Dana Silverbush, Liv Jürgensen, Nelson F. Freeburg, Channing Pooley, Fabio Boniolo, Federico Gaiti, Mario L. Suvà, Volker Hovestadt
We introduce a hierarchical non-negative matrix factorization approach (ITHresolveGBM) to deconvolute bulk DNA methylation profiles, inferring the abundance of glial, immune, and neuronal cells of the microenvironment, and further distinguishing differentiation states of malignant cells.
BACKGROUND: Cellular heterogeneity is a defining feature of glioblastoma (GBM), shaping tumor progression and therapeutic response. While single-cell profiling resolves this heterogeneity, it remains impractical for large-cohort studies and clinical implementation. Conversely, DNA methylation-based classification is widely used for GBM diagnostics but does not provide cellular resolution. METHODS: We introduce a hierarchical non-negative matrix factorization approach (ITHresolveGBM) to deconvolute bulk DNA methylation profiles, inferring the abundance of glial, immune, and neuronal cells of the microenvironment, and further distinguishing differentiation states of malignant cells. RESULTS: Using ITHresolveGBM, we find that low tumor cell content impairs methylation-based classification, most notably linking the mesenchymal subtype with high immune cell infiltration. By integrating multi-omic single-cell data, we show that epigenetic deconvolution captures a malignant differentiation continuum ranging from stem-like to more differentiated tumors. This continuum aligns prior GBM classification systems and is associated with distinct molecular drivers (eg, PDGFRA, TP53, EGFR) and survival outcomes. CONCLUSIONS: Our framework reconciles DNA methylation- and RNA-based classification systems and provides a blueprint for unifying bulk tumor profiles with single-cell biology, thereby refining molecular stratification and enhancing GBM diagnostics.