Genomic mapping and survival prediction in GBM: Molecular sub-classification as an adjunct to

hemodynamic imaging biomarkers - A TCGA Glioma Phenotype Research Group Project

Rajan Jain, MD

Division of Neuroradiology, Department of Radiology Department of Neurosurgery Henry Ford Health System Assistant Professor, WSU School of MedicineDetroit, MI, US

34a- 34a - Adult Brain Neoplasms III Tuesday, April 24, 2012 12-O-876-ASNR

TCGA Glioma Phenotype TCGA Glioma Phenotype Research GroupResearch Group

Rajan Jain,1,2 Laila Poisson,3 Jayant Narang,1 David Gutman,4 Adam Flanders,5 Lisa Scarpace,2 Scott N Hwang,4 Chad Holder,4 Max Wintermark,6 Rivka R Colen,7 Justin Kirby,8 John Freymann,8 Brat Daniel,4 Carl Jaffe,9 Tom Mikkelsen 2

1Division of Neuroradiology, Department of Radiology, 2Department of Neurosurgery and 3Department of Public Health Sciences, Henry Ford Health System, Detroit, MI4Emory University, Atlanta, GA; 5Thomas Jefferson University Hospital, Philadelphia, PA; 6University of Virginia, Charlottesville, VA; 7Brigham & Womens Hospital, Boston, MA; 8SAIC-Frederick, Inc.; 9Boston University, Boston, MA

https://wiki.cancerimagingarchive.net/display/Public/TCGA-GBM

DisclosuresDisclosures

Funded by NCI Contract No. HHSN261200800001E.

Recently, there has been progress in understanding the molecular basis of the tumor aggressiveness and heterogeneity.

Various molecular sub-classifications have been proposed based on the genetic makeup of GBM with the hope that a better understanding of origin of tumor cells and molecular pathogenesis may predict response to targeted therapies.

Phillips HS, et al. Molecular subclasses of high-grade glioma predict prognosis, delineate a pattern of disease progression, and resemble stages in neurogenesis. Cancer Cell 2006;9(3):157-173.

Verhaak RG, et al. Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, VEGFR, and NF1. Cancer Cell 2010;17(1):98-110.

.

Background : Genomic Mapping

Phillips HS, et al. Molecular subclasses of high-grade glioma predict prognosis, delineate a pattern of disease progression, and resemble stages in neurogenesis. Cancer Cell 2006;9(3):157-173.

High-Grade Gliomas: Molecular Sub-classes

Verhaak RG, et al. Integrated genomic analysis identifies clinically relevant subtypes of glioblastoma characterized by abnormalities in PDGFRA, IDH1, VEGFR, and NF1. Cancer Cell 2010;17(1):98-110.

GBM: Molecular Sub-classes

The Cancer Genome Atlas (TCGA) researchers have recently cataloged recurrent genomic abnormalities in GBM providing a platform for better understanding of the molecular basis of these aggressive but heterogeneous tumors.

In parallel the Cancer Imaging Program is retrospectively obtaining radiological imaging data for TCGA patients and making it available via The Cancer Imaging Archive (TCIA).

TCIA is a large and growing archive service providing DICOM images for use in research

TCGA-GBM collection currently includes over 170 de-identified glioblastoma subjects with accrual still ongoing

TCGA and TCIA

http://cancerimagingarchive.net

Radio-genomics

•Integration of this vast genomic information with imaging data (radio-genomics)

•May provide an opportunity to use some of the non-invasive imaging features or parameters as biomarkers.

A limited number of publications on this topic have correlated morphologic imaging features (presence or absence of contrast enhancement) with various gene expression pathways affecting tumor cell mitosis, migration, angiogenesis, hypoxia, edema and apoptosis.

Radio-genomics: Morphologic Features

Proc Natl Acad Sci U S A 2008;105(13):5213-5218.Diagn Mol Pathol 2006;15(4):195-205.Radiology 2008;249(1):268-277.

Barajas et al correlated histologic features with ADC and rCBV estimates, but the authors did not directly correlate physiologic measures with gene expression. Radiology 2010;254(2):564-576.

Pope et al correlated ADC histogram analysis with differential gene expression. AJNR published online

Jain et al correlated perfusion parameters (tumor blood volume and permeability) with angiogenesis related gene expression in GBM. AJNR published online

Radio-genomics: Functional Imaging Markers

Purpose

The purpose of this study was to correlate tumor blood volume, measured using DSC T2* magnetic resonance (MR) perfusion with patient survival and also correlate it with molecular sub-classes of GBM.

Material and Methods: Patient Population

•57 patients with treatment naïve GBM underwent DSC T2* MR perfusion studies at 2 different institutions and were selected from TCIA’s TCGA-GBM collection. •50 patients had gene expression data available from TCGA.

•35 patients at Institute 1 HFH •3.0 T scanner, n=14•1.5 T scanner, n=21

•15 patients at Institute 2 UCSF• 1.5 T scanner, n=15

•According to those TCGA requirements, the pathology was confirmed as GBM using adequate frozen tissue ≥ 0.5 g consisting of ≥70% tumor nuclei and < 50% necrosis. 

https://wiki.cancerimagingarchive.net/display/Public/TCGA-GBM

Material and Methods: MRI Image acquisition

Institute 1 Institute 2

0.1 mmol/kg Gd-DTPA, 5ml/sec 0.1 mmol/kg Gd-DTPA, 5ml/sec

95 phases of GE T2* 60 phases of GE T2*

TR = 1900 msec, TE =40 msec, and flip angle =90°

TR = 2000 msec, TE =54 msec, and flip angle =30°

Temporal resolution 2.0 sec Temporal resolution 2.0 sec

Matrix size128 x 128, 26-cm FOV

Matrix size128 x 128, 26-cm FOV

Slice thickness 5 mm Slice thickness 3-6 mm

Material and Methods: MRP Post-processing

•Studies from both institutions were processed using NordicICE software (NordicImagingLab AS) using the FDA approved DSCT2* perfusion module.

•The module corrects for contrast agent leakage from the intravascular to extracellular space using the method published by Boxerman et al AJNR Am J Neuroradiol 2006;27(4):859-867.

•Normalized relative cerebral blood volume (rCBV) maps with leakage correction were produced by the software, which normalizes the CBV relative to a globally determined mean value.

Material and Methods: Image Analysis

•ROIs were drawn on the rCBV maps fused with post-contrast T1W images and FLAIR images.

•rCBV mean the contrast enhancing portion of the tumor (excluding any areas of necrosis and vessels)

•rCBV max 10 x 10 voxel ROI was placed on the hottest appearing part of the tumor

•rCBVNEL 3, 10 x 10 voxel ROIs on non-enhancing FLAIR abnormality within 1 cm of the edge of the enhancing lesion

Comparison of average rCBV measures between groups was done using two-sample t-tests or one-way ANOVA.

Kaplan-Meier estimation was used to calculate median survival and for some univariate testing.

Survival analysis with Cox proportional hazards models was employed primarily to estimate hazard ratios and for testing multivariable models.

Material and Methods: Statistical Analysis

Potential covariates in the multivariable models– Patient age at diagnosis (years, continuous), – MR scanner type (1.5T, 3T)

– molecular classification (Verhaak or Phillips) Huse J et al. Glia 2011;59(8):1190-1199.

– Karnofsky performance score (KPS, continuous)– level of resection (gross-total, sub-total)

Age and scanner were not significant predictors and did not enhance the models so they were excluded from the presented models for the sake of parsimony given the sample size.

KPS and resection data were only available for samples from institution 1.

Material and Methods: Statistical Analysis

Results: rCBV analysis using molecular sub-classification

rCBVmean rCBVmax rCBVNEL

Verhaak p=0.66 p=0.95 p=0.43

Classical (n=10) 2.66 (0.78) 4.55 (0.76) 0.66 (0.24)

Mesenchymal (n=17) 2.61 (1.26) 4.80 (1.49) 0.88 (0.45)

Neural (n=11) 2.30 (0.84) 4.68 (0.95) 0.81 (0.27)

Proneural (n=12) 2.27 (0.68) 5.06 (3.61) 0.84 (0.26)

Phillips p=0.32 p=0.57 p=0.70

Mesenchymal (n=24) 2.68 (1.16) 4.76 (1.33) 0.83 (0.40)

Proneural (n=20) 2.32 (0.72) 5.03 (2.79) 0.83 (0.25)

Proliferative (n=6) 2.15 (0.59) 4.04 (0.65) 0.70 (0.30)

Results: Survival analysis using Verhaak sub-classification

• Median overall survival (OS) was 1.14 years (IQR: 0.49, 2.11). • When the Verhaak classification scheme was applied to these samples, the classical

sub-class had the best survival, with median of 2.13 years (IQR: 1.53, 2.59) and the proneural sub-class had the worst survival with median 0.41 years (IQR: 0.65, 1.19).

Verhaak et al

Cancer Cell 2010;17(1):98-110Present study

Results: Survival analysis using Verhaak sub-classification

The difference in survival by Verhaak sub-classification was significant between groups with the difference being more prominent earlier during follow-up (Wilcoxon P=0.0445, log-rank P=0.0696).

However, the proneural subclass also had the worst median survival (0.94 years, IQR: 0.78, 1.23) in the publication by Verhaak et al.

Verhaak et al

Cancer Cell 2010;17(1):98-110Present study

Results: Survival analysis using Phillips sub-classification

• There was no evidence that the Phillips classification was associated with survival in our sample (log-rank P=0.6432, Wilcoxon P=0.4548).

• The best median survival is attributed to the mesenchymal sub-class with 1.28 years (IQR: 0.61, 2.22), followed closely by the proneural subclass with 1.12 years (IQR: 0.33, 1.86).

Phillips et al

Cancer Cell 2006;9(3):157-173Present study

Results: Survival analysis using Phillips sub-classification

• The proliferative sub-class had the worst median survival at only 0.54 years (IQR: 0.34, 3.96) but this class was only represented by six patients (five deaths), one of whom was still surviving at 3.96 years

Phillips et al

Cancer Cell 2006;9(3):157-173Present study

Results: Survival analysis using only rCBV measures

Parameters Mean Max NEL

Model 1: rCBV 1.25 (0.1918) 1.24 (0.0131) 2.45 (0.0555)

Model 2: rCBV + 1.46 (0.0212) 1.24 (0.0062) 2.56 (0.0704)

Verhaak (0.0250) (0.0476) (0.0917)

Classical 0.21 0.26 0.30

Mesenchymal 0.43 0.48 0.48

Neural 0.44 0.41 0.55

Proneural 1.00 1.00 1.00

Model 3: rCBV + 1.27 (0.1670) 1.24 (0.0152) 2.51 (0.0566)

Phillips (0.5892) (0.6888) (0.6533)

Mesenchymal 0.72 0.79 0.74

Proliferative 0.98 1.02 0.87

Proneural 1.00 1.00 1.00

Results: Survival analysis using rCBV and molecular sub-classification

Parameters Mean Max NEL

Model 1: rCBV 1.25 (0.1918) 1.24 (0.0131) 2.45 (0.0555)

Model 2: rCBV + 1.46 (0.0212) 1.24 (0.0062) 2.56 (0.0704)

Verhaak (0.0250) (0.0476) (0.0917)

Classical 0.21 0.26 0.30

Mesenchymal 0.43 0.48 0.48

Neural 0.44 0.41 0.55

Proneural 1.00 1.00 1.00

Model 3: rCBV + 1.27 (0.1670) 1.24 (0.0152) 2.51 (0.0566)

Phillips (0.5892) (0.6888) (0.6533)

Mesenchymal 0.72 0.79 0.74

Proliferative 0.98 1.02 0.87

Proneural 1.00 1.00 1.00

Results: Survival analysis using rCBV and molecular sub-classification

Parameters Mean Max NEL

Model 1: rCBV 1.25 (0.1918) 1.24 (0.0131) 2.45 (0.0555)

Model 2: rCBV + 1.46 (0.0212) 1.24 (0.0062) 2.56 (0.0704)

Verhaak (0.0250) (0.0476) (0.0917)

Classical 0.21 0.26 0.30

Mesenchymal 0.43 0.48 0.48

Neural 0.44 0.41 0.55

Proneural 1.00 1.00 1.00

Model 3: rCBV + 1.27 (0.1670) 1.24 (0.0152) 2.51 (0.0566)

Phillips (0.5892) (0.6888) (0.6533)

Mesenchymal 0.72 0.79 0.74

Proliferative 0.98 1.02 0.87

Proneural 1.00 1.00 1.00

Hemodynamic tumor measures (rCBV) using MR perfusion did not have any significant correlation with the various sub-classes using the two most commonly accepted molecular sub-classification systems of GBM.

But rCBV measures did provide important prognostic information, as patients in whom the tumor had higher rCBV showed worse prognosis and poor survival.

Conclusions

Verhaak sub-classification schema remained significant in the survival models providing additional survival information about rCBVmean measurements.

Molecular mapping of GBM can provide important therapy targets by providing insight into the molecular basis for tumor cell origin; however, in vivo imaging markers (such as rCBV measures) can provide important prognostic information that may be used as an adjunct to genomic markers in future.

Conclusions

rajanj@rad.hfh.edu