This study presents an explainable AI framework that reveals how the Heidelberg DNA methylation classifier distinguishes between 82 brain tumor types and 9 normal tissue classes. Using Random Forest models trained on 2,801 samples and 428,799 probes, the researchers found that only 2.3% of probes accounted for over 61% of the classifier’s decisions. These high-usage probes were concentrated in biologically relevant regions such as CpG islands, enhancer elements, heterochromatin, and lamina-associated domains, with distinct usage patterns across tumor types. IDH-mutant gliomas and ETMR relied on hypermethylated CpG island probes, while tumors like LIPN and PITAD predominantly used hypomethylated probes in open sea regions. A streamlined version of the classifier using 10,000 probes revealed 88 distinct probe clusters, over 80% of which were specific to individual tumor classes. These clusters were spread throughout the genome, indicating redundancy that enhances classifier robustness. Unsupervised clustering and dimensionality reduction showed clear class-specific patterns, such as clusters 27 and 78 for ETMR and cluster 1 for IDH-mutant gliomas. To make these findings accessible, the authors created a web application called “shinyMNP,” which allows users to explore class-informative probes and associated gene expression. The study demonstrated this tool’s utility through four tumor-specific examples: SHPRH hypermethylation in ETMR, PWWP3A hypomethylation and overexpression in HGNET_MN1, TBX19 promoter hypomethylation in PITAD_ACTH, and RET hypermethylation with high expression in HGNET_BCOR. The framework’s outputs were consistent with other interpretability methods such as SHAP and generalizable to other cancer types like sarcoma, supporting its broad applicability.