Transcript vipinICCABS_ver4x - University of Minnesota

Discovering Combinatorial Biomarkers
Vipin Kumar
[email protected]
http://www.cs.umn.edu/~kumar
Department of Computer Science and Engineering
ICCABS, Feb 2012
High-throughput technologies
Clinical Data e.g.
brain imaging
SNP
Structural Variation
DNA Methylation
Gene Expression &
non-coding RNA
Metabolites
Proteins
Adopted from E. Schadt

Data mining offers potential solution for analysis of these large-scale datasets
•
•
•
Novel associations between genotypes and phenotypes
Biomarker discovery for complex diseases
Personalized Medicine – Automated analysis of patients history for customized treatment
2
Biomarker Discovery and its Impact
Biomarkers:
Genes:
BRCA1 (breast cancer)
Protein variants
IVS5-13insC (type 2 diabetes)
Pathways/networks:
P53 (cancers)
Clinical Impact:
Diagnosis
Prognosis
Treatment
fMRI
Schizophrenia
vs controls
Lim et al.
Miki et al. 1994
Chiefari et al. 2011
Oren et al. 2010
3
Published Genome-wide Associations through 06/2010
1,904 published GWA at p≤5*10-8 for 165 traits
SNP as an
illustration
NHGRI GWA Catalog
www.genome.gov/GWAStudies
4
Published Genome-wide Associations through 06/2011
1,449 published GWA at p≤5*10-8 for 237 traits
50% increase in one year
SNP as an
illustration
NHGRI GWA Catalog
www.genome.gov/GWAStudies
5
Challenge: Limitations of Single-locus Association Test
High coverage but
low odds ratio (1.2)
High odds ratio (15.9)
but low coverage (7%)
Many other studies
No significant
associations
6
A Example where Single-locus Test Led to No Significant Associations
•
Given a SNP data set of Myeloma patients, find SNPs that are
associated with short vs. long survival.
3404 SNPs
•
•
•
3404 SNPs selected from various
regions of the chromosome
70 cases (Patients survived shorter
than 1 year)
73 Controls (Patients survived
longer than 3 years)
Myeloma
Survival Data
cases
Controls
Van Ness et al 2008
Top ranked SNP:
-log10P-value = 3.8; Odds Ratio = 3.7
Myeloma SNP data has signal  the need of
discovering combinations of SNPs
7
Single-locus Tests Ignore Genetic Interaction
Non-additive effect “Genetic Interaction”
Ripke et al. 2011
Extensively observed in model
organisms, e.g. yeast, C. elegans, fly.
Costanzo et al. 2010
Scholl et al. 2009
Ruzankina et al. 2009
Kamath, 2003
8
The focus of this talk:
Higher-order Combinatorial Biomarker
......
Complex biological system
Complex human diseases
Higher-order genetic buffering
Triple mutations only exist
in disease subjects
Control
Disease
A synthetic pattern
9
Discovering High-order Combinatorial Biomarkers
Challenge I: Computational Efficiency
Given n features, there are 2n candidates!
The Apriori framework for efficient
search of exponential space
How to effectively handle the
combinatorial search space?
Millions of user,
thousands of items
Brute-force search e.g. MDR can only
handle 10~100 SNPs. [Rita et al. 2001]
Support based pruning
null
Disqualified
A
B
C
D
E
AB
AC
AD
AE
BC
BD
BE
CD
CE
DE
ABC
ABD
ABE
ACD
ACE
ADE
BCD
BCE
BDE
CDE
ABCD
ABCE
ABDE
ABCDE
Prune all the supersets
ACDE
BCDE
+
+
[Agrawal et al. 1994]
10
Discovering High-order Combinatorial Biomarkers
Challenge I: Computational Efficiency
• Traditional Apriori-based pattern mining techniques
• Designed for sparse data
• Unique challenges of genomic datasets
• High density
• A SNP dataset has a density of 33.33%
• Three binary columns per SNP  the three genotypes
• High dimensionality
• Makes the search more challenging
• Disease heterogeneity
• Each combination supported by a small fraction of subjects
A novel anti-monotonic objective function designed for
mining low-support discriminative patterns from dense
and high-dimensional data
[Fang et al. TKDE 2010]
11
Discovering High-order Combinatorial Biomarkers
Challenge II: Statistical Power
•
null
A
B
C
D
E
Computational challenges can be addressed by
• Better algorithm design,
•
e.g. Apriori-based
• High-performance computing
AB
AC
AD
AE
BC
BD
BE
CD
CE
DE
•
ABC
ABD
ABE
ABCD
ACD
ABCE
ACE
ADE
ABDE
BCD
ACDE
BCE
BCDE
BDE
CDE
Statistical challenges call for additional efforts
• Limited sample size
• Huge number of hypothesis tests
Many combinations are trivial extensions of their subsets
ABCDE
Myeloma Survival Data
Kidney Rejection Data
Lung Cancer Data
Subsets
having lower
association
Subsets
having higher
association
Targeting patterns with better association than their subsets reduces # of hypothesis tests
[Fang, Haznadar, Wang, Yu, Steinbach, Church, Oetting, VanNess, Kumar, PLoS ONE, 2012]
12
High-order Combinatorial Biomarkers: an example
Patients
Size-5
Best
Best size-4
Best size-3
Best size-2
size-1
Control
All heavy smokers
Lung Cancer Data
Jump
Data from Church et al. 2010
The five genes are
functionally related
www.ingenuity.com
[Fang, Pandey, Wang, Gupta, Steinbach and Kumar, IEEE TKDE, 2012]
[Fang, Haznadar, Wang, Yu, Steinbach, Church, Oettng, Van Ness and Kumar, PLoS ONE, 2012]
13
Insights on High-order Functional Interactions
Patterns with positive Jump are functionally more coherent
Lungcancer
Lung cancer dataset
Size-5
Control
Best
Best size-4
Best size-3
Best size-2
size-1
Kidney Rejection Data
Lung Cancer Data
Jump
Combined
14
[Fang, Haznadar, Wang, Yu, Steinbach, Church, Oettng, Van Ness and Kumar, PLoS ONE, 2012]
High-order Combinations Discovered from Different Types of Data
mRNA: Breast Cancer
Data from Oetting et al. 2008
AE COPD
Metabolites: COPD
Stable COPD
Control
No-rejection
Rejection
Survived (5-year)
SNP: acute kidney rejection
Data from Vijver et al. 2002
Data from Wendt et al. 2010
The proposed framework is general to handle different types of data
[Fang, Pandey, Wang, Gupta, Steinbach and Kumar, IEEE TKDE, 2012]
[Fang, Haznadar, Wang, Yu, Steinbach, Church, Oettng, Van Ness and Kumar, PLoS ONE, 2012]
15
Biomarker Discovery using Error-tolerant Patterns
 True patterns are fragmented due to noise
and variability
 Possible solution: Error-tolerant patterns
•
X
0
1
1
1
0
0
1
0
1
0
1
1
0
0
0
1
1
1
0
1
1
0
0
0
1
1
1
0
0
0
0
0
0
0
1
0
1
1
1
1
1
0
0
1
1
1
1
1
0
0
0
0
1
1
1
1
0
0
0
0
1
1
1
1
√
These patterns differ in the way errors/noise in
the data are tolerated
[Yang et al 2001]; [Pei et al 2001]; [Seppanen et al 2004]; [Liu et al
2006]; [Cheng et al 2006]; [Gupta et al., KDD 2008]; [Poernomo et al
2009]
See Gupta et al KDD 2008 for a survey
16
Error-tolerant pattern vs. Traditional association patterns
 Four Breast cancer gene-expression data sets are used for experiments:
158 cases
+
+
GSE7390
GSE6532
+
GSE3494
GSE1456
433 controls
 Cases: patients with metastasis within 5 years of follow-up;
 Controls: patients with no metastasis within 8 years of follow-up
 Discriminative Error-tolerant and traditional association patterns case/control
are discovered and evaluated by enrichment analysis using MSigDB gene
sets
 Greater fraction of errortolerant patterns enrich at
least one gene set (higher
precision)
 Greater fraction of gene
sets are enriched by at
least one error-tolerant
pattern (higher recall)
Gupta et al. BICoB 2010; Gupta et al. BMC Bioinformatics 2011
Error-tolerant patterns
Traditional patterns
Error-tolerant patterns
Traditional patterns
17
Differential Coexpression Patterns
•
Differential Expression (DE)
– Traditional analysis targets
changes of expression level
[Silva et al., 1995], [Li, 2002], [Kostka & Spang, 2005],
[Rosemary et al., 2008], [Cho et al. 2009] etc.
•
Differential Coexpression
(DC)
– Changes of the coherence of
gene expression
[Eisen et al. 1999] [Golub et al., 1999], [Pan 2002],
[Cui and Churchill, 2003] etc.
•
•
Combinatorial Search
Genetic Heterogeneity
– calls for subspace analysis
18
Subspace Differential Coexpression Analysis
Enriched with the TNF-α/NFkB signaling pathway
(6/10 overlap with the pathway, corrected p value: 1.4*10-3)
≈ 10%
Suggests that the dysregulation of TNF-α/NFkB
≈ 60%
pathway may be related to lung cancer
Three lung cancer datasets [Bhattacharjee et al.
2001], [Stearman et al. 2005], [Su et al. 2007]
[Fang, Kuang, Pandey, Steinbach, Myers and Kumar, PSB 2010]
Selected for highlight talk, RECOMB SB 2010
Best Network Model award, Sage Congress, 2010
Combinatorial Biomarkers: Summary
• Higher-order combinations
• Important for understanding complex human diseases
• A novel framework
• Improved computational efficiency
• Enhanced statistical power
• Naturally handles disease heterogeneity
• Error-tolerance
• Different types of differentiation: coexpression
• General to handle different types of data
• SNP
• Gene expression
• Metabolomic data
• Brian imaging data (e.g. fMRI)
20
References
• G. Fang, R. Kuang, G. Pandey, M. Steinbach, C.L. Myers, and V. Kumar. Subspace differential coexpression
analysis: problem definition and a general approach. Pacific Symposium on Biocomputing, 15:145-156, 2010.
• G. Fang, G. Pandey, W. Wang, M. Gupta, M. Steinbach, and V. Kumar. Mining low-support discriminative patterns
from dense and high-dimensional data. IEEE TKDE, 24(2):279-294, 2012.
• G. Fang, Majda Haznadar, Wen Wang, Haoyu Yu, Michael Steinbach, Tim Church, William Oetting, Brian Van Ness,
and Vipin Kumar. High-order SNP Combinations Associated with Complex Diseases: Efficient Discovery,
Statistical Power and Functional Interactions. PLoS ONE, page in press, 2012.
• R. Gupta, N. Rao, and V. Kumar. Discovery of errortolerant biclusters from noisy gene expression data. In BMC
Bioinformatics, 12(S12):S1, 2011.
• R. Gupta, Smita Agrawal, Navneet Rao, Ze Tian, Rui Kuang, Vipin Kumar, "Integrative Biomarker Discovery for
Breast Cancer Metastasis from Gene Expression and Protein Interaction Data Using Error-tolerant Pattern
Mining", In Proc. of the International Conference on Bioinformatics and Computational Biology (BICoB), 2010
• Gowtham Atluri, Rohit Gupta, Gang Fang, Gaurav Pandey, Michael Steinbach and Vipin Kumar, Association
Analysis Techniques for Bioinformatics Problems, Proceedings of the 1st International Conference on
Bioinformatics and Computational Biology (BICoB), pp 1-13, 2009.
• S. Landman Vipin Kumar Michael Steinbach, Haoyu Yu. Identification of Co-occurring Insertions in Cancer
Genomes Using Association Analysis. International Journal of Data Mining and Bioinformatics, in press, 2012.
• M. Steinbach, H. Yu, G. Fang, and V. Kumar. Using constraints to generate and explore higher order
discriminative patterns. Advances in Knowledge Discovery and Data Mining, pages 338-350, 2011.
• S. Dey, Gowtham Atluri, Michael Steinbach, Angus MacDonald, Kelvin Lim, and Vipin Kumar. A pattern mining
based integrative framework for biomarker discovery. Tech report, Department of Computer Science, University
of Minnesota, (002), 2012.
• G. Pandey, C. Myers, and V. Kumar. Incorporating functional inter-relationships into protein function prediction
algorithms. BMC bioinformatics, 10(1):142, 2009.
• G. Pandey, B. Zhang, A.N. Chang, C.L. Myers, J. Zhu, V. Kumar, and E.E. Schadt. An integrative multi-network and
multi-classifier approach to predict genetic interactions. PLoS computational biology, 6(9):e1000928, 2010 (Cited
as one of the major computational biology breakthroughs of 2010 by a Nature Biotechnology feature article).
• J. Bellay, G. Atluri, T.L. Sing, K. Toufighi, M. Costanzo, P.S.M. Ribeiro, G. Pandey, J. Baller, B. VanderSluis, M.
Michaut, et al. Putting genetic interactions in context through a global modular decomposition. Genome
Research, 21(8):1375-1387, 2011.
21
Acknowledgement
Kumar Lab, Data Mining
Gang Fang
Wen Wang
Vanja Paunic
Yi Yang
Benjamin Oatley
Xiaoye Liu
Sanjoy Dey
Gowtham Atluri
Gaurav Pandey
Michael Steinbach
Myers Lab, FuncGenomics
Jeremy Bellay
Chad Myers
Kuang Lab, Compbio
TaeHyun Hwang
Rui Kuang
Masonic Cancer Center
Tim Church
Bill Oetting
Van Ness Lab, Myeloma
Brian Van Ness
Lim Lab, Brain Imaging
Kelvin Lim
McDonald Lab, Behavior
Angus McDonald
Wendt Lab, Lung Disease
Chris Wendt
Mayo Clinic-IBM-UMR fellowship, Walter Barnes Lang fellowship,
NSF: #IIS0916439, UMII seed grant, BICB seed grant,
Computations enabled by the Minnesota Supercomputing Institute.
BioMedical Genomics Center at University of Minnesota,
International Myeloma Foundation. Etiology and Early Marker Study program of the
Prostate Lung Colorectal and Ovarian Cancer Screening Trial
Thanks!
23