Towards an integrated cellular and molecular classifier for glioblastoma

The recognition of tumor heterogeneity has highlighted the necessity of examining tumor samples through the lens of single-cell genomics. In glioblastoma (GBM), a highly heterogeneous tumor, single-cell analysis is critical to assist in assessing tumor composition and in the longitudinal analysis of response to therapies. However, single-cell genomic approaches face practical challenges for broad implementation, underscoring the importance of developing deconvolution methods that may assist in the interpretation of bulk profiles and can be deployed at scale. Bulk DNA methylation data, a stable and widely used diagnostic tool in gliomas and central nervous system tumors, provides a promising substrate for deconvolution. However, the limited availability of cell state-specific references in DNA methylation, coupled with low-coverage single-cell DNA methylation data, poses significant challenges. We present a hierarchical non-negative matrix factorization approach to deconvolute bulk DNA methylation profiles, initially resolving cell types and subsequently refining cell states within a cell type. By integrating multi-omics single-cell data, we mapped DNA methylation components to their transcriptional counterparts, enabling accurate predictions of transcriptional cellular composition from bulk DNA methylation. This methodology allows the decomposition of GBM bulk DNA methylation into glial, immune, neuronal, and malignant cell types, with further distinction into malignant stem-like and malignant differentiated cell states. Our findings reveal that low cancer cell fractions can distort classification, prompting the development of an in-silico purification method to enhance diagnostic accuracy. Additionally, we provide a framework to assist in quantifying the influences of the immune micro-environment on GBM bulk classification, unmasking the underlying genetic heterogeneity and tumor subtype. Our work provides a blueprint to reconcile DNA methylation, bulk transcription-based and single-cell classifications of GBM. Illumina Infinium MethylationEPIC (850K) for 11 samples of 9 GBM tumors. Out of the 11 samples 4 are FFPE and 7 Fresh, for two tumors we sequenced matched FFPE and Fresh tissue.