Low-input CSF cfDNA shallow whole-genome sequencing for CNV-based diagnosis and monitoring of leptomeningeal metastasis in lung cancer Authors: Xiudong Guan, Yubing Li, Zongkun Shi, Juan An, Yichun Hua, Xiaosheng Ding, Xixi Zheng, Weiran Xu, Dainan Zhang, Zhan Xue, Peng Kang, Dandan Cao, Xiaoyan Li, Wang Jia Published: 2026 ABSTRACT Leptomeningeal metastasis (LM) is a life-threatening complication of lung cancer for which magnetic resonance imaging (MRI) and cerebrospinal fluid (CSF) cytology may miss disease and provide limited molecular guidance. Because the low cfDNA input typical of CSF constrains deep sequencing and mutation-focused approaches may underrepresent the copy-number-rich genomic architecture of CNS metastases, we developed a low-input shallow whole-genome sequencing approach to quantify genome-wide copy-number changes from trace CSF cfDNA. In 57 patients with lung cancer and CNS metastases (28 with LM and 29 with brain metastasis only), CSF showed stronger tumor-DNA signals and a higher burden of genome amplification than matched plasma (54 pairs), despite lower cfDNA concentrations. Among samples with detectable tumor DNA, genome amplification proportion (GAP), defined as the fraction of the genome showing amplification, identified LM with 86% sensitivity and no false positives observed in this cohort, and showed performance comparable to cytology as a complementary quantitative readout. Copy-number profiling further highlighted chromosomal regions and focal gene amplifications enriched within LM, supporting a CNV-enriched molecular state in the CSF compartment. Serial CSF sampling also captured treatment-associated molecular shifts, including progressive TYMS amplification during intrathecal pemetrexed in an exploratory case. Together, these findings support low-input CSF sequencing as a practical molecular approach to complement LM diagnosis and longitudinal monitoring in lung cancer and provide a rationale for prospective validation in independent multi-center cohorts.Supplementary InformationThe online version contains supplementary material available at 10.1186/s40478-026-02350-7. Introduction Central nervous system (CNS) metastases represent a major clinical challenge in lung cancer and account for a substantial proportion of CNS metastatic cases [4, 9, 14]. Brain metastases (BM) and leptomeningeal metastases (LM) are the two most common clinical manifestations. LM, in particular, is associated with dismal outcomes, with median survival typically measured in months. While magnetic resonance imaging (MRI) remains central to the evaluation of CNS metastases and is the diagnostic standard for BM, its sensitivity for LM is limited, ranging from 76 to 83%, with false-negative rates of up to 30% [5, 20, 24]. Cerebrospinal fluid (CSF) cytology remains the diagnostic gold standard for LM, but its sensitivity at the first lumbar puncture is only approximately 50–73% [16, 17]. Importantly, neither MRI nor cytology provides molecular information needed to guide individualized treatment strategies. In the era of precision oncology, tumor molecular profiling has become integral to disease classification and therapeutic decision-making. However, when applied to CNS metastases, this predominantly mutation-centric paradigm faces important biological and technical limitations. Direct biopsy of CNS lesions is often impractical or unsafe, and lesion-to-lesion genetic heterogeneity together with the blood–brain barrier limits the value of extracranial tumor tissue and plasma cell-free DNA (cfDNA) as surrogates for intracranial disease. In addition, because CSF often contains only trace amounts of tumor-derived nucleic acids, cfDNA concentrations may be too low to meet the input requirements for conventional high-depth targeted sequencing [2, 3, 6, 7, 18]. Moreover, the molecular landscape of CNS metastases may diverge from that of the primary lung tumor. Prior studies using high-depth targeted sequencing have linked circulating-tumor DNA (ctDNA) abundance and selected therapy-associated mutations to prognosis in patients with CNS metastases [30]. However, current ctDNA profiling frameworks are largely anchored in primary tumor biology and prioritize single-nucleotide variants (SNVs) and insertions/deletions (InDels), potentially underestimating larger-scale genomic alterations relevant to CNS metastatic progression. Recent exploratory studies suggest that, in contrast to mutation-centered models of primary tumorigenesis, copy number variations (CNVs) may play a prominent role in the progression of CNS metastases and may be undercaptured by mutation-focused assays (Supplementary Table 1) [10, 21]. Beyond EGFR, few recurrent, high-confidence mutations have been consistently identified as specific drivers of CNS metastases. Together, these technical constraints and biological mismatches may limit the detection of clinically relevant actionable alterations that shape CNS metastasis biology. Accordingly, there is a need for profiling strategies that can robustly capture CNVs from trace CSF and more comprehensively define the genomic landscape of CNS lesions. Shallow whole-genome sequencing (sWGS) has emerged as a practical approach for genome-wide CNV detection when cfDNA input is limited and targeted panel testing is constrained, offering a potential solution to both technical and biological barriers. sWGS can detect large-scale genomic alterations, including broad CNVs and high-level amplifications, even from trace amounts of cfDNA [13]. However, most prior sWGS-based studies of lung cancer LM have remained exploratory and largely descriptive, focusing on the presence or absence of tumor-derived DNA rather than deriving quantitative CNV metrics with diagnostic or longitudinal relevance [2, 7, 11, 29]. Here, we shifted the analytical focus from mutations to CNVs to address the specific challenges of genomic profiling of CNS metastases from trace CSF cfDNA. We developed an sWGS-based approach to characterize CNS metastasis genomes through CSF, enabling simultaneous estimation of tumor fraction (TF), chromosomal arm-level CNVs, and gene-level amplifications. By applying sWGS to paired plasma and CSF, we asked whether CNV-derived parameters could support the diagnosis of LM and reveal CNS-compartment alterations with potential relevance to treatment stratification in lung cancer patients with CNS metastases. Methods Patient enrollment This study was approved by the institutional review boards of Beijing Tiantan Hospital, Capital Medical University (KY2021-155-02, KY2023-155-03). All procedures were conducted in accordance with the Declaration of Helsinki. A total of 57 patients with CNS metastases primarily caused by lung cancer were enrolled from December 2021 to May 2023, with 177 corresponding samples collected. Patients were excluded if they had undergone allogeneic organ transplantation, allogeneic bone marrow transplantation, or blood transfusion within 30 days prior to enrollment, were pregnant, or lacked informed written consent. Clinical data curation and reference standard We manually assembled and annotated comprehensive clinical datasets—demographics precisely matched to tumor histopathology reports, CSF cytology results and cfDNA samples collection. CNS metastases were classified as leptomeningeal metastases (LMs) and brain parenchymal metastases (BMs) using integrated CSF cytology and contrast-enhanced MRI as the reference standard. LMs were diagnosed by any of the following: (i) cytologically confirmed malignant cells in centrifuged CSF; (ii) abnormal leptomeningeal, cranial-nerve-root, or ependymal enhancement on MRI; or (iii) both. BMs were defined as radiologically confirmed intraparenchymal metastatic lesions seen exclusively on MRI. Treatment histories were documented immediately before and after blood and CSF sampling, and all sampling events were time-labeled to ensure unambiguous mapping to downstream cfDNA processing and analyses. Sample collection and cfDNA extraction CSF was primarily obtained through lumbar puncture, ventricular drainage or Ommaya reservoir. For some BM patients suitable for surgery, CSF was collected intraoperatively via dural puncture during craniotomy, before opening the dura mater. CSF could be only accepted for research testing if there is no visible blood contamination. Blood samples were collected from peripheral vein. For paired analysis, CSF and blood collected within 24 h of each other were classified as pairs, unless a neurosurgical manipulation or invasive biopsy was performed in between. To preserve cfDNA integrity, samples were collected into cfDNA-stabilizing tubes and processed within 48 h of collection. After two-step sequential centrifugation steps at low and high speeds, cells and debris were removed from the sample, minimizing the risk of in vitro genomic DNA contamination from leukocyte lysis. cfDNA was isolated for subsequent sequencing and analysis. Detailed procedures and reagents are provided in the Supplementary Methods. Library preparation and sequencing sWGS libraries were prepared using 5 ng of cfDNA per plasma sample. For CSF samples, 5 ng was used if input was sufficient, otherwise half of the available cfDNA was used. Library preparation followed standard protocols with end-repair, adapter ligation, and PCR amplification. The libraries were then sequenced on the Illumina NovaSeq 6000 platform to a target yield of ~ 10Gb per sample. For samples with an inferred tumor fraction ≥ 20% in subsequent analyses, additional sequencing was extended to ~ 30Gb for mutation detection. To assess sWGS robustness under trace-input conditions, we performed a cross-factor design technical validation using serially diluted cfDNA from a single CSF sample, testing multiple initial cfDNA amounts and PCR cycles in combination. Full details on kits and PCR conditions are available in the Supplementary Methods. In a few cases, targeted sequencing was additionally performed on CSF samples using the Onco Sonar Cancer Mutation Profiling Liquid Kit (Genetron Health Co., Ltd., China) to validate the reliability of sWGS in detecting small-scale mutations (see Supplementary Methods for details). Sequence alignment and variations detection Reads were trimmed, aligned to the hg19 reference genome, and deduplicated using standard bioinformatics pipelines. Quality metrics including coverage, GC bias, and duplication rates were assessed. The CNV analysis workflow mainly consists of two steps. First, the GC- and mappability-corrected log2 copy ratios were calculated from bin-level read counts using HMMcopy (v1.42.0). Then, the corrected bin-level log2 ratios were segmented by ichorCNA (v0.3.2) [1] to simultaneously infer tumor fraction, ploidy, and segment-level copy numbers. CNV status is defined as gain if CN ≥ 3, loss if CN ≤ 1, or amplification if CN ≥ 5 [21, 26]. However, under ~ 3 × sWGS, read counts per bin are low, inflating sampling variance relative to arm-level aggregation. Consequently, modest gains (e.g., CN 2–3) are difficult to separate from noise, and deletions are further destabilized by sparse counts combined with log-scale normalization. We therefore limited bin-level inference to high-level amplifications (CN ≥ 5), while assessing gains and losses primarily at the arm level. Focal amplifications were determined by the corrected bin-level copy ratios, whereas arm-level gain/loss events were confirmed using segment-level results. A chromosome arm was judged as gain or loss if more than 70% of its segments showed consistent directional change [19, 26]. Clinically relevant SNVs and InDels associated with NSCLC were assessed by manual review using Integrative Genomics Viewer (IGV, v2.4) [27]. A sample was designated mutation-positive at a given site if at least one mutation-supported read was observed. Detailed procedures and the full list of mutations analyzed are provided in the supplementary methods. Modeling for leptomeningeal metastasis diagnosis Because CNV quantification becomes unreliable at low tumor fractions, TF gating was incorporated in model construction to ensure robust classification. The diagnostic model used two sWGS-derived CNV biomarkers, aneuploidy and GAP, as candidate biomarkers. For each sample, let TF denote the inferred tumor fraction and M the value of an sWGS-derived CNV biomarker. The LM classifier was defined as a two-stage decision rule:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ \hat{Y}\left( {t,\theta } \right) = \left\{ {\begin{array}{*{20}l} {0,} \hfill & {TF < t} \hfill \\ {M \ge \theta ,} \hfill & {~~TF \ge t} \hfill \\ \end{array} } \right. $$\end{document}where t is a candidate TF threshold and θ a biomarker cutoff. A predefined TF grid was traversed to identify the minimum TF threshold at which biomarkers-based classification becomes reliable (TF*). For each TF threshold t, diagnostic performance Perf(t) was estimated using leave-one-patient-out cross-validation (LOPO-CV) at the patient level. Within each fold, the training patients were used to determine the biomarker threshold θ(t) that maximized the Matthews correlation coefficient (MCC) as performance, and performance was evaluated on the left-out patient. The resulting fold-specific θ(t) was then applied to all samples from the held-out patient to generate out-of-fold predictions for that fold. Concatenating predictions across all folds yielded, for each TF cutoff t, a single cross-validated MCC value Perf(t) summarizing performance over the entire cohort. The Perf(t) was subsequently smoothed using isotonic regression to characterize the onset of a stable performance plateau. The minimal TF value within this plateau was defined as the minimum effective tumor fraction (TF*). Robustness of TF* was assessed by recomputing TF* while systematically leaving each patient out, yielding a distribution of TF* estimates. After TF* was finalized, biomarker thresholds θ* were optimized through grid search. To enable comparison with CSF cytology—which operates clinically with zero false positives—thresholds were selected under the constraint of perfect specificity, and sensitivity was maximized to obtain θ*. Diagnostic performance was evaluated under two complementary schemes. The first assessed the intrinsic discriminative ability of each biomarker by restricting analysis to samples with TF ≥ TF*. The second reflected the intended clinical workflow, in which samples with TF < TF* are automatically assigned LM − , whereas samples with TF ≥ TF* are classified using θ*. Confidence intervals for all performance metrics were estimated using patient-level bootstrap resampling, in which patients rather than individual samples were resampled to account for repeated measures within individuals and potential changes in CNS disease status over time. Paired comparisons between ctDNA-based markers and cytology were performed using McNemar’s test. Full algorithmic details, including the TF threshold scan, cross-validation structure, threshold optimization, and bootstrap procedures, are provided in the Supplementary Methods. External public data re-analysis For this study, three databases from the Memorial Sloan Kettering Cancer Center (MSK) were utilized to analyze genetic alterations associated with CNS metastasis in lung cancer: MSK-MET [15], MSKCC-LUAD [10], and MSKCC-BM-NSCLC [23]. Lung cancer samples were categorized based on primary, extracranial metastatic, and brain metastatic origin. Comparative analyses were conducted to identify variants and CNVs enriched in brain metastases compared to other groups. The data extraction procedures and detailed statistical methods are provided in the Supplementary Methods. Statistical analysis All statistical tests were two-sided unless otherwise stated. Exact sample sizes (n) are reported in figure legends, table legends and/or the main text. Analyses include paired comparisons (matched CSF–plasma samples), longitudinal repeated sampling within individuals, and case–control comparisons; where multiple samples were available per patient, performance estimation and resampling procedures accounted for within-patient clustering, treating the patient as the primary unit. No multivariable models were fitted and no covariates were tested unless explicitly stated. For group comparisons, nonparametric tests were used when distributional assumptions were not warranted. P values are reported as exact values where suitable; where multiple hypotheses were tested, we specify whether P values were adjusted or interpreted as exploratory. Diagnostic performance metrics (e.g., sensitivity, specificity and AUC) are reported as effect-size measures together with 95% confidence intervals estimated by bootstrap resampling that accounts for repeated measures within patients. Results Patients and samples This study included 57 patients with lung cancer and CNS metastases: 29 with BM only, 9 with LM only, and 19 with both BM and LM. Baseline characteristics differed between patients with LM (LM only or BM + LM) and those with BM only (Table 1). The most pronounced difference was line of therapy (P < 0.001), with LM patients more frequently receiving third-line or later systemic therapy (64.3% vs. 8.3%). Molecular subtype distribution also differed between groups (P = 0.022), driven by enrichment of EGFR-mutant adenocarcinoma in the LM group (89.3%) relative to the BM-only group (51.7%). Extracranial metastases were more common in patients with LM than in those with BM only (82.1% vs. 51.7%, P = 0.031). Histological subtype distribution showed a non-significant trend toward group differences (P = 0.067). Given the baseline imbalances between LM and BM-only groups (Table 1), group-level comparisons were interpreted cautiously. Diagnostic modeling therefore incorporated tumor fraction to mitigate low-signal-related misclassification.Table 1Baseline characteristics of the study populationCharacteristicNo. (%)P value*Brain-only metastasisn = 29Leptomeningeal metastasisn = 28Age/years, median (range)59 (40–73)61.5 (25–79)0.898Gender0.219Female14 (48.3)19 (67.9)Male15 (51.7)9 (32.1)Histology subtype0.067Adenocarcinoma24 (82.8)28 (100.0)Small cell carcinoma5 (17.2)0 (0.0)Molecular subtype0.022EGFR mutation15 (51.7)25 (89.3)ALK rearrangement2 (6.9)1 (3.6)RET rearrangement0 (0.0)1 (3.6)KRAS1 (3.4)0 (0.0)MET amplification1 (3.4)0 (0.0)Not detected10 (34.5)1 (3.6)Extracranial metastases0.031Exist15 (51.7)23 (82.1)Not exist14 (48.3)5 (17.9)Line of therapy < 0.001Treatment naïve6 (20.7)0 (0.0)1st line13 (44.8)3 (10.7)2nd line3 (10.3)6 (21.4) ≥ 3rd line2 (6.9)18 (64.3)Unknown5 (17.2)1 (3.6)*Statistical comparisons were performed using Student’s t-test for age and chi-squared test for all categorical variables Table 1Baseline characteristics of the study populationCharacteristicNo. (%)P value*Brain-only metastasisn = 29Leptomeningeal metastasisn = 28Age/years, median (range)59 (40–73)61.5 (25–79)0.898Gender0.219Female14 (48.3)19 (67.9)Male15 (51.7)9 (32.1)Histology subtype0.067Adenocarcinoma24 (82.8)28 (100.0)Small cell carcinoma5 (17.2)0 (0.0)Molecular subtype0.022EGFR mutation15 (51.7)25 (89.3)ALK rearrangement2 (6.9)1 (3.6)RET rearrangement0 (0.0)1 (3.6)KRAS1 (3.4)0 (0.0)MET amplification1 (3.4)0 (0.0)Not detected10 (34.5)1 (3.6)Extracranial metastases0.031Exist15 (51.7)23 (82.1)Not exist14 (48.3)5 (17.9)Line of therapy < 0.001Treatment naïve6 (20.7)0 (0.0)1st line13 (44.8)3 (10.7)2nd line3 (10.3)6 (21.4) ≥ 3rd line2 (6.9)18 (64.3)Unknown5 (17.2)1 (3.6)*Statistical comparisons were performed using Student’s t-test for age and chi-squared test for all categorical variables Baseline characteristics of the study population *Statistical comparisons were performed using Student’s t-test for age and chi-squared test for all categorical variables A total of 66 blood samples and 111 CSF samples were collected (Supplementary Fig. 1), including 54 matched CSF-blood pairs obtained within 24 h. These matched CSF–blood pairs enabled benchmarking of sWGS performance in low-input cfDNA and compartment-specific comparisons of CNV burden between CSF and blood. Technical robustness of sWGS for low-input CSF cfDNA We first assessed whether sWGS can reliably quantify CNV burden—a dominant class of genomic alterations in CNS metastases—from trace CSF cfDNA. We implemented a low-input workflow comprising cfDNA isolation from CSF or blood, library preparation and shallow (~ 3 ×) whole-genome sequencing (Fig. 1a), followed by an analytical pipeline reports tumor fraction, chromosomal arm-level CNVs and focal amplifications. We focused on two sWGS-derived metrics: aneuploidy, defined as the number of altered chromosomal arms, and genome amplification proportion (GAP), defined as the fraction of the genome exhibiting amplification. For samples above a prespecified tumor-fraction threshold, sequencing was extended to higher depth to enable mutation detection (Methods).Fig. 1Low-input sWGS yields higher tumor fraction and CNV burden in CSF than in plasma despite lower cfDNA concentration. a Schematic of the low-input cfDNA profiling workflow, including cfDNA isolation from CSF or plasma, library preparation, shallow (~ 3 ×) whole-genome sequencing and downstream inference of tumor fraction, chromosomal arm-level CNVs and focal amplifications (Methods). b Distribution of cfDNA concentration categories in plasma (n = 66) and CSF (n = 111), CSF cfDNA concentrations were significantly lower than those in plasma (two-sided Wilcoxon rank-sum test, P < 0.0001); samples below the quantification limit are categorized as undetected (Supplementary Methods). c–e Paired comparison of tumor fractions (c), aneuploidy (d), and genome amplification proportion (e) between matched blood and CSF samples (n = 54 pairs), with statistical significance denoted (two-sided Wilcoxon signed-rank test, ****P < 0.0001). Lines connect matched pairs Low-input sWGS yields higher tumor fraction and CNV burden in CSF than in plasma despite lower cfDNA concentration. a Schematic of the low-input cfDNA profiling workflow, including cfDNA isolation from CSF or plasma, library preparation, shallow (~ 3 ×) whole-genome sequencing and downstream inference of tumor fraction, chromosomal arm-level CNVs and focal amplifications (Methods). b Distribution of cfDNA concentration categories in plasma (n = 66) and CSF (n = 111), CSF cfDNA concentrations were significantly lower than those in plasma (two-sided Wilcoxon rank-sum test, P < 0.0001); samples below the quantification limit are categorized as undetected (Supplementary Methods). c–e Paired comparison of tumor fractions (c), aneuploidy (d), and genome amplification proportion (e) between matched blood and CSF samples (n = 54 pairs), with statistical significance denoted (two-sided Wilcoxon signed-rank test, ****P < 0.0001). Lines connect matched pairs CSF cfDNA testing posed greater technical challenges than plasma because residual CSF volume available for research was limited after routine clinical assays (Supplementary Fig. 2a). CSF cfDNA concentrations were significantly lower than those in plasma (P < 0.0001), and 41.07% of CSF samples were below the quantification limit (Fig. 1b). Nevertheless, sequencing libraries were successfully generated for all CSF samples and could be quantified after amplification, including those with cfDNA concentrations below the quantification limit. In matched CSF-blood pairs, tumor fraction, aneuploidy and GAP were all significantly higher in CSF than in plasma (all P < 0.0001, Wilcoxon signed-rank test; Fig. 1c–e), indicating that CSF provides a richer metastasis-derived CNV signal despite lower cfDNA concentration. Extremely low cfDNA input inevitably reduces library yield, increasing duplication and lowering effective depth after deduplication (Supplementary Fig. 2b–d). However, across serial dilutions down to 100 pg and PCR amplification conditions spanning 12–25 PCR cycles, tumor fraction estimates and CNV profiles remained stable (Supplementary Fig. 3). Together, these results support the robustness of trace-input CSF sWGS for downstream CNV analyses and provide a foundation for subsequent compartment comparisons and diagnostic modeling. CNV-dominant biological features of CNS metastasis are effectively captured by sWGS To corroborate prior evidence that CNS metastasis hotspots are enriched for CNVs relative to mutations (Supplementary Table 1), we reanalyzed lung cancer data from three Memorial Sloan Kettering Cancer Center (MSKCC) cohorts (MSK-MET, MSKCC-LUAD, MSKCC-BM-NSCLC). Across these datasets, CNVs showing significant differences between primary and metastatic lesions were markedly more frequent than SNVs or InDels (Supplementary Fig. 4a), and CNVs enriched in brain metastases substantially outnumbered those enriched in extracranial metastasis (Supplementary Fig. 4b). Moreover, CNV frequencies exhibited large-scale spatial clustering, often spanning neighboring genes or entire chromosomal arms rather than isolated focal changes (Supplementary Fig. 4c). Leveraging the technical robustness of sWGS, we next compared CNV detection between CSF and blood using the 54 paired samples in our cohort (Supplementary Fig. 5). CNVs detected in CSF only were more frequent than those unique to blood or shared in both sample types across nearly all chromosomal arms (Fig. 2a). For genes previously reported as amplification hotspots in CNS metastases (Supplementary Table 1), detection rates were high in both CSF and blood samples from our cohort, with consistently stronger signals in CSF (Fig. 2b). Importantly, when bins were stratified by their relative distance from reported CNV hotspots into five groups (< 25 Mb, 25–50 Mb, and ≥ 50 Mb from amplification hotspots; < 25 Mb from deletion hotspots; and bins on different chromosomes), the patterns of amplification detection concordance between CSF and blood were significantly dependent on this distance stratification (PERMANOVA, P < 0.001; Fig. 2c). Bins nearer amplification hotspot regions exhibited higher detection rates in both plasma and CSF, with a more pronounced increase in CSF. This distance-graded pattern supports the notion that biologically meaningful amplifications often extend over tens of megabases and that CSF-based sWGS effectively aggregates signals across linked loci.Fig. 2CNV Landscape and Distance-Dependent Detection Patterns in CSF and Blood. a, b Frequency of chromosomal arm-level CNVs (a) and reported gene amplification hotspots in lung cancer CNS metastases (b) across the sample pairs (n = 54), classified by detection compartment (blood only, CSF only, or both). c Bin-level amplification and deletion detection rates stratified by distance to reported hotspots (< 25 Mb, 25–50 Mb, and ≥ 50 Mb from amplification hotspots; < 25 Mb from deletion hotspots; and bins on different chromosomes), and shown by detection compartment (n = 54 pairs). Differences across the compartments were tested by PERMANOVA. P-values for pairwise comparisons are annotated as follows: NS., not significant; *P < 0.05; **P < 0.01; ***P < 0.001 CNV Landscape and Distance-Dependent Detection Patterns in CSF and Blood. a, b Frequency of chromosomal arm-level CNVs (a) and reported gene amplification hotspots in lung cancer CNS metastases (b) across the sample pairs (n = 54), classified by detection compartment (blood only, CSF only, or both). c Bin-level amplification and deletion detection rates stratified by distance to reported hotspots (< 25 Mb, 25–50 Mb, and ≥ 50 Mb from amplification hotspots; < 25 Mb from deletion hotspots; and bins on different chromosomes), and shown by detection compartment (n = 54 pairs). Differences across the compartments were tested by PERMANOVA. P-values for pairwise comparisons are annotated as follows: NS., not significant; *P < 0.05; **P < 0.01; ***P < 0.001 Diagnosis of leptomeningeal metastases using CSF cfDNA We first compared CSF characteristics between LM-positive (LM+) and LM-negative (LM−) patients. Because longitudinal samples were available for some patients and CNS disease status could change over time, these analyses were performed at the CSF sample/time-point level and interpreted as descriptive comparisons rather than independent patient-level inference. CSF cfDNA concentrations were significantly higher in LM+ than in LM− samples, and cfDNA fragments were longer in LM+CSF samples (P = 0.0215 and P < 0.01, respectively; Wilcoxon rank-sum test; Supplementary Fig. 6a, b). Tumor fraction, aneuploidy, and GAP were significantly elevated in LM+ relative to the LM− samples (all P < 0.001, Wilcoxon rank-sum test; Supplementary Fig. 6; Fig. 3a, b), suggesting that these sWGS-derived CNV metrics may serve as candidate diagnostic biomarkers for LM.Fig. 3CSF cfDNA sWGS-derived markers for the diagnosis of leptomeningeal metastases. a, b Sample-level distributions aneuploidy and genome amplification proportion compared between LM− (n = 25 samples from 24 patients) and LM+ (n = 87 samples from 29 patients) groups, with statistical significance denoted (two-sided Wilcoxon rank-sum test, ***P < 0.001; **P < 0.01). c Receiver operating characteristic (ROC) curves for LM diagnosis using aneuploidy and GAP. Analyses were restricted to CSF samples meeting the marker-specific minimum effective tumor fraction (TF*) thresholds: TF ≥ 8.5% for aneuploidy (n = 76 samples) and TF ≥ 7% for GAP (n = 79 samples). Colored dots indicate each marker’s optimal marker threshold under the constraint of zero false positives (specificity = 1). The position of CSF cytology is annotated for reference. LM, leptomeningeal metastasis; LM− , patients without LM; LM+ , patients with LM; GAP, genome amplification proportion; AUC, area under the curve CSF cfDNA sWGS-derived markers for the diagnosis of leptomeningeal metastases. a, b Sample-level distributions aneuploidy and genome amplification proportion compared between LM− (n = 25 samples from 24 patients) and LM+ (n = 87 samples from 29 patients) groups, with statistical significance denoted (two-sided Wilcoxon rank-sum test, ***P < 0.001; **P < 0.01). c Receiver operating characteristic (ROC) curves for LM diagnosis using aneuploidy and GAP. Analyses were restricted to CSF samples meeting the marker-specific minimum effective tumor fraction (TF*) thresholds: TF ≥ 8.5% for aneuploidy (n = 76 samples) and TF ≥ 7% for GAP (n = 79 samples). Colored dots indicate each marker’s optimal marker threshold under the constraint of zero false positives (specificity = 1). The position of CSF cytology is annotated for reference. LM, leptomeningeal metastasis; LM− , patients without LM; LM+ , patients with LM; GAP, genome amplification proportion; AUC, area under the curve We next focused on 94 CSF samples from 44 patients with paired cytology to evaluate diagnostic performance. Because CNV-derived metrics become unreliable at very low tumor fraction, and tumor fraction differs between LM+ and LM− samples, we implemented a tumor fraction–gated classifier in which samples below a tumor-fraction cutoff were assigned LM− and the remaining samples were classified using either aneuploidy or GAP thresholds (Methods; Supplementary Methods). Using patient-level internal LOPO cross-validation across a tumor-fraction grid, performance increased and then plateaued, yielding marker-specific minimum effective tumor-fraction thresholds (TF*): 8.5% for aneuploidy and 7% for GAP (Supplementary Fig. 6d). To provide a clinically interpretable benchmark against cytology—a routinely used, high-specificity test in the LM work-up and one component of the composite clinical reference standard used in this study (Methods)—we additionally performed a scenario analysis at an operating point with no false positives observed in this cohort (specificity = 1). Under this constraint, we rederived the optimal biomarker thresholds to maximize sensitivity (Supplementary Table 2). When restricting analysis to samples with TF ≥ TF* (marker-only evaluation), the AUCs were 0.814 for aneuploidy and 0.944 for GAP (Fig. 3c). At specificity = 1, GAP achieved a sensitivity of 0.86 (95% CI 0.73–0.96), compared with 0.56 (0.30–0.77) for aneuploidy and 0.76 (0.62–0.88) for cytology in this cohort (Supplementary Table 2). In a more clinically realistic TF-gated model, in which samples with TF < TF* were classified as LM−, the AUCs of aneuploidy and GAP remained high at 0.888 and 0.922, respectively (Supplementary Fig. 6e). Although the sensitivity difference between GAP and cytology did not reach statistical significance (paired McNemar test, P = 0.211), GAP maintained robust diagnostic performance that was at least comparable to cytology (0.80 [0.61–0.92] vs. 0.76 [0.62–0.88]), and both numerically more sensitive than aneuploidy detection (0.51 [0.26–0.72]). Together, these results support tumor fraction–gated GAP as a quantitative, CNV-derived biomarker that complements conventional cytology for LM diagnostics. Genomic features of leptomeningeal metastasis To delineate the genomic landscape of leptomeningeal metastases, we compared the CSF ctDNA profiles between the LM+ and LM− patients. To minimize false-negative CNV calls in low-frequency samples, we excluded samples with tumor fraction < 7% and selected, for each patient, the CSF sample with the highest tumor fraction for downstream analyses (Methods). Even under these conditions, pronounced differences were observed between LM+ (n = 28) and LM− (n = 13) groups. LM+ patients showed higher ploidy (median, LM + 2.08 vs. LM − 1.98, P = 0.0058), greater aneuploidy (median, LM + 13 vs. LM − 7, P = 0.0031), and markedly elevated GAP (median, LM + 3.45% vs. LM − 0.12%, P = 0.0007). In a further stratified analysis, both LM-only and BM+LM cases showed higher CNV-derived metrics than BM-only cases, suggesting that the LM+ association was not solely attributable to coexisting parenchymal BM (Supplementary Fig. 7). These findings indicate that the genomic differences seen in Fig. 2a and b are not merely a consequence of undetected CNVs in low-tumor fraction samples from patients with brain metastases, but instead reflect large-scale genomic instability characteristic of LM+ group. At the chromosomal arm level, significant differences were observed in 11 arms between LM+ and LM− groups (χ2 test; Fig. 4). To contextualize these arm-level alterations, we re-analysed the MSK-MET metastatic cancer cohort. Because leptomeningeal metastases are rarely resected, CNS/brain metastasis specimens in MSK-MET predominantly represent parenchymal brain metastases (Supplementary Note 1). At the focal level, between-group differences in high-level amplification (CN ≥ 5) were concentrated predominantly on 7p, with a smaller subset near the telomeric region of 5p; among focal events, EGFR amplification showed the strongest statistical support (two-sided Fisher’s exact test, P = 0.0010), whereas amplifications involving several resistance-associated loci showed non-significant enrichment in LM + (Fig. 4; Supplementary Fig. 8). In a subset of samples with tumor fraction > 20% sequenced at higher depth (Methods), hotspot mutation profiling suggested differences in EGFR L858R and KRAS G12D between groups; however, this analysis was underpowered and potentially confounded by histological imbalance. Detailed per-event frequencies and external contextualization using MSK-MET are provided in Supplementary Note 1.Fig. 4Comparison of genomic profiling between brain metastases only and leptomeningeal metastases groups. Each column represents an individual patient-level CSF sample (LM− , n = 13; LM+ , n = 28, including 18 LM+ with BM), utilizing the CSF sample with the highest tumor fraction per patient after excluding those with TF < 7%. Each row represents a genomic variant, including arm-level CNVs, gene amplifications, and key gene mutations. Sample annotations at the top illustrate the histological subtype, lesion status, tumor fraction, and ploidy, with statistical significance indicated where applicable. Variant annotations on the right side comprise a bar plot indicating the frequency of each variant in the two cohorts and a forest plot displaying OR to compare the likelihood of variant occurrence in LM+ versus LM− patients. Statistical significance for continuous variables (TF and ploidy) was determined using the two-sided Wilcoxon rank-sum test, and categorical comparisons were evaluated using the two-sided Fisher’s exact test or Chi-squared test. P-values are explicitly denoted (*P < 0.05; **P < 0.01; ***P < 0.001). LM− , patients without LM; LM+ , patients with LM; OR, odds ratios Comparison of genomic profiling between brain metastases only and leptomeningeal metastases groups. Each column represents an individual patient-level CSF sample (LM− , n = 13; LM+ , n = 28, including 18 LM+ with BM), utilizing the CSF sample with the highest tumor fraction per patient after excluding those with TF < 7%. Each row represents a genomic variant, including arm-level CNVs, gene amplifications, and key gene mutations. Sample annotations at the top illustrate the histological subtype, lesion status, tumor fraction, and ploidy, with statistical significance indicated where applicable. Variant annotations on the right side comprise a bar plot indicating the frequency of each variant in the two cohorts and a forest plot displaying OR to compare the likelihood of variant occurrence in LM+ versus LM− patients. Statistical significance for continuous variables (TF and ploidy) was determined using the two-sided Wilcoxon rank-sum test, and categorical comparisons were evaluated using the two-sided Fisher’s exact test or Chi-squared test. P-values are explicitly denoted (*P < 0.05; **P < 0.01; ***P < 0.001). LM− , patients without LM; LM+ , patients with LM; OR, odds ratios Collectively, these data highlight that LM is characterized by diffuse chromosomal instability with arm-level CNVs and focal amplifications, superimposed on a background of relatively limited recurrent point mutations. Serial cfDNA analyses reveal therapeutic responses and resistance We further explored whether longitudinal CSF cfDNA profiling could track treatment responses and CNV evolution in CNS metastases. Longitudinal CSF monitoring showed that total cfDNA concentration, sWGS-derived ctDNA fraction, and cytology-derived tumor cell fraction provided related but non-identical information. In patients who achieved durable intracranial responses (P043 and P058), both CSF cfDNA concentrations and ctDNA fractions declined steadily, accompanied by parallel decreases in tumor cell fraction, total protein and CEA levels, while glucose levels increased (Supplementary Fig. 9a, b). These cases represent a concordant pattern in which total cfDNA concentration and ctDNA fraction changed in the same direction during treatment response. In contrast, during progression or more complex clinical transitions, total CSF cfDNA concentration was often more dynamically concordant with LM status than ctDNA fraction or cytology. In patient P055, rising total cfDNA concentration and tumor cell fraction reflected transitions in LM status from PR to SD and then PD, whereas the ctDNA fraction remained relatively stable and showed limited dynamic variation (Supplementary Fig. 9c). Similarly, in patient P032, cfDNA concentrations tracked clinical response and progression across multiple lines of therapy, whereas the high tumor cell fraction showed a paradoxical decrease immediately before treatment switch during second-line treatment and cytology showed little change during the fourth-line PR phase; the corresponding ctDNA fraction remained relatively stable without corresponding dynamic changes (Supplementary Fig. 9d). In patient P053, total cfDNA concentration was concordant with two SD-to-PD transitions during third- and fourth-line therapies whereas cytology fluctuated and the ctDNA fraction remained high but showed limited dynamic variation across these transitions (Fig. 5a). Notably, cfDNA sensitively captured LM-specific responses but, like cytology, did not reflect the emergence of new brain parenchymal lesions, underscoring the compartment-specific nature of CSF-based monitoring. These observations suggest that total CSF cfDNA may serve as an empirical longitudinal readout of CSF disease activity in selected LM cases, whereas ctDNA fraction and cytology provide complementary but not always dynamically equivalent information.Fig. 5Longitudinal CSF cfDNA profiling reveals treatment response and progressive TYMS amplification in patient P053. a Temporal dynamics of clinical evaluations and CSF biomarkers across multiple lines of therapy. Panel display, from top to bottom: (i) clinical multi-line treatment regimens and efficacy evaluations; (ii) cellular metrics, including total cell concentration and tumor cell fraction; (iii) cfDNA concentration and ctDNA fraction; (iv) biochemical parameters, including total protein, CEA, and glucose concentrations; and (v) EGFR mutations detected by sWGS and by targeted panel. b Schematic diagram of the pemetrexed mechanism illustrating the inhibition of thymidylate synthase (encoded by TYMS). c Contrast-enhanced MRI, LM lesion efficacy evaluation, and corresponding genome-wide CNV profiles at representative CSF sampling time points. For each time point, a genome-wide copy number map is shown, with TYMS and other detected previously reported amplification and homozygous deletion explicitly marked. CEA, carcinoembryonic antigen Longitudinal CSF cfDNA profiling reveals treatment response and progressive TYMS amplification in patient P053. a Temporal dynamics of clinical evaluations and CSF biomarkers across multiple lines of therapy. Panel display, from top to bottom: (i) clinical multi-line treatment regimens and efficacy evaluations; (ii) cellular metrics, including total cell concentration and tumor cell fraction; (iii) cfDNA concentration and ctDNA fraction; (iv) biochemical parameters, including total protein, CEA, and glucose concentrations; and (v) EGFR mutations detected by sWGS and by targeted panel. b Schematic diagram of the pemetrexed mechanism illustrating the inhibition of thymidylate synthase (encoded by TYMS). c Contrast-enhanced MRI, LM lesion efficacy evaluation, and corresponding genome-wide CNV profiles at representative CSF sampling time points. For each time point, a genome-wide copy number map is shown, with TYMS and other detected previously reported amplification and homozygous deletion explicitly marked. CEA, carcinoembryonic antigen Beyond quantitative response assessment, longitudinal sWGS enabled direct visualization of evolving genomic events. In patient P053, who received continuous intrathecal pemetrexed across third- and fourth-line therapies, serial sWGS revealed progressive amplifications that paralleled clinical progression. Pharmacologically, pemetrexed inhibits thymidylate synthase (Fig. 5b), encoded by TYMS, thereby blocking the conversion of dUMP to dTMP and impairing downstream DNA replication[25]. During progression on third-line therapy, TYMS copy number increased dramatically from 5.0 to 27.2, accompanied by an increase in CCND1 (encoding Cyclin D1) from 6.3 to 11.1 (Fig. 5c). During fourth-line therapy, when disease was assessed as SD, TYMS copy number rose modestly to 31.7 but then escalated to 42.7 at subsequent PD. In contrast, other amplified genomic regions remained stable throughout follow-up, indicating locus-specific amplification rather than global genome doubling. The treatment-stage-specific increases in TYMS copy number during both PD events are consistent with treatment-associated selection and nominate TYMS amplification as a resistance-associated CNV candidate. Notably, sWGS effectively captured progressive TYMS amplification consistent with CNV-mediated resistance, whereas paired deep targeted sequencing (172-gene panel; Methods) detected low-frequency SNVs such as EGFR C797S that were not reliably detected by shallow sWGS, underscoring the complementary strengths of CNV-centric sWGS and targeted panels. Together, these longitudinal analyses illustrate the clinical value of CSF cfDNA for LM monitoring. In addition, progressive TYMS amplification in P053 provides an exploratory proof-of-concept that longitudinal CSF sWGS may nominate candidate CNV-mediated resistance events during treatment progression. Discussion In this study, we show that shallow whole-genome sequencing (sWGS) of trace CSF cfDNA can help address key technical and biological barriers to molecular profiling of CNS metastases. By prioritizing genomic breadth over depth, sWGS enables quantitative assessment of copy number variation (CNV) burden from low-input CSF samples and provides CNV-derived readouts that support LM diagnosis and longitudinal disease monitoring. These findings are consistent with emerging evidence that CNS metastases frequently harbor large-scale structural alterations and that CNVs may contribute substantially to their molecular architecture relative to recurrent focal point mutations. From a technical perspective, our data highlight a fundamental constraint of mutation-centric CSF cfDNA assays: the limiting factor is input DNA rather than sequencing depth. Deep targeted sequencing presumes that sufficient unique cfDNA molecules exist to support high post-deduplication depth. However, in real-world practice, CSF from patients with CNS tumors often has cfDNA concentrations near or below the quantification limit [8]. Previous studies have noted that a substantial fraction of CSF samples fail library preparation, yet most published work has focused on quality-passed subsets, potentially underestimating the impact of low-input samples [3]. Even when libraries are successfully constructed, the absolute scarcity of input molecules limits the effective sequencing depth achievable after deduplication, imposing an absolute limit of detection (LOD) that renders mutation detection unreliable despite large nominal sequencing yields [2, 18]. In contrast, we show that sWGS libraries can be successfully prepared for all CSF samples in our cohort and that tumor fraction estimates and CNV profiles remain stable across a broad range of input amounts down to 100 pg. This robustness arises because CNV inference in sWGS depends on aggregated shifts across many genomic bins rather than depth at any single locus. While sWGS has intrinsically lower sensitivity for rare SNVs and very small InDels, CSF from CNS metastases typically contains relatively high tumor fractions, mitigating dependence on ultra-deep mutation detection in many clinical contexts. Despite this analytical robustness, the trace nature of CSF cfDNA makes the assay particularly susceptible to pre-analytical variables. Therefore, successful clinical translation necessitates stringent standardization of sample collection and processing to prevent in vitro genomic DNA contamination. Biologically, our analyses confirm and refine observations that CNS metastases, and LM in particular, are frequently characterized by extensive CNV superimposed on relatively limited recurrent point mutations. Reanalysis of MSK cohorts showed that CNVs more frequently distinguish brain metastases from extracranial lesions than do SNVs or InDels, and that these CNV events often cluster over contiguous genomic regions spanning neighboring genes or entire chromosomal arms rather than isolated loci. In our cohort, amplification detection exhibited a clear distance-graded pattern relative to known CNS CNV hotspots: genomic bins located closer to reported amplification regions showed higher detection rates in both CSF and plasma, with a more pronounced gradient in CSF. This distance dependence supports the concept that biologically meaningful amplifications extend across linked genomic domains spanning tens of megabases. Together, these observations provide a mechanistic rationale for adopting a breadth-based, CNV-focused strategy using shallow sequencing, in preference to conventional mutation-centric designs that concentrate depth on predefined loci. These biological and technical insights support a CNV-derived diagnostic framework for LM. Because CNV calling becomes unreliable at very low tumor fractions, we implemented a tumor-fraction gate to distinguish low-input LM− samples from CNV-informative samples. Although prior studies have suggested that aneuploidy can discriminate LM-positive from LM-negative cases in various solid tumors [7, 29]. In advanced lung cancer, however, we found that aneuploidy alone did not improve diagnostic specificity beyond cytology in our cohort, likely because chromosomal CNVs are already common in earlier disease stages. In contrast, GAP emerged as a more discriminative metric: among tumor-fraction–adequate samples and under a zero-false-positive constraint mirroring clinical cytology, GAP achieved higher sensitivity than both aneuploidy and cytology. In a TF-gated model approximating real-world practice, GAP maintained cytology-comparable sensitivity and outperformed aneuploidy. Taken together, these results support a diagnostic paradigm in which LM is approached as a quantitative structural variant–dominated disease, with GAP providing a complementary structural readout that extends, rather than replaces, traditional microscopy-based cytology. LoD considerations help explain why concordance is not expected to be perfect across modalities. For sWGS–cytology concordance, the key boundary is the relative LoD of CNV-based sWGS, determined by tumor-derived DNA fraction and broad CNV burden; for sWGS–targeted panel concordance, the key boundary is the absolute LoD imposed by input-molecule scarcity. Thus, low-input CSF sWGS should be integrated with cytology, MRI, CSF biochemical markers, targeted sequencing and clinical assessment when resolving ambiguous diagnostic cases. By design, GAP and aneuploidy are inherently macroscopic indicators of genome instability and do not pinpoint specific clinical actionable loci. We therefore also explored LM-associated genomic features at the level of chromosomal arms, focal amplifications, and selected mutations. Although the limited number of high–tumor fraction LM− cases restricts statistical power to define definitive LM-specific signatures, our data suggest recurrent patterns of arm-level CNV and amplification hotspots—including alterations at loci previously implicated in CNS metastasis—that merit evaluation in larger cohorts. Future work could integrate more refined analytical approaches, such as assessing telomere- or centromere-bounded CNV length distributions in expanded datasets to identify genomic regions under selective pressure [22], followed by functional validation in appropriate in vivo models. Such studies will be critical to delineate how specific CNVs contribute to LM initiation and progression and to uncover structural-variant–based therapeutic vulnerabilities. Our longitudinal analyses underscore the integrated clinical value of CSF cfDNA in LM management. Serial CSF profiling may support longitudinal monitoring by jointly assessing total cfDNA concentration, sWGS-derived ctDNA fraction, CNV burden, cytology and biochemical markers. Longitudinal interpretation also requires distinguishing total CSF cfDNA concentration from sWGS-derived ctDNA fraction. ctDNA fraction is useful for confirming the presence of tumor-derived DNA in the CSF DNA pool. However, in selected longitudinal cases, total CSF cfDNA concentration was more consistently concordant with dynamic changes in LM status than ctDNA fraction or cytology. These observations support total CSF cfDNA concentration as an empirical, clinically informative marker of CSF disease activity in LM cases, although it is not tumor-specific and may reflect both tumor-derived DNA and non-tumor-specific processes such as neuroinflammation, immune responses, or treatment-induced normal cellular death. In addition, sample/time-point-level comparisons may be influenced by repeated sampling, changes in CNS disease status, coexisting BM and CSF collection route, which require validation in larger prospective cohorts. At the same time, the lack of signal for new brain parenchymal lesions highlighted the compartment specificity of CSF-based monitoring: highly sensitive for leptomeningeal disease yet relatively insensitive to isolated parenchymal progression. Beyond quantitative response assessment, sWGS captured adaptive CNV evolution, exemplified by progressive TYMS amplification during continuous intrathecal pemetrexed in one patient. Given that TYMS overexpression has been associated with pemetrexed resistance [28], the stage-specific increases in TYMS copy number across successive progressive-disease events in this case are consistent with a resistance-associated CNV pattern under therapeutic pressure. Long-term serial CSF sampling was available for only a limited number of patients, limiting our ability to systematically characterize resistance evolution. Nevertheless, progressive TYMS amplification in P053 during intrathecal pemetrexed exposure provides an exploratory proof-of-concept that longitudinal CSF sWGS may nominate candidate CNV-mediated resistance events. Larger longitudinal cohorts are needed to determine whether such CNV dynamics can reproducibly inform resistance monitoring or mechanism discovery. Paired targeted panel sequencing further emphasizes the complementary nature of these approaches: while deep panels retain superior sensitivity for rare SNVs such as EGFR C797S, sWGS provides genome-wide information on large-scale CNV evolution. Notably, sWGS remains suboptimal for low-frequency SNVs, small InDels, and very focal CNVs, and targeted sequencing will remain essential at critical time points, particularly around progression. Although this study provides clinically relevant proof-of-concept evidence for low-input CSF sWGS in LM, several limitations should be acknowledged. First, our study cohort has limitations in size and composition. The moderate sample size and limited number of high-tumor-fraction LM− cases reduce statistical power to define LM-associated CNV patterns. The LM group was enriched for EGFR-mutated adenocarcinoma, consistent with the established clinical epidemiology of lung cancer LM [12], but this composition may limit the generalizability of GAP and other CNV-derived markers to EGFR-wild-type NSCLC, LUSC, SCLC and other underrepresented subgroups. Second, the tumor-fraction and GAP thresholds used in our diagnostic framework were derived by internal cross-validation within this cohort. Therefore, their broader use will require external validation and, if necessary, recalibration across technical workflows, clinical settings and biologically diverse patient populations. Third, although our comparative analyses suggest that LM may represent a more CNV-enriched phenotype within the spectrum of CNS involvement, we did not directly test the functional contributions of individual CNVs to LM initiation, progression or treatment response. Future studies integrating CSF sWGS-based CNV mapping with biological investigation and prospective clinical evaluation will be important to clarify the functional relevance of these alterations and refine their clinical utility in LM. In summary, our study supports the utility of low-input CSF sWGS for genomic profiling and longitudinal monitoring of CNS metastases, with particular relevance for LM. By aligning assay design with the CNV-rich architecture of CNS involvement, we present a CNV-centric framework in which LM can be evaluated through quantitative structural readouts from CSF, while recognizing that CSF cfDNA is relatively insensitive to isolated parenchymal progression and should be complemented by other modalities in the assessment of brain metastases. Future studies should validate this framework in larger, multi-center, and histologically diverse cohorts, assess its applicability across other CNS metastatic contexts, and further integrate biological and clinical studies to refine the deployment of sWGS-guided diagnostics and longitudinal surveillance. Supplementary Information Supplementary Material 1 Supplementary Material 2 Supplementary Material 1 Supplementary Material 2