BIOM-36. NONINVASIVE CLASSIFICATION OF CENTRAL NERVOUS SYSTEM TUMORS USING CFDNA METHYLATION PROFILING: DEVELOPMENT AND VALIDATION OF A MULTICLASS RANDOM FOREST MODEL

Abstract BACKGROUND Accurate classification of central nervous system (CNS) tumors using tissue samples has been well-established. However, a noninvasive approach utilizing liquid biopsy samples offers significant advantages for patient diagnosis and monitoring. This study aims to develop a multiclass classification model for CNS tumors based on circulating cell-free DNA (cfDNA) methylation profiles obtained from blood samples. METHODS We performed genome-wide methylation profiling (EPIC array v1.0) on 227 serum liquid biopsy specimens from patients with CNS tumors, including glioma (n=109), meningioma (n=81), ependymoma (n=23), and pituitary tumors (n=14). A random forest classifier was trained on 155 samples, while 72 samples were reserved for independent testing (held-out set). Publicly available methylome data from CNS tumor tissue samples were used to further validate the tumor-specific signatures derived from the blood samples. RESULTS Our random forest model demonstrated a classification accuracy of 91% on the held-out test set, confirming the presence of tumor-specific methylation signatures in blood samples from patients with CNS tumors. Tumor-specific methylation signatures detected in liquid biopsy samples were able to discriminate between related tumors in an independent set of tissue samples. CONCLUSIONS These findings suggest that liquid biopsy-derived cfDNA methylation signatures are a promising noninvasive tool for the classification and surveillance of CNS tumors, offering an alternative or complement to tissue-based and imaging methods.