diagnosed extracranial metastatic disease from tumor types at high risk for CNS involvement, namely triple-negative breast cancer (TNBC), HER2-positive breast cancer, NSCLC, SCLC, and melanoma. The objective of this phase is to prospectively follow patients prior to the development of CNS metastases; II) Part B (i.e., at the time of first CNS metastases diagnosis): this cohort includes patients with a first diagnosis of CNS metastases (BM and/or LM), including those already enrolled in Part A. Moreover, two additional cohorts can be included directly in this phase, with patients with other solid tumors and those with LM, irrespective of the primary tumor origin. This part of the study was specifically designed for intensive biospecimen collection, including CNS metastatic tissue (when available from clinically indicated surgery or biopsy), plasma, and CSF samples. Finally, III) Part C (i.e., after first CNS metastases diagnosis): all patients entering Part B (directly or from part A) enter longitudinal follow-up, with clinical and radiological evaluations every three months. This phase allows for outcome assessment, serial plasma and CSF sampling, and monitoring of long-term endpoints such as survival outcomes and quality of life reporting (Supplementary Fig. 1).

The BrainStorm trial was approved by the Ethics Committee of the Institut Jules Bordet (CE approval number CE3040; date: 23 September 2019). All patients provided written informed consent before entering the study. The study was conducted in accordance with the Declaration of Helsinki and all applicable international, national, and local laws and regulations. For the present translational analysis, only patients enrolled in Part B with matched tumor tissue at screening and with plasma and CSF samples collected at the time of CNS metastasis diagnosis were included. Clinical and pathological data, including demographics, primary tumor subtype, and presence of LM, were prospectively recorded in a centralized electronic database via standardized electronic case report form (eCRF).

2.2. Sample processing and genomic profiling

Sample collection was conducted according to standardized operating procedures to ensure uniform handling and minimize pre-analytical variability across participating centres. Tumor tissue was obtained either from archival diagnostic specimens or from newly resected or biopsied lesions. Plasma was collected at baseline and according to the study schedule, isolated immediately after blood collection, and stored at −80 °C within 1 h after collection. CSF samples were acquired by lumbar puncture, immediately centrifuged to remove cells, and stored under the same conditions.

Genomic profiling of tumor tissue was performed by OncoDNA® using the OncoDEEP® assay, a comprehensive next-generation sequencing (NGS) panel covering 638 cancer-associated genes. This platform enables the detection of single-nucleotide variants (SNVs), small insertions and deletions (indels), copy-number variations (CNVs), and selected gene fusions.

Matched plasma and CSF ctDNA were analysed using the OncoFOLLOW® panel, which integrates up to 15 patient-specific variants identified from tumor sequencing with a 50-gene hotspot panel targeting recurrent driver mutations of diagnostic or prognostic relevance.

Sequencing reads were aligned to the human reference genome (hg19), and variants were called using validated bioinformatics pipelines with a minimum coverage depth of 100×. For SNVs and indels, a variant allele frequency (VAF) threshold of 1% was applied, while copy-number variations were identified using dedicated copy-number calling algorithms. Common polymorphisms and sequencing artifacts were filtered out to ensure high-confidence variant calls. All detected alterations were classified according to their molecular nature as SNVs, indels, CNVs, splice-site variants, or gene fusions, based on the variant annotation available within each compartment (tissue, plasma, and CSF). Pathogenic and likely pathogenic alterations were subsequently categorized as shared or compartment-specific, and mutational overlap across tumor tissue, plasma, and CSF was quantified for each patient. Detected alterations were also considered according to their potential biological and clinical relevance within the underlying tumor type, including established oncogenic drivers that may contribute to molecular patient stratification. However, no formal therapeutic actionability framework or treatment assignment based on the detected alterations was applied in this exploratory analysis.

For each patient, three variant sets were generated corresponding to tumor tissue, plasma, and CSF. Mutations were categorized as either shared across compartments or unique to one, with particular attention to pathogenic variants detected exclusively in the CSF ctDNA. For alteration-level analyses, each distinct genomic alteration identified in an individual patient was counted as a single alteration event.

The main objectives of the present analysis were to assess the detection rate of CSF ctDNA and to compare its mutational profile with matched tumor tissue and plasma profiles. Further aims were to assess the concordance of mutational profiles across compartments, both in the overall study population and in patients with LM, to determine the ability of CSF ctDNA to capture CNS-specific alterations.

2.3. Statistical analysis

Descriptive statistics were used to summarize baseline clinical and molecular characteristics. Continuous variables are reported as medians with interquartile ranges (IQRs), while categorical variables are presented as counts and percentages with explicit denominators when data were missing. Mutational concordance between compartments was quantified using the Jaccard index, defined as the ratio between the number of shared variants and the total number of unique variants detected in either of the two compared samples. A genomic alteration was considered shared when the same alteration was detected in both samples of a given pairwise comparison, whereas alterations detected in only one compartment were considered compartment specific. Thus, the Jaccard index reflects the proportion of shared alterations relative to the overall set of alterations detected across the two compared compartments. Pairwise Jaccard indices were calculated for tumor tissue versus plasma, tumor tissue versus CSF, and plasma versus CSF, and were summarized descriptively in the overall cohort and in the exploratory subgroup of patients with leptomeningeal metastases. Variant allele frequencies in plasma and CSF were also summarized using medians and IQRs. Given the exploratory nature of the study and the limited sample size, no formal hypothesis testing was performed and no adjustment for multiple comparisons was applied.

3. Results

3.1. Baseline patients and disease characteristics

From July 3rd, 2020 to the database cut-off of April 1st, 2025, a total of 105 patients were included in 17 centres in Belgium, France and Luxembourg in Part B. Breast cancer represented the largest subgroup (n = 35) of included patients, followed by lung cancer (n = 21), other solid tumors (n = 20), and melanoma (n = 1), while 28 patients were also included in Part B based on a diagnosis of LM as first CNS event. Among the 105 patients enrolled, 7 were excluded from this analysis due to the lack of available tumor tissue, 9 due to missing plasma samples, and 56 due to absence of CSF samples. After these exclusions, 33 patients had fully matched tumor, plasma, and CSF samples and were assessed for the present analysis. Of these, 8 patients were further excluded, of whom 6 for technical reasons and 2 due to pending analysis results at data cut-off, resulting in a final evaluable cohort of 25 patients: 22 (88.0%) patients with breast cancer (n = 11 HER2-positive, n = 7 luminal-like, and n = 4 triple-negative BC), 2 (8.0%) patients with NSCLC, and 1 (4.0%) patient with melanoma. The overall screening, inclusion, and exclusion process is summarized in the CONSORT diagram (Fig. 1).

Fig. 1.

Study flow diagram

Abbreviations: CSF, cerebrospinal fluid.

At inclusion, the median age was 53.0 years (IQR 45.0-62.0). Fourteen patients (56.0%) had brain metastases (BM), five (20.0%) had both BM and LM, and six (24.0%) had LM only. Nearly half of the patients included (47.1%) had a metastatic disease at initial diagnosis, and the median number of metastatic sites prior to study inclusion was 2.0 (IQR 1.0-3.0), while the median time between initial tumor diagnosis and CNS metastasis development was 3.2 years (IQR 1.6-7.6). The median time from tumor tissue sampling to plasma collection was 35.0 months (IQR 10.5–53.8), and to CSF collection was 40.7 months (IQR 19.0–65.5) (Table 1).

Table 1.

Baseline clinical characteristics of patients included in the present analysis.

	Total (N = 25)
Age At Enrolment
Median (IQR)	53.0 (45.0-62.0)
Primary Tumour Site, N (%)
Breast	22 (88.0%)
Lung	2 (8.0%)
Skin	1 (4.0%)
CNS Metastases Characteristics, N (%)
Brain	14 (56.0%)
Brain and Leptomeningeal	5 (20.0%)
Leptomeningeal	6 (24.0%)
Tumor Stage at Diagnosis, N (%)
I	1 (6.7%)
II	9 (60.0%)
III	1 (6.7%)
IV	4 (26.7%)
Missing	10
Nodal Stage at Diagnosis, N (%)
N0	2 (11.8%)
N > 0	15 (88.2%)
Missing	8
Metastases Stage at Diagnosis, N (%)
M0	9 (52.9%)
M1	8 (47.1%)
Missing	8
Number of Metastatic Sites Identified Before Inclusion
Median (IQR)	2.0 (1.0-3.0)
Number of Patients with only CNS Metastases at Inclusion in part B, N (%)	3 (12.0%)
Mental Status/Cognitive Function at Inclusion, N (%)
Abnormal	3 (13.6%)
Normal	19 (86.4%)
Missing	3
Motor Deficit at Inclusion, N (%)
Abnormal	3 (13.6%)
Normal	19 (86.4%)
Missing	3
Sensory Deficit at Inclusion, N (%)
Abnormal	3 (13.6%)
Normal	19 (86.4%)
Missing	3
Time Between Initial Diagnosis and CNS Metastasis Development (Years)
Median (IQR)	3.2 (1.6-7.6)
Time Between Tumor Tissue Sampling and Liquid Biopsy Collection (Months)
Tissue → Plasma, Median (IQR)	35.0 (10.5-53.8)
Tissue → CSF, Median (IQR)	40.7 (19.0-65.5)
Plasma → CSF, Median (IQR)	0 (0.0-0.3)

Abbreviations: BM, Brain Metastases; CNS, Central Nervous System; CSF, Cerebrospinal Fluid; IQR, Interquartile Range; LM, Leptomeningeal Metastases.

3.2. Mutational findings and concordance across compartments

CtDNA was detectable in CSF cfDNA as tumor-derived variants in 21/25 (84%) evaluable CSF samples in the overall cohort, and in 10/11 (91%) patients with LM. At the single alteration level, 91% of variants were detected in tumor tissue, while 47% and 49% were detected in plasma and CSF, respectively. As illustrated in the Sankey plot, a substantial fraction of mutations changed their detection status across compartments. Approximately half of the mutations detected in tissue were not detected in either plasma or CSF, while a subset of variants remained consistently detected across all compartments (Fig. 2, Supplementary Table 1).

Fig. 2.

Changes in mutation detection across compartments

The Sankey diagram illustrates how individual mutations change their detection status across tissue, plasma, and CSF samples.

Each vertical bar represents the proportion of mutations within a given compartment, categorized as detected (green), not detected (yellow), or technically failed (red).

The connecting ribbons indicate how mutations transition between these categories across compartments, showing patterns of concordance and loss of detection between tissue, plasma, and CSF.

Abbreviations: CSF, cerebrospinal fluid.

Among the 48 alterations detected in tumor tissue, TP53 and PIK3CA were the most frequently involved genes, each accounting for 9 alterations (18.8%), followed by ERBB2 in 4 (8.3%) and GATA3 in 3 (6.2%) tissue samples (Supplementary Table 1–2).

Among the variants detected across compartments, SNVs represented the most frequent type of alteration (65% of all alterations), followed by CNVs (20%), indels (10%), and splice-site alterations (5%). No recurrent gene fusions were identified. SNVs mainly involved canonical driver genes such as TP53 and PIK3CA, while CNVs were primarily observed as gene amplifications in ERBB2, MYC, and CCND1. Frameshift or deletion-type indels were identified in genes such as BRCA2 and NF1, consistent with loss-of-function mechanisms (Supplementary Fig. 2). This distribution was similar across tumor tissue, plasma, and CSF compartments, supporting the analytical robustness of the assay (Supplementary Table 2).

Looking at the differences in VAF levels between plasma and CSF, the median VAF detected in plasma was 22% (IQR 1%-51%) and 28% (IQR 5%-55%) in CSF in the overall population, while among patients with LM, the median VAF in plasma and CSF were 7% (IQR 3%-28%) and 18% (IQR 5%-50%), respectively (Fig. 3A–B).

Fig. 3.

Variant Allele Frequency in plasma and CSF across individual mutations in the A) overall population and B) in patients with leptomeningeal metastases

Boxplots illustrate the distribution of variant allele frequency (VAF) values in plasma and CSF, with each data point representing an individual mutation.

Lines connect the same mutation across compartments, depicting mutation-level changes in VAF between plasma and CSF. Green lines indicate mutations showing an increase in VAF from plasma to CSF, whereas red lines indicate a decrease. Boxes display the median and interquartile range (IQR) of VAF values within each compartment.

Abbreviations: CSF, cerebrospinal fluid; VAF, variant allele frequency.

Mutational concordance was assessed both at the patient-level and at the mutation-level to evaluate the overlap and compartment-specific differences among tumor tissue, plasma, and CSF profiles.

At the patient-level, concordance between paired compartments was 52% for tissue versus plasma, 45% for tissue versus CSF, and 75% for plasma versus CSF.

At the single alteration level, approximately one-third of all detected alterations were shared across all three compartments, whereas a substantial fraction remained tissue-specific. In addition, five alterations identified in PIK3CA, BRCA2, ERBB2, GATA3, and CHEK2, were concordant between tissue and CSF but undetectable in plasma, indicating that certain intracranial clones may preserve molecular features of the primary tumor while remaining compartmentalized within the CNS and underrepresented in the blood. Conversely, three TP53 alteration events were identified in three patients: two were detected exclusively in CSF, whereas one was detected in both plasma and CSF; none of these alterations was identified in the corresponding tumor tissue (Fig. 4).

Fig. 4.

OncoPlot visualization of biomarker concordance across tissue, plasma, and cerebrospinal fluid compartments

Abbreviations: CSF, cerebrospinal fluid, NSCLC, non-small cell lung cancer.

In the subgroup of 11 patients with LM, the overall pattern of mutational concordance mirrored that observed in the full cohort: at the patient level, concordance was 39% for tissue versus plasma, 42% for tissue versus CSF, and 91% for plasma versus CSF. At the mutation level, most shared alterations were detected in both plasma and CSF, while a limited number of variants remained exclusive to the CNS compartment. Specifically, the two TP53 mutations were detected in CSF (one case) and in plasma (one case) but were not identified in tumor tissue (Supplementary Fig. 3).

4. Discussion

To our knowledge, this translational analysis from the BrainStorm study is among the first prospective, multicenter efforts to compare, within the same population, the genomic profiles across matched tumor tissue, plasma, and CSF from metastatic solid tumors with confirmed CNS involvement. Despite the relatively small cohort, our results confirm and reinforce several key findings.

Firstly, we focused on the analytical yield of CSF liquid biopsy: CSF ctDNA was detectable in 84% of evaluable samples overall and in 91% of patients with LM, supporting the high feasibility and sensitivity of CSF-based testing in this setting. Our findings appear even more robust than prior reports, which showed detection rates of actionable alterations ranging between 60.5% and 67.8% in CSF [25,27,29], with marked differences between patients with BM and LM [27]. Similarly, in our analysis, the detection rate of CSF ctDNA was higher among patients with LM but also remained high across the overall study population (84%). At the alteration level, technical failure rates were relatively low in both liquid-biopsy compartments, accounting for 4% of plasma assessments and 9% of CSF assessments, whereas 49% and 42% of alterations, respectively, were not detected despite technically evaluable testing. This distinction is relevant, as absence of detection may reflect biological factors such as limited ctDNA shedding or low tumor-derived DNA abundance rather than assay failure. Pre-analytical standardization is also critical for reliable CSF ctDNA profiling. In the BrainStorm study, CSF samples were processed immediately after collection, centrifuged to remove cellular components, and stored at −80 °C according to standardized procedures across participating centres. Such harmonized handling is particularly important to minimize variability related to sample processing and preserve the quality of cfDNA for downstream molecular analyses [30].

Recurrent alterations in our series reflected the expected biology of the underlying tumor histologies, with TP53, PIK3CA, ERBB2, and BRCA2 being the most frequently mutated genes across compartments. These findings are in line with previous reports in patients with CNS metastases from breast cancer, where TP53 and PIK3CA mutations are recurrent and often coexist with DNA repair pathway alterations [31]. Large-scale genomic analyses, such as the study by Giannoudis and colleagues further confirmed an enrichment of TP53, MYC, and homologous recombination-related genes (i.e., BRCA1/2, PALB2) in BM compared with primary and extracranial metastatic lesions, highlighting CNS-specific evolutionary trajectories [32].

Even within our relatively small cohort, CSF ctDNA showed detection of canonical drivers while also revealing compartment-specific variants that were not captured in matched tumor tissue. TP53 variants emerged exclusively in CSF ctDNA from three patients whose primary tumors were all breast cancers (one TNBC, one HER2-positive, and one luminal-like breast cancer) and among them, two had LM. This observation is biologically plausible, as TP53 loss is a hallmark of metastatic breast cancer and is frequently associated with aggressive clinical behaviour and CNS dissemination [33,34]. The detection of TP53 mutations in CSF but not in the corresponding tissue may therefore reflect clonal evolution under therapeutic pressure, with selection of resistant subclones capable of CNS colonization, and similar findings have been described in previous works using paired CSF and tissue sequencing [35,36]. De Mattos-Arruda and colleagues reported that mutations exclusive to CSF ctDNA often mapped to resistance pathways (i.e., PIK3CA, TP53, PTEN), supporting the concept of CNS-specific sub-clonal expansion [25,29].

A key objective of our analysis was to assess the concordance between mutational profiles derived from tumor tissue and those detected in liquid compartments. In our cohort, the mean overlap among patients was 52% between tumor and plasma, 45% between tumor and CSF, and 75% between plasma and CSF. Although the tissue–plasma concordance may appear lower than expected for recurrent driver alterations, this finding should be interpreted considering the use of archival tumor tissue, the temporal mismatch between tissue and liquid sampling, and the different sequencing approaches used across compartments. These factors, together with treatment-driven clonal evolution and variable ctDNA shedding from intracranial and extracranial disease sites, may have contributed to the observed discordance [37,38].

An additional factor that may have influenced the observed concordance across compartments is the heterogeneity of extra-CNS tumor burden in our cohort. Nearly half of the patients included in the translational analysis had metastatic disease at initial diagnosis, and the median number of extracranial metastatic sites before CNS involvement was two. Patients with limited or absent systemic tumor burden, such as those presenting with LM-only or brain-dominant disease, typically exhibit low plasma ctDNA levels, which may reduce the sensitivity of plasma-based detection and partly explain why several alterations identified in CSF were not retrieved in plasma. This phenomenon has been reported in prior studies, where the absence of active extracranial disease was associated with reduced plasma ctDNA shedding and lower tissue-plasma concordance. In contrast, CSF ctDNA directly reflects intracranial tumor activity, supporting its added diagnostic value particularly in patients with minimal systemic involvement [26]. Considering extra-CNS disease burden is therefore essential for interpreting liquid biopsy concordance and for guiding decisions on when CSF profiling may be most informative.

This observation is also consistent with similar studies which demonstrated that CSF ctDNA more accurately represents the intracranial mutational landscape than plasma ctDNA, frequently revealing potentially actionable alterations undetectable in matched plasma samples.^28-30 Brastianos and colleagues reported that over half of CNS metastases harboured targetable alterations absent from primary tumors, underscoring the genomic divergence between CNS and systemic disease compartments [11].

Our findings therefore reinforce the notion that plasma and CSF provide complementary yet distinct windows into the disease biology, with CSF reflecting CNS-specific molecular evolution while plasma captures systemic tumor dynamics [1]. The relatively high plasma and CSF mutational profile overlapping observed in our series may suggest that contemporaneous sampling of both liquid compartments can capture the full spectrum of clinically relevant mutations, while the lower tissue-liquid concordance underscores the limited representativeness of archival tumor material, which may predate major evolutionary events leading to CNS dissemination [27,39].

When focusing on the small subgroup of 11 patients with LM diagnosis, we observed an even higher concordance between plasma and CSF (91%), whereas overlap with tumor tissue remained only moderate. This pattern reinforces the unique molecular representativeness of CSF in LM, where extensive meningeal infiltration facilitates direct ctDNA release into the cerebrospinal compartment. Similar results have been reported in lung cancer cohorts, where nearly all patients with LM exhibited detectable ctDNA in CSF but only partial detection in plasma [26,27,29].

The literature supports three non-mutually exclusive explanations for why plasma-CSF concordance may exceed that between either plasma or tissue. First, tissue samples often exhibit temporal and spatial heterogeneity, as archived tumor specimens may originate from different sites and time points than the active CNS disease [11,40]. Second, tumor evolution under therapeutic pressure can lead to reduced concordance between tissue and liquid biopsies [25,29,41]. Third, the compartmentalization imposed by the blood-brain barrier may lead to partially independent evolutionary trajectories in the CNS. In patients with active CNS disease but limited extracranial tumor burden, plasma ctDNA often reflects only the most dominant systemic clones, whereas CSF captures the intracranial ones. When both compartments detect the same alteration, it likely corresponds to a clone that is biologically dominant across sites at the time of sampling, which may explain the relatively high plasma-CSF concordance observed in our cohort [26,42].

It should also be noted that in our cohort the median time between the initial cancer diagnosis and the development of CNS metastases was 3.2 years, which may have contributed to substantial molecular divergence between the primary tumor and the samples collected at the time of CNS progression. This temporal gap is particularly relevant given that most tissue specimens were archival samples obtained at diagnosis or from early extracranial disease and therefore may not accurately reflect the genomic landscape at CNS progression. In contrast, plasma and CSF liquid biopsies were collected contemporaneously with intracranial involvement and are therefore more likely to capture the currently dominant tumor clones.

The higher sensitivity of CSF in this setting likely reflects the direct communication between tumor cells and the CSF space, which contrasts with the limited permeability of the blood-brain barrier that restricts ctDNA shedding into the systemic circulation. Consistent with this, in our analysis the VAF observed in CSF was generally higher than those measured in plasma, further supporting the notion that CSF ctDNA provides a more concentrated and biologically relevant snapshot of LM disease [43]. Median VAF values were numerically higher in CSF than in plasma in both the overall and LM cohorts; given the exploratory design and limited paired observations, these descriptive differences should not be interpreted as evidence of enrichment.; however, the trend toward higher levels in CSF suggests an enriched tumor-derived DNA fraction, in keeping with prior studies showing that CSF ctDNA quantity correlates with disease burden and neurological involvement [44].

The detection of mutations exclusive to the CSF compartment or present in both CSF and primary tumor tissue but absent in plasma, supports the concept of independent clonal evolution within the CNS. In our cohort, several mutations were identified solely in CSF or in both CSF and tumor tissue but not in plasma. Notably, TP53 alterations were the most frequent CSF-specific mutations in breast cancer cases, suggesting the emergence of CNS-restricted subclones shaped by therapeutic pressure and the unique brain microenvironment. Large-scale molecular and epigenetic analyses further confirm distinct evolutionary trajectories in the CNS, highlighting the unique molecular landscape of intracranial disease [45,46]. These observations indicate that the CNS acts as a partially isolated evolutionary niche, where selective pressures such as the blood-brain barrier and limited drug penetration may favour the persistence of resistant clones. CSF ctDNA therefore offers a unique and clinically relevant window into these CNS-specific processes, complementing tissue and plasma profiling for a more complete view of tumor evolution.

Despite the reliable results, the present analysis has several limitations that should be acknowledged. The small sample size, predominance of breast cancer, and inclusion of both parenchymal brain and leptomeningeal metastases limit statistical power and generalizability; therefore, all findings should be considered exploratory. Potential selection bias should also be acknowledged, as only 25 of 105 patients enrolled in BrainStorm Part B were evaluable, mainly because CSF was unavailable. In addition, archival tumor tissue was collected substantially earlier than plasma and CSF, so temporal tumor evolution may partly explain the observed discordance. Different assays were used across compartments, and the tumor-informed design of the liquid-biopsy panel partly incorporates tissue-derived variants, limiting both independent concordance assessment and the detection of novel CSF-specific alterations. Finally, the study did not assess procedure-related outcomes, treatment changes, or clinical benefit. Since CSF collection requires lumbar puncture, these results support analytical feasibility but do not establish clinical utility or justify routine CSF sampling for molecular profiling. These aspects are currently being addressed as patient accrual and follow-up continue and will be evaluated with a larger sample size and longer follow-up.

5. Conclusions

The ongoing BrainStorm program represents one of the first international, prospective, multicenter efforts to evaluate the concordance of CSF ctDNA in patients with CNS metastases with the mutational profiles of the other biological compartments. The consistent findings across tumor, plasma, and CSF provide robust evidence for the analytical feasibility and biological relevance of CSF ctDNA analysis. These results have significant clinical implications, as the high detection rate of CSF ctDNA supports its integration into future diagnostic and research workflows as a minimally invasive tool for molecular profiling of CNS disease. Despite the similar concordance between plasma and CSF profiles, our analysis revealed clinically relevant mutations uniquely detectable in the CSF or CSF and tissue but not in plasma, indicating that CNS-specific drivers may be missed when relying solely on systemic biopsies or plasma ctDNA. This observation has important therapeutic implications, as targeted treatments guided by extracranial profiling may not adequately address molecular alterations confined to the CNS compartment. These results contribute to a growing body of evidence positioning CSF liquid biopsy as a valuable complement to standard diagnostic tools (such as MRI and cytology), which often lack sensitivity particularly in leptomeningeal involvement. Beyond its diagnostic role, CSF ctDNA profiling offers a promising opportunity to deepen our understanding of brain metastasis biology, monitor intracranial molecular evolution, and ultimately inform the design of CNS-tailored therapeutic strategies.

Ethics declaration

Written informed consent to take part in the study and to publish the article has been obtained from all participants or their legal representatives. The privacy rights of participants have been observed.

This study included organ or tissue donors. This study includes human biological material and consent was obtained by donors, or their next of kin or legal representatives, for use in this study and for publication of the article. The samples used in this research were not sourced from executed prisoners or prisoners of conscience.

This study was performed in compliance with relevant laws, regulatory frameworks and guidelines where the research took place. This study was approved by the Institut Jules Bordet - HUB. (Approval No. CE3040)

The results of this clinical trial and any associated work have been posted in a registry. This clinical trial was registered with number NCT04109131 (Clinicaltrials.Gov).

Data availability statement

The datasets generated and analysed during the current translational study are part of the ongoing prospective, international, multicenter BrainStorm (OncoDistinct-006) trial (NCT04109131). Due to the ongoing nature of the trial, patient privacy concerns, and ethical restrictions, the full raw data are not publicly currently available. De-identified participant data supporting the findings of this analysis are available from the corresponding author upon reasonable request, subject to approval by the trial sponsor, the Ethics Committee, and in compliance with applicable international, national, and local regulations. The study protocol and additional trial information are publicly accessible on ClinicalTrials.gov (NCT04109131).

Funding source

BrainStorm study is sponsored by Institut Jules Bordet with the financial support of: Fondation Luxembourg, Association Jules Bordet, Fondation Contre le Cancer, Fondation Roi Baudoin, Fonds Gaston Ithier.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Nuria Kotecki reports financial support was provided by Fondation Luxembourg. Nuria Kotecki reports financial support was provided by Association Jules Bordet. Nuria Kotecki reports financial support was provided by Foundation Against Cancer. Nuria Kotecki reports financial support was provided by Fondation Roi Baudoin. Nuria Kotecki reports financial support was provided by Fonds Gaston Ithier. Luca Arecco reports a relationship with AstraZeneca that includes: travel reimbursement. Luca Arecco reports a relationship with Gilead Sciences Inc that includes: funding grants. Luca Arecco reports a relationship with Pfizer that includes: funding grants. Luca Arecco reports travel support from AstraZeneca and research funding to his institution from Gilead and Pfizer. Andrea Gombos reports consulting/advisory roles for Novartis, MSD, AstraZeneca, Gilead Sciences, and Daiichi Sankyo (all paid to her institution), and research funding to her institution from MSD, Roche/Genentech, AstraZeneca, Sanofi/Aventis, Daiichi Sankyo/AstraZeneca, and Novartis. Florian Clatot reports advisory roles for Merck, Novartis, AstraZeneca, Gilead, MSD, and Danone; speaker honoraria from Roche, Lilly, Novartis, AstraZeneca, Gilead, Merck, and MSD; and travel grants from Merck, MSD, Nutricia, and Novartis, all outside the submitted work. Francois P. Duhoux reports a postdoctoral research grant from Fondation Belge contre le Cancer; consulting fees paid to his institution from Amgen, AstraZeneca, Daiichi Sankyo, Eli Lilly, Gilead Sciences, Merck Sharp & Dohme, Novartis, Pfizer, Pierre Fabre, and Roche; and travel support from Amgen, AstraZeneca, Daiichi Sankyo, Gilead Sciences, Pfizer, Roche, and Teva. Paul M. Clement reports research funding from AstraZeneca and compensation to his institution for advisory board participation from Bristol Myers Squibb, MSD, Servier, Leo Pharma, Novocure, and Merck. Evandro de Azambuja reports honoraria and/or advisory board participation from Roche/GNE, Novartis, Seagen, Zodiac, Libbs, Pierre Fabre, Lilly, AstraZeneca, MSD, and Gilead Sciences; travel grants from AstraZeneca and Gilead; institutional research grants from Roche/GNE, AstraZeneca, GSK/Novartis, Gilead Sciences, and Seagen/Pfizer; and non-financial roles as ESMO Director of Membership (2023–2025) and BSMO President (2023–2026). Nuria Kotecki reports honoraria, travel support, accommodation, and expenses from AstraZeneca. Hannelore Denys reports consulting fees (advisory role) paid to her institution by PharmaMar, AstraZeneca, Eli Lilly, Novartis, Amgen, GSK, MSD, Seagen, Gilead, AbbVie, Menarini, Incyclix, Daiichy Sankyo, Tubulis and BioteqPartner and other fees paid to her institution (travel, accommodations, expenses) by, AstraZeneca, MSD, GSK, Gilead; all of these outside the submitted work. Dr. Awada's disclosures: Advisory role: Amgen, AstraZeneca, Bayer, Daiichi, EISAI, Genomic Health, Hengrui, Innate, Ipsen, Leo Pharma, Lilly, Merck, MSD, Novartis, Pfizer, Seattle Genetics, Menarini Speaker fees : Amgen, AstraZeneca, Bayer, Daiichi, EISAI, Genomic Health, Ipsen, Leo Pharma, Lilly, Merck, MSD, Novartis, Pfizer, Seattle Genetics, Hikma, Sothema, Biocon Research grants to my Institute : BMS, Roche.

All remaining authors declare no competing interests.

Appendix A. Supplementary data

The following is the Supplementary data to this article: