Assessing CSF ctDNA to Improve Diagnostic Accuracy and Therapeutic Monitoring in Breast Cancer Leptomeningeal Metastasis

Metastasis to the leptomeninges (LM) occurs in 5% to 10% of patients with metastatic breast cancer (1) and, despite advances in breast cancer treatments, the median survival after the diagnosis of BCLM remains only 3 to 4 months (2–4). Clinical case series have shown the strongest negative prognostic factor in the LM is poor performance status (4) therefore diagnosing earlier with prompt initiation of LM-directed therapy, before significant decline in neurological function, may lead to improved survival. Currently, rapid and robust diagnosis remains a significant challenge in the LM. Diagnosis has traditionally centered on the cytologic assessment of cerebrospinal fluid (CSF) obtained by lumbar puncture, which has 100% specificity upon the finding of malignant cells in CSF (5). However, it is hampered by a high false-negative rate, with sensitivity of just 45% to 67% at the first lumbar puncture (6–8). Consequently, around one-third of patients undergo repeated lumbar puncture (5), with sensitivity rising to 80% to 89% (6, 7, 9), albeit at the cost of not only additional morbidity but delays in diagnosis and initiation of therapy.

Although improvements in LM diagnosis have been achieved with contrast-enhanced high-field magnetic resonance imaging (MRI), it remains an imperfect diagnostic tool. Sensitivity of MRI for LM diagnosis is 76% to 86% (5, 8, 10, 11) with specificity of 77% (5). Furthermore, neuroimaging findings in LM (12) can be subjective with difficulty in standardizing findings between radiologists and heterogeneity in the imaging protocols used (13).

The recent EANO-ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up of patients with LM from solid tumors (14) continue to recommend CSF cytology alongside MRI imaging as part of the diagnostic workup and recommend that repeat lumbar puncture should be performed upon initial negative or equivocal CSF studies. Further, the LM treatment algorithm recommends a therapeutic approach, which includes intrathecal chemotherapy when CSF cytology is positive. Intrathecal chemotherapy remains the standard-of-care treatment in LM despite a lack of randomized trials showing benefit (15). More recently, BCLM patients with HER2+ breast cancer have been successfully treated with intrathecal trastuzumab (16, 17). These invasive therapies require close monitoring and judicious response assessment to ensure continued benefit; however, there are no LM quantitative biomarkers, and MRI findings are not easily measurable. Intrathecal treatment protocols recommend reduction in dosing intensity when CSF cytology “clears”; however, given the high false-negative cytology rate, this is not always a reliable marker of response.

CSF, similar to plasma and other body fluids, contains circulating cell-free DNA (cfDNA) and in the presence of malignancy a fraction is tumor-derived (ctDNA). In a study of 12 patients with primary brain tumors or brain metastases, ctDNA was demonstrated to be more abundant in CSF than plasma when tumor involvement was predominant in the central nervous system (CNS; ref. 18). Studies determining presence of CSF ctDNA have mostly employed droplet digital PCR (ddPCR), detecting highly prevalent hotspot mutation, such as BRAF^V600E in melanoma (19) or EGFR^T790M in lung adenocarcinoma (20–22). However, in patients with breast cancer, there are few hotspot mutations, and broad sequencing of the matched primary tumor is required for the identification of patient-specific somatic mutations (23). This process can be both time-consuming and costly, making it a suboptimal approach when a rapid diagnosis of LM is required.

In this proof-of-concept study, we demonstrate the potential utility of using ultra-low-pass whole genome sequencing (ulpWGS), a relatively low-cost and rapid tool, which does not require prior knowledge of the tumor mutational landscape to detect ctDNA in CSF of patients in BCLM for diagnosis and monitoring of response to intrathecal therapy.

CSF was prospectively collected by lumbar or ventricular route (Ommaya reservoir or ventriculoperitoneal shunt), from patients with breast cancer at initial evaluation for known or suspected BCLM (n = 30). In 27 patients, baseline blood samples were collected concurrently. Written informed consent was obtained from all individuals in accordance with the Declaration of Helsinki under the following research ethics committee approved studies (REC ID 13/LO/1248, South East London Cancer Research Network, UK, and REC ID 14/LO/0292, Royal Marsden NHS Foundation Trust, UK, and local ethics committee GZA Hospital Sint-Augustinus, Antwerp).

Baseline neuroaxis MRI was reviewed by an independent consultant neuroradiologist to evaluate cases where initial reporting was either borderline or negative for BCLM. CSF cytology reports were categorized as equivocal if findings were suspicious but not confirmatory for the presence of malignant cells and otherwise classified as positive or negative.

Serial CSF and blood samples were collected immediately prior to administration of intrathecal therapy in 12 patients. In patients who did not undergo intrathecal therapy, no further samples were obtained.

Following defined standard operating procedures, samples were handled as follows: CSF samples were collected in standard universal containers and transported on wet ice to the laboratory, for centrifugation within 1 hour of collection at 300 × g for 10 minutes. CSF cell-free supernatant was stored at −80°C. Peripheral blood was collected in EDTA tubes and centrifuged within 4 hours, at 1,600 × g for 20 minutes to obtain plasma and buffy coat, both were stored at −80°C until extraction. CSF total protein levels were recorded from hospital records. For cfDNA extraction, CSF supernatant and plasma were thawed at 4°C, extraction performed using the Qiagen Circulating Nucleic Acid Kit following the manufacturer's instructions with the following optimization: elution buffer volume of 50 μL remained on the column for 20 minutes at room temperature prior to final centrifugation. The extracted cfDNA was quantified by Qubit Fluorometer 2.0 using the Qubit dsDNA High Sensitivity Assay Kit, and thereafter stored in DNA LoBind tubes at −20°C. Germline DNA was extracted from buffy coat using the Qiagen DNeasy Blood and Tissue Kit. Genomic DNA (gDNA) was obtained from archival primary tumor blocks using the QIAamp DNA AllPrep FFPE Kit, following macrodissection of Nuclear Fast Red–stained tissue sections to obtain maximum tumor cellularity. DNA was quantified using a Qubit Fluorometer 2.0 using the Qubit dsDNA Broad Range Assay Kit.

Using the NEBNext Ultra II DNA Library Prep Kit for Illumina with six to 15 cycles of PCR enrichment, 0.5 to 20 ng CSF/plasma cfDNA or buffy coat gDNA underwent ulpWGS library preparation. Samples were indexed using NEBNext Multiplex Oligos for Illumina. Resultant libraries were pooled and sequenced on a NovaSeq 6000, acquiring 2- × 100-bp reads. No samples failed library preparation or sequencing, and depth of coverage across all samples was 0.26× (median). ulpWGS FASTQ files were aligned against the human assembly build GRCh38 using (BWA) v0.7.17, duplicate reads were removed with Picard v2.23.8 and mean coverage of aligned reads were derived from the deduplicated BAM file using the CollectWgsMetrics (Picard v2.23.8) tool. Using ichorCNA (24), estimates of tumor purity and copy-number aberrations were generated. HMMcopy readCounter v1.1.0 was used to divide the genome into nonoverlapping bins of either 50 kb or 1 Mb. Centromeres were filtered on the basis of chromosome gap coordinates obtained from UCSC for GRCh38. Aligned reads from the overlap within each window were normalized to correct for GC-content and mappability biases. The log₂ ratio copy-number profiles were generated for each window relative to a panel of matched germline DNA that were sequenced concurrently. The ichorCNA parameters were set as follows: n (normal fraction) values of 0.6, 0.7, 0.8, and 0.9; ploidy initial values of 2 or 3; and subclone fraction estimation not enabled.

To identify patient-specific somatic variants for CSF and plasma cfDNA ddPCR analysis, primary tumor and matched germline DNA were subjected to hybridization capture to either Human All Exome V6 panel for WES or a targeted custom-designed capture panel. For WES, library preparation was performed using the SureSelect XT Library Preparation Kit. Postcapture libraries were run on the NovaSeq 6000 acquiring 2- × 100-bp paired-end reads. Median coverage across all samples was 223×. Reads were aligned against GRCh38 using Burrows-Wheeler Aligner (BWA) v0.7.15, deduplication by Picard v2.0.1. SNV and indel variant calling on paired tumor/normal was performed with Mutect2 (25), and all variants annotated by ANNOVAR (2016–02–01 version; ref. 26). For target panel sequencing, DNA samples were sequenced using a targeted, custom-designed capture panel (RMH200 v1.3.1 panel) consisting of 233 cancer-related genes (27). NGS libraries were prepared using the KAPA HyperPlus Kit and Integrated DNA Technologies UDI 8-bp adapters, followed by target enrichment by hybridization to a custom DNA bait library (Nimblegen), and underwent 2- × 100-bp paired-end sequencing on NovaSeq 6000. Mean coverage across all samples was 280×. Analysis was performed using an in-house developed pipeline Molecular Diagnostic Information Management System version 3.0 (MDIMSv3) using the following bioinformatic software and versions: demultiplexing by bcl2fastq 2.17.1.14 to isolate reads for each sample, read alignment against GRCh37 by BWA v0.7.15, SNVs, and indel calling by Mutect2 from the GATK 3.5.0 suite. Variants were annotated by Personal Cancer Genome Reporter v.0.6 (PCGR). In cases where primary tumor was not obtainable, sequencing was performed on plasma cfDNA (n = 3), metastasis tissue (n = 1), or CSF tumor cells (n = 1).

Patient-specific candidate somatic driver mutations for ddPCR assay development were identified from targeted panel/WES of matched tumor and germline samples by the following criteria (i) cancer-associated hotspot variant (listed on https://www.cancerhotspots.org) and/or (ii) variant with high allele frequency in a Cancer Gene Census gene (28). Candidate variants (Supplementary Table S2) were manually inspected on IGV prior to selection for ddPCR assay design. Custom ddPCR assays for PIK3CA p.H1047R/H1047L/E545K/E542K, ERBB2 L755S, and TP53 p.R196*/R273H were designed in-house using primer3 software (https://primer3.ut.ee/). ddPCR assays for variants HIF1A p.A565G, KNL1 p. I654V, PIK3R1 p.450_452del, RB1 p.S829X, ANK1 p.A717T, PIK3CA p.G118D, and TP53 p.158fs/Q104*/R158C/T253fs, were designed using Custom TaqMan SNP Genotyping Design Tool (Thermo Fisher Scientific). Primers and probes, ordered from Thermo Fisher Scientific and detailed in Supplementary Table S3, were used at a final concentration of 900 and 250 nmol/L, respectively. Commercially available assays for AKT1 p.E17K and ATM p.R337H were purchased from Bio-Rad (Supplementary Table S4). ddPCR was performed as described previously (23). The exome-sequenced tumor DNA was used to validate the mutation by ddPCR and determine the optimum cycling conditions prior to assaying the matched cfDNA samples. There was no minimum cfDNA input in the ddPCR assays, however where sufficient cfDNA was available, 10 ng of DNA was assayed. All runs included tumor DNA (positive control) and matched germline DNA (negative controls) and wells containing assay without DNA (nontarget controls).

ddPCR analysis was performed on QuantaSoft Analysis Pro Software, version 1.7.4. Wells with total droplet count <10,000 were excluded, and a minimum of two mutant-FAM positive droplets per well was required to define the presence of tumor DNA. No sample was called positive for tumor content on the finding of only double (FAM and VIC) positive droplets because these might represent false positives due to the introduction of polymerase-driven errors during amplification.

Mutant allele fraction (MAF) was calculated as follows:

• ddPCR MAF = FAM (mutant) copies/μL[FAM (mutant) copies/μL + VIC (wild-type) copies/μL]

To assess the number of mutant copies per milliliter, the number of mutant-FAM positive droplets was adjusted for the number of wells run for the sample, the total number of droplets generated, the median volume of a droplet (0.89 pL), and volume equivalent of CSF/plasma per run, using the following formula:

• ddPCR mutant copies per milliliter = (total number of droplets positive for FAM) × 20,000 × (number of wells run/volume of plasma or CSF equivalents)/(total number of droplets generated × 0.89)

When using ddPCR to determine MAF, the lower limit of detection (LoD) is determined by the total number of DNA molecules screened, that is, the total DNA input. Because the DNA input varied per sample due to cfDNA yield, the minimum MAF detectable by ddPCR (LoD) for each sample was calculated using the following formula, and samples not achieving a ddPCR LoD of 0.10 were marked as failed:

• ddPCR LoD = 3/[haploid genome equivalents (FAM copies/μL + VIC copies/μL) × 20]

All statistical analysis was performed with GraphPad Prism version 9.1.0 and R 3.6.2. All P values are two-sided and considered significant if P < 0.05. All comparisons between two groups were made using two-tailed, unpaired Student t test. If the analysis did not pass normality test (Shapiro–Wilk test), groups were analyzed by Mann–Whitney U test. If more than two groups were compared, one-way ANOVA analysis was performed with post hoc Tukey's test for multiple comparison. Kaplan–Meier analysis using log-rank (Mantel–Cox) test was used to assess BCLM overall survival.

This study comprised 30 patients presenting with typical symptoms of CNS involvement (Table 1; Supplementary Table S1). Twenty-four out of 30 cases were diagnosed with confirmed BCLM by the treating oncology team, based on standard diagnostics and clinical assessments (BCLM+ group), and in two cases, synchronous brain metastases (BCBMs) were detected. The remaining six patients had suspected but nonconfirmed BCLM, and at median 39.7-month follow-up, had no evidence of BCLM (BCLM−). All six BCLM− patients were free of CNS metastasis at the time of CSF sampling; however, three out of six later developed BCBM at 9.5, 9.8, and 22.5 months, while remaining BCLM-free. In keeping with their known predilection for LM (8, 29), lobular breast cancers were enriched in the BCLM+ group, comprising 46% (11/24) of cases. In line with this enrichment, the predominant IHC subtype was hormone receptor-positive (HR+) HER2−, comprising 67% (16/24), meanwhile HER2+ cases comprised 12.5% (3/24) of the BCLM+ cohort.

Standard diagnostics performed were: CSF cytology in 30/30 cases; cranial ± spinal MRI in 29/30 cases. In nine out of 30 (30%) cases, multiple diagnostic lumbar punctures were performed. In BCLM+ cases, CSF cytology was suspicious (n = 2) or negative (n = 10) in 12/24 (50%) on initial lumbar puncture. Final CSF cytology (following repeat) was positive in 17/24 (71%) of BCLM+ cases. All six out of six BCLM− cases were negative on both initial and repeated cytology. In BCLM+ cases, MRI imaging was diagnostic in 20/23 (87%) with three scans remaining borderline following independent neuroradiologist review (Table 1). MRI imaging was negative in all BCLM− cases (1/6 had generalized dural thickening but no leptomeningeal enhancement).

CSF and plasma were collected at initial evaluation for BCLM (n = 30) and longitudinally during the course of intrathecal therapy in a subset of 12 BCLM+ patients (median 7.5 time points, range, 2–11; Fig. 1A). A median of 1.8 mL CSF (range, 0.3–26.0 mL) and 3.8 mL plasma (range, 1.0–12.0 mL) across all time points were obtained. CSF and plasma cfDNA were quantified (Fig. 1B) and fragment length assessed by capillary electrophoresis (Supplementary Fig. S1A). cfDNA concentration was significantly lower in CSF versus plasma (median, CSF cfDNA 21.9 ng/mL; range, 3.98–1363 ng/mL vs. plasma cfDNA 38.8 ng/mL; range, 8.45–1773; P = 0.005; Fig. 1B) and coupled with the smaller volumes obtained, this resulted in lower total cfDNA yield from CSF samples, in line with previous reports (30, 31).

To detect and quantify ctDNA fraction in CSF and plasma, we sought a tumor-agnostic approach requiring no prior sequencing of the patient's primary tumor, and a method suitable for low DNA input. To this aim, paired-end, ulpWGS was performed on 141 cfDNA samples (95 CSF and 46 plasma) with sequencing libraries created from as little as 0.5 ng cfDNA (Fig. 1C). The equivalent sample volumes used for library preparation were median 0.54-mL CSF (range, 0.05–3.96) and 0.72-mL plasma (range, 0.01–2.40). ulpWGS provided an accurate assessment of ctDNA fragment length, revealing differing cfDNA fragmentation patterns, with the highest fraction of reads corresponding to a shorter fragment length of 152 bp in CSF versus 166 bp in plasma (P < 0.0001 by Mann-Whitney U test; Supplementary Fig. S1B).

ichorCNA (1-Mb bin size) was used to identify large-scale copy-number aberrations and through probabilistic modeling to estimate tumor purity (ctDNA fraction) and ploidy (24). Representative examples in BCLM+ and BCLM− settings are shown in Fig. 1D. Further, performing ichorCNA analysis using a 50-kb bin size, focal gene amplifications were identified. Examples of ERBB2 amplification in CSF are shown in Supplementary Fig. S2.

Overall, the ctDNA fraction in CSF was significantly higher in patients with a BCLM+ diagnosis (n = 24, median, 0.57; range, 0.27–0.89) compared with BCLM− patients (n = 6; median, 0.03; range, 0.00–0.08; P < 0.0001; Fig. 2A). Adalsteinsson and colleagues determined that ichorCNA analysis of ulpWGS data has a high degree of accuracy for ctDNA fraction ≥0.10 (24); consequently, a cut-off of 0.10 was used in this study for ctDNA detection. CSF ctDNA fraction was ≥0.10 (detectable) in 100% of BCLM+ cases, and <0.10 in all BCLM− cases. Comparing CSF ctDNA fraction to standard BCLM diagnostics, ctDNA was above the detection level in all BCLM+ patients regardless of initial negative or suspicious CSF cytology (50% of BCLM+; Fig. 2A) and in those cases where MRI imaging was borderline (17% of BCLM+; Fig. 2B). CSF collection site (lumbar vs. ventricular) did not impair CSF ctDNA detection (Supplementary Fig. S3A).

Of the three out of six BCLM− cases who later developed BCBM, there was no difference in CSF ctDNA fraction compared with those who remained free of CNS disease (Fig. 2C).

These data suggest that determining the ctDNA fraction in CSF has greater diagnostic sensitivity in BCLM than CSF cytology or MRI and correctly identifies patients who are BCLM-free. Furthermore, ulpWGS determination of tumor content was successful in all CSF samples, and there was no requirement for further CSF sampling, contrary to the 33% of cases where repeated lumbar puncture was indicated for definitive CSF cytology. Of note, there was no correlation between extracted CSF cfDNA concentration (ng/mL) and CSF ctDNA fraction, indicating that cfDNA concentration alone cannot act as a surrogate for tumor fraction (Fig. 2D). Similarly, there was no correlation between the CSF volume collected and either cfDNA concentration or CSF ctDNA fraction (Supplementary Figs. S3B and S3C).

Consistent with previous reports (18), ctDNA fraction was significantly lower in plasma than CSF in BCLM+ cases (median ctDNA fraction in plasma 0.071 vs. in CSF 0.566, P < 0.0001) with 12/22 plasma samples below ctDNA detection level. In contrast, in BCLM− cases, there was no significant difference in ctDNA fraction between the five paired plasma and CSF samples (Fig. 2E). Plasma ctDNA was found to be detectable (≥0.10 ctDNA fraction) only in those with progression of extracranial disease or who had previously received whole brain radiotherapy (Fig. 2F).

ddPCR is an established methodology to determine ctDNA abundance through direct measurement of tumor-specific somatic mutations in cfDNA. For validation of ichorCNA tumor purity estimates, we assessed the concordance between ddPCR MAF and ulpWGS ctDNA fraction on 107 baseline and serial samples (73 CSF and 34 plasma). Patient-specific somatic variants, identified from targeted or whole exome sequencing of matched tumor samples (Fig. 3A; Supplementary Table S2), were used to design custom ddPCR assays (Supplementary Tables S3 and S4). Figure 3B illustrates findings for two patients in this study. For BCLM+ patient RMH006 (top), ulpWGS estimated a ctDNA fraction of 0.67 and 0.01 in the matched CSF and plasma samples, respectively. ddPCR identified abundant mutant PIK3CA E542K copies in the primary tumor and CSF, with a mutant CSF allele fraction (MAF) of 0.696; however, no plasma PIK3CA E542K mutant copies were detected within 746 haploid genome equivalents. In BCLM− patient KCL663 (bottom), ctDNA was undetectable in both CSF and plasma (ctDNA fraction 0) by ulpWGS. By ddPCR, the TP53 T253fs variant was found at high MAF (0.596) in the primary tumor, but no mutant TP53 T253fs copies were detected within 150 and 3,300 haploid genome equivalents screened in CSF and plasma, respectively.

There was a strong correlation between ddPCR MAF and ulpWGS ctDNA fraction for the 107 samples undergoing both methods (Fig. 3C, Spearman r = 0.852, P < 0.0001; Supplementary Fig. S4A–S4C) showing baseline samples only). The majority of samples (101/107) achieved high-confidence mutation detection at sufficient ddPCR LoD (as defined in Materials and Methods); however, in the remaining six samples (one baseline CSF; five serial CSF), there was insufficient cfDNA yield, attributable to lower sample volumes collected (median 0.55 mL vs. 2.0 mL; Supplementary Fig. S4D).

Of the 24 BCLM+ patients, 19 received intrathecal methotrexate and/or trastuzumab therapy, and six patients received cranial or spinal radiotherapy (Fig. 4A). In 12 patients undergoing intrathecal therapy, serial CSF samples were collected immediately before treatment administration. Baseline ctDNA fraction showed only a weak nonsignificant correlation with BCLM survival in all 24 BCLM+ patients (Fig. 4B), and there was no significant difference in median overall survival associated with receipt of intrathecal therapy (5.7 months with vs. 3.4 months without intrathecal therapy, P = 0.55; Fig. 4C). Despite this, achieving ctDNA suppression on intrathecal therapy (CSF ctDNA fraction <0.10) was associated with a significantly lengthened survival (Fig. 4D, median survival 11 months vs. 3.25 months, P = 0.0337). Furthermore, a reduction in ctDNA fraction from baseline to the second time point (at a median of 3 weeks) was observed more frequently in those achieving PFS ≥6 months on intrathecal therapy (Fig. 4E).

In the 12 patients undergoing longitudinal sample collection on intrathecal therapy, ctDNA fraction, whether measured by ulpWGS or ddPCR, showed dynamic and quantitative changes on therapy, which associated with treatment response, failure to achieve response, and early notification of therapy failure (Fig. 5A and B; Supplementary Fig. S5A–S5L). CSF cfDNA concentration and CSF protein levels were not robust indicators of disease status. ddPCR provides two quantitative measures: MAF and mutant copies per milliliters of CSF/plasma volume. Both ddPCR measures showed similar trends to ulpWGS in ctDNA fraction monitoring (Supplementary Fig. S5A–S5L).

Patient KCL553 (Fig. 5A; Supplementary Fig. S5A) commenced intrathecal therapy during week 0 of BCLM diagnosis. The ctDNA fraction of 0.88 by ulpWGS, reduced during treatment to 0.16 at week 15, and was undetectable at week 29, in line with marked improvement in clinical symptoms. MRI at baseline showed decompensated communicating hydrocephalus, which improved on serial imaging at week 28. At week 52, although the patient remained well with stable MRI and negative CSF cytology, the ctDNA fraction rose to 0.27. Twelve weeks later, the patient developed marked confusion and vomiting, the MRI scan performed showed new abnormal leptomeningeal enhancement, and intrathecal methotrexate was discontinued. The patient succumbed to BCLM 9 weeks later. In patient RMH006 (Fig. 5B; Supplementary Fig. S5B), CSF sample collection started at week 12, at the time of worsened hydrocephalus in the setting of MRI-positive BCLM. At this time point, the ctDNA fraction was 0.66 and reduced to undetectable levels during the next 8 weeks of intrathecal methotrexate therapy. At week 18, neurological symptoms were moderately improved, and MRI showed reduction in leptomeningeal enhancement and hydrocephalus. Although CSF cytology remained negative, ctDNA fraction rose from the nadir of 0 at week 20, to 0.16 at week 22, and 0.28 at week 28. Malignant cells were later detected on CSF cytology at week 30. MRI imaging at week 25, on retrospective review, showed more conspicuous leptomeningeal enhancement in the left internal auditory meatus. Neurologic symptoms worsened at week 34, and the patient succumbed to BCLM at week 46. These two cases illustrate the role of ctDNA fraction monitoring for earlier prediction of therapy failure, which opens up an opportunity window for switching to alternative treatment approaches or, importantly, timely cessation of futile intrathecal therapy and earlier focus on best supportive care.

The other 10 cases (Supplementary Figs. S5C–S5L) provide further examples of the potential clinical utility of ctDNA monitoring. For example, monitoring ctDNA levels could differentiate intracerebral methotrexate toxicity from BCLM progression [Supplementary Figs. S5G (KCL625) and S5I (KCL650)] and detect failure of adequate treatment response, as measured by persistently elevated ctDNA fraction, despite apparent improvements in CSF cytology and/or MRI [Supplementary Figs. S5F (KCL622), S5H (RMH011), and S5L (5197)].

This proof-of-concept study demonstrates that ulpWGS-assessed CSF ctDNA fraction is a promising tumor-agnostic marker with potential to improve both current diagnostics and therapeutic response monitoring in BCLM. Notably, CSF ctDNA detection correctly identified all BCLM+ cases, whereas only 50% were detectable through initial CSF cytology. Using ctDNA detection in CSF during diagnostic work-up could therefore avoid the need for multiple lumbar punctures to reach, or confidently refute, BCLM diagnosis.

We demonstrate that both ulpWGS and ddPCR are suitable methods to determine ctDNA fraction in CSF liquid biopsies, with good correlation between approaches. However, ulpWGS is the preferable method because it (i) can be performed without prior primary tumor sequencing, (ii) is not constrained by low DNA yields or small-volume CSF samples, and (ii) provides a more time- and cost-effective approach than ddPCR. Furthermore, the genomic landscape of breast cancer, with relatively few cancer driver mutations but a highly copy-number aberrant architecture, is suited to a genome-wide approach for tumor-derived DNA detection (32). Furthermore, ulpWGS was able to detect the acquisition of ERBB2 amplification, which has an increased occurrence in CNS metastatic disease (33). Although no cases of ERBB2 amplification occurring in CSF of hitherto HER2-negative cases were found in this study, detection of such events would have clinical use in guiding pharmaco therapeutic options.

This study provides further evidence that CSF is an abundant source of ctDNA in CNS-predominant malignancy, with minimal detection in the plasma unless there was coexistent progression in extracranial metastatic sites. CSF ctDNA fraction in BCLM+ patients (median, 0.57; range, 0.27–0.89), was at much higher abundance than typically found in plasma during progression of systemic metastases (median ∼0.05; ref. 34), in keeping with other reports on CSF ctDNA analysis in primary and metastatic CNS cancer (18, 31). Likely explanations for this are the smaller circulating volume of CSF versus plasma [150 mL (35) vs. 2,800 mL (36), respectively], reduced levels of nontumor, leukocyte-derived cfDNA in CSF (37) and reduced CSF turnover compared with plasma (38). Therefore, although CSF sampling is a more invasive method to obtain liquid biopsy than by simple blood test, there is a clear rationale to sample CSF, which is a rich source of ctDNA for the study of CNS disease.

During the preparation of this article, White and colleagues published findings on a cohort of 22 patients with solid tumor LM and, using a similar WGS methodology, demonstrated improved sensitivity and accuracy of ulpWGS-assessed CSF ctDNA for presence of LM when compared with diagnostic cytology (39). This supports the translatability of CSF ctDNA assessment by ulpWGS to other leptomeningeal metastases settings and primary and secondary brain malignancies, where ctDNA is known to be shed into CSF (18, 31, 40, 41).

A major obstacle in treating BCLM is the variability in response to the leptomeningeal-directed therapy of intrathecal chemotherapy. In this study, 10 of 19 patients receiving intrathecal therapy survived with BCLM for greater than 6 months. However, the invasive nature of this treatment, and difficulty in selecting patients likely to benefit, often discourages its administration. Currently, the only response predictors in routine clinical practice during intrathecal therapy are: (i) CSF cytology clearance by 8 weeks (42), and (ii) serial MRI imaging, generally assessed 2 to 3 months following initiation of treatment (12). This study explored the value of early ctDNA monitoring, at a median of 3 weeks into intrathecal therapy. ctDNA suppression (to ctDNA fraction <0.10) during treatment was a significant predictive marker of longer BCLM survival. Conversely, in patients with markedly rising ctDNA fraction at this early time point, there was a lack of treatment response (failure to achieve PFS >6 months on treatment). Therefore, CSF ctDNA serial monitoring has the potential to identify patients early in treatment for whom invasive intrathecal therapy is unlikely to benefit, thereby allowing a switch to alternative therapeutic approach such as systemic therapy, radiotherapy or novel approaches such as immunotherapy (43). Similarly, rising ctDNA fraction in patients who initially responded to treatment anticipated BCLM progression by up to 12 weeks. If these findings are validated in an appropriately designed clinical biomarker study, ctDNA measurement in BCLM may allow the option for a timely switch in therapy, prior to patient's performance status declining and precluding further treatments, or an earlier focus on best supportive care.

In conclusion, within the limitations of a small study without predefined study design, the data presented provide proof of concept for ulpWGS-assessed CSF ctDNA monitoring in BCLM as a potential biomarker to improve current diagnostics and evaluation of treatment response and failure. Prospective assessment in a larger, appropriately designed clinical biomarker validation study (44, 45) is warranted, and this should include an adequate panel of control including patients with brain metastasis without BCLM and patients with nonmalignant brain conditions; minimum sample requirements; predefined clinical assessments including standardized neurologic and radiologic evaluations by RANO criteria (12); and establishment of optimal ctDNA diagnostic threshold, to enable a detailed analytical interpretation of the clinical utility of this promising biomarker.

A.F. Okines reports grants and personal fees from Pfizer and Roche; personal fees from AstraZeneca and Seagen; other support from Leo Pharmaceuticals; and personal fees and other support from Eli Lilly and Diachi Sankyo outside the submitted work. N.C. Turner reports advisory board honoraria from AstraZeneca, Bristol Myers Squibb, Eli Lilly, Merck Sharpe & Dohme, Novartis, Pfizer, Roche/Genentech, GlaxoSmithKline, Zentalis Pharmaceuticals, Repare Therapeutics, and Arvinas and research funding from AstraZeneca, BioRad, Pfizer, Roche/Genentech, Merck Sharpe & Dohme, Guardant Health, Invitae, Inivata, Personalis, and Natera. A.N. Tutt reports other support from Tesaro/GlaxoSmithKline, AstraZeneca, Merck KGaA, Medivation, Myriad Genetics, MD Anderson, Medscape, VJ Oncology, SABCS mini symposium, EM Partners, Gilead, Inbiomotion, and CRUK; personal fees from Pfizer, Vertex, Artios, Prime Oncology, and Gilead; grants from Breast Cancer Now, CRUK, and AstraZeneca outside the submitted work; and has a patent with AstraZeneca with royalties paid to the Institute of Cancer Research with royalties paid from AstraZeneca. No disclosures were reported by the other authors.

A. Fitzpatrick: Conceptualization, resources, data curation, formal analysis, funding acquisition, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing. M. Iravani: Resources, data curation, formal analysis, investigation, visualization, writing–original draft, project administration, writing–review and editing. A. Mills: Software, formal analysis, writing–review and editing. L. Childs: Resources, data curation, investigation, writing–review and editing. T. Alaguthurai: Resources, writing–review and editing. A. Clifford: Resources, writing–review and editing. I. Garcia-Murillas: Resources, investigation, methodology, writing–review and editing. S. Van Laere: Resources, writing–review and editing. L. Dirix: Resources, writing–review and editing. M. Harries: Resources, writing–review and editing. A. Okines: Resources, writing–review and editing. N.C. Turner: Resources, writing–review and editing. S. Haider: Software, formal analysis, writing–review and editing. A.N. Tutt: Conceptualization, resources, supervision, funding acquisition, methodology, writing–original draft, project administration, writing–review and editing. C.M. Isacke: Conceptualization, resources, supervision, funding acquisition, visualization, methodology, writing–original draft, project administration, writing–review and editing.

We thank the patients who participated in the studies, along with the study and clinical staff who assisted with the sample collection. For patient enrollment and sample collection in South East London Cancer Research Network, we thank clinicians S. Irshad, H. Kristeleit, J. Mansi, M. Nathan, E. Sawyer, A. Swampillai, and the study nurses and staff, and R. Liccardo, R. Marlow, and V. Shah for patient enrollment and sample collection. For patient enrollment and sample collection in Royal Marsden NHS Foundation Trust, we thank clinicians P. Barry, S. Johnston, M. Parton, M. Robert, and L. Ulrich and the study nurses and staff, A. Bambra, D. Kelly, S. Thompson. We thank the study nurses and staff in GZA Hospital Sint-Augustinus. We thank M. Beaney and S. Hrieben for laboratory advice and ddPCR assays, J. Seoane and R. Mayor at Vall d'Hebron Institute of Oncology for advice on the extraction of CSF cfDNA, the ICR Tumour Profiling Unit, and A. Gao for bioinformatics input. This work was supported by Medical Research Council Clinical Research Training Fellowship (MR/P001564/1 to A. Fitzpatrick, C. Isacke), NIHR Biomedical Research Centre (BRC) at the Royal Marsden and the ICR postdoctoral support funding (W94500; to A. Fitzpatrick), Breast Cancer Now Programme Funding to the Breast Cancer Now Toby Robins Research Centre (CTR-Q4-Y1-5, CTR-Q5-Y1 to C. Isacke, N. Turner, and A. Tutt), Breast Cancer Now Research Unit at Kings College London (KCL-Q2-Y1-5, KCL-Q3-Y1 to A. Tutt), and Cancer Research UK Centre Grant funding to the Kings College London (CRUK RE15196-9 to A. Tutt).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.