Path-12. Robust Cell-Free Dna Methylation-Based Classification of Pediatric Brain Tumor Liquid Biopsies

Abstract BACKGROUND DNA methylation-based classification has been transformative in the molecular diagnostics of pediatric brain tumors, while additionally providing foundational insights into disease biology. Current diagnostic pipelines integrate histopathology and molecular tools in accordance with the WHO classification of CNS tumors which rely on the availability of tissue specimens. However, sometimes tumors are not amenable to neurosurgical resection due to their high-risk locations. Likewise, tissue collection is not routinely performed longitudinally. Liquid biopsies (LBs) have emerged as a minimally invasive, serially collectable source of tumor-derived material that have shown promise in detecting minimal residual disease (MRD) in pediatric neuro-oncology. Still, success rates of LB profiling have varied between studies, warranting the deployment of more reliable and reproducible assays. METHODS In the current study, we implemented a novel cell-free DNA (cfDNA) methylation sequencing (EM-seq) workflow in a large cohort of cerebrospinal fluid (CSF) samples (n200 patients) representing most pediatric brain tumor entities. Optimization of EM-seq for CSF specimens enabled the generation of genome-wide cfDNA methylation profiles from minimal starting material. Utilizing well-established CNS tumor DNA methylation datasets as references, we trained a deep neural network specializing in the classification of low tumor burden and sparse methylation profiles. Implementation of a novel tumor enrichment algorithm, coupled with DNA methylation imputation and subsequent deconvolution, facilitated detection and classification of low tumor burden samples, including those with balanced genomes that would have been overlooked using previous generation assays. RESULTS Our LB-based classifier allowed for highly accurate tumor detection and entity prediction (AUC=0.9) across the cohort. Case examples highlight the potential of this approach in establishing tumor classifications in inoperable cases and in performing serial disease monitoring. CONCLUSIONS Collectively, this study provides a blueprint for developing a CSF-based tumor classifier from cfDNA methylation profiles, and the motivation for prospective validation in future clinical trials.