Abstract Introduction: Our goal is to develop interpretable, prognostic models for Group 3/4 medulloblastoma (G3/4 MB) and PFA ependymomas (PFA-EP) using brain tumor DNA methylomes. Our models use prior knowledge of genome regulation and may identify cellular processes that help develop molecular therapies. We previously developed a classifier algorithm, netDx, which uses similarity networks to integrate heterogeneous patient data and predict outcome1. netDx demonstrates excellent performance and allows measures to be grouped into biologically meaningful features (e.g. grouping genes into pathway features). Ependymoma Methods: We predicted survival in PFA-EP using 569 methylomes2 (Illumina 450K). Using netDx, we evaluated a model that grouped CpG-level methylation into sets reflecting 25 cell types from the developing human cerebellum, 3 chromatin states, EZH2 binding sites, and brain super enhancers (80:20 train/test split, feature selection >=8/10; 10 splits). This model was compared to a baseline lacking prior knowledge. Preliminary Results: Organizing methylomes by prior knowledge significantly improves prognostic prediction (Table 1, p < 5x10-3, one-sided WMW). Features that predict prognosis are consistent with known dysregulation in PFA-EP2 (Table 1). G3/4 MB Methods: We predicted binarized survival using 285 methylomes3 (Illumina 450K), using identical methods as above. Preliminary results: Tumor methylomes carry predictive signal for survival prediction (Table 1), consistent with previous findings. We are currently evaluating the effect of including prior knowledge. Interim Conclusion: Prior knowledge can improve survival prediction in PFA-EP and identifies features reflecting tumor biology. We are interested in extending interpretable modeling to other tumours.1. Pai et al. (2019) Mol Sys Biol. 15. 2. Pajtler et al. (2018) Acta Neuropathol. 136. 3. Northcott et al. (2017) Nature 547. Table 1. Average model performance and top features. Dataset Predictor design AUPR (mean+/- SD, 10 train/test splits) Predictive features PFA-EP; 252 good/317 poor survivors2 DNAm, no prior knowledge 64.2 +/- 2.3 n/a * Covariates: CXOrf67 mutation status, sex DNAm, prior knowledge* 67.5 +/- 3.5(p < 5x10-3; one-sided WMW test compared to baseline model) CXOrf67mut, methylation in: {H3K9me3 sites; marker genes for ependymal cells of choroid plexus; marker genes for multiple interneuron classes} G3/4 MB: 146 good/139 poor survivors3 DNAm, no prior knowledge 0.61 +/- 6.3 n/a ** Covariates: age, sex DNAm, prior knowledge** In progress Citation Format: Indy Ng, Alexander Fricke, Shraddha Pai. Predicting prognosis in PFA ependymoma and group 3/4 medulloblastoma methylomes using interpretable epigenetic models [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 2505.