1. A graph-based integration of multiple layers of cancer genomics data (Progress Report)
Do Kyoon Kim 1

2. Outline Introduction Databasing TCGA Data
Graph-based Semi-Supervised Learning with Gene Expression Data

3. Introduction

4. Introduction
Although microarray technology allows the investigation of the transcriptomic make-up of a tumor
the transcriptome does not completely reflect the underlying biology due to alternative splicing, post-translational modification
This increases the importance of integration more than one source of genome-wide data, such as the genome, transcriptome, proteome, and epigenome
The current increase in the amount of available omics data emphasizes the need for a methodological integration framework

5. Introduction Data integration: different point of view
Heterogeneous data from different sources were analyzed sequentially
The term data integration has also been used as synonym for data merging in which different data sets are concatenated at the database level by cross-referencing the identifiers
Integrate multiple layers of experimental data into one mathematical model for the development of more homogeneous classifiers in clinical decision support
Daemen et al., 2009, Genome Medicine

6. The Cancer Genome Atlas (TCGA)
Mission
The Cancer Genome Atlas (TCGA) is a comprehensive and coordinated effort to accelerate our understanding of the molecular basis of cancer through the application of genome analysis technologies, including large-scale sequencing
Goal
To improve our ability to diagnose, treat and prevent cancer
A pilot project developed and tested the research framework needed to systematically explore the entire spectrum of genomic changes involved in human cancer
Focus on three selected cancer types
Serous cystadenocarcinoma (ovarian)
Squamous carcinoma (lung)
Glioblastoma multiforme (brain)
500 samples per tumor type 6

7. TCGA data
How to integrate?
TCGA research network., (2008), Nature 7

8. The second page of TCGA project8

9. Specific Goal
Problem: Prediction of recurrence in GBM patients using multiple types of genomic data
PHENOTYPE
SNP
SEQUENCE
EXPRESSION
METHYLATION
COPY NUMBER
miRNA 9

10. Biological Organization
TF binding
SNP
methylation
CNV,LOH, Del
TFbs
TFbs
TFbs
TRANSCRIPTION
CNV,LOH, Del
Gene
Gene
Gene
alternative
splicing
EXPRESSION
microRNA
microRNA
mRNA
mRNA
mRNA
TRANSLATION x
post modification
glucosylation
phosphorylation
Protein
TF Protein
FUNCTION
TF: transcription factor
TFbs: transcription factor binding site
Phenotype

11. Graph-based Learning
Recently, to integrate multiple data sources, a simidefinite programming (SDP) based SVM method was introduced
In SDP/SVM, multiple kernel matrices corresponding to each of data sources are combined
However, when trying to apply SDP/SVM to large problems, the computational cost can become prohibitive, since both converting the data to a kernel matrix for SVM and solving the SDP are time and memory demanding 11

12. Graph-based Learning
Significant progress of graph-based semi-supervised learning methods in the machine learning community
One important problem in graph-based learning, which has not yet been addressed, is the combination of multiple graphs
Each vectorial data can be incorporated after conversion into a network
Due to the sparsity of network edges, the computation time is nearly linear in the number of edges of the combined network 12

13. Graph-based integration
expression
miRNA
Methylation
CNV 13

14. Databasing TCGA data

15. Data release
Data Levels I and II correspond to raw and processed data, respectively, for each sample
Level III data are the output of basic analyses of Level I/II data, such as mutational calls of sequenced genes, copy number and LOH calls of genomic regions of aberrations, and expression level of a gene for each sample
Level IV data represent interpretations of the data, such as what genes are significantly mutated, or altered in copy number, DNA methylation, or expression across multiple samples and data types
For protection of patient privacy, access to Level I and/or II data for certain platforms (e.g. SNP genotyping) or data types (e.g. germ-line mutations) is restricted to qualified researchers and requires approval of a TCGA Data Access Committee 15

16. 16

17. Download directory structure and URL construction17

18. Retrieving available TCGA Data: Done
Cancer type: GBM
Time: about 10 days
Size: About 230 GB 18

19. Databasing TCGA data 19

20. Python scripts for inserting data into database20

21. Databasing annotation files from multiple types of platforms
Multiple types of Annotation Files -> Database
Annotation (each platforms)
ADF files (same genome build) 21

22. Insert new platforms and Experiment data
Column wise queries 22

23. Row wise queries Theoretically, queries are possible
Select all data with level 3 where gene symbol is ‘ERBB2’
Select expression, CGH, methylation with level 2 where chromosome position is ‘17q36.1’ 23

24. Statistics of data Overlap samples with tumor type = ‘solid tumor’
expression
miRNA
Methylation
CNV P:
P: 1510
P: 1498 P:
S: 273
S: 276
S: 238
S: 441 24

25. Graph-based Semi-Supervised Learning with Gene Expression Data

26. Gene expression data Data reduction
Class (Procudure_Type from Phenotype)
1: Surgical Resection
-1: Secondary Surgery for tumor recurrence/progression: locoregional procedure
expression
expression
expression
with output variable
Gene summarization P: P: P:
S: 273
S: 258
S: 258 26

27. Data plot 27

28. Graph-based SSL Without Feature selection: (258 x 12043)
W matrix: K-NN + exp-weighted graphs
K = 5
SSL
Mu = 1
5-fold cross validation
ROC score: 28

29. Feature Selection
Identify differential expressed genes from two phenotypes
T-test
Using mattest in MATLAB:
p_value < 0.05: 768
p_value < 0.01: 181
p_value < 0.001: 23 29

30. Graph-based SSL With Feature selection (p_value < 0.05)
258 x 768
W matrix: K-NN + exp-weighted graphs
K = 20
SSL
Mu = 10
5-fold cross validation
ROC score: 30

31. Future work Systematically control parameters
Other methods for making W matrix
Correlation
Tanh-weighted graphs
Any good method with large-scale features ?
Experiment with other data types
miRNA
Methylation
CNV
Combine multiple types of genomics data
ROC score improved? 31