Method of diagnosing or treating cancer

Abstract

Methylation arrays and a method of diagnosing or treating a cancer in a subject using differential methylation levels. In one implementation, the method includes the steps of obtaining a cfDNA sample from a subject, probing the cfDNA sample with a plurality of CpG probes, and comparing information from the probed cfDNA sample to a methylation array. The methylation arrays are used to determine a differential methylation level at each of a plurality of limited CpG sites within the cfDNA sample. The plurality of limited CpG sites correspond to the plurality of CpG probes, and two or more of the CpG sites are located on chromosome (2), two or more are located on chromosome (6), and two or more are located on chromosome (16).

Description

[0001] METHOD OF DIAGNOSING OR TREATING CANCER

[0002] TECHNICAL FIELD

[0003] The present disclosure relates generally to systems and methods for treating cancer, more particularly, tumor generating central nervous system (CNS) cancers. BACKGROUND

[0004] The inventors of the present application have worked to establish epigenetic liquid biopsy scores to detect, diagnose, and help treat cancer. While prior publications generally describe potential methods for how this could be done, the present disclosure details the results of those investigations and particular differential methylation array s / pattems that improve diagnosis and treatment outcomes for cancers, mainly for tumor causing CNS cancers. The following three publications and their supplementary materials are accordingly incorporated by reference herein in their entireties:

[0005] (1) Sabedot et al., “A serum-based DNA methylation assay provides accurate detection of glioma,” Neuro-Oncology, 23(9), 1494-1508, 2021. (2) Herrgott et al., “Detection of tumor-specific DNA methylation markers in the blood of patients with pituitary neuroendocrine tumors,” NeuroOncology, 24(7), 1126-1139, 2022.

[0006] (3) Herrgott et al., “Detection of diagnostic and prognostic methylation-based signatures in liquid biopsy specimens from patients with meningiomas,” Nature Communications, 14:5669, 1-19, 2023.

[0007] The disclosure herein is focused more substantially on subjects with meningiomas, but as will be understood, are applicable to other cancer types as well, particularly CNS cancers that result in tumors (e.g., meningiomas, pituitary neuroendocrine tumors, and gliomas). Meningiomas are the most common primary tumors of the central nervous system (CNS). According to the World Health Organization (WHO), meningiomas classified as grades 2 and 3 account for 20-30% of cases. These tumors present an estimated rate of recurrence of 20-75% across grade 2 and an observed universal rate of recurrence across grade 3. within 10 years of patient follow-up. Additionally, some cases have potential for malignization, metastasizing and may even prove lifethreatening. An immediate challenge following meningioma identification lies in determining whether temporal surveillance through imaging or a tailored interventional approach, such as surgery or radiation, is the most appropriate patient management. Off-label investigational therapeutics have been attempted in clinical trials; however, no widely approved systemic therapies for this disease currently exist.

[0008] One of the principal hindrances to meningioma treatment advancement is the paucity of standardized assessment criteria or adequate biomarkers to measure success in clinical trials. In tandem, detection of genomic and epigenomic biomarkers has become standard practice in oncology and proven valuable for classification, prognostication and appropriate management of CNS tumors, including meningiomas. Specifically, stratification of meningiomas according to DNA methylation patterns in tumor tissue has proven to be an independent and reliable outcome predictor across all meningioma subtypes, and has outperformed the WHO grading system alone, across retrospective and prospective cohorts. Furthermore, integration of DNA methylationbased groups with complementary molecular features (e.g., copy number variations, WHO grading, specific mutations: NF2. TERT, etc.) exhibited marked improvements in predicting the recurrence risk in patients with meningiomas. Currently, these histologic and molecular characterizations are contingent on the profiling of meningioma tissue obtained through surgery7. However, this approach may be infeasible for surgically inaccessible tumors, for patients with complicative comorbidities, or delayed, when tumors detected by imaging are mistakenly considered benign meningioma based on whether they are small, asymptomatic or discovered incidentally through imaging approaches. Additionally, multiple surgeries are impractical and pose inherent cumulative risks for serial assessment of these tumors. Therefore, development of minimally- or noninvasive approaches to detect established or novel molecular markers which reflect real-time tumor biology and behavior is warranted. Imaging techniques are the current noninvasive approach used to guide diagnosis and management of meningiomas; however, its associated prognostic value is still unclear and longitudinal assessment may prove costly, unavailable and cumbersome for some patients. Moreover, consensus standard radiographic criteria for inclusion and outcome evaluation for use across interventional trials has illustrated that limitations exist in the application of imaging criteria alone to characterize this heterogenous disease.

[0009] Liquid biopsy (LB) is a non- or minimally invasive approach that allows for detection of material shed by tumors (e.g., circulating tumor cells and cell-free or tumor genomic elements) in biofluids (e.g.. blood, cerebrospinal fluid, stool, urine, saliva and others). Several studies have described the feasibility of applying blood-based LB to screen mutations and DNA methylation abnormalities using serum- or plasma-cell free (cf) DNA from patients with CNS tumors. However, current methylation-based prognostication models have been reported solely across surgically obtained tissue, but not LB specimens, from patients harboring these tumors.

[0010] SUMMARY

[0011] In accordance with one embodiment, there is provided a method of diagnosing or treating a cancer in a subject, comprising the steps of: obtaining a cfDNA sample from a subject, probing the cfDNA sample with a plurality of CpG probes, and comparing information from the probed cfDNA sample to a methylation array to determine a differential methylation level at each of a plurality of limited CpG sites within the cfDNA sample. The plurality of limited CpG sites correspond to the plurality of CpG probes, and two or more of the CpG sites are located on chromosome 2, two or more are located on chromosome 6, with two or more are located on chromosome 16. The differential methylation level is used to diagnose and / or treat the cancer.

[0012] In various embodiments, the plurality of limited CpG sites includes three or more located on chromosome 6. The plurality7of limited CpG sites may not include CpG sites on chromosome 8, on chromosome 9, on chromosome 18, or chromosome 21. The cancer is a central nervous system (CNS) tumor causing cancer, and the CNS tumor causing cancer causes one or more meningiomas, one or more pituitary neuroendocrine tumors, or one or more gliomas.

[0013] In various embodiments, the CNS tumor causing cancer causes one or more meningiomas, and the plurality of limited CpG sites includes a majority7or all of the CpG sites of Table 2. The CNS tumor causing cancer may cause one or more pituitary7neuroendocrine tumors, and the plurality of limited CpG sites includes a majority or all of the CpG sites of Table 3. The CNS tumor causing cancer causes one or more gliomas. and the plurality of limited CpG sites includes a majority or all of the CpG sites of Table 4.

[0014] In various embodiments, the method includes the step of determining a score from the differential methylation level. The score has one or more CpG sites of the plurality of limited CpG sites weighed differentially than one or more other CpG sites of the plurality of limited CpG sites. Machine learning can be used to adjust the score on a case-by-case basis. A differential methylation level can be determined for each of the plurality of limited CpG sites and each methylation level has an associated score.

[0015] In various embodiments, the cfDNA sample is a serum-based liquid biopsy sample. The differential methylation level can be a hypermethylation status determined for each of the plurality of limited CpG sites, and a treatment regime for the subject can be altered based on the differential methylation level to adjust an active interv ention or change a disease surveillance strategy. A non-transitory, computer-readable storage medium can store instructions thereon, which when executed by one or more electronic processors causes the one or more electronic processors to cany' out the method.

[0016] In some embodiments, a methylation array is used to determine a differential methylation level at each of a plurality of limited CpG sites within a cfDNA sample. The methylation array comprises a plurality of CpG probes, each CpG probe of the plurality' of CpG probes corresponding to each CpG site of the plurality of limited CpG sites. The methylation array is configured to diagnose or treat meningiomas, and the plurality of CpG probes includes a majority of the CpG probes in Table 2, in Table 3, and / or in Table 4.

[0017] It is contemplated that any number of the individual features of the abovedescribed embodiments and of any other embodiments depicted in the drawings or description below can be combined in any combination to define an invention, except where features are incompatible.

[0018] BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Preferred example embodiments will hereinafter be described in conjunction with the appended drawings, wherein like designations denote like elements, and wherein: FIG. 1 is a principal component analysis (PCA) depicting the genome-wide mean methylation levels of serum cfDNA derived from patients with meningioma (MNG; n = 63) and non-MNG conditions (other CNS entities and non-neoplastic diseases; n = 141).

[0020] FIG. 2 shows mean methylation levels of the differentially methylated CpG probes (DMP, n = 98) across comparisons between MNG and non-MNG (Wilcoxon rank sum test; Kruskal-Wallis; *p<0.05, **p < 0.01, ***p<0.001). Box plots - data are presented as median and upper (75%) and lower (25%) quartiles. Whiskers represent minimum to maximum values, excluding outliers. Exact p-values: Meningioma vs Non- neoplastic Disease: p = 0.018; Meningioma vs Glioma: p=4.4e-16; Meningioma vs pituitary neuroendocrine tumors: p = 0.017; Glioma vs Other CNS tumors: p = 0.012. Note: DMP: differentially methylated probes.

[0021] FIG. 3 includes t-distributed stochastic neighbor embedding (t-SNE) plots displaying clustering of meningioma-specific DMPs across MNG and non-MNG tissue specimens.

[0022] FIG. 4 shows a subset of the DMPs of FIG. 3 that are detected in the serum and also distinguish equivalent groups.

[0023] FIG. 5 is a t-SNE plot displaying dimension-reduced diagnostic-Meningioma Epigenetic Liquid Biopsy (d-MeLB) probes (similarly methylated probes (SMP): n = 18k CpGs) across CNS tumor tissue, liquid biopsy (LB) (serum and plasma) and tumor tissue from patients with MNG.

[0024] FIG. 6 shows distribution of the d-MeLB scores across independent cohorts (original liquid biopsy serum, n=93).

[0025] FIG. 7 shows additional MNG serum, n = 19 on the left, and on the right, additional MNG plasma, n = 10 (Dashed line: MeLB cutoff score). Box plots - data are presented as median and upper (75%) and lower (25%) quartiles. Whiskers represent minimum to maximum values, excluding outliers. Upper left comer: performance measures. For designation: ACC Accuracy, SE Sensitivity7, SP Specificity, CUI Clinical Utility Index, MCC Matthew’s Correlation Coefficient, IT initial treated, IU initial untreated, RT recurrent treated, RU recurrence untreated. FIG. 8 is a methylation heatmap displaying the 1000 most variable methylated probes (P-values) across serum meningioma unsupervised k-clusters (n = 63). Samples are sorted into methylation based clusters and annotated with clinicopathological / molecular features. Vertical tracks (right) genomic annotations. For designation: LI Labeling Index, EOR Extent of Resection, NI non-informed, MRI magnetic resonance imaging.

[0026] FIG. 9 shows distribution of the prognostic-Meningioma Epigenetic Liquid Biopsy (p-MeLB) scores across (on the left) the original independent cohort (n=23) and (on the right) additional validations (n =50) from patients with meningiomas presenting different outcomes (confirmed recurrence or no recurrence; dashed lines: p-MeLB score cutoff). Box plots - data are presented as median and upper (75%) and lower (25%) quartiles. Whiskers represent minimum to maximum values, excluding outliers. Upper left comer: performance measures.

[0027] FIG. 10 is a scatterplot displaying the relationship between p-MeLB and the nomogram recurrence risk prediction across the primary meningioma tissue subset. Linear relationship is depicted with 95% confidence interval (lower and upper limits). Measurements of concordance are displayed (Cohen’s unweighted kappa / Spearman’s p, p < 0.05). Table comparing the accuracies of p-MeLB and the nomogram-based classifier across an independent subset of primary meningioma tissue (n = 69). For designation: CR Confirmed Recurrence, CNR Confirmed No Recurrence, RR recurrence risk.

[0028] FIG. 11 includes Kaplan-Meier survival curves displaying meningioma tumor tissue samples stratified by their predicted recurrence risk (n = 127, vertical ticks: censorship). Survival curves are depicted with 95% confidence intervals (lower and upper limits) for point estimates; comparisons of median survival time in both recurrence risk groups were conducted using log-rank tests (p < 0.0001).

[0029] FIG. 12 shows clinicopathological feature proportions and associated odds ratios (p-values: two-sided Fisher's Exact test; error bars: 95% confidence interval estimates) derived from the comparison between meningioma serum samples predicted to present high or low recurrence risks. Reference column depicts the mean proportion of each feature across the whole cohort. For designation: SB Skull-base, NSB Non- Skull Base. Y Yes, N No, GTR Gross Total Resection, STR Subtotal Resection, PD Progressive Disease, SD Stable Disease, NED Non-Enhancing Disease; Bolded features are those with observed statistical significance.

[0030] FIG. 13 shows immune cell proportions and associated mean differences derived from the comparison between meningioma serum samples predicted to present high or low recurrence risks (error bars: mean difference 95% confidence interval; p- values: two-sided t-test). Reference column depicts the mean proportion across the whole cohort. For designation: NLR Neutrophil-Lymphocyte Ratio. Bolded features are those with observed statistical significance.

[0031] FIG. 14 is a schematic summarization of observed clinicopathological and molecular features across samples (LB-serum and / or tissue) from patients with meningiomas predicted to present high or low risk of recurrence through p-MeLB. For designation: RFS Recurrence Free Survival, RR Recurrence Risk, MNG C Bayley Meningioma C group, CNV copy number variation, PRC Polycomb Repressive Complex.

[0032] FIG. 15 is a schematic representation — clinical application of liquid biopsy DNA methylation-based diagnostic and prognostic classifiers in patients suspected to present meningioma. For designation: MeLB Meningioma epigenetic Liquid Biopsy, d-MeLB and p-MeLB diagnostic- and prognostic- Meningioma Epigenetic Liquid Biopsy, respectively. MRI magnetic resonance imaging.

[0033] FIG. 16 is a principal component analysis depicting the tissue-derived recurrence risk group DMPs (High vs Low) and respective outcomes (Confirmed Recurrence vs. Confirmed No Recurrence) across the tumor tissue cohort.

[0034] FIG. 17 is a principal component analysis depicting the tissue-derived recurrence risk group DMPs (High vs Low) and respective outcomes (Confirmed Recurrence vs. Confirmed No Recurrence) across the original liquid biopsy serum cohort and an additional serum cohort of samples from patients with meningiomas.

[0035] FIG. 18 is a heat map displaying methylation and expression levels of differentially methylated probes and differentially expressed target genes that are negatively correlated across high and low recurrence risk groups identified in an external molecular meningioma tissue dataset (Choudhury et al., 2022). For designation: *PGP Promoter-linked probe-gene pair, PGP probe-target gene pair.

[0036] FIG. 19 is a principal component analysis of liquid biopsy serum samples using the identified and concordantly methylated probe-gene pairs as input.

[0037] FIG. 20 is a scatter plot depicting serum-derived risk-specific probes that are also detected in meningioma tissue and respective target genes expression changes between high and low risk sample cohorts.

[0038] DETAILED DESCRIPTION

[0039] Definitions

[0040] The terms '‘cell free DNA” and “cfDNA” reference to DNA may include fragments of DNA and / or otherwise enriched or processed DNA samples derived from cfDNA. In an advantageous embodiment, cfDNA is derived from serum or plasma, more particularly serum, in a liquid biopsy approach that allows for detection of material shed by tumors (e.g.. circulating tumor cells and cell-free or tumor genomic elements) in biofluids (e.g., blood, cerebrospinal fluid, stool, urine, saliva, and others).

[0041] As used herein. “CpG sites’' include regions of DNA having 5'-C-phospohate- G-3’ (cytosine and guanine separated by a single phosphate group) that is methylated or unmethylated. Once methylated, the cytosines form 5-metyhlcytosine. In some embodiments, the CpG site can comprise uracil that is generated by the conversion of the unmethylated cytosine. “CpG probes'’ are probes that bind to cfDNA or cfDNA fragments at the CpG sites specified herein. A “methylation array” or interchangeably a “methylation signature” or “methylation profile” may be a limited set of CpG probes that are used to strategically identify a corresponding limited set of CpG sites.

[0042] A “differential methylation level” refers to hypomethylated or hypermethylated CpG sites that are present in a cancerous control group as compared to a non-cancerous control group. As detailed herein, methylation at particular CpG sites has been shown to be indicative of the presence or absence of certain central nervous system tumors, such as meningiomas, pituitary neuroendocrine tumors, and gliomas. Generally, hypermethylated refers to a maj only of designated CpG sites in a methylation array that are methylated in a particular cancer population as compared to a particular non- cancerous population (e.g., greater than 50%, or more particularly, greater than 80%, greater than 90%, or greater than 95% in some implementations). These CpG sites may also be interchangeably referred to as “differentially methylated probes’’ or “DMPs”. Groups of CpG sites that have similar methylation patterns in a particular cancer population as compared to a non-cancerous population may be referred to as “similarly methylated probes” or “SMPs.”

[0043] The terms "tumor", “tumor causing cancer,” and "cancer" are used interchangeably and refer to a cell or population of cells whose grow th, proliferation or survival is greater than growth, proliferation or survival of a normal counterpart cell. The cell or population of cells in a tumor or cancer possess abnormal growth, and typically the growth is uncontrolled.

[0044] As used herein, the terms "treatment" or "treating” refer to an approach for obtaining a beneficial or desired result, preferably a beneficial or desired clinical result. Such beneficial or desired clinical results include, but are not limited to: (1) curing, healing, alleviating, relieving, altering, remedying, ameliorating, improving, interfering with, or affecting a condition (e.g., a disease), the symptoms of the condition, or the biological manifestations of the condition; (2) interfering with one or more points in the biological cascade that leads to or is responsible for the condition; (3) preventing, delaying, or slowing the onset or progression of the symptoms, complications, biological manifestations, and / or biochemical indicia of a disease or condition; or (4) otherwise arresting or inhibiting further development of the disease, condition, or disorder. The treatment regime may include adjusting an active intervention or changing a disease surveillance strategy. The compositions and methods of the present disclosure are suitable for obtaining beneficial or desired results such as reducing the proliferation of (or destroying) cancerous cells or other diseased cells, reducing metastasis of cancerous cells found in cancers, shrinking the size of the tumor, decreasing symptoms resulting from the disease, increasing the quality of life of those suffering from the disease, decreasing the dose of other medications required to treat the disease, delaying the progression of the disease, and / or prolonging survival of treatment subjects. Treatment can include administration of one or more therapeutic agents with the purpose to achieve beneficial or desired results that are at least partially based on the methods described herein. The term "prevention" as used herein refers to a prophylactic approach intended to substantially diminish the likelihood or severity of a condition or biological manifestation thereof, or to delay the onset of such condition or biological manifestation thereof.

[0045] As used herein, the term "subject" is intended to include human and non-human animals. The terms "subject" and "patient" are used interchangeably and can refer to human patients, as well as non-human primates or experimental animals such as rabbits, dogs, cats, rats, mice, and other animals. Preferred subjects of the present disclosure include mammals, or more particularly, human patients in need of a treatment for a disease or disorder. A subject of the present disclosure may be a patient suffering from cancer, such as a tumor causing central nervous system cancer. The methods herein are particularly suitable for diagnosing or treating meningiomas, pituitary neuroendocrine tumors, and / or gliomas.

[0046] As used herein, "about" means within acceptable error range for the particular value as determined by one of ordinary skill in the art, which will depend in part on how the value is measured or determined, i.e., the limitations of the measurement system. For example, "about" can mean within 1 or more than 1 standard deviation per the practice in the art. Alternatively, "about" can mean a range of up to 20%. When particular values are provided in the application, unless otherwise stated, the meaning of "about" should be assumed to be within acceptable error range for that particular value.

[0047] As used herein, an "effective amount" is defined as the amount required to confer a therapeutic effect on the treated patient, and is typically determined based on age, surface area, w eight, and condition of the patient. The interrelationship of dosages for animals and humans (based on milligrams per meter squared of body surface) is described by Freireich et al.. Cancer Chemother. Rep., 50: 219 (1966). Body surface area may be approximately determined from height and weight of the patient. See, e.g., Scientific Tables, Geigy Pharmaceuticals, Ardsley , New- York, 537 (1970).

[0048] The term '‘processor’’ or ‘'electronic processor” as used herein may include any ty pe of device capable of processing electronic instructions, including microprocessors, microcontrollers, host processors, controllers, and application specific integrated circuits (ASICs). It can be a dedicated processor used only for a specially programmed computer used to carry out one or more of the methods herein, or it may be shared or used with other diagnostic or assessment related devices. A processor executes various types of digitally-stored instructions, such as software or firmware programs stored in memory. For example, a processor can execute programs or process data to carry out at least part of the diagnostic and treatment methods discussed herein. Memory may be a temporary powered memory, and any non-transitory computer readable medium, or other type of memoiy. For example, memory can be any of a number of different types of RAM (random-access memory), ROM (read-only memory ), solid-state drives (SSDs), hard disk drives (HDDs), magnetic or optical disc drives, etc.

[0049] The terms ‘'for example,” "‘for instance,” "‘such as,” and ’‘like,” and the verbs “comprising,” “having,” “including,” and their other verb forms, when used in conjunction with a listing of one or more compounds or other items, are each to be construed as open-ended, meaning that the listing is not to be considered as excluding other, additional compounds or items. Other terms are to be construed using their broadest reasonable meaning unless they are used in a context that requires a different interpretation. In addition, the term “and / or” is to be construed as an inclusive OR. Therefore, for example, the phrase “A, B, and / or C” is to be interpreted as covering all the following: “A”; “B”; “C”; “A and B”; “A and C”: “B and C”: and “A, B, and C”. Any numerical range disclosed herein encompasses the and lower limits and each intervening value, unless otherwise specified. Other than in working examples, or where otherwise indicated, numerical values (such as numbers expressing quantities of ingredients, reaction conditions) as used in the specification and claims are modified by the term "about". Accordingly, unless indicated to the contrary7, such numbers are approximations that may vary7depending upon the desired properties sought to be obtained. At the very least, and not as an attempt to limit the application of the doctrine of equivalents to the scope of the claims, each numerical parameter should be construed in light of the number of significant digits and ordinary rounding techniques.

[0050] While the numerical parameters setting forth the scope of the disclosed subject matter are approximations, the numerical values set forth in the working examples are reported as precisely as possible. Any numerical value, however, inherently contains certain errors necessarily resulting from the standard deviation found in its respective testing measurements.

[0051] Unless defined otherwise, the meanings of technical and scientific terms as used herein are those commonly understood by one of ordinary skill in the art to which the disclosed subject matter belongs.

[0052] Description

[0053] Described herein are methylation arrays and methods of diagnosis and treatment that can be used for a subject having cancer. The discussion herein is focused on meningiomas, but the teachings may be applicable to other cancer types, in particular for pituitary' neuroendocrine tumors and gliomas. Particular methylation arrays have been developed to particularly target these various tumor types, and accordingly, uses particular CpG probes that correspond to CpG sites. The methods and arrays described herein were developed to allow for more accurate serum-based liquid biopsy screening, which can present unique challenges. While less invasive, serum-based liquid biopsy screening may be less accurate than tissue-based analysis. The methods and arrays described herein can provide for an improved treatment regime as compared with other diagnostic and treatment methods, particularly, those for meningiomas, pituitary neuroendocrine tumors, and gliomas.

[0054] DNA methylation-based signatures were surveyed and identified in serum which allowed for the development of machine learning classifiers able to accurately distinguish meningioma from controls and other CNS entities and predict recurrence risk, which may also be applied across tissue specimens. These findings lay the foundation for the implementation of a presurgical detection of meningioma and assessment of its recurrence risk prediction (and possibly progression surveillance) using a noninvasive approach such as a blood draw, ultimately impacting the management and outcomes of patients harboring these tumors.

[0055] Glioblastoma (GBM). in particular, is the most common and aggressive malignant brain tumor in adults, characterized by rapid progression, poor prognosis, and considerable intratumoral heterogeneity. Despite advances in molecular classification and imaging, the current standard-of-care for monitoring GBM progression relies heavily on radiologic interpretation of serial MRI scans. This method is often subjective and prone to misidentifying treatment-related changes as true progression, delaying critical therapeutic decisions. To address this unmet need, new blood-based biomarkers are described herein, for example the glioma epigenetic liquid biopsy (GeLB) score, which leverage cfDNA methylation profding to detect GBM recurrence months before radiologic evidence becomes apparent. These epigenetic signals can be integrated wi th advanced imaging and Al tools to enable accurate, non- invasive, and timely assessment of tumor progression. GeLB elevations can also be correlated with brain needle biopsy results to establish definitive ground truth, addressing a key translational barrier. Additionally, an Al-driven MRI surrogate for the score can be developed, allowing broader clinical adoption in settings without access to methylation profiling. An Al pathology' model can also be used to predict progression-free survival using histological data, empowering pathologists and clinicians with early prognostic insights. To ensure scalability and generalizability, the models can be deployed through a federated learning (FL) network already established across five continents and release all code, models, and data through the NIH’s Cancer Research Data Commons (CRDC), in partnership with the NCI’s Center for Biomedical Informatics and Information Technology’ (CBIIT). This can generate unprecedented multi-modal GBM data, including paired GeLB. MRI, pathology, molecular, and clinical information. The scores can be used as a clinically actionable tool, which can enable broader access through Al-powered imaging surrogates, and set the stage for precision treatment management and adjustment of treatment based on the results. The results can be used to influence surveillance protocols, inform treatment adaptation, support clinical trial stratification, and ultimately improve outcomes for patients facing devastating CNS cancers. In some implementations, the blood-based biomarker or score (e.g., GeLB score) and complementary Al tools using MRI and pathology can enable earlier, more accurate detection of tumor recurrence. By improving surveillance and personalizing care, at least partially through the work of integrating imaging with genomics, this w ork can significantly enhance outcomes for patients and reduce the societal burden of these conditions. Preliminary' findings suggest that the approaches herein can detect tumor response or recurrence earlier than traditional imaging features alone, which can accordingly inform treatment strategy. The arrays and methods herein can also be used as a complementary tool to support imaging analysis, radiotherapy planning, and diagnostic workflows.

[0056] Results

[0057] Meningioma cohort features — Demographic and clinicopathological features of patients with meningiomas (MNG) and other CNS entities (non-MNG) were treated at Henry Ford Health (HFH) and the University of Sao Paulo (USP). Additionally, cohorts were retrieved from literature.

[0058] Methylation data features across liquid biopsy specimens — The preprocessing and quality assessment of the methylation arrays described herein showed that all liquid biopsy samples, excluding one, met expected quality control standards. No batch effects related to sample collection or extraction dates were observed across the liquid biopsy specimens’ methylomes.

[0059] Serum circulating cfDNA concentration (ng / pL) from patients with meningiomas were significantly lower than gliomas (Wilcoxon rank sum test; p < 0.001) and pituitary tumors (Wilcoxon rank sum test; p < 0.01). No significant differences in cfDNA concentration (ng / pL) across meningioma WHO grades, or recurrence risk predictions were observed. The serumk3 cluster, identified through an unsupervised approach further described in Methods, presented the lowest concentration of serum cfDNA compared to other k-means clusters (k2 and k4).

[0060] Paired serum- and plasma or tissue presented similar DNA methylation profiles — The comparison between paired serum and DNA plasma methylomes (n = 10 pairs) demonstrated that genome-wide DNA methylation levels and estimated immune cell profiles were highly correlated (Pearson’s p = 0.89-0.96). The diagnostic and prognostic classifications results from both serum and plasma were mostly concordant (80 and 70%, respectively). Across the comparison between paired serum and tissue methylomes (n = 25), there was a significant and positive correlation in relation to genome-wide DNA methylation levels (Pearson’s p = 0.694-0.907). The correlation across immune proportions between both sources was positive but nonsignificant (Pearson’s p=0. 132-0.695). Serum cfDNA methylation levels distinguish meningioma from other CNS entities — It was observed that genome-wide cfDNA methylation levels in serum only partly distinguished MNG from non-MNG conditions as depicted in the Principal Component Analysis (PCA) (FIG. 1). However, through supervised methods, we identified 98 meningioma-specific differentially methylated probes (DMPs; 0.15<diffmean < -0.175) which significantly separated both groups (Wilcoxon rank sum test: p value FDR < 0.05). Notably, the mean DNA methylation levels in MNGs serum were significantly lower compared to controls and non-MNG samples, such as gliomas (Wilcoxon rank sum test; p < 0.001) and pituitary' tumors (Wilcoxon rank sum test; p < 0.05) (FIG. 2).

[0061] To investigate whether similar supervised methods would allow for translatability between tumor tissue and liquid biopsy methylomes, MNG and non- MNG tissue collections were compared and a subset of meningioma-specific probes were identified (n = 221 DMPs; |diffmean | >0.55; pFDR < 0.001) from which some signatures were detectable across the serum methylome and also distinguished MNG and non-MNG across serum specimens (n = 24 DMPs; Wilcoxon rank sum test: pFDR < le-04) (FIGS. 3 and 4).

[0062] The diagnostic-Meningioma Epigenetic Liquid Biopsy classifier (d-MeLB methylation array) accurately classifies samples independently of the specimen source — 256,447 tumor-specific CpG probes were identified, termed meningiomaspecific DMPs, which exhibited significant differential methylation between MNG tumor tissue (n = 31) and publicly available nontumor control collections (epileptic brain; n =21) and were utilized as input into diagnostic classifier construction. Within the algorithm, these DMPs were further filtered to those with high DNA methylation level similarities between paired meningioma serum and tissue specimens collected at the time of surgery’ (origin: Henry Ford Health), namely similarly methylated probes (SMPs: n = 7659). T-distributed stochastic neighbor embedding (t-SNE) was applied to visualize the behavior of these signatures across internal and external tissue cohorts of meningiomas and other central nervous system (CNS) entities. Interestingly, it was observed that the SMPs clustered meningioma tissue samples together with serum and plasma specimens, and effectively distinguished meningiomas from other CNS entities (FIG. 5). By filtering these SMPs through a serum-based supervised analysis (untreated MNG vs non-MNG cohorts; Wilcoxon rank-sum test), a signature methylation array was derived that is particularly applicable to liquid biopsy samples.

[0063] The application of the d-MeLB classifier across the model selection serum cohort showed that a score threshold >0.48 had the highest classification accuracy (AUC: 1.00). Validation of the diagnostic classifier was conducted across an independent cohort of MNG and non-MNG including the samples randomized (n=30) and excluded (n = 63) during model construction, namely the independent ‘original’ serum cohort (n = 93). Across this cohort, an 84.9% accuracy was observed in identification of meningioma / non-meningioma, with satisfactory performance measures, i.e., Matthew correlation coefficient (MCC: 0.56) and clinical utility index (CUI + : 0.405) (FIG. 6).

[0064] Two rounds of additional MNG collections (validation i & ii) were profiled and incorporated as validation independent cohorts after the initial model derivation (serum: n = 19; plasma: n = 10) in which d-MeLB displayed classification accuracies of 68.4% and 70.0%, respectively (FIG. 7). For comparison, performance across the entire liquid biopsy independent cohort (n = 122) was considered.

[0065] In an effort to investigate whether the developed diagnostic MeLB methylation array presented a spurious immune-related bias due to potential contamination of serum with immune-cell signature released by white blood cells during the clotting process, the generated classifier was applied across an independent cohort of fluorescence activated cell sorting (FACS) purified immune cell and whole blood profiles (n =59). The d-MeLB had an overall accuracy of 93.2% to classify these samples as nonmeningiomas, including neutrophils and whole blood.

[0066] Of note, formulation of the d-MeLB classifier was not conducted with tissue classification in mind and did not include tissue specimens within discovery' or independent validation sets; so, it was expected that tissue application would be limited (ACC: <10%). To address this limitation, d-MeLB signatures were used as coefficients for a simple linear-based discriminant algorithm to classify tissue-based collections composed of meningioma and non-meningioma. Summarily, we observed an accuracy of 94.3% to classify an independent cohort into their correct memberships (n = 176). Confirmation of the detection of our diagnostic signatures (d-MeLB: n = 25 CpGs) across 10 cfDNA samples profiled through whole genome bisulfite sequencing (WGBS) was conducted. Their methylation levels determined by P-values (EPIC Array), or percentage values (WGBS) were significantly correlated across these samples (Pearson’s p=0.6, p < 2.2e-16).

[0067] The d-MeLB classifier outperforms other classifier approaches — We compared the performance of the random forest (RF) approach used to develop the d-MeLB with other classification methods using our internal methylome cohort data (n = 239), with identical discovery (n = 117 MNG & Non-MNG) and validation (N = 122; randomized=30; excluded=63; additional serum=19; additional plasma=10) cohorts as those used throughout the d-MeLB. Compared to the results obtained with the d-MeLB classifier based on random forest algorithm, other approaches including dimension reduction RF, linear discriminant analysis (LDA), extreme gradient boosting [package: XGBoost vl.7.4] and logistic regressions (univariate and multivariate analyses of mean methylation values) presented lower accuracies in classification of independent liquid biopsy MNG and non-MNG.

[0068] Serum cfDNA methylation clusters are associated with distinct clinicopathological features, outcomes, and immune composition across meningioma specimens — Unsupervised consensus clustering analysis revealed four main k-clusters with distinct cfDNA methylation profiles across MNG serum specimens (FIG. 8). The annotation of these serum-derived molecular groups with clinicopathological and molecular features showed that the clustering occurred independently of sex, age, and race and were enriched with features associated with MNG outcomes and prognosis. For instance, compared to kl-and k2-cl usters, k3- and k4-clusters presented an enrichment of WHO grades 2 and 3, and confirmed recurrence during patient follow up (FIG. 8).

[0069] Through cfDNA methylation-based deconvolution analysis, we discovered that the k4 cluster was enriched with neutrophil cell signatures and possessed the highest neutrophil-lymphocyte ratio (NLR) and depleted of the majority of immune cell types included in the analysis. In contrast, the kl -cluster was depleted in neutrophils while enriched with almost all immune cell type proportion estimates (B- and T-cells, natural killer (NK) and monocytes) compared to other clusters. The prognostic-Meningioma Epigenetic Liquid Biopsy classifier (p-MeLB methylation array) predicts risk of recurrence (RR) of meningiomas using serum or tissue specimens — The p-MeLB classifier presented an overall 87.7% accuracy and satisfactory performance measurements (CUI + : 86.4%; MCC = 0.577) in predicting true recurrence in an independent validation tissue- and liquid biopsy-based cohort, as confirmed during established follow-up (original cohort: ACC = 82.6%; additional validations: ACC = 90%; FIG. 9).

[0070] The application of the p-MeLB classifier to primary' meningioma tissue collections (n = 69), derived from all three separate profilings (original, validation i & ii), demonstrated significant agreement with the classifications obtained from a previously published nomogram (Cohen’s unweighted kappa; K = 0.269, pK=0.01). Interestingly, across a subset of primary MNG independent from p-MeLB classifier derivation, p-MeLB demonstrated higher sensitivity (SE) to predict true recurrence confirmed during follow-up, compared to the nomogram-derived results (<5 yrs: SE = 76.5 vs. 47. 1%; >5 yrs: SE=88.9 vs. 55.6%) (FIG. 10).

[0071] Across our total cohorts of MNG-tissue (n = 123) and serum (n = 80) specimens which possessed attributed person-time (mean: 3.7 and 2.7 years, respectively), the 5- year recurrence-free survival probability was significantly lower in MNG classified as having a high risk than those classified as low risk of recurrence (tissue: 20% vs 73%; serum: 35% vs 75%; log-rank p < 1.0e-4) (FIG. 11).

[0072] The detection of our prognostic methylation signatures (p-MeLB: n = 13 CpGs, 70 high-risk related DMPs) was also confirmed across ten cfDNA samples profiled through whole genome bisulfite sequencing (WGBS). Similar to the observed for d- MeLB signatures, the DNA methylation levels of the p-MeLB signatures as determined by [3-values (EPIC array) or percentage values (WGBS) were highly correlated across these samples (p-MeLB: Pearson’s p=0.73, p=3.1 e-16; risk-related DMPs: Pearson’s p=0.68, p = 2.2e-16). These findings indicate that these EPIC-based results are further supported by WGBS, which serves as a secondary benchmark profiling method.

[0073] High and low recurrence risk meningioma groups present differential clinicopathological features and estimated immune landscapes in serum and tissue specimens — To further characterize predicted risk groups, differences in the distribution of relevant clinical features associated with prognosis were estimated between samples classified as high and low risk for recurrence (FIGS. 12 and 13). Summarily, in serum specimens, a significantly higher odds ratio (OR) of a confirmed recurrence was observed during follow-up occurring in high risk compared to low-risk specimens (OR = 15.45, 95% CI: [1.45, 844.77]; p < 0.05). No significant differences were observed in relation to MNG location (skull base / non-skull base), extent of resection (gross total resection [GTR] / subtotal resection [STR]). WHO grades (2&3 / 1). progression in post-surgical MRI reports (progressive / non-enhanced and stable disease) or vital status (deceased / alive), among others (FIG. 12). Additionally, in high-risk specimens we observed significant enrichment in the estimated proportions of neutrophils and depletion of B-cells (p = 0.002). NK (p=4.00E-04) and CD4-T cells (p = 0.07) and high NLR compared to their low-risk counterparts (FIG. 13).

[0074] In tissue specimens, similar to serum findings, significantly higher odds of a confirmed recurrence during follow-up were observed in high- compared to low-risk specimens; no differences regarding tumor location or grade (OR = 32.2, 95% CI: [5.71, 351.06]; p < 0.05); estimated NLR (p = 0.08), neutrophils (p=0.07), and NK (p = 0.1) proportions. In contrast to serum findings, compared to low risk, high risk samples presented decreased odds of having a gross total resection (OR = 0.15, 95% CI: [0.01, 0.76]; p < 0.05) and higher odds of progressive disease postsurgical MRI reports (OR = 1 1.91, 95% CI: [2.63, 77.84]; p < 0.05). p-MeLB classifications are validated across external meningioma tissue cohorts — In order to further assess the robustness of the p-MeLB predictor disclosed herein, the performance of p-MeLB was compared with other tissue methylome based prognostic classifiers using a common external meningioma tissue cohort. The p-MeLB recurrence risk predictions across external meningioma tissue-methylome cohorts aligned with prognostic and survival differences reported previously. For instance, within the Choudhury hypermitotic group characterized by the poorest 5-year recurrence-free survival probability (-35%), a majority of samples was classified as high risk for recurrence by p-MeLB; while the Merlin-intact group with the most favorable 5-year recurrence-free survival (~85%)was largely classified as low risk (FIG. 14). The Bayley malignant MenG-C group with the poorest recurrence-free survival probability compared to their more benign counterparts (MenG-A and -B groups), was unanimously classified as high recurrence risk by p-MeLB (FIG. 14). Overall, the assessment of p-MeLB’s ability to detect true recurrence during follow-up across these external cohorts was limited due to the lack of sufficient longitudinal information.

[0075] Tissue-derived differentially methylated probes are detectable and moderately differentiate recurrence risk groups in serum specimens — A subset of tissue-derived prognostically -relevant DMPs was identified from the companson between confirmed recurrence (CR) and confirmed non-recurrence (CNR) specimens (pFDR < 0.001 & |diff.mean | >0.55; n = 260 DMPs) which presented congruent DNA methylation levels with serum and distinguished a majority of high and low risk in our original (n = 63) and additional (validation i & ii; n = 18) independent serum cohorts (pFDR < 0.001 & |diffmean | > 0.27; n= 39 DMPs), with minor intermingling of risk-groups (FIGS. 16 and 17).

[0076] Prognostic probes located in gene regulatory elements potentially control the expression of target genes associated with tumor development and growth (in silico functional analysis) — Through the intersection between Choudhury and p-MeLB classifications, two prognostic groups in tissue specimens were identified and compared, i.e., high-risk hypermitotic vs low-risk merlin MNG to identify prognosticspecific DMPs. By performing an integrative analysis of paired methylome and transcriptome meningioma tissue data, prognostic DMPs were identified in regulatory regions which were differentially methylated and targeted genes which were differentially expressed between these prognostic groups (probe-gene pairs [PGPs]). For downstream analysis, to identify putative epigenetically regulated genes, PGPs were selected which possessed a negative correlation between DNA methylation and expression levels (n=65 PGPs; FIG. 18). Among these PGPs identified in tissue, twelve CpGs presented concordant DNA methylation between serum and tissue specimens and also differentiated recurrence risk groups across serum specimens (e.g., hypermethylated in both tissue and serum in high-risk specimens) (FIG. 19).

[0077] Additionally, across serum specimens, 70 risk related DMPs were identified through the supervised analysis between high and low recurrence risk meningioma. Mapping these serum derived DMPs to tissue sample methylomes and to their putative target genes in the Choudhury dataset, PGPs were selected which exhibited negative correlation between gene expression and CpG probe DNA methylation levels across risk group comparisons (n = 25 PGPs). CpG probes with concordant DNA methylation levels between tissue-serum PGP are highlighted (FIG. 20).

[0078] Finally, the potential biological functions and diseases associated with these PGPs were explored through gene set enrichment analyses. It was identified that tissue- or serum-derived prognostically relevant PGPs were related to tumorigenesis processes (n=38 genes), specifically related to meningioma (n=5 genes), cell growth / proliferation / movement (n=26 genes), cell cycle (n=7 genes), and immune response (n=21 genes), amongst others, see Table 1 below:

[0079] Table 1: Gene set enrichment analysis results using Ingenuity Pathway Analysis (IP A)

[0080] Identified prognostic DMPs exhibited overall DNA hypermethylation in CR or high-risk samples compared to NCR or low-risk samples. Strong DNA hypermethylation was also observed, particularly in the regulatory regions of gene promoters associated with Polycomb repressive complexes (PRC). This DNA hypermethylation was detected across liquid biopsy specimens (serum and plasma), as well as tumor tissue specimens.

[0081] Discussion of Results

[0082] Genome-wide DNA methylation assessment provides an objective, robust and unbiased approach to define discrete molecular groups of CNS tumors. This approach overcomes the limitations and subjective biases associated with histopathological and grading classification approaches. Detection of distinct DNA methylome patterns is reproducible and stable within and across diseases and allows for the fine-tuning of molecular subtyping associated with distinct recurrence and growth-prone behaviors in many tumors. Capitalizing on this knowledge, several reports have shown that specific DNA methylation signatures identified in tumor tissue specimens are amenable to the development of machine learning classifiers able to accurately diagnose and prognosticate several tumor types and subtypes, including meningiomas. However, we observed that these previous classifiers were not able to classify liquid biopsy samples into their diagnostic or prognostic memberships, possibly due to their formulation being solely based on tissue-derived methylomes. To circumvent this limitation, herein, we developed machine learning classifiers using meningioma-specific DNA methylation markers suitable to diagnose and prognosticate these tumors using either liquid biopsy or tissue specimens (FIGS. 6, 7, and 9). Confirming their tumor-of-origin specificity, these detected markers clustered together liquid biopsy (serum and plasma) and tissue specimens from patients with meningioma, while simultaneously distinguishing these tumors from other CNS entities, when applied to external and independent tumor tissue cohorts (FIG. 5). The d-MeLB signatures generated during the diagnostic model development presented an overall accuracy of -85% to classify serum samples according to meningioma or nonmeningioma memberships (FIG. 6). These signatures were also able to correctly classify tissue samples with -94% accuracy using additional linear machine-learning methods. Altogether, the current findings corroborate our previous reports showing the viability to use LB-oriented classifiers for diagnosing CNS tumors. Notably, this classifier allowed for the accurate identification of recurrent meningioma using serum samples (FIGS. 6 and 7), which can be useful as a standalone or complementary noninvasive tool along with imaging to monitor these tumors.

[0083] Through p-MeLB classifier development, we identified risk specific DNA methylation markers in serum useful for the stratification of serum or tissue specimens according to their recurrence risk, with an observed total accuracy of 87.7% across independent cohorts (FIG. 9). In addition to the noninvasive application of p-MeLB, its accuracy is comparable to or even surpasses other individual benchmark methods evaluated in surgical specimens, such as Ki-67 / MIBl immunoexpression (AUC: 87.7%), transcriptome-base markers (AUC: 0.81) or composite scores involving multiple risk factors (AUC: 0.849) with or without consideration for imaging features (AUC: 0.75-0.78).

[0084] The observed agreement between the prognostic classification using p-MeLB or existing classifiers, further reinforced the validity and robustness of p-MeLB in assessing the likelihood of recurrence. However, compared to the Nassiri nomogram, the p-MeLB model excelled in predicting the risk to recur in an independent meningioma tissue cohort (i.e., p-MeLB vs nomogram accuracies - within 5 years: 76.5 vs. 48%; after 5 years: 88.9 vs. 56.6%) (FIGS. 9 and 14). Additionally, external samples predicted as high risk through p-MeLB classification exhibited significantly poorer overall survival, with an approximate 20% probability without a tumor recurrence after 5 years’ time. These observed survival trends are concordant with recurrence-free survival rates reported for subtypes with higher propensities for progression by other authors (e.g., Bayley’s MNG-C: -45%; Choudhury’s Hypermitotic:~35% and Nassiri high risk / grade 3-25%) (FIG. 10).

[0085] Additionally, considering the differing sensitivities of Choudhury’s Hypermitotic and Merlin-intact subty pes to cytotoxic agents in preclinical studies (with decreased and increased vulnerabilities, respectively), and the consistent alignment of p-MeLB classifications with high and low risk for recurrence in these subty pes, p- MeLB has the potential to guide experimental therapeutic decisions.

[0086] Notably, the p-MeLB classifier presents some unique advantages compared to these existing models: (1) in contrast to the cross-sectional information utilized as input in existing models, p-MeLB signatures disclosed herein have been derived from longitudinal data. This was accomplished by comparing cases with confirmed recurrence or no recurrence over a minimum 5-year period of clinical and radiographic surveillance follow-up; (2) p-MeLB requires solely DNA methylation data as input, in contrast to other tissue-based meningioma classifiers which rely on the integration of methylomic data with clinicopathological features prone to subjectivity (e.g., extent of resection) or multiomic profiling, which may be financially detrimental for a potential clinical application; and (3) it is able to accurately predict outcomes when applied across different specimen sources (tissue, serum, and potentially plasma).

[0087] Overall, these findings suggest that the application of d-MeLB and / or p-MeLB classifiers can be a valuable noninvasive approach for diagnosing and distinguishing meningiomas from other mimicking diseases in preoperative assessments and possible monitoring tumor progression and treatment response through a blood draw (FIG. 15). Additionally, they may complement traditional and advanced imaging approaches, such as radiomics, to provide a more comprehensive and accurate evaluation of meningioma status.

[0088] Through the integration of paired methylome and transcriptome data derived from meningioma tissue generated by Choudhury et al., genes whose expressions are possibly regulated by epigenetic control were identified (FIG. 18). Interestingly, many of the CpG probes associated with these genes in meningioma tissue were also detected in serum specimens (FIG. 19). Among these gene sets, yve found enrichments for biologically relevant terms such as CNS tumor development, cellular growth / proliferation and movement, and prognosis (Table 1). Distinct hypermethylation was also detected within regulatory regions of gene promoters associated with PRC across high-risk samples. DNA hypermethylation in promoter regions of this complex, as observed across multiple sample sources (serum, plasma and tissue), has been previously linked to malignancy in meningiomas. Altogether these results indicate that the identified methylation arrays could be mechanistically involved in the recurrence risk of these tumors and could be used as prognostic markers detectable in serum specimens.

[0089] Most reported LB-oriented studies have used plasma instead of serum as a source of cfDNA to perform omics analysis. Herein, serum was mainly profiled, the sole blood component available at the time, but other cfDNA sources are certainly possible. Although certain molecular results (e.g., detection of somatic mutations) could be impacted by the use of serum profiling due to potential dilution or contamination with genomic DNA derived from blood and other cells during the coagulation process, it does not seem to interfere with the detection of cell-specific cfDNA methylation markers as shown herein. Additionally, even after considering potential dilution of tumor derived cfDNA in serum specimens, DNA methylation array platform (EPIC) or whole-genome sequencing are sensitive approaches to detect abnormalities in minute amounts of intact or fragmented cfDNA in liquid biopsy specimens (e.g., <1 ng). Notably, among the detected risk-specific probes disclosed herein, whether derived through serum- or tissue-based analyses, at least some of the targeted genes are implicated in immune response pathways. In serum, this enrichment could arguably reflect contamination with cfDNA from lysed white blood cells potentially introducing an immune bias into our signatures (Table 1). However, several lines of evidence were gathered to support the hypothesis that the immune-related findings discussed herein are genuine and associated with the presence of meningiomas, as well as pituitary endocrine tumors and gliomas. To formally address concerns about genomic DNA contamination in serum, the methylomes of paired serum and plasma samples were profiled and compared from an additional meningioma cohort. There was a high correlation in the genome-wide DNA methylation levels, estimated immune cell proportions, and diagnostic and prognostic classifications between the two blood elements in most samples. Additionally, meningioma-specific CpG probes detected in plasma clustered together with their matching serum and tissue counterparts, confirming the specificity of these CpG probes regardless of the specimen source (FIG. 5). Furthermore, a significant difference was observed in several estimated proportions of immune cells in comparison of whole blood and meningioma liquid biopsy (LB) serum samples.

[0090] Additionally, our d-MeLB model accurately classified whole blood and purified immune cells samples as non-meningiomas. indicating that the d-MeLB signatures described herein are not biased towards spurious immune enrichment. It was observed that samples predicted to have a high risk of recurrence exhibited higher neutrophil-to- lymphocyte ratios (NLR), increased neutrophil levels, and reduced proportions of B- cells and natural killer cells (FIG. 13). These alterations have been associated with poor prognosis in other tumors. It was also found that the immune compositions between serum and matching tissue were not significantly correlated, suggesting that the systemic immune or inflammatory response to the presence of meningiomas is distinct from the local immune response in the tumor microenvironment.

[0091] Altogether, these immune-related findings in serum specimens appear to be authentic markers of a systemic inflammatory response to the presence of meningiomas with varying recurrence risk, rather than a spillover of the local immune response or contamination with DNA from white blood cells (FIG. 13). While confirmation with a gold standard approach such as flow cytometry may also be desirable, the results herein suggest that this DNA methylation-based deconvolution approach could offer additional insight into proposed prognostic classifications. It has the potential to stratify patients with meningiomas based on their immune landscape and guide future immunotherapy strategies. Consequently, the treatment and / or monitoring status of the patient can be adjusted accordingly, as shown for example in FIG. 15.

[0092] A user-friendly platform containing the diagnostic and prognostic classifiers may be employed, similar to available tissue-based web tools, but instead specifically developed for serum and preferred embodiments using one or more of the methylation arrays taught and described with respect to Tables 2-4. The work from publications (l)-(3) disclosed in the Background has been refined, aggregating more serum and tissue methylome data along with clinicopathological information, to help optimize the models for potential clinical application. In summary', it was showed that blood-based specimens, specifically serum, are amenable for the detection of tumor-specific DNA methylation signatures. The identified signatures in the methylation arrays described herein not only enabled differentiation between meningiomas and other intracranial entities but also showed accuracy in identifying meningiomas with distinct recurrence risks. Similar results were uncovered with pituitary neuroendocrine tumors and gliomas. Potentially, the signatures in these arrays may serve as a valuable surveillance tool for detecting CNS tumor recurrence during follow-up. The successful clinical implementation of these DNA methylation-based classifiers will refine cancer recurrence risk stratification at the time of diagnosis and possibly during follow-up, ultimately impacting management and outcomes of these patients. The machine learning classification approach disclosed herein, based on methylome analysis, has the potential to be extended for the diagnosis and prognostication of a broader spectrum of tumors using liquid biopsy-derived specimens. Methods

[0093] Archival serum was collected from 204 patients who underwent resection of meningiomas (MNG group) and other CNS entities and controls (non-neoplastic diseases) at the Neurological Surgery Department at Henry Ford Health from 06 / 2011 through 08 / 2019, namely ‘original’ cohort. Meningioma tissue methylomes generated internally at Henry Ford Health were also retrieved and analyzed (n =31), and provided by the Department of Neurosurgery of the Universify of Sao Paulo (n = 72), or from publicly available repositories (n=900). Longitudinal follow-up information was available for 50 tissue (Henry Ford Health and University of Sao Paulo) and 25 liquid biopsy serum collections (Henry Ford Health). Serum specimens collected at recurrence were available for 19 meningioma collections (two paired with serum collected at first / initial surgery).

[0094] Besides these cohorts, an “additional” MNG cohort (namely, validation i & ii) consisting of 69 archival tissue and blood derived liquid biopsy collections obtained between 12 / 1999 and 07 / 2021 at Henry’ Ford Health were collected: 8 paired tumor tissue and liquid biopsy (serum and plasma) pairs, 9 pairs of tissue and serum, 2 paired serum and plasma, and 23 individual tissue collections. Longitudinal followup data attributed to this cohort was available for a subset of samples (tissue, n = 27; liquid biopsy serum, n = 8; liquid biopsy plasma, n=4). Simultaneous collections of tissue and serum samples were available for 25 MNG patients. Additional collections of non- neoplastic diseases (n=6) were also profiled for expansion of the control arm of this study. Information characterizing internal and external cohorts was obtained.

[0095] Congruent to definitions reported by others, meningioma recurrence was defined as tumor growth / progression or additional surgery following gross or subtotal resection through review of the immediate postoperative imaging and / or information found in medical records during follow-up (person-time), across initial or recurrent collections (namely Confirmed Recurrence (CR) group, Henry Ford Health and University of Sao Paulo tissue and liquid biopsy serum cohorts; n = 84). Non-recurrent MNG was defined as the absence of growth / progression in any post-surgical MRI or medical reports or absence of further tumor resection across a minimum attributed follow-up of 5-years (namely Confirmed No Recurrence [CNR]; Henry Ford Health and University of Sao Paulo tissue: n= 17; Henn’ Ford Health liquid biopsy: n=9; Table 1). In order to ensure precise categorization, serum samples or publicly available tissue samples without available follow-up information were labeled as either high risk or low risk for recurrence through the prognostic classifier (p-MeLB). We also retrieved publicly available methylomes, some paired with transcriptome data (RNA-sequencing or microarray) from meningiomas. To perform correlative analyses between our prognostic (p-MeLB) classifier with others, we excluded samples from the Bayley cohort which were not correctly classified by their algorithm (n = 15), at the request from the authors. Updated clinical and follow-up information was kindly provided by the Bayley group. Within the Choudhury cohort, we excluded samples which were not fully annotated across provided clinical data (n = 60). Prognostic classification of our internally generated samples through Nassiri et al. nomogram w as kindly performed by the Nassiri group.

[0096] DNA isolation, quantification, quality control, DNA methylome data generation and preparation — Extracted DNA from meningioma-derived specimens (serum, plasma and tissue) were bisulfite converted using the Zymo EZ DNA methylation kit as specified by the manufacturer (Zymo Research, Irvine, CA, USA) and profiled using an Illumina Human EPIC array (EPIC) at the USC Norris Molecular Genomics Core Facility. Prior to profiling, the isolated DNA was restored using a restoration kit provided by Illumina. This allow ed restoration of fragmented DNA and concentration of the low yield. Bisulfite converted DNA samples were recovered in a 1 Opl volume, and 1 pl was used to evaluate bisulfite conversion completeness and recovery’. DNA methylation was profiled using the Illumina Human 850k (EPIC) and a matching subset using whole genome bisulfite sequencing (WGBS). Publicly available tissue methylome was profiled using 450k (HM450k; tissue) or 850k (EPIC) arrays. Data quality of the methylome data was assessed with the R-based graphical user interface shinyMethyl v3.16.

[0097] DNA methylation preprocessing — DNA methylation array (EPIC) data were preprocessed (e.g. removal of masked probes and SNP) using the minfi package. Before downstream analysis, liquid biopsy methylome data was examined for potential batch effects regarding plate number and extraction dates using commonplace methodologies (ComBat v3.20.0- sva). Tissue-based MNG methylome data was corrected for batch effect by institution prior to t-distributed stochastic neighbor embedding (t-SNE) visualizations. For prognostic classifier derivation, only probes common between 450 K and 850 K arrays were selected to maintain retrospective cohort applicability (n = ~339k CpGs). For technical validation, the methylation levels of a subset of meningioma LB samples were profiled using whole genome bisulfite sequence (WGBS).

[0098] DNA methylation exploratory’ analysis / unsupervised and supervised analyses — methylome patterns were explored across all serum samples (CNS tumor types and nonneoplastic controls) using standard unsupervised approaches and visualizations. To identify MNG molecular groups, hierarchical consensus k-means clustering was applied to the most variant methylated probes (n = 1000) across MNG serum specimens and the optimal number of clusters was selected based on statistical parameters such as the Cumulative Distribution Function (CDF) and the Calinski- Harabasz curve.

[0099] Supervised analyses — In order to identify Methylation Epigenetic Liquid Biopsy (MeLB) probes (serum-derived) or tissue-derived MNG specific differentially methylated probes (DMPs), supervised comparisons were performed between MNG- and non-MNG group methyl omes (serum: 44 primary and 19 recurrent MNG, 141 non- MNG; tissue: 326 MNG; 367 non-MNG). To identify prognostically relevant probes, a comparison was conducted between predicted high and low risk for recurrence groups (31 high risk and 32 low risk MNG) in the serum or confirmed recurrence (CR: n =35) and non-recurrence (CNR: n = 15) groups in the tissue.

[0100] To reduce the potential for capturing background noise, volcano plot visualization techniques were used to guide selection of the DMPs which presented significant FDR adjusted p-values and mean methylation differences across comparisons within a variable range reported in the literature (differential mean DNA methylation >10-15%; Wilcoxon rank sum tests; p-value FDR < 0.05).

[0101] Each DMP was mapped to their CpG genomic location previously defined as CpG islands (CGI), shores, shelves, and open sea regions and to their putative target gene using EPICmanifest (hg38). Enhancer elements were defined using the GeneHancer database (hg38) provided by the UCSC Genome Browser. Promoter elements were defined using GENCODE v.31 annotations, with consideration of CpGs located 200 bp up / downstream from the target gene.

[0102] Methylome-based predictions / Random Forest machine learning approach — To investigate the potential diagnostic and prognostic applications of MeLB, the random forest machine-learning (ML) approach was used to generate a binary classifier to differentiate MNG vs non-MNG specimens and low vs high recurrence risk groups using serum cfDNA- and / or tissue-derived DNA methylation-based signatures.

[0103] Cohorts were randomized through machine-driven processes into sets encompassing representative and proportional samples of each comparison group: 1) the discover^' set used to construct the classifier, further subdivided into la) a training set for the identification of relevant signature sets and algorithm training and lb) a model selection set, used to evaluate the resulting classifier’s performance, or 2) an independent validation set. not involved in any of the previous development steps, used to validate the finalized classifier. Following the completion of classifier formulation and independent validation, an additional set of MNG & non-MNG samples was included to further validate the generalization of the diagnostic classifier (additional validation set).

[0104] Supervised feature extraction processes were automated within machine learning construction and specialized for each classification task, as detailed below, in efforts to isolate diagnostic or prognostically relevant methylation arrays, and reduce the potential for the training classifiers using noisy signals.

[0105] The performances of both diagnostic and prognostic prediction models were assessed using Matthew's correlation coefficient (MCC), which measures the quality of a binary classification by the agreement between predicted and actual (observed) values (ranging from+1 to -1, i.e. perfect agreement [perfect prediction] to total disagreement [poor prediction] and the Clinical Utility Index (CUI), interpreted as excellent, good, satisfactory or poor when values are >0.81, >0.64, >0.49 or <0.49, respectively. This index can be primarily used to express the relative benefit of using the classifiers disclosed herein, compared to use of an optimal test, when making clinical decisions (CUI positive [+], and CUI negative [-]).

[0106] Meningioma diagnostic classification / Diagnostic MeLB (d-MeLB) — In diagnostic classifier construction, following exclusionary measures (recurrent gliomas, inflammatory non-neoplastic diseases), serum samples were randomly assigned to discovery’ (n = 117) or independent validation cohort sets (n = 30), both encompassing analogous proportions of meningioma (initial and recurrent) and non-meningioma serum specimens. The discovery cohort was further randomly partitioned into training (80%) and model selection sets (20%). To instill inherent MNG-specificity to the identified signature, dimensionality reduction of the entire genome w as performed through two distinct methods: 1) conducting genome-wide supervised analysis between internally-profiled MNG tissue specimens (n=31) and publicly available nontumor control specimens (n=21), selecting CpG probes which exhibit differential methylation between the two groups according to a randomly selected significance measure, namely tumor specific DMPs (n= -256 k; Wilcoxon rank sum test p-value range: le-4-0.05); and 2) among the tumor-specific DMPs, selecting those CpG probes which possessed the greatest DNA methylation level similarities between matching training set serum and tissue samples, namely similarly methylated probes (SMPs, n = 7659 CpGs; range: mean difference < 0.1-0.2%). To explore the efficacy of our reduction technique and further solidify the MNG-specificity built into our SMPs, unsupervised t-distributed stochastic neighbor embedding (t-SNE) was employed across a wide array of tumor types, some not included within the studied liquid biopsy cohort. The entire cohort included w ithin the t-SNE was derived from both external (n =2038) and internal (n = 229) collections. Further specialization of the DNA methylation signature was completed through comparison of untreated MNG (i. e.. no presurgical radiation; n = 27) and non- MNG specimens (excluding glioblastomas, which were appreciably distinct across genome-wide visualizations; n = 32) across the aforementioned SMPs, to identify meningioma-specific DMPs with no potential for treatment-related epigenetic modifications. This final set of signatures for the array was named diagnostic- Meningioma Epigenetic Liquid Biopsy (d-MeLB. range number: 20-30 CpGs; Wilcoxon rank sum test, p-valueFDR). Notably, variability was introduced across the machine-generated classifiers through randomization of parameters across iterations, including tissue-based nontumor differential significance, matching serum / tissue similarity score, sampling of non-MNG groups and final signature set size.

[0107] Using d-MeLB, we generated a classifier using the function ‘train’ (caret, v6.0.94) with 1000 decision trees and 10-fold cross validation conducted across the training set. To guide selection of the classifier, the classifier was applied to the model selection set and the score cutoff was selected which optimized the relationship between true positive and false negative rates, defined through inspection of associated receiveroperating characteristic (ROC) curves, namely diagnostic or d-MeLB score. The selected classifier was validated using our original independent validation set (n = 93), including the excluded recurrent gliomas and inflammatory non-neoplastic diseases (n=63). Further validations of the classifier were conducted across additional meningioma liquid biopsy -based validations, including both serum (n = 19) and plasma (n = 10) collections. Finally, the diagnostic random forest-based classifier was compared with other machine learning methods.

[0108] To investigate whether our d-MeLB signature was suitable for the diagnosis of tumor tissue samples (profiled through EPIC array), a simplified linear discriminant analysis (LDA) based machine learning classifier was constructed using MNG (n = 138; Henry Ford Health & Universify of Sao Paulo), intracranial mesenchymal tumors (n= 20; GSE164994), and non-neoplastic collections (n =21; GSE111165). These samples were randomized into discovery (80%) and independent validation (20%) sets, with discovery samples used to train and construct the LDA classifier, and the validation set further expanded with external collections of gliomas (n = 16; GSE147391), MNG (n = 1 10; GSE189521), and internally profiled pituitary' tumors (n = 14). Clustering of the total tumor tissue independent validation set (n = 176) across the d-MeLB signature was visualized using principal component analysis, and performance measures were calculated, as described.

[0109] Meningioma prognostic prediction / Prognostic MeLB (p-MeLB) — To investigate the application of MeLB as a prognostic tool (p-MeLB), MNG tissue methylome data was compiled from confirmed recunent (CR, n = 35) and confirmed non-recurrent (CNR, n = 15) collections. Machine-driven randomization assigned the total tissue cohort into training (CR, n=24; CNR, n = 10) and independent validation sets (CR, n = 11; CNR, n = 5). Then, a series of machine-driven and sequential supervised analyses comparing methylomes across specific meningioma tissue prognostic groups (CR and CNR) and / or serum k-cl usters (kl, k2, k3 and k4) were implemented to define serum-applicable prognostic methylation signatures. The defined signatures were utilized as input to a machine learning random forest algorithm, generated through use of the function ‘train’ (caret, v6.0.94) with 1000 trees and 10- fold cross validation conducted across the training set. Signature set derivations and subsequent generation of the random forest algorithms were repeated across 1000 iterations, with each concluding in storage of the signature, training set classifications, and associated out-of-bag (OOB) error.

[0110] Following 1000 iterations, a p-MeLB classifier was selected with an optimal score cutoff based on the smallest OOB error. Finally, the chosen classifier was validated, using the independent validation cohort (n = 16) and an additional set of internal and external samples meeting our criteria for definition of CR (Henry Ford Health: 29 tissue, 8 serum; Bayley et al.: 20 tissue). The classifier was then applied across the liquid biopsy sample cohort for stratification by recurrence risk.

[0111] Comparison between published prognostic and p-MeLB classifiers — The agreement of p-MeLB predictions was evaluated with prognostic groups involving the analysis of methylomes combined with additional molecular and / or clinicopathological features defined by previously reported tissue-based classifiers. Through the provision of the DNA methylation data developed and disclosed herein, the extent of resection and WHO Grade, the internal primary meningioma tissue cohort (n = 69) was classified according to a 5-year recurrence risk prediction nomogram and resulting classifications were compared with p-MeLB ’s conclusions. The p-MeLB classifier was also applied across publicly available methylome data, following appropriate exclusions, and the relationship between p-MeLB predictions and purported DNA methylation prognostic groups was explored. To circumvent the general lack of attributed person time for each sample in the publicly available cohorts, local freedom from recurrence (LFFR) Kaplan -Mei er curves were generated across Choudhury methylation-based groups and the resulting p-MeLB risk scores were correlated with their reported outcome data (Cohen’s unweighted kappa coefficient [K] and Spearman's correlation coefficient [p]).

[0112] Identification of biologically relevant and equivalent DNA methylation markers between tissue and serum — To identify biologically relevant and equivalent DNA methylation markers between tissue and serum, an in silico functional analysis approach was used, capitalizing on the availability of tissue-derived and paired methylome and transcriptome analysis of meningioma provided by Choudhury et al. (n = 185).

[0113] Methylation-based Deconvolution — Previously described and validated DNA methylation-based methodologies were applied to deconvolute the relative contribution of immune cell types to a given liquid biopsy and tissue specimen (package:

[0114] MethylCIBERSORT v0.2).

[0115] Clinicopathological and molecular features across MNG subgroups — The distribution of clinicopathological and molecular features were analyzed (categorical and continuous) across serum- and tissue-derived MNG specimens according to k- clusters and high and low recurrence risk groups.

[0116] Comparison between serum and plasma specimen-derived DNA methylation profiles — To compare results obtained through serum and plasma-derived specimens, genome-wide DNA methylation levels (EPIC Array) and the estimated immune and nonimmune cell proportions were correlated across a paired serum and plasma samples set (n = 10 pairs).

[0117] Statistical analysis — All data processing and statistical analyses were completed using R (3.6. 1 ). Non-parametric two-sided Kruskal -Wallis and Wilcoxon rank sum tests and multiple testing adjustments (e.g., FDR) were used to identify' significant DMPs and discrete variable differences across binary group comparisons. Machine-learning classifiers were formulated using a random forest (RF) algorithm. Receiver operating characteristic (ROC) curves were utilized to estimate the predictive power for each iteration of the diagnostic RF classifier (false positive rate (1 -specificity [SP]), true positive rate (sensitivity [Se]). Concordance across classifiers was estimated using Cohen’s unweighted kappa coefficient (K); correlative relationships were quantified using Pearson’s correlation coefficient (p). Statistical significance related to differences in survival probability for estimated risk groups were established through p-values (logrank test; p < 0.05). K-nearest neighbor imputation machine learning was used in the event of missing DNA methylation P-values found within tumor tissue methylomes across downstream analyses (t-SNE).

[0118] Methylation Arrays Through in vivo research and strategic refinement of machine learning classifiers, the following methylation arrays were developed for meningiomas (MeLB; Table 2), pituitary neuroendocrine tumors (PeLB; Table 3), and gliomas (GeLB; Table 4). The CpG probes listed in Tables 2-4 can be used to probe a cfDNA sample of a patient and determine whether the CpG sites corresponding to the CpG probes are differentially methylated. The results can be used to inform and / or adjust a prognosis, a monitoring plan, or a treatment plan. For example, active interventions may be sought out if a certain differential methylation level is assessed for certain CpG probes, for example, whereas disease monitoring may be maintained without an active intervention (chemotherapy, radiation, surgery, etc.) below that level. The actual adjustments to the prognosis process and / or treatment plans may be assessed at the clinical level, informed at least partially by the methodologies and methylation arrays described herein.

[0119] Table 2: MelB CpGs

[0120] Table3:PeLB CpGs

[0121] Table 4: GeLB CpGs

[0122] These methylation arrays are particularly useful for the diagnosis and treatment of CNS tumors and have been particularly efficacious for meningiomas, pituitary neuroendocrine tumors, and gliomas, respectively. With all three methylation arrays shown in Tables 2-4, two or more of the CpG sites are located on chromosome two (chr2), two or more are located on chromosome six (chr6), and two or more are located on chromosome sixteen (chrl6). Advantageously, the array includes there or more CpG sites on chr6, whereas for meningiomas and pituitary neuroendocrine tumors in particular, there are no CpG sites on chr8, chr9, chrl8, or chr 21. Narrowing the scope of analysis in this fashion can help improve detection, particularly with serum-based liquid biopsy, which is more challenging than tissue-based testing methods, particularly with respect to meningiomas, pituitary neuroendocrine tumors, and gliomas.

[0123] In at least some embodiments, for meningiomas, the methylation array contains a majority of the CpG probes in Table 2, and preferably all of the CpG probes in Table 2. For pituitary neuroendocrine tumors, the methylation array contains a majority of the CpG probes in Table 3, and preferably all of the CpG probes in Table 3. For gliomas, the methylation array contains a majority of the CpG probes in Table 4, and preferably all of the CpG probes in Table 4. It may be possible to add other probes, for example, to the methylation arrays described herein.

[0124] Limiting the CpG probes of the analysis to focus on the arrays disclosed in Tables 2-4 can allow for scoring or weighting of different probes to help hone the diagnostic analysis. For example, certain CpG probes in one of Tables 2-4 may be 80- 90% more likely to indicate a tumor if methylated (hypermethylation), so the differential methylation level for that CpG probe may be ranked or scored higher than another CpG probe in the array that may only be 60-70% more likely to indicate a tumor if methylated. In some embodiments, when a differential methylation level is determined for each of the limited CpG sites (limited in that not all CpG sites are analyzed), this methylation level can be a score that is factored into an overall analysis (e.g., CpG sites located on chr2, chr6, and / or chr 16 are more likely to indicate a tumor so these sites are weighted more heavily than sites on other chromosomes). These scores, parameters, etc. may be adjusted depending on the clinician’s judgment and factors such as cancer type, treatment stage, chance of recurrence, to cite a few examples.

[0125] In an advantageous embodiment, one or more machine learning models, such as those disclosed herein, can be used with the methylation arrays and potential scores overtime to adjust diagnostic and treatment parameters. The machine learning model may be executable on or with an electronic processor that is configured to cany' out the diagnosing / treatment methods described herein. This can provide refinement of the methods on a case-by-case basis overtime.

[0126] It is to be understood that the foregoing description is of one or more preferred exemplary embodiments of the invention. The invention is not limited to the particular embodiment(s) disclosed herein, but rather is defined solely by the claims below. Furthermore, the statements contained in the foregoing description relate to particular embodiments and are not to be construed as limitations on the scope of the invention or on the definition of terms used in the claims, except where a term or phrase is expressly defined above. Various other embodiments and various changes and modifications to the disclosed embodiment(s) will become apparent to those skilled in the art. All such other embodiments, changes, and modifications are intended to come within the scope of the appended claims.

Claims

PCT / US25 / 43090 22 August 2025 (22.08.2025)1059-3057-WO2 [2024-075]CLAIMS1. A method of diagnosing or treating a cancer in a subject, comprising the steps of: obtaining a cfDNA sample from a subject; probing the cfDNA sample with a plurality of CpG probes;5 comparing information from the probed cfDNA sample to a methylation array to determine a differential methylation level at each of a plurality of limited CpG sites within the cfDNA sample, wherein the plurality of limited CpG sites correspond to the plurality of CpG probes, and wherein two or more of the CpG sites are located on chromosome 2, two or more are located on chromosome 6, and two or more are located10 on chromosome 16; and using the differential methylation level to diagnose and / or treat the cancer.

2. The method of claim 1, wherein the plurality of limited CpG sites includes three or more located on chromosome 6.

3. The method of claim 1, wherein the plurality of limited CpG sites includes no CpG sites on chromosome 8, on chromosome 9, on chromosome 18, or chromosome 21.

4. The method of claim 1, wherein the cancer is a central nervous system (CNS) tumor causing cancer.

5. The method of claim 4, wherein the CNS tumor causing cancer causes one or20 more meningiomas, one or more pituitary neuroendocrine tumors, or one or more gliomas.

6. The method of claim 4, wherein the CNS tumor causing cancer causes one or more meningiomas, and the plurality of limited CpG sites includes a majority of the CpG sites of Table 2.25 7. The method of claim 6, wherein the plurality of limited CpG sites includes all of the CpG sites of Table 2.

8. The method of claim 4, wherein the CNS tumor causing cancer causes one or more pituitary neuroendocrine tumors, and the plurality of limited CpG sites includes a majority of the CpG sites of Table 3.PCT / US25 / 43090 22 August 2025 (22.08.2025)1059-3057-WO2 [2024-075]9. The method of claim 8. wherein the plurality of limited CpG sites includes all of the CpG sites of Table 3.

10. The method of claim 4, wherein the CNS tumor causing cancer causes one or more gliomas, and the plurality of limited CpG sites includes a majority of the CpG sites of Table 4.

11. The method of claim 10, wherein the plurality of limited CpG sites includes all of the CpG sites of Table 4.

12. The method of claim 1, comprising the step of determining a score from the differential methylation level, wherein the score has one or more CpG sites of the plurality of limited CpG sites weighed differentially than one or more other CpG sites of the plurality of limited CpG sites.

13. The method of claim 12, wherein machine learning is used to adjust the score on a case-by-case basis.

14. The method of claim 12, wherein a differential methylation level is determined for each of the plurality of limited CpG sites and each methy lation level has an associated score.

15. The method of claim 1, wherein the cfDNA sample is a serum-based liquid biopsy sample.

16. The method of claim 1, wherein the differential methylation level is a hypermethylation status determined for each of the plurality of limited CpG sites.

17. The method of claim 1, wherein a treatment regime for the subject is altered based on the differential methylation level to adjust an active intervention or change a disease surveillance strategy.

18. A non-transitory, computer-readable storage medium storing instructions thereon that when executed by one or more electronic processors causes the one or more electronic processors to carry out the method of claim 1.

19. A methylation array to determine a differential methylation level at each of a plurality of limited CpG sites within a cfDNA sample, wherein the methylation arrayPCT / US25 / 43090 22 August 2025 (22.08.2025)1059-3057-WO2 [2024-075] comprises a plurality of CpG probes, each CpG probe of the plurality' of CpG probes corresponding to each CpG site of the plurality of limited CpG sites, wherein the methylation array is configured to diagnose or treat meningiomas and wherein the plurality of CpG probes includes a majonty of the CpG probes in Table 2.5 20. A methylation array to determine a differential methylation level at each of a plurality of limited CpG sites within a cfDNA sample, wherein the methylation array comprises a plurality of CpG probes, each CpG probe of the plurality' of CpG probes corresponding to each CpG site of the plurality of limited CpG sites, wherein the methylation array is configured to diagnose or treat pituitary neuroendocrine tumors10 and wherein the plurality of CpG probes includes a majority' of the CpG probes in Table 3.

21. A methylation array to determine a differential methylation level at each of a plurality of limited CpG sites within a cfDNA sample, wherein the methylation array comprises a plurality of CpG probes, each CpG probe of the plurality' of CpG probes 15 corresponding to each CpG site of the plurality of limited CpG sites, wherein the methylation array is configured to diagnose or treat gliomas and wherein the plurality of CpG probes includes a majority of the CpG probes in Table 4.

Images