KELLER: Estimating Time Evolving

ocelotgiantΤεχνίτη Νοημοσύνη και Ρομποτική

7 Νοε 2013 (πριν από 3 χρόνια και 9 μήνες)

70 εμφανίσεις

Le Song

lesong@cs.cmu.edu

Joint work with Mladen Kolar and Eric Xing

KELLER: Estimating Time Evolving
Interactions Between Genes

2

Transient Biological Processes

3




3

PPI Network

4

Time
-
Varying Interactions

5

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)?



6

Regulation of cell response to stimuli is
paramount, but we can usually only measure
(or compute)
steady
-
state interactions

Transcriptional
interactions
Protein

protein
interactions
Biochemical
reactions

Chromatin IP

Microarrays

Protein
coIP

Yeast two
-
hybrid

Metabolic flux
measurements
Transcriptional
interactions
Protein

protein
interactions
Biochemical
reactions

Chromatin IP

Microarrays

Protein
coIP

Yeast two
-
hybrid

Metabolic flux
measurements
7



t=1

2

3

T

Current Practice

Static Networks

Microarray Time Series

Dynamic
Bayesian
Networks

8

Our Goal


Reverse engineer temporal/spatial
-
specific “
rewiring

gene networks







Time

t
*

n
=1

---

what are the difficulties?

9

Two Scenarios

Smoothly evolving networks

Abruptly changing networks

10

Scenario
I (This paper)


Kernel reweighted L1
-
regularized logistic regression
(KELLER)







Key Idea I: reweighting observations



Key Idea II: regularized neighborhood estimation













11

Key Idea


Weight temporally adjacent observations
more

than
distal observations










12

Key Idea


Estimate the neighborhood of each gene separately
via L1
-
regularized logistic regression










Kernel
Reweighting

Log
-
likelihood

L1
-
regularization

13

Consistency


Theorem 1
: Under certain verifiable conditions
(omitted here for simplicity),
KELLER

recovers the
true topology of the networks:










14

Synthetic data

DBN and static networks do not benefit
from more observations

Number of Samples

15


Key idea:
Temporally Smoothing












Tesla (Amr and Xing, PNAS 2009)

TESLA:



Senario
II

16

Drosophila Life Cycle

Larva

Embryo

Pupa

Adult




66 microarrays across
full life cycle



588 genes related to
development


17

molecular

function

biological process

cellular

component

18

19

20

21

22

23

24

25

26

27

28

29

30

31

32

33

34

35

36

37

38

39

40

Network Size vs. Clustering Coefficient

mid
-
embryonic

mid
-
pupal

41

Network Size vs. Clustering Coefficient

mid
-
embryonic stage

tight local clusters

mid
-
pupal stage

loose local clusters

42

Interactivity of Gene Sets

27 genes based on ontology

43

Interactivity of Gene Sets

25 genes based on ontology

44

Transient Gene Interactions

Time

Gene Pairs

Active

Inactive

msn


dock

sno


Dl

45

Transcriptional Factor Cascade

Summary networks 36 transcription factors

Node size its total activity


46

TF Cascade


mid
-
embryonic stage

47

TF Cascade


mid
-
larva stage

48

TF Cascade


mid
-
pupal stage

49

TF Cascade


mid
-
adult stage

50

Transient Group Interactions

51

Conclusion


KELLER for reverse engineering “rewiring” networks



Key advantages:



Computational
ly efficient (scalable to 10
4

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/

52

The End


Thanks



Travel fellowship:


Office of Science (BER), U.S. Department of
Energy, Grant No. DE
-
FG02
-
06ED64270



Funding: Lane Fellowship,



Questions?

53

Interactivity of Gene Sets

30 genes based on ontology

54

Timing of Regulatory Program

Galactose

55

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
!