Genetic profiling of cancer has the potential to identify epigenetic changes that predict response to chemotherapy. For example local vascular endothelial growth factor (VEGF) expression, as quantified by increased CBV has shown to reflect changes in response to VEGF inhibition¹. The difficulty in predicting response to genomic targeted chemotherapy using biopsy derived genomic data is the marked heterogeneity of tumor phenotype (i.e. local variation of gene expression). In this work we develop a methodology for analyzing multi-contrast MRI to predict local gene expression in histologically confirmed GBM.This study was performed in order to find correlations between MR imaging parameters—T1, T1 post Gd, T2 and apparent diffusion coefficient (ADC)—, and O⁶-Methylguanine-DNA methyltransferase promoter (pMGMT) and the four subtypes of GBM. The four subtypes are the neural, proneural, mesenchymal, and classical subtypes². Any correlation found could possibly make MRI a valid method of identifying certain types of tumor and predicting response to treatment. Prior to this analysis, the variations in the source of the MR scans required for a normalization of the data. Therefore, a part of this research was conducted to develop a method of normalizing cross-site MR data.There are databases publicly available that contain the genetic information and MR images of interest, namely the Cancer Genome Atlas (TCGA) and the Cancer Imaging Atlas (TCIA). However, the TCIA database containing the MR images consists of scans from multiple sites that use different pulse sequences and protocols, which stymies any attempt to use the data for analysis. A sorting algorithm was developed to collect T1, T1 post Gd, T2 and ADC images. The collected images were then coregistered and segmented using Statistical Parametric Mapping (SPM Version 8 Wellcome Trust Centre for Neuroimaging, London, UK), and a region of interest (ROI) was drawn on the tumors excluding any necrotic areas. Grayscale pixel values were converted to normalized quantitative values using a linearized model derived from the MR contrast signal equations constrained by historical reference values. Normalized values were correlated to the subtype classification, which were sourced from Brennan et al³, and to pMGMT methylation status, dichotomized by methods in Bady et al⁴. The mean and standard deviation of the converted T1, T1 post Gd, T2 and ADC values of the tumor were compared to the GBM subtypes for 25 patients, and to pMGMT methylation status for 20 patients. In addition for the pMGMT group, t-test, ROC analysis, and Kaplan-Meier survival curves were performed to further study the correlations.The variability within a tumor and between image types are shown in Figure 1. The mean and standard deviation of the normalized T1, T1 post Gd, T2 and ADC values for each GBM subtype and pMGMT methylation group (+ for methylated and – for unmethlyated) are shown in Table 1. This result does not show any clear trends for any of the subtypes, and there is a wide variability for neural and proneural subtypes. Neural subtypes seem to be correlated with high T1, T1 post Gd, T2 and ADC, while mesenchymal and classical subtypes show low T2 and ADC. On the other hand, a significant difference was present for pMGMT groups and T1 and 1/ADC. Performing a t-test returned p-values < 0.05 for both T1 and 1/ADC, and the area-under-the-curve for ROC curves were 0.89 for both. Kaplan-Meier survival curves calculated using 1/ADC derived and T1 derived pMGMT methylation status showed close resemblance to the actual (Figure 2). An example of a probability map of pMGMT methylation as predicted by T1 and 1/ADC is shown in Figure 3; this particular case was reported from biopsy as pMGMT+.Preliminary results show that the normalization process works as intended. Clear correlations could not be found between the MRI parameters and the four subtypes of GBM. This may be due to the rather broad classification of the subtypes. It could also be the small sample size and the limited number of contrast parameters studied here. For pMGMT methylation status, however, T1 and 1/ADC were found to be good predictors. Although 1/ADC was reported to be more significant than ADC, it should be noted that the p-value for ADC was very low (0.06). Future research will include a larger sample of data, and a validation study to compare to actual local genomic variations.Different tumor phenotypes of GBM necessitate different prognosis and treatments. With more analysis, MRI can emerge as a noninvasive method of identifying gene expression signatures and phenotypes of tumors.