Top Banner
5/21/2013 1 Statistical methods for inferring the gene regulatory networks – Part II Lecture 2 – May 16 th , 2013 GENOME 541, Spring 2013 SuIn Lee GS & CSE, UW [email protected] 1 Outline (5/14, 5/16) Basic concepts on Bayesian networks Probabilistic models of gene regulatory networks Learning algorithms Evaluation Recent probabilistic approaches to reconstructing the regulatory networks 2 Today Known structure, complete data Network structure is specified Learner needs to estimate parameters Data does not contain missing values ? ? H L L ? ? ? ? ? ? H H L H L B E P(A | E,B) E B A E, B, A <H,L,L> <H,L,H> <L,L,H> <L,H,H> . . <L,H,H> E B A Learner .9 .1 H L L .7 .3 .99 .01 .8 .2 H H L H L B E P(A | E,B) samples 3 Learn the parameters based on D Training data P(G1) P(G3|G1) P(G2|G1) P(G4|G2) P(G5|G1,G2,G3) θ G1 θ G2|G1 θ G3|G1 θ G4|G3 θ G5|G1,G2,G3 G1 G2 G3 G4 G5 n instances D= c1 c2 c3 c4 c5 c6 c7 c8 4
12

Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

Jul 24, 2020

Download

Documents

dariahiddleston
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

1

Statistical methods for inferring the gene regulatory networks – Part II

Lecture 2 – May 16th, 2013GENOME 541, Spring 2013

Su‐In LeeGS & CSE, UW

[email protected]

Outline (5/14, 5/16)

Basic concepts on Bayesian networks Probabilistic models of gene regulatory 

networks Learning algorithms Evaluation Recent probabilistic approaches to 

reconstructing the regulatory networks

2

Today

Known structure, complete data

Network structure is specified Learner needs to estimate parameters

Data does not contain missing values

? ?

H

L

L

? ?

? ?? ?

HH

L

H

L

BE P(A | E,B) E B

A

E, B, A<H,L,L><H,L,H><L,L,H><L,H,H>

.

.<L,H,H>

E B

ALearner

.9 .1

H

L

L

.7 .3

.99 .01.8 .2

HH

L

H

L

BE P(A | E,B)

samples

3

Learn the parameters based on D

Training data

P(G1)

P(G3|G1)

P(G2|G1)

P(G4|G2) P(G5|G1,G2,G3)

θG1

θG2|G1

θG3|G1

θG4|G3 θG5|G1,G2,G3

G1

G2

G3

G4

G5

n instances

D = c1 c2 c3 c4 c5 c6 c7 c8

4

Page 2: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

2

LET’S CONSIDER THE SIMPLEST EXAMPLE.

5

Parameter estimation for a single variable

Variable X ‐ an outcome of a thumbtack toss Val(X) = {head, tail}

Data A set of thumbtack tosses: x[1] … x[M]

The Thumbtack example

X

6

Say that P(x=head) = Θ, P(x=tail) = 1‐Θ P(HHTTHHH…<Mhheads, Mt tails>; Θ) = 

Definition: The likelihood function L(Θ : D) = P(D; Θ)

Maximum likelihood estimation (MLE) Given data D=HHTTHHH…<Mhheads, Mt tails>, find Θ that maximizes the likelihood function L(Θ : D).

Maximum likelihood estimation

7

Likelihood function

8

Page 3: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

3

Given data D=HHTTHHH…<Mh heads, Mt tails>  MLE solution θ* = Mh / (Mh+Mt ).

Proof:

MLE for the Thumbtack problem

9 MtMhMh

θ

ˆ

Bayesian Network with table CPDs

Difficulty

GradeX

Intelligence

D: {H…x[m]…T} D: {(i1,d0,g1)…(i[m],d[m],g[m])…}

The Thumbtack exampleThe Student example

Data

Likelihood function

Parameters  

MLE solution

Joint distribution

vs

θI, θD, θG|I,D

P(X) P(I,D,G) = 

L(θ:D) = P(D;θ)

θ

θMh(1‐θ)Mt 1,1|1

1110

01

10

01

1 ,|dDiIgGdDdDiIiI

M

dDiIgG

M

dD

M

dD

M

iI

M

iI

01

011

011

,

,,,|

dDiI

dDiIgGdDiIgG M

M

10

Unknown structure, complete data

Network structure is not specified Learner needs to estimate both structure and parameters

Data does not contain missing values

? ?

H

L

L

? ?

? ?? ?

HH

L

H

L

BE P(A | E,B) E B

A

E, B, A<H,L,L><H,L,H><L,L,H><L,H,H>

.

.<L,H,H>

E B

ALearner

.9 .1

H

L

L

.7 .3

.99 .01.8 .2

HH

L

H

L

BE P(A | E,B)

11

Score‐based learning Define scoring function that measures how well a certain structure fits the observed data.

Search for a structure that maximizes the score.

E, B, A<Y,N,N><Y,Y,Y><N,N,Y><N,Y,Y>

.

.<N,Y,Y>

E B

A

E

B

AE

B A

G1

Score(G1) = 10 Score (G2) = 1.5 Score (G3) = 0.01

G2  G3 

12

Page 4: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

4

Structure score Likelihood score: 

Bayesian score Average over all possible parameter values

Penalized likelihood score

)θP(D|S, Sˆ

dSPSDPSDP )|(),|()|(

Maximum likelihood parameters

Likelihood Prior distribution over parameters

Marginal likelihood

),(S,complexity modellog DθC)P(D|S,θ SS

13

Decomposability of scores Likelihood score

(see slide 11)

Bayesian score

i

ii DLDL ):():(

dSPSDPSDP )|(),|()|(

dSPmPamxPk

ii m

iii...1

):():][|][(

iii

miii

i

dSPmPamxP ):():][|][(

i

i D:oreBayesianSc

14

Search for optimal network structure Start with a given network structure.

Empty network Best simple structure (e.g. tree) A random network

At each iteration Evaluate all possible changes Apply change based on score

Stop when no modification improves the score.

15

Search for optimal network structure Typical operations:

S C

E

D

S C

E

D

S C

E

D

S C

E

D16

Page 5: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

5

Search for optimal network structure Typical operations:

S C

E

D

S C

E

D

S C

E

D

score = S({C,E} D) - S({E} D)

Score decomposability:At each iteration only need to score the site that is being updated !

17

Outline

Basic concepts on Bayesian networks Probabilistic models of gene regulatory networks

Learning algorithms Parameter learning Structure learning Structure discovery

Evaluation Recent probabilistic approaches to reconstructing the regulatory networks 

18

Structure discovery Task: Discover structural properties

Is there a direction connection between X and Y? Does X separate between two “subsystems”? Does X causally affect Y?

Example: scientific data mining Disease properties and symptoms Interactions between the expression of genes

19

Model averaging

There may be many high‐scoring models Answer should not be based on any single model Want to average over many models

E

R

B

A

C

E

R

B

A

C

E

R

B

A

C

E

R

B

A

C

E

R

B

A

C

P(S|D)

20

Page 6: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

6

Define a structural feature f(S) of a model S. For example:

We are interested in computing

E

R

B

A

C

E

R

B

A

C

E

R

B

A

C

E

R

B

A

C

E

R

B

A

C

P(S|D)

S

DSP DSPSfSfE )|()()]([)|(

otherwise0

has Sgraph a if1)(

CASf

f(S) 0 0 1 0 0

21

Bootstrapping Sampling with replacement

Original data Bootstrap data 1 data 2 data N…

gene

s

experiments

22Inferring sub‐networks from perturbed expression profiles, Pe’er et al. Bioinformatics 2001

Bootstrap data 1 data 2 data N…

Bootstrapping Sampling with replacement

(N) networks # totaledge the containthat networks #

Estimated confidence of each edge i 

0.8

0.31

0.430.56

0.26

0.75

0.73

23Inferring sub‐networks from perturbed expression profiles, Pe’er et al. Bioinformatics 2001

Outline

Basic concepts on Bayesian networks Probabilistic models of gene regulatory networks

Learning algorithms Evaluation

Predicted co‐regulated groups of genes Putative regulator‐regulatees

Recent probabilistic approaches to reconstructing the regulatory networks

24

Page 7: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

7

Functional coherence of gene clusters Gene Ontology (GO) [http://www.geneontology.org/]

The GO database provides a controlled vocabulary to describe gene and gene product attribute in any organism. 

Set of biological phrases (GO terms) which are applied to genes

Organized as three separate ontologies Molecular functions Biological processes Cellular components

Each gene may Have more than one in molecular function. Take part in more than one biological process. Act in more than one cellular component.

25

Structure of ontologies Shows the relationship between different terms

One term may be a more specified description of another more general term.

Shows hierarchies of the terms (directed acyclic graph). Each child‐term is a member of its parent‐term

26

Predicted regulatory interaction I Say that your network suggests:

If HAP4 is a transcription factor, Targets should have a binding site for HAP4. Or there should be different kind of evidence that HAP4 binds to 

genes in Module A (chip‐chip or chip‐seq data).

PHO5PHM6

PHO3PHO84

VTC3GIT1

PHO2

HAP4

PHO4

HAP4Module A

AGTCTTAACGTTTGACCGCTAATT

Module A

27

Predicted regulatory interaction II Say that your network suggests:

If HAP4 really regulates module A, deletion (or overexpression) of HAP4 should lead to significant up/down‐ regulation of genes in module A. There are many publicly available gene expression data that measure expression of genes after deleting/over‐expressing a certain gene. 

PHO5PHM6

PHO3PHO84

VTC3GIT1

PHO2

HAP4

PHO4

Module A

28

Page 8: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

8

Create functional categories For each GO term,

Genes that have the same GO term form a functional category

Other gene annotation systems KEGG: Kyoto Encyclopedia of Genes and Genomes 

[http://www.genome.jp/kegg/]

Molecular Signature Database [http://www.broadinstitute.org/gsea/msigdb/index.jsp]

Functional categories29

Functional coherence

How significant is the overlap? Calculate P(# overlap ≥ k | m, n, N; two groups are independent)

based on the hypergeometric distribution

Modules Known functional categories

Gene ontology (GO)http://www.geneontology.org/

Predicted targets of regulatorsSharing TF binding sites

:

km  n 

Module 1 Cholesterol synthesis

genes

30

Examples Say N=1000, m=100, n=200 genes

If k = 40 genes in the intersection, p‐value = 2.7410e‐07. If k = 30, p‐value = 0.0039 If k = 20, p‐value = 0.4394.

31

km  n 

Module 1 Cholesterol synthesis

How significant is the overlap? Calculate p-value = P(# overlap ≥ k | m, n, N; two groups are

independent), based on the hypergeometric distribution What p-values are considered to be significant?

Multiple hypothesis testing Say that there are 200 modules and 3000 functional categories

32

How many hypotheses are we testing? 200 x 3000 = 600,000 Is p‐value of 0.001 significant?  (p‐value=0.001: frequency of 

observing the # genes in intersection by random.) P‐values should be “corrected”

Bonferroni correction: min(1, p‐value x # hypotheses) FDR correction: control false discovery rate

Modules Known functional categories

genes

Page 9: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

9

Outline

Basic concepts on Bayesian networks Probabilistic models of gene regulatory networks

Learning algorithms Evaluation

Predicted co‐regulated groups of genes Putative regulator‐regulatees

Recent probabilistic approaches to reconstructing the regulatory networks

33

Challenges Too large search space

For a network with n genes, what is the number of possible structures?

Computationally costly

Heuristic approaches may be trapped to local maxima.

Biologically motivated constraints can alleviate the problems Module‐based approach Only the genes in the candidate regulators list can be parents of 

other variables

2/2

3~ n

34

X1

X3 X4

X5 X6

X2Module 2

Module 3

Module 1X1

X3 X4

induced

repressed

The Module networks concept

=

Activator X3

Module3

Repressor X4-3 x+

0.5 x

Linear CPDs

Activator X3

Repressor X4

truefalse

truefalse

Mod

ule

3ge

nes

regu

latio

n pr

ogra

m

target gene expression

repressor expression

activator expression

Tree CPDs

Segal et al. Nat Genet 03, JMLR 0535

Feature selection via regularization

Assume linear Gaussian CPD

xN…x1 x2

w1w2 wN

Y

Candidate regulators (features)Yeast: 350 genesMouse: 700 genes

P(Y|x:w) = N(Σwixi , ε2)

Problem: This objective learns too many regulators

MLE: solve maximizew ‐ (Σwixi ‐ Y)2

parametersw1

w2 wN

Regulatory network

36

Page 10: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

10

L1 regularization “Select” a subset of regulators

Combinatorial search? Effective feature selection algorithm: L1 regularization (LASSO)

[Tibshirani, J. Royal. Statist. Soc B. 1996]

minimizew (Σwixi ‐ Y)2+  C |wi|: convex optimization! Induces sparsity in the solution w (Many wi‘s set to zero)

xN…x1 x2

w1w2 wN

Y

Candidate regulators (features)Yeast: 350 genesMouse: 700 genes

x1 x2

P(Y|x:w) = N(Σwixi , ε2)

37

Iterative procedure Learn a regulatory program for each module Cluster genes into modules

Linear module network

38

PHO5PHM6

SPL2

PHO3PHO84

VTC3GIT1

PHO2

TEC1

GPA1

ECM18

UTH1MEC3

MFA1

SAS5SEC59

SGS1

PHO4

ASG7

RIM15

HAP1

PHO2

GPA1MFA1

SAS5PHO4

RIM15

Lee et al., PLoS Genet 2009

M1

M120

M22

M1011

M321M321

M120=

MFA1

Module

GPA1-3 x+

0.5 x+

-1.2 x

Better? L1 regularized optimizationminimizew (Σwixi - ETargets)2+ C |wi|

38

LEARNING 

Let’s consider the module network with tree CPDs…

39

Learning module networks Score‐based learning – Find the structure that maximizes Bayesian score log P(S|D) (or via regularization)

“Hidden” variables How genes are organized into modules is not known.

Expectation Maximization (EM) algorithm: Repeat E‐step: filling in hidden variables

M‐step: parameter estimation

40

Page 11: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

11

Learning module networks Learning algorithm

Initialization: Group genes by (k‐means) clustering into modules

M‐step: Given a partition of the genes into modules, learn the best regulation programs (tree CPD) for modules.

E‐step: Given the inferred regulatory programs, we reassign genes into modules such that the associated regulation program best predicts each gene’s behavior.

Repeat until convergence.

41

Iterative procedure (EM‐steps) Cluster genes into modules (E‐step) Learn regulatory programs for modules (tree CPD) (M‐step)

Learning module networks

PHO5PHM6

SPL2

PHO3PHO84

VTC3GIT1

PHO2

GPA1

ECM18

UTH1MEC3

MFA1

SAS5SEC59

SGS1

PHO4

ASG7

RIM15

HAP1

TEC1

M1

MSK1

M22

KEM1

MKT1DHH1

PHO2

HAP4MFA1

SAS5PHO4

RIM15

Maximum increase in the Bayesian score

Candidate regulators

42

M‐step: Learning regulatory programs

Combinatorial search over the space of trees

Arrays sorted in original order

HAP4

Arrays sorted according to expression of HAP4

Segal et al. Nat Genet 2003

PHO5PHM6

PHO3PHO84

VTC3GIT1

PHO2

HAP4

PHO4

43

M‐step: Learning regulatory programs

HAP4 SIP4

Segal et al. Nat Genet 2003

PHO5PHM6

PHO3PHO84

VTC3GIT1

PHO2

HAP4

PHO4

SIP4

44

Page 12: Outline (5/14, 5/16) for inferring the networks –Part IIhomes.cs.washington.edu/~suinlee/genome541/lecture2...5/21/2013 1 Statistical methods for inferring the gene regulatory networks

5/21/2013

12

M‐step: Learning regulatory programs

HAP4

0 0

Score:log P(M | D)

log ∫P(D| M,, )P(,) dd

Score of HAP4 split:log P(M | D)

log ∫P(DHAP4| M,, )P(,) dd

+ log ∫P(DHAP4| M,, )P(,) dd

0

Segal et al. Nat Genet 2003

45

M‐step: Learning regulatory programsHAP4

0 0

Split as long as the score improves

HAP4

0 0

YGR043C

0

Score of HAP4/YGR043C split:log P(M | D) ∝

log ∫P(DHAP4|M,,)P(,) dd+ log ∫P(DHAP4DYGR043C|M,,)P(,) dd

+ log ∫P(DHAP4DYGR043C |M,,)P(,) dd

Score of HAP4 split:log P(M | D) ∝log ∫P(DHAP4| M,, )P(,) dd+ log ∫P(DHAP4| M,, )P(,) dd

46

Iterative procedure Cluster genes into modules (E‐step) Learn regulatory programs for modules (M‐step)

Learning module networks

PHO5PHM6

SPL2

PHO3PHO84

VTC3GIT1

PHO2

GPA1

ECM18

UTH1MEC3

MFA1

SAS5SEC59

SGS1

PHO4

ASG7

RIM15

HAP1

TEC1

M1

MSK1

M22

KEM1

MKT1DHH1

PHO2

HAP4MFA1

SAS5PHO4

RIM15

Maximum increase in Bayesian score

Candidate regulators

HAP4

MSK1

truefalse

truefalse

47

Summary

Basic concepts on Bayesian networks Probabilistic models of gene regulatory networks

Learning algorithms Evaluation Recent probabilistic approaches to reconstructing the regulatory networks

48