Transcript PPT - Stockholm Bioinformatics Center

Expression signatures as
biomarkers: solving
combinatorial problems with
gene networks
Andrey Alexeyenko
Department of Medical Epidemiology and
Biostatistics, Karolinska Institute
FunCoup is a data integration framework to discover
functional coupling in eukaryotic proteomes with
data from model organisms
Bmouse
Human
Rat
Fly
Yeast
High-throughput
evidence
?
Find orthologs
Amouse
Andrey Alexeyenko and Erik L.L. Sonnhammer. Global networks of functional
coupling in eukaryotes from comprehensive data integration. Genome Research.
Published in Advance February 25, 2009
FunCoup
•
•
•
•
Each piece of data is evaluated
Data FROM many eukaryotes (7)
Practical maximum of data sources (>50)
Predicted networks FOR a number of
eukaryotes (10…)
• Organism-specific efficient and robust
Bayesian frameworks
• Orthology-based information transfer and
phylogenetic profiling
• Networks predicted for different types of
functional coupling (metabolic, signaling
etc.)
http://FunCoup.sbc.su.se
TGFβ <-> cancer pathway cross-talk
FunCoup was queried for any links between members of
TGFβ pathway (left blue circle) and habituées of known
cancer pathways (members of at least 7 out of 18
groups; right blue circle). MAPK1 and MAPK3 belonged
to both categories.
http://FunCoup.sbc.su.se
FunCoup: recapitulation of known cancer
pathways
Figure 5 from:
The Cancer Genome Atlas Research Network
Comprehensive genomic characterization defines human
glioblastoma genes and core pathways.
Nature. 2008 Sep 4. [Epub ahead of print]
The same genes submitted to FunCoup
No TCGA data were used.
Outgoing links are not shown.
×
Outcome,
Optimal treatment,
Severity/urgency
etc.
Single molecular markers are (often) far from perfect.
Combinations (signatures) should perform better.
The problem:
How to select optimal combinations?
Biomarker discovery in network context
The idea:
Construct multi-gene predictors with
regard to network context
• Reduce the computational complexity
• Make marker sets biologically sound
Accounting for network context is taking either:
a) network neighbors or
b) genes at remote network positions
Procedure
“Rotterdam” dataset (Wang et al., 2005): 286 patients
Clinical data:
Expression:
~22000 probes
×
Estrogen receptor status: +/ –
Lymph. node status: all –
Relapse : yes/no and time (days)
Individual probe p-values (~22000):
Estrogen receptor-specific ability to predict relapse
Select most significant probes (1000):
Candidate members for marker signatures
Compile set of probes:
N probes at a time (e.g. N=20 or N=50)
1.
Split data: 75% to train, 25% to test.
2.
Produce a linear regression equation (weight terms step-wise, reward for
performance, penalize for complexity) on the train sub-set.
3.
Apply the equation to the test set to predict outcome (relapse yes/no).
Repeat m times4. Record the specificity/sensitivity (Type I/II error rates) as ROC curve.
RELAPSE = γ1g1 + γ2g2 + γ3g3 + … + γNgN
Procedure
Select most significant probes (1000):
Candidate members for marker signatures
Compile set of probes:
N probes at a time (e.g. N=20 or N=50)
Test X randomly
retieved sets
Take the best ones
Account for the
network context
1.
Split data: 75% to train, 25% to test.
2.
Produce a linear regression equation (weight terms step-wise, reward for
performance, penalize for complexity) on the train sub-set.
3.
Apply the equation to the test set to predict outcome (relapse yes/no).
Repeat m times4. Record the specificity/sensitivity (Type I/II error rates) as ROC curve.
RELAPSE = γ1g1 + γ2g2 + γ3g3 + … + γNgN
Candidate signature in the network
Biomarker candidates
Ready signature in the network
RELAPSE = γ1EIF3S9+ γ2CRHR1 + γ3LYN + … + γNKCNA5
Testing “top”, “free”, and “network”
approaches
Estrogen receptor status:positive
Frequency
netw
free
Top
Estrogen receptor status:negative
netw
free
91%
92%
93%
94%
Quality of prognosisrelapse/no relapse
Top
95%
96%
97%
(area under ROC curve)
Frequency
90%
93%
94%
95%
96%
97%
Quality of prognosisrelapse/no relapse
98%
99%
(area under ROC curve)
Signature involves genes mutated in cancer
Cancer individuality:
each tumor is unique in its molecular state and set of
mutated/disordered genes
Tumour tcga-02-0114-01a-01w
Partial correlations:
a way to get rid of spurious links
0.7
0.6
0.4
Cancer individuality via network view
Functional coupling
transcription ? transcription
transcription ? methylation
methylation
? methylation
mutation

methylation
mutation

transcription
mutation
? mutation
+
mutated gene
is a framework for biomarker discovery:
• Markers can be discovered and presented in the
network dimension.
• Choice of data types to incorporate is unlimited –
from metabolite profiling to patient phenotypes.
Useful features:
• Web-based resource ready for further expansion
and presenting new research results in an interactome
perspective;
• Cross-species network comparison of human and
model organisms.
• Efficient query system to retrieve network
environments of interest.
http://FunCoup.sbc.su.se
Thank you for attention!
Decomposing biological context
Develomental
Common
rPLC = 0.95
rPLC = 0.88
ANOVA
(Analysis Of VAriance):
Look at F-ratios:
Signal of interest /
Residual (“error”) variance
Dioxin-enabled
rPLC = 0.76
Accounting for edge features:
dioxin-enabled vs. dioxin-sensitive links
Andrey Alexeyenko, Deena M Wassenberg, Edward K Lobenhofer, Jerry Yen, Erik LL
Sonnhammer, Elwood Linney, Joel N Meyer Transcriptional response to dioxin in
the interactome of developing zebrafish. submitted.
a