Abstract BACKGROUND DNA methylation profiling has become a pivotal tool in neuro-oncology. However, little is known about the impact of methylation profiling on radiological imaging in IDH-wildtype glioblastoma (GBM). This study explored the potential of radiomics to non-invasively predict molecular characteristics based on the methylation profiling of GBM. METHODS This study included a multi-sampling cohort of 32 patients with 113 samples and a single-sampling cohort of 87 patients with 87 samples, which underwent genome-wide methylation analysis and were classified as GBM by the DKFZ/Heidelberg CNS tumor classifier. We performed two different methylation-based deconvolution analyses. Deconvolution 1 estimated the fractions of tumor subtypes (RTK_I, RTK_II, MES_TYP, and MES_ATYP) and cell types in the microenvironment. Deconvolution 2 inferred the abundance of malignant cell states (stem-like and differentiated cell components) and cell types in the microenvironment. We derived 6084 radiomic features from preoperative multi-parametric MRI in each patient. RESULTS In the multi-sampling cohort, Deconvolution 1 and Deconvolution 2 demonstrated that tumor subclass components and immune components exhibited heterogeneous distribution across samples within patients; however, the stem-like to differentiated cell ratio was preserved across samples within patients. In the entire cohort, patients with tumors exhibiting a high proportion of the stem-like component, defined as stem-like tumors, had significantly shorter overall survival. Stem-like tumors exhibited significantly higher levels of the RTK_I component and lower proportions of the RTK_II component compared to non-stem-like tumors. A machine learning model that utilizes support vector machines to classify stem-like tumors using radiomic features achieved a mean AUC of 0.71 (95% confidence interval: 0.56–0.85) in nested cross-validation. CONCLUSIONS We showed that radiomethylomics can be used to predict tumor composition and patient survival from preoperative MRI in GBM. Based on these promising results we will increase our power with additional samples to facilitate a personalized approach to GBM management.