DNA Methylation-Based Classification of CNS Tumors: Comparable Performance Between Nanopore and EPIC Technologies

Abstract

Abstract Background DNA methylation profiling enables precise classification of pediatric central nervous system (CNS) tumors. Oxford Nanopore Technologies (ONT) offers same-day, single-sample methylation readouts, but its concordance with Illumina EPIC arrays in routine diagnostic tasks remains incompletely defined. Methods We profiled 23 pediatric tumors (18 CNS, 5 non-CNS) by EPIC (v1/v2; FFPE w/o FF) and ONT (FF). CNS tumors from both platforms were classified with the crossNN_brain model; ONT data were additionally classified with Rapid-CNS2 and Sturgeon. Non-CNS tumors were classified with the crossNN pan-cancer model. We compared (i) classifier agreement with integrated histology (w/o NGS) at family/class levels, (ii) pass-rate above platform-specific score cutoffs, (iii) cross-platform concordance of copy-number variation (CNV), and MGMT promoter methylation status. Results In CNS cases, ONT and EPIC methylation profiles demonstrated strong correlation, except for a single low-cellularity outlier (P2), which was excluded from further analysis. Comparative assessment of the two platforms showed that: (a) Molecular classification of CNS tumors using the crossNN classifier was consistent with histology (w/o NGS) at the family level in all cases (100%, 17/17). At the class level, classification agreement was 100% (17/17) for EPIC arrays and 88% (15/17) for ONT data. (b) Copy-number profiles showed high concordance between platforms. (c) MGMT promoter methylation status matched in 94% of cases (16/17). When comparing ONT-specific analysis pipelines using the ONT data, the Rapid-CNS2 pipeline yielded the most reliable class level assignments with 94% (16/17) concordance with the histopathological diagnosis, which marginally exceeded the crossNN classifier, while a third tool (sturgeon) underperformed. In non-CNS tumors, the pan-cancer model produced low-confidence outputs with poor agreement with histology (w/o NGS) (only 1/5 concordant), indicating limited readiness for these entities. Conclusions ONT enables same-day, clinically reliable family-level CNS tumor classification with high concordance to arrays, while EPIC retains a modest class-level edge. High concordance for MGMT and CNV further supports an ONT-first workflow in most CNS cases. Limitations of our study include mostly the cohort size. A key limitation of ONT is its reliance on fresh-frozen DNA and on classifiers originally built around array-derived CpG sites, rather than on models developed natively from ONT data. Building ONT-specific models could further improve class-level accuracy and confidence.