Detection of human brain cancers using genomic and immune cell characterization of cerebrospinal fluid through CSF-BAM

Abstract

Patients with radiographically detectable lesions in their brain or other symptoms compatible with brain tumors pose challenges for diagnosis. The only definitive way to diagnose such patients is through brain biopsy, an invasive and dangerous procedure. Here we present a new workflow termed “CSF-BAM” that simultaneously identifies B cell or T cell receptor sequences, Aneuploidy, and Mutations using amplification of both strands of the DNA from cerebrospinal fluid (CSF) samples. We applied CSF-BAM to a validation set of 209 samples from patients with brain cancers. Among the 129 samples from patients with the most common aggressive cancer types, the sensitivity of detection was 81%. None of 30 CSF-BAM assays were positive in CSF samples from patients without brain cancers (100% specificity). CSF-BAM provides an integrated approach to identify neoplasia in the central nervous system, provides information about the genetics and immune environment, and has the potential to inform patient management.

Introduction

Brain cancers represent a heterogenous but highly aggressive class of neoplasms. They can be broadly categorized as primary or metastatic. Glioblastoma and medulloblastoma represent the most common types of primary brain cancers in adults and children respectively (1–3). Brain metastases occur in over 200,000 individuals in the United States every year and can be classified as parenchymal, which occur in the substance of the brain and are more common, or leptomeningeal, which occur in the lining of the brain in up to 10% of all cancer patients (4,5). Lung, breast, colon, melanoma, and renal cancers represent the most common cancer types to metastasize to the brain. Despite aggressive multi-modality treatments, both primary and metastatic brain cancers are associated with abysmal long-term survival, with most forms of brain cancer being incurable (6–9).

The primary and almost exclusive methodology to diagnose brain cancers remains neurosurgical biopsy, which has significant inherent risks and expense. A brain biopsy is unlike tissue sampling in any other organ. It typically requires general anesthesia with inpatient hospitalization, and it carries a 5–10% risk for neurological decline, a 1% risk for catastrophic hemorrhage, a 5–10% risk for a non-diagnostic result and is susceptible to sampling bias as only a very small fragment of the tumor is extracted (10–16). In select cases, such as with suspected leptomeningeal disease (LMD) or in primary central nervous system lymphoma (PCNSL), CSF cytology can aid in diagnosis but the sensitivity is relatively low (17–22). To date, about 1 in 4 patients with LMD are diagnosed based on imaging or clinical findings while having negative cytology (23). In addition, outside of tissue sampling, there is no methodology to understand the molecular and cellular composition of CNS cancers. However, it is becoming increasingly evident that understanding the genetic and epigenetic drivers of CNS neoplasms is essential for establishing a diagnosis for both primary and metastatic brain cancers, enabling prognostication and for informing therapeutic decision making (2,6,7,24–26). The primary diagnostic criteria based on the WHO 2021 guidelines incorporate chromosomal copy number alterations and disease defining somatic mutations in addition to DNA methylation and protein expression (27). Interrogating CSF through liquid biopsies, which has demonstrated superior performance compared to evaluating peripheral blood, represents a promising means of accessing such genetic and epigenetic information for CNS neoplasms without the need for potentially morbid tissue biopsy (28–37). Additionally, improved CSF liquid biopsy will provide essential information about response and resistance to investigational agents in clinical trials, particularly because repeat tissue sampling is not commonly performed (38).

In recent years, there has also been a burgeoning understanding of the central relationship between the immune microenvironment and susceptibility to immune modulating anti-cancer therapies (39–42). For example, the advent of immune therapies such as checkpoint inhibitors and CAR T cell approaches has increased the need for understanding the relationships within immune cell populations associated with CNS cancers (43–46). In addition, some of the rare but challenging brain cancers are primary central nervous system lymphomas, which are lymphomas thought to originate in the brain (47,48). Understanding the cancerous cell clones and their molecular phenotype has important biological and therapeutic ramifications (49–53). CSF is essential to interrogate for this information as expanded clones may not be represented in peripheral blood (54). Similarly, CSF is often obtained in patients with CNS disease other than neoplasia, and the analysis of the B cell and T cell repertoire in those patients can provide meaningful information for diagnosis and management (55–60). This is particularly relevant to autoimmune, inflammatory and infectious diseases affecting the CNS.

With these considerations in mind, we sought to establish an approach that could supplement neurosurgical biopsy by identifying and characterizing brain tumors through interrogation of genetic and immune cell profiles in CSF. There are several key capabilities that would be required for such an assay: 1) identification of driver mutations across a wide array of primary and metastatic cancers, 2) genome wide identification of chromosomal copy number alterations, 3) characterization of B and T cell populations, 4) robust compatibility with a relatively small amount of CSF DNA, 5) a uniform workflow that could use the same starting material for all analyses to minimize cost while maximizing ease, yield, sensitivity, and reproducibility, and 6) an informatics pipeline capable of analyzing and integrating these heterogenous datasets. The design and execution of such an approach is described in this manuscript (Fig. 1). We show that the assay developed here is able to identify brain tumors and provide clinically actionable information from CSF liquid biopsy rather than neurosurgical tissue biopsy.

Figure 1.

CSF-BAM overview. CSF is obtained and DNA is extracted from the entire sample. CSF-BAM examines three analytes simultaneously: B cell or T cell receptor rearrangements, Aneuploidy, and Mutations using PCR-mediated amplification of both strands of the DNA.

Results

Overview

Once DNA was purified from CSF, a DNA library was generated through a modified version of a protocol described previously, named SaferSeqS, which produces a relatively high conversion efficiency of the original DNA template molecules into library DNA molecules (see Methods) (61). This conversion efficiency was particularly important when the quantity of CSF fluid was limited or when the DNA concentration in that fluid was low. Equally importantly, the SaferSeqS library preserved DNA from both the Watson and Crick strands of the original DNA templates. The ability to independently assess both strands of DNA exponentially increases the accuracy of the resulting sequencing data when the fraction of aberrant DNA molecules is low (61–63). SaferSeqS libraries contain ~200 copies of each of the original template molecule strands, and therefore can be used for multiple downstream assessments of DNA (61). For CSF-BAM, we chose to analyze the clonal composition of DNA derived from malignant or normal B or T cells, as well as chromosome copy number alterations and somatic mutations derived from the cancer cells (Fig. 2). The paradigm for evaluation of each of these three components was identical and was performed in distinct stages to maximize reproducibility and minimize overfitting (Supplementary Table S1):

1. Analytical stage: optimize the experimental procedures and bioinformatic analysis using DNA from peripheral blood leukocytes or plasma cell-free DNA from healthy individuals.

2. Training stage: use the optimized procedure to evaluate DNA from CSF or blood from a different cohort of patients with and without cancer to establish thresholds for specificity and estimate sensitivity at the chosen thresholds.

3. Validation stage: use the optimized procedure to evaluate CSF from an independent cohort of patients to determine sensitivity and specificity at the pre-defined thresholds for positivity.

Figure 2.

Schematic of CSF-BAM. A, Independent libraries are generated from each strand of the original DNA template molecules in a manner where unique molecular identifiers (UIDs) allow the strands from the two libraries to be mapped back to their original duplex. B, BCRs are evaluated with SafeBSeqS. Illustrated here is amplification of BCRs with a primer targeting the IGHJ1 segment, among the multiplex set of 4 total IGHJ primers. C, Aneuploidy is evaluated with WGS through amplification of total libraries using adapter-specific primers. D, Mutations are evaluated with SaferSeqS. Illustrated here is amplification of IDH1 using an IDH1-specific primer, among the multiplex set of 120 total gene-specific primers.

Development of the experimental procedures and bioinformatic pipelines for CSF-BAM

The workflow for all three components of CSF-BAM begins with creation of libraries from original DNA template molecules (Fig. 2A). Copies of the DNA templates in the library, which are mappable to the original templates, are then split for analysis with each of the “B,” “A,” and “M” components (Supplementary Table S2A–D). Because essentially unlimited copies of DNA templates can be generated in the amplified library, material is not limiting after the initial library generation step and there are no input constraints on splitting copies between the different analytical components.

The “B” component of CSF-BAM queries the B cell (BCR) and T cell (TCR) receptor genes that are integral to the adaptive immune system. In any single individual, millions of BCR and TCR receptors in normal B and T cells, respectively, are generated through imprecise joining of variable (V), diversity (D), and joining (J) segments of the BCR or TCR genes. The nature of these sequences and the degree of clonality provides a wealth of information about the adaptive immune system in that particular patient. In comparison to flow cytometry, TCR or BCR sequencing provides comprehensive repertoire descriptions (64). Moreover, because any B cell or T cell cancer is derived from a single B cell or T cell, respectively, a neoplastic clone is characterized by a single VDJ rearrangement. The presence of malignant cells in CSF can thereby be detected by the over-representation of a single sequence in the CSF, implying a predominant clone. Although such clones can be detected through previously published methods that sequence either RNA or DNA templates, accurate detection and quantification of clonotypes is challenging. Among the reasons for this is that sequencing from DNA templates has generally required multiplex combinations of primers to amplify all possible V and J gene segment pairs (65–67).

SafeBSeqS and SafeTSeqS overcome this challenge by requiring primers for only the J segments of one of the BCR (IGH) or one of the TCR (TCRB) genes, respectively (Fig. 2B). Though these data cannot be used to analyze the constant region, they can be used to re-construct the entire VDJ sequence of the TCR or BCR. They are therefore adequate to determine the clonal representation of any population of B cells or T cells as well as to identify certain characteristics of the rearrangements associated with cancers or autoimmune disease (68).

Experimentation in the analytical stage was conducted to empirically test DNA shearing conditions, library preparation methods, primer sites, primer sequences, primer combinations, primer ratios, and reaction conditions using DNA from different primary cell types, cell lines, and synthetic constructs with a goal to maximize yield and uniformity of representation while minimizing errors. We found that amplification using four primer sets were sufficient to assess the entire BCR repertoire with SafeBSeqS (Supplementary Table S2A). One primer amplified gene segments IGHJ1, IGHJ4, and IGHJ5, and one primer each amplified IGHJ2, IGHJ3, and IGHJ6. These four primers were mixed together and used for hemi-nested amplification of the SaferSeqS libraries. They yielded uniform amplification of all the queried gene segments as tested on a sample of DNA derived from fibroblasts with a uniform representation of gene segments (Supplementary Fig. S1A–B). For the training stage, we then applied SafeBSeqS to CSF samples from 25 individuals with primary and secondary CNS lymphomas and peripheral blood samples from 95 healthy control individuals without cancer. Peripheral blood samples were used as negative controls because they enabled setting conservative thresholds to ensure high specificity while reserving the more limited number of CSF samples for use in the validation stage. The summaries of results are listed in Supplementary Table S3 and Supplementary Table S4. Based on these results, we defined the positive criteria for clonality as total UIDs ≥20 and top clone UIDs / total UIDS ≥0.3 (30%). In total, 1/95 samples from healthy individuals and 13/25 CSF samples from patients with CNS lymphoma in our set met these criteria (Supplementary Tables S3 and Supplementary Table S4). For the validation stage, when SafeBSeqS was applied to CSF from a different cohort of individuals without known cancers, we found that 0 met the positive criteria for clonality (Supplementary Table S4). Similarly, 1 of the 205 CSF samples from patients with cancers other than CNS lymphomas scored positively (Supplementary Table S4). Of 4 CSF samples from patients with B cell lymphomas of the CNS, 2 scored positively in this assay with clonal fractions of 32% and 86% (Supplementary Table S4).

In the analytical stage for SafeTSeqS, we found that 13 primers sets, one for each TRBJ gene segment, were sufficient to assess the TCR repertoire (Supplementary Table S2B). Uniform amplification of all 13 J segments was obtained in control fibroblast DNA (Supplementary Fig. S2A–B). Supplementary Table S5 lists the summary data of SafeTSeqS data obtained from the evaluation in the training stage of DNA from peripheral white blood cells of 95 healthy control individuals. Using the same criteria as for SafeBSeqS, we defined positive clonality as total UIDS ≥20 and top clone UIDs / total UIDs ≥0.3 (30%). Zero samples met these criteria. When SafeTSeqS was applied for the validation stage to CSF from a different cohort of 30 samples from individuals without known cancers, we found that 0 samples met the clonality criteria (Supplementary Table S6). Three of the 205 CSF samples from patients with cancers other than CNS lymphomas scored positively in this assay (Supplementary Table S6).

The initial validation stage for CSF-BAM was limited in that it included only 4 samples derived from patients with B cell cancers and 0 samples derived from patients with T cell cancers, the types most relevant for the SafeBSeqS and SafeTSeqS components. We therefore applied the SafeBSeqS and SafeTSeqS assay components to a second validation set of CSF samples. The samples in this set could not be used for validation with full CSF-BAM because they lacked matched peripheral white blood cell samples necessary for the mutation analysis component. This second validation set consisted of CSF samples from 74 patients with peripheral B cell cancers and 14 patients with peripheral T cell cancers (Supplementary Table S7). The set was divided into three groups. The first group consisted of samples that were interpreted to be true positives for CNS disease because the patients had concurrent evidence for active CNS disease on cytopathology, flow cytometry, or MRI. For these, 18/29 samples from patients with B cell cancers and 9/10 samples from patients with T cell cancers were positive for clonality on CSF-BAM for sensitivities of 62% and 90% respectively. The second group consisted of patients with no evidence of active CNS disease on cytopathology, flow cytometry, or MRI and who had no documented CNS relapse subsequent to CSF sampling. These samples were interpreted to be true negatives. Of these, 0/41 samples from patients with B cell cancers and 0/3 samples from patients with T cell cancers were positive on CSF-BAM for specificities of 100%. The third group consisted of patients with no evidence for active CNS disease at the time of CSF collection but who did have subsequent CNS relapse, meaning that there was possible but uncertain low level CNS involvement at the time of CSF sampling. For these, 0/4 B cell cancer and 0/1 T cell cancer samples were positive on CSF-BAM. Finally, for additional validation of SafeTSeqS, we assessed peripheral white blood cell DNA from 3 patients with peripheral T cell neoplasms who had clonal T cells present in the blood on clinical flow cytometry testing. Of 3 such patients, 3 scored positively in this assay, with clonal fractions of 55%, 57% and 95% (Supplementary Table S5).

The “A” component of CSF-BAM queries aneuploidy throughout the genome. SaferSeqS libraries are converted to a form suitable for whole genome sequencing (WGS) by the addition of primers whose sequences match those of the Illumina NovaSeq flow cells (Fig. 2C and Supplementary Table S2C). The WGS data are mapped to the human genome through standard methods. A modified version of the ichorCNA algorithm (69) is used to assess gains or losses on 34 chromosome arms. Sex chromosomes, acrocentric chromosomes, and arms with high background are excluded in the analysis (see Methods). Though Watson and Crick strands can be identified from the sequencing data, there is no need to couple the reads to ensure single base pair accuracy to assess copy number alterations of entire arms.

For the training stage we applied this WGS assay to SaferSeqS libraries from CSF of 31 individuals without cancer (training cohort, Supplementary Table S8). We used two metrics of aneuploidy to derive a threshold for scoring samples as aneuploid. First, we evaluated the number of arms altered. In our training cohort, two of the 31 samples had exactly one arm altered while the remaining 29 samples had no arms altered. As a result, any sample with more than one arm altered would be scored as positive for aneuploidy. Next, we evaluated the estimated tumor fraction as predicted by ichorCNA. In our training cohort of individuals without cancer, the largest predicted tumor fraction was 1.4%. Any sample predicted to have a tumor fraction >1.5% would be scored as positive.

When the same assay was applied for the validation stage to WGS data derived from SaferSeqS libraries of CSF from a different cohort of individuals without active cancers, we found that 0 of 30 scored positively (100% specificity, credible interval 89% to 100%) (Supplementary Table S9). 121 of 209 CSF samples from patients with CNS cancers of various types had detectable aneuploidy. Of these, 55 of 83 samples from patients with high grade gliomas were positive (66%, credible interval 56% to 76%). Patients with medulloblastomas or cancers outside the CNS that had metastasized to the brain also were often positive for aneuploidy using this measure (77%, credible interval 56% to 90% and 88%, credible interval 69% to 95%, respectively, Supplementary Table S9). In addition to overall positivity or negativity, the aneuploidy component also described chromosome arm changes, details that can further inform diagnosis, prognostication, and management (Supplementary Table S9).

The reproducibility of the approach was assessed through the evaluation of independent aliquots of CSF DNA as technical replicates from the same patient (n = 104). Each aliquot had an independent SaferSeqS library. For the non-cancer controls (n=25) with a technical replicate, all arms were concordant while 98.5% of the arms in the cancer technical replicates (n=80) were concordant (Supplementary Table S10; Supplementary Fig. S3A–B; Methods). A subset of samples (n=30) were previously described and evaluated using the Repetitive Element Aneuploidy Sequencing System (RealSeqS)(33). RealSeqS uses a single PCR primer to concomitantly amplify ~350,000 loci spread through genome in order to evaluate aneuploidy (70). We compared the chromosome arm level calls between the two assays and found 94.1% of the arms were concordant (Supplementary Table S11; Methods). Of the discordant calls, a majority could be explained as falling just below the threshold for one of the two assays.

The “M” component of CSF-BAM identifies subtle somatic mutations such as single base substitutions (SBS) or small insertions or deletions (indels). For this component, extremely high specificity is required to minimize errors during the experimental or bioinformatic components of this assay. The workflow is identical in principle to that described above for SafeTSeqS and SafeBSeqS (Fig. 2B and D) but different primers are used. Instead of 4 primers for SafeBSeqS and 13 primers for SafeTSeqS, 120 primers were used for mutation analysis, (Fig. 2D). These primers (Supplementary Table S2D) were chosen after extensive experiments in the analytical stage to maximize the uniformity of representation of the amplicons as well as minimize the number of off-target reads upon sequencing. The genomic regions for evaluation were selected based on a previously described algorithm that optimizes the number of cancers detectable with the fewest amplicons (71). The cancers of interest here included primary CNS tumors and common human cancers including those that frequently metastasize to the CNS (71). In sum, the panel consisted of 120 amplicons, encompassing regions in 40 genes. These span coding regions in oncogenes and tumor suppressor genes, splice sites in tumor suppressor genes, and promoter hotspots for TERT. The precise coordinates for each amplicon are listed in Supplementary Table S2D. The uniformity of amplification of the 120 amplicons queried by the 120-plex is shown in Supplementary Fig. S4A–E.

It is well known that mutations in cell-free DNA (cfDNA) from peripheral blood largely arise from either tumors or CHIP (clonal hematopoiesis of indeterminate potential) (72). To help ensure mutations identified in the CSF were not a result of CHIP, we analyzed matched WBC DNA in available cases as previously reported (73).

For the training stage we evaluated CSF from a different cohort of 300 individuals without cancer (Supplementary Table S12). These samples do not have matched peripheral blood to eliminate CHIP mutations and were not used for the evaluation of performance metrics. We used this cohort to tune our somatic mutation calling approach and determine thresholds for positivity (training set). We report 20 mutations in TP53, 3 in KRAS, 5 NRAS, 8 mutations in FBXW7 in codon 505, and 7 in other amplicons. Given the abundance of non-canonical mutations in KRAS and NRAS in the non-cancers, we restricted future mutation calls within these genes to only KRAS codon 12 and NRAS codon 61, which represent the most commonly mutated hotspots in cancer.

The FBXW7 codon 505 mutations in the non-cancer samples were unexpected. This specific codon is not typically mutated in CHIP and other codons throughout FBXW7 are typically mutated in cancer. Upon closer inspection, every molecule with a mutant FBXW7 codon 505 also had mutations at codons 289, 299, 300, and 314. We performed blat analysis to all possible non-human genomes (74). The sequence from the observed mutated FBXW7 molecules perfectly matched the bovine genome. We hypothesized that bovine-derived hemostatic agents frequently used in neurosurgery (75), when CSF was collected for the trigeminal neuralgia samples, may have contributed minor amounts of DNA that amplified and incorrectly aligned to FBXW7.

On the basis of these data, we chose to positively score mutations that were present in more than one original template molecule and found in cancer patients in the COSMIC database (see Methods). Given the importance of TERT promoter mutations in CNS neoplasms, we relaxed this metric to score samples with even one mutant template molecule as positive.

Using these thresholds for the validation stage, we identified somatic mutations in 0 of 30 individuals without cancer (credible interval 89% to 100%) (Supplementary Table S13). 79 of 209 CSF samples from patients with CNS cancers of various types had detectable mutations (38%, credible interval 31% to 45%). Of these, 44 of 83 samples from patients with high grade gliomas were positive (53%, credible interval 42% to 63%). Of the 5 high grade gliomas that were scored as positive based on the presence of only one mutant TERT molecule, one had the canonical gain on chr7 above the threshold for aneuploid positivity. Given the heterogenous nature of medulloblastoma driver mutations, only 4 of 22 (18%, credible interval 8% to 39%) samples scored positive for mutations. 18 of 24 metastatic cancer samples (75%, credible interval 56% to 88%) had mutations detected.

Application of CSF-BAM to CSF samples (validation stage)

We evaluated 239 CSF samples from 222 patients; for each sample, peripheral blood WBC DNA was available to exclude any mutations due to CHIP. Clinical information including demographics are described in Supplementary Table S14, Supplementary Table S15, and Supplementary Table S16. The sample categories presented here represent specific groupings where the number of samples for that category was sufficient to make aggregated conclusions. More detailed diagnostic information for each sample is presented in Supplementary Table S16. The amount of CSF available for these studies averaged 3.5 mL and ranged from 0.5 to 14 mL. The amount of DNA recovered from CSF averaged 25 ng (IQR 2.6 to 29.6 ng) (Supplementary Table S16). Supplementary Table S16 also includes summaries of the sequencing data obtained from all patients, and whether they scored positively in the B, A, or M components of CSF-BAM. If a patient scored positively in at least one of these assays using the pre-defined thresholds described above, the patient was considered positive for CSF-BAM. Of note, 76% (93/122) of samples derived from individuals with grade 3 or 4 primary brain tumors were positive.

30 samples from patients without cancer:

None of the CSF samples from these patients scored positively in any of the three components, yielding 100% specificity (credible interval 89% to 100%) (Fig. 3 and 4A–C and Supplementary Table S16).

Figure 3.

Performance of CSF-BAM and each analyte. The sensitivity of each analyte within CSF-BAM is demonstrated across the major classes of tumors tested. The composite sensitivity of CSF-BAM is demonstrated in black.

Figure 4.

A, BCR clonality (clonality for non-evaluable samples with total UIDs ≤20 defined as 0), B, estimated tumor fraction by aneuploidy, and C, mutant allele frequency for each sample evaluated with CSF-BAM.

83 samples from patients with high grade glioma:

62 of these samples scored positively for at least one of the three components, yielding a sensitivity of 75% (credible interval 64% to 83%) (Fig. 3 and 4A–C and Supplementary Table S16). Notably, 79% (54/68) of glioblastomas were detected via CSF-BAM. No sample scored positively with the B component, as expected. The aneuploidy component was the predominant basis for sensitivity, with 66% positive (credible interval 56% to 76%). The mutation component scored positively in 44 samples, including 7 that were not scored positively by aneuploidy.

22 samples from patients with medulloblastomas:

19 of these samples scored positively for at least one of the three components, yielding a sensitivity of 86% (credible interval 66% to 95%) (Fig. 3 and 4A–C and Supplementary Table S16). One sample scored positively with the B component. The aneuploidy component was the predominant basis for sensitivity, with 77% positive (credible interval 56% to 90%). The mutation component scored positively in 4 samples, including 2 that were not scored positively by aneuploidy.

24 samples from patients with metastatic lesions to the brain:

23 of these samples scored positively for at least one of the three components, yielding a sensitivity of 96% (credible interval 80% to 99%) (Fig. 3 and 4A–C and Supplementary Table S16). No sample scored positively with the B component, as expected. The aneuploidy component was the predominant basis for sensitivity, with 88% positive (credible interval 69% to 95%). The mutation component scored positively in 18 samples, including 2 that were not scored positively by aneuploidy.

4 samples from patients with CNS lymphomas:

4 of these samples scored positively for at least one of the three components, yielding a sensitivity of 100% (credible interval 48% to 100%) (Fig. 3 and 4A–C and Supplementary Table S16). Two samples scored positively for the B component. All 4 samples scored positively with the aneuploidy component. The mutation component scored positively for 2 samples.

76 samples from patients with other tumor types:

These patients included those with gliomas other than high grade, ependymomas, and various other primary brain tumor types (Supplementary Table S16). 29 of these samples scored positively for at least one of the three components, yielding a sensitivity of 38% (credible interval 28% to 49%) (Fig. 3 and 4A–C and Supplementary Table S16). As with high grade gliomas, medulloblastomas, and metastatic cancers, the aneuploidy component was the predominant basis for sensitivity, with 32% positive (credible interval 22% to 43%). The mutation component scored positively in 11 patients, including 5 that were not scored positively by aneuploidy.

Other genetic observations of interest:

When detectable by the A component, the median neoplastic DNA fraction based on the analysis of aneuploidy was 8.2% (IQR 2.5% to 26.9%). When detectable by the M component, the median mutant allele fraction based on the analysis of mutations was 3.9% (IQR 0.7% to 21.9%). The correlation between the two genetically altered fractions was high (R=0.55, P<5e-6; Supplementary Fig. S5).

The nature of the genetic alterations provided some insight into the type of tumor present in the CNS. Mutations in IDH1 at codon 132 or 172 were observed in 8 samples, and all (100%) of these were from patients with gliomas (2 with oligodendroglioma WHO Grade 3, 1 with astrocytoma WHO grade 3, and 5 with astrocytoma WHO grade 4). In all of these cases, standard sequencing of the resected gliomas revealed IDH mutations. In 3 additional samples from individuals with IDH mutant gliomas, no IDH mutations were observed in the CSF. Histone H3F3A mutations at codon 28 were detected in the CSF of 6 patients (3 with diffuse midline gliomas, 3 with GBM). Two of the six patients had their tumor sequenced, and in both cases the identical mutation was found in the matching CSF. However, no H3F3A mutations in the CSF were described in 5 other individuals who were diagnosed with H3.3 mutated gliomas. KRAS mutations at codon 12 were noted in 3 patients which harbored metastatic lesions from outside the CNS.

Relationship of CSF-BAM results to clinical characteristics

Cytology is not routinely performed on all CSF at our institutions given its poor performance but was available in 53 cases, 48 of which were from individuals with cancer (Supplementary Table S16). The five cases without cancer were negative by cytology as well as by the CSF-BAM assay. Cytology was positive in nine (19%) of the cancer cases, and eight of these (89%) were also scored as positive by the CSF-BAM assay. Of five cases recorded as “suspicious” on cytology, four (80%) were scored as positive by the CSF-BAM assay. Of 34 cases with cancers diagnosed as negative by cytology, 21 (62%) were scored as positive by CSF-BAM. Overall, CSF-BAM was positive in 25 cancer cases where cytology was negative or inconclusive, while only missing one cancer case of 9 that were positive by cytology.

We additionally evaluated whether tumor contact with the CSF space was related to positivity by CSF-BAM (Supplementary Table S17). We compared all samples from patients with cancer for which clinical data was available. Using MRI contrast enhancement as a tumor marker, tumors abutting the CSF space (cistern, ventricle or cortical surface) were more likely than those not abutting the CSF space to be positive by CSF-BAM (P = 0.003 by Fisher’s exact test, odds ratio = 4.4, 95% confidence interval 1.7–11.2). Using MRI T2 signal as a tumor marker, CSF-BAM positivity as a function of tumor’s abutment to a CSF space did not reach statistical significance (P = 0.08 by Fisher’s exact test, odds ratio = 3.7, 95% confidence interval 0.9–13.6).

Immune receptor repertoire profiling

We evaluated a total of 352 CSF samples using SafeBSeqS for the B component of CSF-BAM in addition to SafeTSeqS for TCRs (Supplementary Tables S4, S6, S7, and S16). We identified a mean of 446 (range 0–11666) total UIDs per sample representing original TCR DNA template molecules and a mean of 232 (range 0–13162) total UIDs per sample representing original BCR DNA template molecules (Supplementary Fig. S6A–B). Because each unique UID can only be derived from a single cell in a SaferSeqS library (61), these data can be used to determine the number of mature B or T cells present in the CSF when accounting for input amounts of CSF. T cell clonality varied significantly by cancer type, with T cell cancers demonstrating the greatest number of distinct clonal populations (P < 0.001 by Kruskal-Wallis test) (Supplementary Fig. S7A). The sequencing data also revealed 173 TCR specificity groups, composed of at least 4 unique TCRs from patients with cancer, as determined by TCR amino acid sequence motifs, suggesting possible shared activity (Supplementary Table S18A) (76). The top TCR clone in only 2 samples had a match with high or very high confidence for sequences of previously identified TCRs with annotated specificity (Supplementary Table S18B) (77).

B cell clonality also varied substantially by cancer type (P < 0.001 by Kruskal-Wallis test) (Supplementary Fig. S7B). Using clonality as a metric to classify B cell cancers including primary and secondary CNS lymphomas against all other cancer types for evaluable samples produced a ROC curve with an AUC of 0.95 (95% confidence interval 0.89–1.00, P < 0.001) (Supplementary Fig. S8A). The proportion of samples for which the most frequently observed clone contained the IGHV4–34 gene segment was significantly enriched in B cell cancers including CNS lymphomas compared to all other sample types (P < 0.0001 by Fisher’s exact test) (Supplementary Fig. S8B). This degree of IGHV4–34 gene segment representation was similar to that observed in prior studies of both peripheral and CNS lymphomas and is important because IGHV4–34 could potentially serve as a therapeutic target (78–82).

As a comparison to neoplastic disease and to validate the utility of CSF-BAM for the study of other conditions, particularly with the SafeTSeqS and SafeBSeqS components, we applied SafeTSeqS and SafeBSeqS to analyze 4 CSF samples from patients with multiple sclerosis. Despite limited sample volumes of 1 ml, SafeTSeqS and SafeBSeqS recovered a mean of 523 (range 75–1204) total TCR UIDs and a mean of 21 (range 0–42) total BCR UIDs per sample (Supplementary Table S19). For one sample, SafeBSeqS demonstrated the presence of a clonal population of B cells expressing IGHV4–34, which is a potential therapeutic target as in CNS lymphomas (80,81). This suggests that SafeTSeqS and SafeBSeqS could produce useful information when applied to investigate autoimmune, inflammatory, and other conditions as well as cancers.

Case reports demonstrating clinical applicability

We highlight three cases within this cohort that demonstrate the potential clinical applicability of CSF-BAM. GLIA 793 is a sample from a patient with grade 3 oligodendroglioma who underwent previous resection followed by adjuvant radiation and chemotherapy. Repeat resection was performed when there was a concern for tumor progression. Pathological examination of the surgical specimen revealed treatment effect without evidence of recurrent disease, a common finding in post-treatment gliomas(83). CSF-BAM was negative on CSF obtained three weeks prior to the repeat resection, consistent with the diagnosis of treatment effect and absence of cancer, although a false negative assay result could not have been excluded as a possibility at that time (Supplementary Fig. 9A). Another example is GLIA 914, an individual with metastatic breast cancer to the brain with suspicion of leptomeningeal disease (LMD). The patient had over five lumbar punctures for evaluation of LMD via cytology over a 12-month period and each was negative. The final CSF sample was tested via cytology and also CSF-BAM. The cytology was indeterminant with only rare atypical cells. However, this sample was robustly positive via CSF-BAM with a tumor fraction estimated to be approximately 40% with both aneuploidy and mutations detected. Because this patient did likely have residual parenchymal tumor, follow up investigation would be required to distinguish whether CSF-BAM positivity was reflective of parenchymal disease, LMD, or both. A third case is GLIA 886, derived from an individual with a spinal cord ganglioglioma. The patient had a slowly recurrent spinal cord tumor which was initially resected many years ago prior to the advent of routine tumor sequencing and treated with carboplatin. The tumor was not responsive to chemotherapy and the individual underwent a repeat resection which demonstrated a BRAF mutant ganglioglioma (Supplementary Fig. 9B). CSF-BAM identified the same disease defining canonical BRAF p.V600E mutation in CSF obtained at the time of surgery.

Discussion

Minimally invasive methods to detect brain cancers is a major unmet need in neurology and neuro-oncology. The current standard of care requires invasive neurosurgical biopsy with associated risk and cost. Our group and others have previously demonstrated that CSF is enriched for tumor associated genetic and epigenetic alterations (33,34,84–86). Other studies have shown that the evaluation of B cells and T cells in CSF can provide valuable information about the pathogenesis of neoplastic and non-neoplastic diseases (53,55,87).

We here describe a unified approach that can comprehensively assess CSF DNA for somatic mutations and aneuploidy as well as enumerate clonal compositions of both B cells and T cells. The workflow incorporates several novel features and can be applied to relatively small amounts of CSF. The entire assay can easily be performed within a week of obtaining CSF because a single DNA library, including amplification of both strands of the entire genome, is used for all its component assays. The utility of the CSF-BAM approach with integrated evaluation of multiple analytes was supported by our observation that 34% (79/209) samples from patients with cancer were positive with only one assay component. Had all three components not been integrated, cancer cases would have been missed. The integrated components altogether revealed B or T cell clones that can be a therapeutic target, somatic mutations that direct therapy, and diagnosis defining copy number changes and somatic mutations. Without all three components, less clinically relevant information would have been available.

When considering the components evaluated with CSF-BAM, comparisons to other approaches for cancer evaluation from CSF are also useful. Many biomarkers in CSF with potential utility related to CNS malignancies have been studied, including cells, extracellular vesicles, metabolites, proteins, RNA, and DNA (37). In addition to CSF-BAM, methods for evaluating DNA include digital droplet PCR (ddPCR), WGS, methylation profiling with techniques such as cell-free methylated DNA immunoprecipitation and high-throughput sequencing (cfMeDIP-seq), and simultaneous genetic and epigenetic base sequencing platforms (30,31,88,89). These techniques have demonstrated detection of cancer markers using nanogram level DNA input at detection limits of at least 0.001%. Results have indicated encouraging sensitivity and specificity with each, although relative performance is difficult to assess because studies have been conducted on heterogenous cohorts and few direct comparisons have been reported. Certain limitations are inherent to the current forms of some of these techniques, such as the ability of ddPCR to simultaneously evaluate only a small number of known mutations and the relative sequencing cost of deep WGS. Ultimately, clinical utility will depend on the further optimization of these nascent technologies as well as scalability and other logistical considerations (30,37). CSF-BAM appears promising from this standpoint as the components are cost-effective and high-throughput (61,90).

Practicality for implementation is important because the clinical scenarios in which CSF-BAM could be applied are vast. The results of our study reveal the multiple ways in which CSF-BAM could impact clinical care. One way is by providing potential for diagnosis of brain cancers before tissue biopsy, with information added beyond what can be obtained from imaging. For the 66% of samples from patients with cancer that were positive by CSF-BAM in our study, neoplasia could have potentially been diagnosed based on CSF sampling and imaging prior to neurosurgical biopsy. In important situations patients could have potentially been spared from tissue biopsy, as illustrated for GLIA 793. In this case, CSF-BAM could have provided support for pseudoprogression despite increasing contrast enhancement on MRI scan. Pseudoprogression is a common challenge in both primary and metastatic brain tumors treated with chemotherapy and/or radiation therapy. Having additional molecular tools to better identify pseudoprogression without tissue sampling would be clinically meaningful. Because a false negative result on CSF-BAM could be possible, CSF-BAM would contribute predictive value to support decision making within the overall clinical context rather than rendering a definitive diagnosis. In other situations, CSF-BAM could provide diagnostic information when tissue biopsy is impossible. This is illustrated by the GLIA 936 sample, which was obtained from a patient who presented with a lesion in the medulla that was suspected to represent either low grade glioma or DMG. Because of the lesion’s location, a biopsy could not be safely performed and thus tissue diagnosis could not be obtained. Based on initial imaging and clinical presentation, the tumor was favored to represent low grade glioma and management was directed based on that working diagnosis. However, with subsequent clinical evolution, a diagnosis of DMG was eventually considered to be more likely. CSF-BAM identified an H3F3A mutation in this sample, supporting a diagnosis of DMG. Had the CSF-BAM results been available at initial presentation, they could have informed an alternate diagnosis and management strategy.

A second way in which CSF-BAM could improve upon current clinical care is by providing more sensitive and detailed diagnoses than can be obtained through cytology. Although cytology is often used to assess CSF samples for the presence of malignant cells and is highly specific, it requires large volumes of CSF (>8 mL) (91), its sensitivity is relatively low, and results of uncertain significance (“suspicious for malignancy”) are often reported. Flow cytometry can add diagnostic value in some settings(36). In the current study, we show that most (25 of 33) cases of CNS cancers that were detected with CSF-BAM were not detected by cytology. Conversely, CSF-BAM detected the great majority (8 of 9) cancers positive by cytology, and was positive in 4 of 5 cases which were classified as “suspicious” by cytology. Although the sensitivity of CSF-BAM is considerably higher than that of cytology, it would be prudent to include both types of assessments for diagnosis in future studies. More work will also be required to determine to what extent CSF-BAM can specifically distinguish LMD in patients who have concurrent parenchymal involvement. The potential utility of CSF-BAM is illustrated by the case of GLIA 914, a patient with leptomeningeal metastasis who was positive with CSF-BAM. Prognoses for leptomeningeal metastases are extremely poor in part due to the difficulties in diagnosis. Indeed, this patient had multiple negative cytologic examination over many months. CSF-BAM in this instance could have potentially provided swifter diagnosis, acknowledging that parenchymal disease could have confounded conclusions. Earlier detection can enable more rapid initiation of treatment to preserve neurological function and potentially improve survival.

A third impact from CSF-BAM is from the detailed genetic characterization provided by the assay. The molecular composition of CNS neoplasms can provide opportunities for rational and biologically guided treatments (92). For example, targeted therapies are showing increasing success in neuro-oncology. Having an assay like CSF-BAM to diagnose the BRAF mutation in GLIA 886 could have enabled treatment with BRAF inhibitors, which have shown promise in gangliogliomas and if successful could have delayed or obviated the need for a repeat resection (93). In addition, neurosurgical biopsies are often small and don’t provide definitive diagnosis as in the case of GLIA 58, an adult with a spinal cord tumor. The patient underwent surgical biopsy, but the pathological analysis returned as glial hypercellularity with features of a diffuse astrocytoma, an equivocal finding. CSF-BAM was positive and if combined with tumor tissue testing could have enabled a definitive diagnosis.

The sensitivity of CSF-BAM was high for the major cancer types that would prompt evaluation of CSF during the diagnostic work-up. In particular, more than 80% of GBM, medulloblastomas, CNS lymphomas, and metastases from primary tumors outside the CNS could be detected (Fig. 3 and Supplementary Table S16). As ours was a retrospective study based on available samples rather than a prospective study, the fraction of the various tumor types evaluated was different than what would be seen in practice. If corrected for incidence according to the Surveillance, Epidemiology and End Results (SEER) Database, we would have about 10–15 times the number of metastatic cases compared to the primary CNS cancers, which would impact overall test performance metrics given that brain metastases were correctly detected in over 90% of tested samples.

The limits of sensitivity for detecting malignancy in the CSF are not yet known. From studies of pancreatic, colorectal, breast, and other cancer types in plasma, it is known that some advanced cancers do not release (“shed”) DNA into the plasma (94). Analogously, some CNS cancers may not release DNA into the CSF, especially those that are not anatomically adjacent to a CSF space (34). Other variables such as the CSF collection method and the patient treatment status may also affect sensitivity (Supplementary Table S16). Improvements in CSF-BAM could increase sensitivity further. One simple improvement would be the analysis of a higher volume of CSF; many of the samples in our retrospective collection included <1 mL of CSF, which limits the number of molecules that can be assessed, particularly for mutations or B or T cell clonality. In addition, applying metrics for minimum input DNA would help ensure sufficient starting material is available to reliably detect the three BAM analytes. In this study, CSF-BAM assays were positive in samples from patients with cancer using as low as 0.1 ng of DNA for library input, which could represent a lower limit of detection (Supplementary Table S16). Another improvement would be the inclusion of additional clinically relevant oncogenes and tumor suppressor genes in the mutational analysis.

A third potential improvement could be the inclusion of epigenetic alterations into the analysis. This would technically be straightforward, as SaferSeqS library preparation can be easily modified to permit evaluation of DNA methylation as well as of aneuploidy and mutations (90). DNA methylation could also help determine the CNS tumor type present, analogous to plasma cfDNA used to determine the tumor of origin of non-CNS cancers (26,95). A caveat to using methylation is the high analyte sensitivity required to detect cases. In our study, most of the cancers that were detected by CSF-BAM were detected through aneuploidy rather than mutations or immune cell clonality. To be scored as positive for aneuploidy generally required a 1.5% tumor fraction. In other words, if < 1.5% of the DNA in the CSF did not arise from cancer aneuploid cells, then it would generally not be scored as positive. 16 cancers not detected by the aneuploidy component were detected by mutations (Fig. 3 and 4 A–C and Supplementary Table S16) and the limit of detection for mutation was as low as 0.01% in this cohort. If methylation is to meaningfully add sensitivity to CSF-BAM, the methylation assay would have to be capable of detecting differentially-methylated regions of the genome present in <1% of the template molecules, while retaining a specificity close to 100%, which is challenging.

Although there is strong and increasing interest in sequencing of TCRs and BCRs to understand the immune components of disease, current technologies to sequence immune receptors suffer from limitations in throughput, sensitivity, reproducibility, and abundance quantification (65,66). SafeBSeqS and SafeTSeqS, the new technologies described in this manuscript, use DNA rather than RNA as the analyte. Combined with the high conversion of template molecules to library molecules and the sequencing of both strands of DNA to limit errors made during library preparation and sequencing, these tools provide a way to quantify the number of B cells or T cells in any bodily compartment, as well as the repertoire of the TCRs and BCRs encoded by those cells. The number of B cells or T cells in a sample cannot be reliably assessed with RNA-based technologies because the amount of RNA can vary substantially among cells, especially in complex environments such as the CNS. In contrast, each mature B cell or T cell generally contains a single rearranged BCR or TCR gene, so the number and fraction of UIDs in the sequencing data provides a direct estimate of the number of B cells and T cells in the original sample.

Previous studies of immune receptor repertoires in CSF have generally focused on aging, neurodegeneration, and neuroinflammatory conditions (59,60,96). In patients with cancer, studies have primarily reported TCR and BCR sequences derived from cells in tumor tissue with relatively limited data reported on immune receptor repertories derived from CSF (87,97–101). We substantially add to this information through the analysis of BCRs and TCRs in 329 CSF samples from patients with neoplasms. These TCRs and BCRs identified in the CSF could reflect characteristics of immune cell infiltrates in corresponding tumor tissue (87). In this study, integrating immunoreceptor information into the CSF-BAM assay was important for detecting and describing T cell and B cell cancers, understanding the repertoire architectures of immune cell populations, and identifying potential therapeutic targets such as IGHV4–34. In future use, the SafeTSeqS and SafeBSeqS components of CSF-BAM could provide insights through serial CSF sampling in immunotherapy clinical trials to better predict treatment response and elucidate treatment mechanisms (38). As increasingly sophisticated methods to leverage immunologic data come into practice, such as those using machine learning, the ability of CSF-BAM to generate information about TCR and immunoglobulin sequences may become increasingly impactful (102–105).

Limitations of our study include the relatively small sample size for several tumor types. It is conceivable that the pre-operative results of CSF-BAM, and its degree of positivity, provide information about prognosis for some cancer types, but our study was not powered to do so. Moreover, we studied only a single sample of CSF for most patients. Serial, longitudinal studies of patients post-treatment would be required to determine whether the analysis of CSF will provide information about the effectiveness of treatment on patients with CNS tumors, which is often challenging based on radiographic findings alone. A third limitation of our study is the restricted validation of the CSF-BAM mutation and aneuploidy results through comparison to sequencing from primary tumor tissue. The tissue from the patients in our study was not available to us for research purposes. Because of this, we could compare results only for 15 patients for whom tumor sequencing was performed as standard of care (Supplementary Table S13). In sum, 14 mutations identified with CSF-BAM were identical to those identified in the tumor tissue. Thirteen mutations identified with CSF-BAM were not reported in the clinical records. However, information regarding coverage of the genes and gene regions assessed clinically was not available for every patient and thus we cannot determine if these 13 mutations were assessed in the tumor tissue. For example, CSF-BAM identified H3F3A mutations for samples CGLI 184 and GLIA 522. The tumors from these patients were H3K27 mutant by immunohistochemistry but H3 mutation was not reported in the tumor sequencing data. For context, the previous generation of the core assay used for mutations in CSF-BAM was shown to have greater than 99% specificity for multiplex mutation detection in plasma in a cohort of 812 healthy control individuals, with newer generations of the technology used in this study demonstrating greater than 100-fold improvement in specificity at the nucleotide level (61,71). For validation of aneuploidy in CSF-BAM, we additionally compared the aneuploidy component results with results from an orthogonal method for evaluating aneuploidy, RealSeqS (33). Greater than 94% of chromosome arm level calls were concordant between the samples (Supplementary Table S11). Future studies to evaluate CSF-BAM would benefit from inclusion of more patients with paired CNS tumor tissue analysis to measure congruence with CSF-BAM. Finally, CSF samples from many more patients without cancer will be required to obtain a more precise estimate of specificity than the one we currently have, i.e., 100%, but with a lower bound of the credible interval at 89%.

One of the strengths of our study was the segregation of available samples into independent analytical, training, and validation cohorts, permitting pre-defined thresholds to be applied to the validation cohort (106). Another strength is that the libraries used to implement CSF-BAM provide a limitless source of DNA from the CSF, given that there is a 1:1 correspondence between both the Watson and Crick strands of the template DNA molecules and the molecules in the amplified library. PCR-mediated re-amplification of the original libraries would thereby permit analysis of the same precious CSF samples with improved versions of CSF-BAM, as well as permit sharing of samples among laboratories. For these reasons, as well as its encouraging performance on CSF samples to date, CSF-BAM provides a promising foundation for the molecular genetic evaluation of the DNA from CSF samples. The potential for CSF-BAM integration into clinical workflows will be the subject of future prospective studies.

Methods

Study design

This study was approved by the Institutional Review Boards for Human Research at Johns Hopkins Hospital in compliance with the principles of the Declaration of Helsinki and the Health Insurance Portability and Accountability Act. All patients provided written informed consent. All samples were deidentified immediately following collection. All available samples were included in the study. Clinical data for the individuals included in the study are presented in Supplementary Table S11. Some samples have been previously studied in prior publications using different technologies (33,34). All samples in this manuscript had new CSF-BAM libraries generated as described below from CSF derived DNA.

Sample processing and DNA purification

CSF samples were collected into standard CSF collection tubes. Volumes used for DNA purification are reported in Supplementary Tables S7, S16, and S19. Blood samples were collected into Streck Cell-Free DNA BCT (#230469). DNA from CSF, plasma, or leukocytes was purified using the BioChain Cell-free DNA Extraction Kit (#K5011625). Control primary dermal fibroblasts were obtained from ATCC (#PCS-201–012) and DNA was purified using Qiagen QIAamp DNA Mini Kit (#51304).

Library construction

We developed a library preparation workflow that can efficiently recover input DNA and simultaneously incorporate double-stranded molecular barcodes (61). In brief, libraries were prepared using an Accel-NGS 2S DNA Library Kit (Swift Biosciences, #21024) with the following critical modifications: 1) DNA was pretreated with 3 U of USER enzyme (New England BioLabs #M5505L) for 15 min at 37 °C to excise uracil bases; 2) the SPRI bead/PEG NaCl ratios used after each reaction were 2.0×, 1.8×, 1.2× and 1.05× for end repair 1, end repair 2, ligation 1 and ligation 2, respectively; 3) a custom 50 μM 3′ adapter was substituted for reagent Y2 and 4) a custom 42 μM 5′ adapter was substituted for reagent B2. Libraries were subsequently PCR amplified in 50-μl reactions using primers targeting the ligated adapters. The following reaction conditions were used: 1× NEBNext Ultra II Q5 Master Mix (New England BioLabs #M0544L), 2 μM universal forward primer and 2 μM universal reverse primer. Libraries were amplified with 8 or 11 cycles of PCR, depending on how many experiments were planned, according to the following protocol: 98 °C for 30 s, cycles of 98 °C for 10 s, 65 °C for 75 s and hold at 4 °C. If eight cycles were used, the libraries were amplified in single 100-μl reactions. If 11 cycles were used, the libraries were divided into eight aliquots and amplified in eight 50-μl reactions, each supplemented with an additional 0.5 U of Q5 Hot Start High-Fidelity DNA Polymerase (New England BioLabs #M0493L), 1 μl of 10 mM dNTPs (New England BioLabs #N0447L) and 0.4 μl of 25 mM MgCl2 solution (New England BioLabs #B9021S). The products were purified with 1.8× SPRI beads (Beckman Coulter, #B23317) and eluted in EB buffer (Qiagen #19086).

The “B” component of CSF-BAM (BCR)

TCR and BCR sequences were amplified from libraries using gene-specific primers for TRBJ and IGHJ segments as previously described for sequencing of somatic mutations (61,90). A total of 13 unique primers were used to cover all TRBJ gene segments in each amplification step and 4 unique primers were used to cover all IGHJ gene segments in each amplification step (Supplementary Table S2A and S2B). For clonotype analysis, demultiplexed reads were used to generate clonotype tables using the MiXCR 4.6.0 package (RRID:SCR_018725) (107) with the function “analyze generic-amplicon-with-umi” and specifications “--species hsa --dna --rigid-left-alignment-boundary --floating-right-alignment-boundary J.” The specification “--tag-pattern ^(R1:*)\^(UMI:N{14})(R2:*)” was used for Watson strands and the specification “--tag-pattern “^(UMI:N{14})(R1:*)\^(R2:*)” was used for Crick strands. Clones with duplex support were assembled using the function “exportClonesOverlap” with the specifications “--criteria CDR3|NT|V|J” for each Watson and Crick sample pair. Total UIDs are reported as the sum of UIDs from Watson and Crick samples for clonotypes with duplex support. Clustering was performed using GLIPH2 (RRID:SCR_025690) (76) for productive TCR sequences from CSF samples from patients with cancer. Comparison of TCR sequences to a database of TCRs with known specificities was performed using VDJdb (77). To determine the relative yield for IGHJ and TRBJ gene segments, we evaluated DNA from primary fibroblasts. Non-rearranged IGHJ and TRBJ gene segments were analyzed using the computational pipeline as described for the “M” component below, with the modification of mapping to hg38. IGHJ1, IGHJ4, and IGHJ5 were computationally distinguished using primer sequence inputs GAGGAGACGGTGACCAGGGTGCCCTGGCCCCAGTG, GAGGAGACGGTGACCAGGGTTCCCTGGCCCCAGTA, and GAGGAGACGGTGACCAGGGTTCCCTGGCCCCAGGG respectively in the analysis pipeline.

The “A” component of CSF-BAM (aneuploidy)

Library DNA was amplified in 50 μL reactions in Ultra Q5 with primers at 2 μM for seven cycles with the following conditions: 98 °C for 30 s, then seven cycles of 98 °C for 10 s to denature, and 65 °C for 75 s to anneal and extend. WGS libraries were sequenced on a NovaSeq 6000 (RRID:SCR_016387) with paired-end 2×100bpreads. The depth of sequencing averaged 40.0 M reads of 100 bp (IQR 33.8M to 45.1M, Supplementary Table S9). Cutadapt (RRID:SCR_011841) was used to trim 27 base pairs from both reads (108) and BWA-MEM was used to align reads to the hg19 genome (arXiv 1303.3997v2). Duplicate molecules were marked and removed using Samtools (RRID:SCR_002105)(109). Reads with a quality >10 were binned into 500kb intervals and counted. IchorCNA (RRID:SCR_024768) was then used to perform GC correction and call the estimated tumor fraction using the following parameters: “—chrs “c(1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,17,18,20,21,22) —normal “c(0.9,0.95,0.98)” –estimateScPrevalence FALSE –scStates “c()” –maxCN 3 –ploidy “c(2)” –normalPanel 20240722.CSF_PON_median.rds.” The panel of normal was based on 8 CSF sample without the presence of cancer and were not used in the study. Chr16 and chr19 were excluded due to the high variance on each of chromosomes. We and others have reported that increased GC content on these chromosome can produce higher numbers of false positives (110). Next, we wanted to generate chromosome arm level calls. However, because ichorCNA only produces chromosome level not chromosome arm level calls, we then generated arm level calls using the median across each the GC corrected ichorCNA 500kb intervals. We discarded chr6p which produced numerous false calls associated with alignment artifacts surrounding the MHC regions. In line with previous work, chr 13p, 14p, 15p, 21p, and 22p were excluded because in healthy individuals there is high variability in size and copy numbers of repeated sequences in these arms, making gain and loss calls unreliable (69,70,111–113).

The “M” component of CSF-BAM (mutations)

Following library creation, two separate PCRs were designed to selectively enrich the Watson or Crick strand. Both PCRs used the same gene-specific primer, but each used a different anchoring primer. PCR duplicates derived from each strand could be distinguished by the orientation of the insert relative to the exogenous UID. Sequencing reads underwent initial processing by extracting the first 14 nucleotides as the exogenous barcode sequence (UIDs) and masking adapter sequencing with Picard (RRID:SCR_006525) IlluminaBasecallsToSam (http://broadinstitute.github.io/picard). Reads were then mapped to the hg19 reference genome using BWA-MEM (RRID:SCR_010910) (arXiv 1303.3997v2) and sorted by barcode sequence using Samtools (109). Duplex mutations were defined as mutations present in >80% of both the Watson and Crick families with the same UID. We had several metrics for the interpretation of mutations: only genomic positions with at least 2 or more reported annotations in the Catalogue of Somatic Mutations in Cancer (COSMIC) (RRID:SCR_002260) in genome-wide studies and were confirmed somatic mutations were considered (114); only positions with at least two observations of mutant molecules were considered; only positions with at least 5x coverage were considered; only positions at least 30 bp away from the end of the molecule were considered.

During the analysis, we noticed two major outliers: FBXW7 codon 505 and KRAS codon 61. The frequency of mutations at these two codons was much higher than expected. Upon closer inspection, every molecule with a mutant FBXW7 codon 505 also had mutations at codons 289, 299, 300, and 314. We then performed blat analysis to compare the observed matches to any non-human genome (74). The observed sequence perfectly matched the bovine genome. We hypothesized that bovine-derived hemostatic agents frequently used in neurosurgery (75) may have contributed minor amounts of DNA that amplified and incorrectly aligned to FBXW7. We also found KRAS codon 61 mutations in two patients with trigeminal neuralgia. The significance of these mutations is unknown. To ensure high specificity of the assay and given the low abundance of this particular mutation in CNS cancers, we excluded all mutations at this position in our study.

Statistical analysis

Cohort sample size was not selected for statistical power but rather was based on sample availability. The posterior probability that the assay was the most and least sensitive (or specific) was calculated using Bayesian methods and we report credible intervals for the sensitivity and specificity of CSF-BAM. A uniform Beta(1,1) prior distribution was used and we assumed the number of correct assays followed a binomial distribution. The prior distribution was used based on the observed data. Therefore, the resulting posterior distribution was Beta(1+# of success, 1+# failures). Statistics were calculated using GraphPad Prism (RRID:SCR_002798) 10.4 and R (RRID:SCR_001905) 4.4. Statistical tests are specified in figure legends.

Statement of significance.

There is a paucity of technologies beyond surgical biopsy that can accurately diagnose central nervous system neoplasms. We developed a novel, sensitive, and highly specific assay that can detect brain cancers by comprehensively identifying somatic mutations, chromosomal copy number changes, and adaptive immunoreceptor repertoires from samples of cerebrospinal fluid.

Data availability

Data generated in this study are available in the paper and supplementary files. Sequence data are available in the European Genome-phenome Archive (RRID:SCR_004944) under accession EGAS00001008199.