1. Radiogenomics in glioblastoma multiforme
Olivier Gevaert, PhD
Stanford University
Cancer Center for Systems Biology
Information Sciences in Imaging Program

2. Overview Multi-omics data integration
Quantitative image features in GBM
Integrating quantitative image features and module networks

3. Multi-omics data integration
Discovering cancer driver genes and their targets

4. Integrative methods
Need to develop statistical methods that allow to integrate multi-omics cancer data
Create mechanistic models of cancer
how is gene expression influenced by genomic events
how to identify cancer drivers and their targets

5. Public domain data Gene & miRNA expression Copy number DNA Methylation
Agilent & Affy microarray
RNA sequencing
Copy number
Affy SNP 6.0
DNA Methylation
Agilent Infinium (27k)
Mutation
DNA sequencing
Medical Images (MRI)

6. Method Two step algorithm Generating the List of Candidate Drivers
Associating Candidate Drivers with their Downstream Targets
Step 2: Using an idea that was developed at Stanford by Daphne called module networks

7. Step 1 Generating the List of Candidate Drivers
If gene expression can be explained by genomic events
Copy Nr
Methylation
Mutation
Gene
candidate driver gene
Expression
Rationale:
Genes driven by multiple genomic events in a significant subset of samples are unlikely to be randomly deregulated.
De redenering hier is dat het heel onwaarschijnlijk is dat als een gen significant wordt beinvloed door zowel copy number, methylatie als mutatie, dat dit toevallig gebeurd.

8. Step 1 Generating the List of Candidate Drivers
Gene expression – DNA methylation correlation
Gene expression – Copy number correlation
Model gene expression as a function of
Copy number
DNA methylation
Mutation data
Incorporating prior knowledge
β1 has to be positive
β2 has to be negative
Pairwise Spearman correlation

9. Step 2: Associating Candidate Drivers with their Downstream Targets
genes
patients
Driver
genes from Step 1
Known transcription factors
clustering
clusters
Potential
Regulators Ri
Module
FOXM1
Cluster 37
E2F8
Linear Regression +
Lasso regularization
Lee et al. PLoS Genetics, 2009

10. Results Step 2: Module network
GBM
Gevaert O and Plevritis S. PSB 2013

11. Results Step 2: Module network
GBM
RBMS1, PLAGL1, MAML2, PNRC1 are part of module 74 and inhibiting module 18
Gevaert O and Plevritis S. PSB 2013

12. Results Step 2: Module network
GBM
Gevaert O and Plevritis S. PSB 2013

13. GBM Network Top DNA repair module DNMT1 PARP1 CHAF1B
Key DNA methylation driver gene
PARP1
binds DNMT1
known function in DNA damage
CHAF1B
putative gene involved in DNA repair
cross-cancer
Top regulator in ovary & breast for DNA repair
Gevaert O and Plevritis S. PSB 2013

14. Summary Integration of multi-omics GBM data incorporating
gene expression
DNA methylation
copy number data
Reduces molecular data dimensions
reduces the multiple testing correction problem
Rich model going beyond clustering
downstream targets allow to check for enrichment analysis and annotate the regulator genes
Gevaert O and Plevritis S. PSB 2013

15. Quantitative image features
Tissue level data

16. GBM Medical Image Data 152 TCGA GBM patients have image data
Imaging data is very heterogeneous
different planes
noisy annotation
different pulse sequences
T1 pre gadolinium
T1 post gadolinium …
Microarrary allows to measure the expression in human genome. (25000 genes)

17. GBM Medical Image Data

18. GBM Medical Image Data AXIAL images
T1 pre gadolinium
T1 post gadolinium
T2 FLAIR
Highlight different parts of the tumor
necrosis
enhancement
edema
Manual annotation of ROIs for these concepts 2D
largest slice for that lesion
for multifocal lesions ROI for each

20. GBM Medical Image Data Results
at least one ROI for 55 patients
~40 patients have all ROI types
Three ROI types are matched according to location to create a super ROI
If multiple super ROIs, features are combined
Two readers + some redundancy to estimate intra-reader variability

21. Quantitative image features
Create a set of quantitative image features for each ROI
Same feature set as Lung Cancer project
Gevaert et al. Radiology 2012
Computational features
iPAD

22. Quantitative image features
Texture Features
Shape Features
Edge Sharpness Features
Used curvature along the boundary
difference in intensities of tissue surrounding the tumor and the tumoritself
Window (W)
Scale (S)
Gabor Filter Bank
150 Computational features
Jiajing Xu, Sandy Napel

23. Quantitative image features
Computational features
iPAD
iPAD
Computational features
computational features
Texture Feature
Shape
Feature
Edge Sharpness
150-element feature vector

24. Quantitative image features
Focused on 28 highly interpretable quantitative image features
Compactness of ROI
Edge sharpness
Edge Shape (LAII)

25. Explorative analysis

26. Size metrics of necrosis, enhancement and edema
Created ratios of the size of each ROI vs. larger ROI
necrosis/enhancement
necrosis/edema
enhancement/edema

27. Size metrics of necrosis, enhancement and edema
Comparison with the VASARI features

28. Size metrics of necrosis, enhancement and edema
Comparison with the VASARI features

29. Size metrics of necrosis, enhancement and edema
Comparison with the VASARI features

30. Size metrics of necrosis, enhancement and edema
Correlated these with overall & progression free survival
no significant correlation with any survival outcome
Potential problems
small data set, not enough power (<55 samples)
how to combine multi-focal lesions
multi-focal necrosis
multi-focal enhancement
2D vs. 3D
Interesting
size of edema is weakly correlated with progression free survival (p-value 0.03)

31. Univariate survival analysis
Quantitative image feature
ROI type
Wald Test HR
HR lower
HR upper
RDS std
Enhancement
1.6809
1.1925
2.3692
Edge sharpness window kurtosis
Necrosis
1.4889
1.0976
2.0196
Compactness
1.4306
1.0632
1.925
Edge sharpness window skewness
1.4155
1.0485
1.9111
LAII std-5R
1.4792
1.0541
2.0757
LAII std-8R
1.5828
1.06
2.3636
Edge Sharpness scale max
Edema
1.49
1.0509
2.1125
Edge sharpness scale mean
1.4692
1.0484
2.059
1.4695
1.0448
2.0669
RDS mean
1.4878
1.029
2.1514
1.4827
1.0239
2.147
1.4493
1.0215
2.0562

32. Creating a radiogenomic map

33. Radiogenomics map Necrosis

34. Radiogenomics map Necrosis
Compactness of Necrosis ROI
high = irregular shape
low = spherical shape
Correlated with Module 64
P-value (Spearman rho, FDR 4%)
Inverse correlation
Low
compactness
High
compactness

35. Radiogenomics map Necrosis
Edge sharpness window necrosis
high = blurry edge
low = sharp edge
Correlated with Module 10
P-value
Inversely correlated
Sharpe
edge
Blurry
edge

36. Overall summary
Developed module network method that integrates/summarizes multi-omics data
Gathered quantitative image features from MRI image data
Correlated quantitative image features with modules

37. Acknowledgements Sylvia Plevritis Greg Zaharchuk Sandy Napel
Lex Mitchell
Caroline Yu
Jiajing Xu
Chris Beaulieu

38. Questions Pulse sequence annotations
T1 Post is missing in many samples
Readers interested in annotating ROIs