Document

advertisement
Breast cancer is a complex and
heterogeneous disease
Tumor samples
Protein expression
Transcriptional
Subtype
Clinical features
Mutational status
Adapted from TCGA, Nature 2012
Breast cancer is a complex and
heterogeneous disease
Tumor samples
Protein expression
Transcriptional
Subtype
Clinical features
Mutational status
Adapted from TCGA, Nature 2012
HPN-DREAM Challenge: How are
signaling pathways deregulated across
breast cancers?
•
Genomic and epigenomic aberrations (mutations, copy
number changes, etc) influence cancer development
•
Collection of aberrations in an individual sample create a
unique “biological context” that influences cell signaling
•
Improved understanding of network function will lead to the
development of more effective therapies
Patient
Tumor
Cell Line
High-throughput screen of protein
Data generated by Reverse
signaling dynamics
Phase Protein Array (RPPA)
MCF7
~200 Proteins
~200 Proteins
…
~200 Proteins
……
Proteins
~45 …
…
UACC812
8 Stimuli
BT20
8 Stimuli
BT549
8 Stimuli
8 Stimuli
……
…
……
…
…………
……
……
…
Time
Inhibitors
DMSO
FGFR1/3i
AKTi
AKTi+MEKi
.
.
Inhibitor N
4 cell lines x 8 stimuli = 32 biological
contexts for network prediction
Stimuli
Serum
PBS
EGF
Insulin
FGF1
HGF
NRG1
IGF1
7 Timepoints
0 5
Inhibitor
15
Stimulus
…
4
Creating a “Gold Standard” for
assessment of predictions
Training Data (4)
treatments)
FGFR1/3i
AKTi
AKTi+MEKi
DMSO
45
All Data (N treatments)
Test Data (N-4)
treatments)
45
Test1
Test2
….
TestN-4
Hold out a
subset of
inhibitor data
for assessment
of network
inference and
timecourse
predictions
Companion in silico challenge
Mimics key aspects and characteristics of the
experimental data
Generated from a dynamical signaling network model
Inferred networks can be assessed against against a
true gold standard with known network structure
Task: Create a network where nodes represent
phosphoproteins and directed edges represent
causal relationships between the nodes
Assessment: Score against held-out test data
Training Data
Predict
1A Experimental data: predict 32 context-specific networks
1B In silico data: predict 1 network
Complete submission requires both A and B parts
Using inhibitor data to infer network structure
Causal edges:
1. predict that perturbing (ie, inhibiting) parent node A will
induce change in child node B
2. are context-specific, and vary with cell line and stimulus
Cell Line 2,
Stimulus 1
Cell Line 1,
Stimulus 1
B
Node B
abundance
Control
With A inhibitor
A
B
Node B
abundance
A
Time
Time
Task: Build a dynamical model to predict
phospho-protein trajectories following inhibition
of test nodes
Assessment: Score against held-out test data
Predict
Protein
Abundance
Training Data
Time
2A Experimental data
2B In silico data
Task: Devise novel approaches to represent
high-dimensional timecourse datasets
Assessment: Crowd-based peer-review
Training Data
Submit
HPN-DREAM Challenge: Participation
 237+ registered participants
 Complete final submissions:
 SC1 Network Inference: 59
 SC2 Timecourse Pred: 10
 SC3 Visualization: 14
 Collaborative Bonus Round to foster exchange of ideas
and development of hybrid models
 Some details of assessment and test data will not be
released until after the close of the collaborative round
Analysis and scoring
Steven Hill
Thomas Cokelaer
*Sach Mukherjee
In silico data generation
Michael Unger
*Heinz Koeppl
Experimental data generation
Nicole Nesser
Katie Johnson-Camacho
Gordon Mills
Joe Gray
*Paul Spellman
Challenge organizers
Laura Heiser
Julio Saez-Rodriguez
Thea Norman
*Gustavo Stolovitzky
Synapse development
Jay Hodgson
Bruce Hoff
Mike Kellen
*Steven Friend
Heritage Provider Network
Jonathan Gluck
Poster: DREAM03
synapse.org/#!Challenges:DREAM8
Sustained response
Transient response
Inhib 3
Serum
PBS
EGF
Insulin
FGF1
HGF
NRG1
IGF1
Inhib 2
Serum
PBS
EGF
Insulin
FGF1
HGF
NRG1
IGF1
Inhib 1
Serum
PBS
EGF
Insulin
FGF1
HGF
NRG1
IGF1
DMSO
Serum
PBS
EGF
Insulin
FGF1
HGF
NRG1
IGF1
Serum
PBS
EGF
Insulin
FGF1
HGF
NRG1
IGF1
Proteins
An information rich timecourse
Inhib 4
Download