Malignant brain tumor affects a large number of patients and often have poor prognosis and low response to therapeutics. An indicator of the progress of brain tumor and its response to treatment is the DNA repair protein, O6-methylguanine-DNA methyltransferase (MGMT). As such, accurate assessment of MGMT is of great clinical significance. Biospy is not only invasive but also has the risk of undersampling the tumorous tissue. We presented a novel deep learning model that uses multi-MRI modalities to assess the expression of MGMT in glioblastoma (GBM) patients. Results showed that the model can achieve good performance.