Transcript ISMB powerpoint presentation.

KELLER: Estimating Time Evolving
Interactions Between Genes
Le Song
[email protected]
Joint work with Mladen Kolar and Eric Xing
Transient Biological Processes
2
PPI Network
3
3
Time-Varying Interactions
4
The Big-Picture Questions
 What are the interactions?
 What pathways are active at a particular time
point and location?
 How will biological networks respond to
stimuli (eg. heat shot)?
5
Regulation of cell response to stimuli is
paramount, but we can usually only measure
(or compute) steady-state interactions
Transcriptional
interactions
▲ Chromatin IP
▲ Microarrays
Protein—protein
interactions
▲ Protein coIP
▲ Yeast two-hybrid
Biochemical
reactions
▲ Metabolic flux
measurements
6
Current Practice
Static Networks
Microarray Time Series
t=1
2
3
T
…
Dynamic
Bayesian
Networks
7
Our Goal
 Reverse engineer temporal/spatial-specific “rewiring”
gene networks
--- what are the difficulties?
t*
n=1
Time
8
Two Scenarios
Smoothly evolving networks
Abruptly changing networks
9
Scenario I (This paper)
 Kernel reweighted L1-regularized logistic regression
(KELLER)
 Key Idea I: reweighting observations
 Key Idea II: regularized neighborhood estimation
10
Key Idea
 Weight temporally adjacent observations more than
distal observations
11
Key Idea
 Estimate the neighborhood of each gene separately
via L1-regularized logistic regression
L1-regularization
Kernel
Reweighting
Log-likelihood
12
Consistency
 Theorem 1: Under certain verifiable conditions
(omitted here for simplicity), KELLER recovers the
true topology of the networks:
13
Synthetic data
Number of Samples
DBN and static networks do not benefit
from more observations
14
Senario II
 Key idea: Temporally Smoothing
…
TESLA:
 Tesla (Amr and Xing, PNAS 2009)
15
Drosophila Life Cycle
 66 microarrays across
Adult
full life cycle
Embryo
 588 genes related to
development
Pupa
Larva
16
biological process
molecular
function
cellular
component
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
Network Size vs. Clustering Coefficient
mid-embryonic
mid-pupal
40
Network Size vs. Clustering Coefficient
mid-embryonic stage
tight local clusters
mid-pupal stage
loose local clusters
41
Interactivity of Gene Sets
27 genes based on ontology
42
Interactivity of Gene Sets
25 genes based on ontology
43
Transient Gene Interactions
Time
Gene Pairs
Active
Inactive
msn  dock
sno  Dl
44
Transcriptional Factor Cascade
Summary networks 36 transcription factors
Node size
its total activity
45
TF Cascade – mid-embryonic stage
46
TF Cascade – mid-larva stage
47
TF Cascade – mid-pupal stage
48
TF Cascade – mid-adult stage
49
Transient Group Interactions
50
Conclusion
 KELLER for reverse engineering “rewiring” networks
 Key advantages:
 Computationally efficient (scalable to 104 genes)
 Global optimal solution is attainable
 Theoretical guarantee
 Glimpse to temporal evolution of gene networks
 Many interactions are rewiring and transient
 Availability: http://www.sailing.cs.cmu.edu/
51
The End
 Thanks
 Travel fellowship:
Office of Science (BER), U.S. Department of
Energy, Grant No. DE-FG02-06ED64270
 Funding: Lane Fellowship,
 Questions?
52
Interactivity of Gene Sets
30 genes based on ontology
53
Timing of Regulatory Program
Galactose
54
Challenges
 Very small sample size
 Experimental data are scarce and costly
 Noisy measurement
 More genes than microarrays
 Complexity regularization needed to avoid over-
fitting
 Observations no longer iid since the networks are
changing!
55