Childhood Cancer Data Initiative Annual Symposium (Abstract Registration): Submission #38

Abstract: Central nervous system (CNS) tumors are the most common solid tumors diagnosed in children, making up about 17% of diagnosed childhood cancer cases. Most pediatric CNS tumor cases require a combination of imaging, histopathological, and molecular diagnoses to determine the tumor type/subtype. My project utilizes an emerging molecular diagnostic technology -- nanopore sequencing -- to quickly and accurately generate high depth whole genome DNA methylation profiles for a diverse cohort of 150 pediatric CNS tumors, and this data will be used to train and validate a pediatric-specific deep learning classifier for intraoperative use. As part of a clinical research collaboration, I tested the previously-developed machine learning classifier Sturgeon on high depth nanopore DNA methylation data for 8 matched pairs of standard patient tissue and tumor aspirate (16 total samples). For each sample, I demonstrated between 4-34x genomic coverage, classified the sample’s tumor type/subtype, validated key subtype-defining features, and generated a digital karyotype to visualize large tumor-specific copy number variation(s). Through this initial analysis, I was able to identify key shortcomings in the existing software, including inappropriate labeling for pediatric samples, insufficient adjustment for low tumor purity samples, and confident misclassification of rare/unique tumors. After making adjustments to the classifier's data processing, all 16 samples were classified correctly with high confidence. Together, I have used nanopore DNA methylation profiling to demonstrate fast and accurate classification of CNS tumor aspirate, and in the future, I will generate a similar machine learning classifier for pediatric-specific CNS tumor classification.