PATH-63. Analysis of histological, molecular and preanalytic features of unclassifiable CNS specimens: comparison of the Heidelberg v12.8 and Bethesda v3 methylation-based brain tumor classifiers

DNA methylation analysis is an important diagnostic method in neuropathology and often enables a precise WHO diagnosis; however, some cases remain unclassifiable by current methylation-based brain tumor classifiers. Factors influencing classifiability of samples are incompletely understood and performance evaluations of new machine learning tools in challenging specimens are limited. We retrospectively analyzed a cohort of 1,853 CNS tissue samples diagnosed between 2017 and 2024 using genome-wide DNA methylation profiling and identified a cohort of 393 cases (21%) unclassifiable by the Heidelberg Brain Tumor Classifier v12.8. We examined preanalytic, histological, and molecular features of these samples and analyzed them using the new Bethesda Classifier v3 (Bv3). Younger age (P = 1.352 × 10⁻⁶), lower tumor purity (P = 2.539 × 10⁻¹¹) and reduced DNA input (P = 3.458 × 10⁻¹⁶) negatively affected the v12.8 classifier accuracy, while unclassifiable sample rates were identical for the 850k and 935k Illumina EPIC arrays (22% each; p = 0.9505). Bv3 was able to classify 213/393 unclassifiable cases (54%), including 196/325 neuroectodermal tumors (60%), 6/21 mesenchymal tumors (29%), 6/20 metastases (30%) and 5/27 non-tumor specimens (19%). A WHO diagnosis was established in 288/325 neuroectodermal tumors (89%; NEC/NOS: n = 37); the Bv3 achieved a classifier score above the threshold of 0.9 in 186 cases (186/288;65%): in 163/186 cases (88%) the classification matched the WHO diagnosis, in 23/186 cases the Bv3 suggested an alternative diagnosis (12%). Notably, the Bv3 correctly identified 117/174 glioblastomas, IDH-wildtype (67%), including 86/102 histological GBM and 16/102 molecular GBM. The RF_purify estimated tumor cell content of 100 GBM samples correctly classified by the Bv3 was significantly lower compared to 458 GBM samples correctly classified by the v12.8 (P = 0.01086). The Bv3 classifier is a valuable tool in currently unclassifiable, diagnostically challenging brain specimens and especially useful in low tumor cell content samples.