Development and validation of a DNA methylation-based classifier for CNS tumors using a large Chinese cohort (n = 1,581)

Abstract

Abstract The clinical implementation of DNA methylation profiling for central nervous system (CNS) tumors faces significant challenges in China, particularly regarding the development and validation of locally applicable classifiers. We developed MethAI-CNS, a locally executable DNA methylation-based classifier for CNS tumors that strictly follows the established DKFZ classification framework. The classifier was trained on global databases augmented with 1,368 local Chinese samples, covers 122 subclasses. The model was independently validated on 213 Chinese samples, which included challenging pathological consultation cases and medulloblastoma molecular subtyping cases, and further validated on an independent public cohort (GSE289137, n = 687), with comparisons against DKFZ v12.8. MethAI-CNS achieved an overall accuracy of 0.988 and an AUC of 0.993 in 5 × 5 cross-validation. At a threshold of 0.9, the sensitivity and specificity were 0.976 and 0.964, respectively. In the Chinese validation cohort, high-confidence predictions from MethAI-CNS showed 99% concordance with the DKFZ classifier v12.8, demonstrating high fidelity to the DKFZ reference framework. On the independent GSE289137 cohort, MethAI-CNS achieved a concordance rate of 99% (543/546) with DKFZ v12.8 at the 0.9 threshold, with an accuracy of 0.956 and an AUC of 0.949. Among diagnostically challenging consultation cases and medulloblastoma cases, methylation-based classification led to revision rates of 32% and 5% of the original histopathological diagnoses, respectively. MethAI-CNS is a locally validated methylation classifier for CNS tumors in a Chinese cohort that demonstrates robust performance in diagnosing complex neuroepithelial tumors and medulloblastoma. It provides a reliable tool for supporting pathological practice, particularly in regions with restricted access to reference models.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗