Kyle S. Smith, Tom T. Fischer, Katie Han, Anna Kostecka, Hong Lin, Daniel Senfter, Taha Soliman, Natalia Stepien, Stefanie Volz, Nathalie Schwarz, Tatjana Wedig, Sibylle Madlener, Christine Haberler, Sandeep K Dhanda, Santhosh A. Upadhyaya, Patrick R. Blackburn, Maria T. Schmook, Judith M. de Bont, H. Haapasalo, Justina Dargvainiene, Frank Leypoldt, Stefan M. Pfister, Esther Hulleman, Brent A. Orr, Amar Gajjar, Giles Robinson, Joonas Haapasalo, Kristiina Nordfors, Johannes Gojo, Kristian W. Pajtler, Kendra K. Maaß, Paul Northcott
Cerebrospinal fluid (CSF) liquid biopsies serve as a rich source of tumor-derived cell-free DNA (cfDNA) for evaluating persons with central nervous system (CNS) tumors. However, challenges stemming from trace cfDNA yields and low mutational burden have hindered sensitivity, whereas first-generation clinical assays have relied on genetic alterations as biomarkers. Leveraging the diagnostic utility of DNA methylation classification in CNS tumors, we developed M-PACT (methylation-based predictive algorithm for CNS tumors), a robust deep neural network that accurately classifies tumors from subnanogram-input cfDNA methylomes. Across embryonal CNS tumor benchmarking (n = 79) and validation (n = 58) cohorts, M-PACT achieved 92% and 88% accuracy, respectively. We further showcase M-PACT utility in nonembryonal CNS tumors, balanced tumor genomes and nonmalignant CSF. Beyond classification, this workflow enables methylation-based cellular deconvolution and sensitive copy-number variation detection. Altogether, we provide a blueprint for CNS tumor classification from low-input cfDNA methylomes, motivating prospective validation for future clinical implementation. Smith et al. present M-PACT, a deep neural network leveraging low-input cell-free DNA methylation profiles to achieve robust classification of pediatric brain tumors and enable cellular deconvolution and sensitive copy-number variation detection.