Understanding Gene Regulation:
From Networks to Mechanisms
Daphne Koller
Stanford University
Gene Regulatory Networks
Controlled by diverse mechanisms
Modified by endogenous and
exogenous perturbations
http://en.wikipedia.org/wiki/Gene_regulatory_network
Goals

Infer regulatory network and mechanisms that
control gene expression

Identify effect of perturbations on network

Understand effect of gene regulation on phenotype
Outline




Regulatory networks for gene expression
Individual genetic variation and gene regulation
Cell differentiation and gene regulation
Expression changes underlying phenotype
Regulatory Network I



mRNA level of regulator can indicate its activity level
Target expression is predicted by expression of its regulators
Use expression of regulatory genes as regulators
Transcription factors, signal transduction
proteins, mRNA processing factors, …
ECM18
UTH1
ASG7
MEC3
GPA1
MFA1
TEC1
HAP1
PHO3
SGS1
SEC59
PHO5
RIM15
PHO84
PHM6
PHO2
PHO4
SAS5
SPL2
GIT1
Segal et al., Nature Genetics 2003; Lee et al., PNAS 2006
VTC3
Regulatory Network II



Co-regulated genes have similar regulation program
Exploit modularity and predict expression of entire module
Allows uncovering complex regulatory programs
module
ECM18
UTH1
ASG7
MEC3
GPA1
MFA1
TEC1
“Regulatory Program” ?
Targets
HAP1
PHO3
SGS1
SEC59
PHO5
RIM15
PHO84
PHM6
PHO2
PHO4
SAS5
SPL2
GIT1
Segal et al., Nature Genetics 2003; Lee et al., PNAS 2006
VTC3
Module Networks*
Activator
Gene1
Repressor
Gene2


Gene3
Genes in module
repressor
expression

activator
false
true
repressor
false
true
regulation
program
activator
expression
repressed
module
genes
target gene
expression
induced
Learning quickly runs
out of statistical power


Poor regulator selection
lower in the tree
Many correct regulators
not selected
Arbitrary choice among
correlated regulators
Combinatorial search

Multiple local optima
* Segal et al., Nature Genetics 2003
Regulation as Linear Regression
minimizew (Σwixi - ETargets)2
x1
w1
w2
…
xN
-3 x
GPA1
0.5 x
MFA1
+
wN
ETargets
=
parameters
x2
Module
ETargets= w1 x1+…+wN xN+ε


But we often have hundreds or thousands of regulators
… and linear regression gives them all nonzero weight!
Problem: This objective learns too many regulators
Lasso* (L1) Regression
minimizew (w1x1 + … wNxN - ETargets)2+  C |wi|
x1
w1
parameters

w2
xN
L2
L1
wN
ETargets
Provably selects “right” features when many features are irrelevant
Convex optimization problem



…
Induces sparsity in the solution w (many wi‘s set to zero)


x2
Unique global optimum
Efficient optimization
But, arbitrary choice among correlated regulators
* Tibshirani, 1996
Elastic Net* Regression
minimizew (w1x1 + … wNxN - ETargets)2+  C |wi| +  D wi2
x1
w1
x2
w2
…
xN
L2
L1
wN
ETargets

Induces sparsity
But avoids arbitrary choices among relevant features

Convex optimization problem



Unique global optimum
Efficient optimization algorithms
Lee et al., PLOS Genetics 2009
* Zhou & Hastie, 2005
Learning Regulatory Network


-3 x GPA1
Cluster genes into modules
0.5 x MFA1
Learn a regulatory program for each module
+
=
Module
ECM18
UTH1
ASG7
MEC3
GPA1
This is a Bayesian network
MFA1
TEC1
HAP1
• But
multiple genes share samePHO3
program
PHO5
PHO84
• Dependency RIM15
model is linear regression
PHM6
SGS1
SEC59
PHO2
PHO4
SAS5
SPL2
GIT1
VTC3
Lee et al., PLoS Genet 2009
Outline


Regulatory networks for gene expression
Individual genetic variation and gene regulation




Effect of genotype on expression
Regulatory potential
Cell differentiation and gene regulation
Expression changes underlying phenotype
Genotype  phenotype
Different
sequences
Perturbations to
regulatory network
…ACTCGGTTGGCCTAAATTCGGCCCGG…
…ACCCGGTAGGCCTTAATTCGGCCCGG…
:
…ACTCGGTAGGCCTATATTCGGCCGGG…
??
Different
phenotypes
Genotype  Regulation
Different
sequences
Perturbations to
regulatory network
…ACTCGGTTGGCCTAAATTCGGCCCGG…
…ACCCGGTAGGCCTTAATTCGGCCCGG…
:
??
…ACTCGGTAGGCCTATATTCGGCCGGG…

Goals:


Infer regulatory network that controls gene expression
Identify mechanisms by which genetic variation affects
gene expression
eQTL Data
BY
[Brem et al. (2002) Science]
RM
112
progeny
Genotype data
×
Markers
0101100100…011
1011110100…001
0010110000…010
:
0000010100…101
0010000000…100
112 individuals
6000 genes
3000 markers
112 individuals
:

Expression data
Traditional Approach: Single Marker

Expression
quantitative trait loci (eQTL) mapping
Marker

Gene
1 2gene,
3
4find
5 the
… marker that is most predictive of its
For each
expression level [Yvert G et al. (2003) Nat Gen].
mRNA
individuals
markers
0101100100100…011
1011110111100…001
0010110001000…010
:
Marker j
0000010110100…101
individuals
genes
Gene i
induced
1110000110000…100
Genotype data
repressed
Expression data
LirNet Regulatory network


E-regulators: Activity (expression) of regulatory genes
G-regulators: Genotype of genes

Measured as values of chromosomal markers
marker
M1011
M321
M1
M120
ECM18
UTH1
M321
M22
ASG7
MEC3
GPA1
MFA1
TEC1
HAP1
PHO3
SGS1
SEC59
PHO5
RIM15
PHO84
PHM6
PHO2
PHO4
SAS5
SPL2
GIT1
Lee et al., PNAS 2006; Lee et al., PLoS Genetics 2009
VTC3
The Telomere Module


40/42 genes in telomeres
Enriched for telomere maintenance (p < 10-11) &
helicase activity (p < 10-18) Includes Rif2 –
Chr XII: 1056097
•
•
•
•
control telomere length
establishes telomeric silencing
6 coding & 8 promoter SNPs
Binds to Rap1p C-terminus
Enrichment for Rap1p targets (29/42; p < 10-15)
Lee et al., PNAS 2006
Some Chromatin Modules
• Locus containing Sir1
• 4 coding SNPs
Chr XI: 643655
4/5 consecutive genes
Known Sir1 targets
• Locus containing uncharacterized
Sir1 homologue
• 87(!) coding SNPs
Chr X: 22213
5/7 consecutive genes
Lee et al., PNAS 2006
Mechanism I
Chromatin as Mechanism
module #genes Runs
728
16
732
7
742
24
743
6
744
5
755
14
757
3
770
35
783
23
790
122
848
15
855
42
862
20
869
48
876
51
890
114
897
15
904
56
917
31
941
110
968
218
972
7
991
4
993
5
1018
88
1025
217
1030
15
1054
6
1061
46
1078
90
1085
42
1120
8
1126
26
1133
19
1145
7
1155
75
Dom
DEG Enrich Chrom Reg-E Reg-G
Telo
Telo
Telo

Telo
Telo

23 modules (out of 165) with
“chromosomal features”
16 have “chromatin regulators”
(p < 10-7)
“Evolutionary strategy” to make
coordinated changes
in gene expression
 Chromatin modification explains
significant
part ofof
variation
by modifying small
number
hubs in
Telo
TY,LTR
Telo
gene expression between strains
Telo
Telo
Telo
Telo
Telo
Lee et al., PNAS 2006
The Puf3 Module
weight
BY RM
112 segregants
147/153 genes (P ~ 10-130)
are pulldown targets of
mRNA binding protein Puf3
HAP4
TOP2
KEM1
GCN1
GCN20
DHH1PUF
P-body
components
family:
 Sequence specific
mRNA binding proteins
(3’ UTR)
 Regulate degradation
of mRNA and/or repress
translation
+1
0
-1
PUF3
expression
genotype
Lee et al., PLOS Genetics 2009
P-Bodies
mRNAs stored in P-bodies
are translationally repressed
Translation
mRNAs can exit P-bodies
and resume translation
GFP of Dhh1**
RNA degradation
RNA stored in P-bodies can be
degraded (via decapping)
Dhh1 regulates mRNA decapping in p-bodies
But what regulates sequence-specific localization of
P-body
Puf3
mRNAs to P-bodies?
?
protein
* Sheth and Parker (2003) Science 300:805
** Beliakova-Bethell , et al. (2006) RNA 12:94
cell
Microscopy experiment


Fluorescent microscopy: Puf3 localizes to P-body
Supports hypothesis of Puf3 involvement in regulating
mRNA degradation by P-bodies
Puf3 protein
?
P-body
Dhh1
Dhh1
Puf3
cell
Joint work with David Drubin and Pam Silver
Lee et al., PLOS Genetics 2009
What Regulates the P-bodies?



A marker that covers a large region in Chr 14.
Region contains ~30 genes and ~318 SNPs.
Experiments for all 30 genes not feasible!
RM
BY
ChrXIV:449639
DHH1
BLM3
GCN20
KEM1
GCN1
Lee et al., PLOS Genetics 2009
Outline


Regulatory networks for gene expression
Individual genetic variation and gene regulation




Effect of genotype on expression
Regulatory potential
Cell differentiation and gene regulation
Expression changes underlying phenotype
Motivation

Not all SNPs are equally likely to be causal.
ChrXIV: 449,639-502,316
Gene
TACGTAGGAACCTGTACCA … GGAAAATATCAAATCCAACGACGTTAGCCAATGCGATCGAATGGGAACGTA
SNP 1:
Conserved residue in a
gene involved in RNA
degradation
SNP 2:
In nonconserved
intergenic region
“Regulatory features” F
SNPs
1. Gene region?
2. Protein coding region?
3. Nonsynonymous?
4. Create a stop codon?
5. Strong conservation?
:

Idea: Prioritize SNPs that have “good” regulatory features

But how do we weight different features?
Lee et al., PLOS Genetics 2009
Bayesian L1-Regularization
M
 log P( y
m 1
m
| x m , w m )   C | wm,k |
m,k
Prior
variance
wi
Module m
higher prior variance
 weight can more easily deviate from 0
 regulator more likely to be selected
Regulator k
xmk
wmk
ym
wi
~ P(w)
= Laplacian(0,C)
~ P(ym|x;w)
= N (k wmkxmk,ε2)
Lee et al., PLOS Genetics 2009
Metaprior Model (Hierarchical Bayes)
YES
Module m
Regulatory features
Inside a gene?
Protein coding region?
Strong conservation?
TF binds to module genes
:
Regulator k
fmk
Cmk
xmk
wmk
Regulatory
prior ß
= g(ßTfmk)
~ Laplacian (0,Cmk)
Module 1NO
:
x1
w11
~ P(ym|x;w)
= N (k wmkxmk,ε2)
Lee et al., PLOS Genetics 2009
regulators
…
x2
w12
NO
NO
:
xN
w1N
Emodule 1
“Regulatory potential” =
ß1 x Inside a gene? + ß2 x Protein
coding region? + ß3 x Conserved? …
:
Module M
x1
ym
YES
NO
Potential
:
wN1
x2
wN2
Emodule M
…
xN
wMN
Metaprior Method
BY(lab)
MVLT L ELVQ A VSDASKQLWDI
RM(wild) MVLT D ELVQG VSDASKQLWDI
0
Non-synonymous
Conservation
AA small  large
0.1
0.2
0.3
1
2
3
4
5
0


0.1
0.2
0.3
Regulatory weights ß
1
0
=
3. Compute regulatory
potential of SNPs
Empirical hierarchical Bayes
6
1
1
1
:
=

Cell cycle
1
0
0
:
0.3
0.9
Regulatory
features F
Use point estimate of model parameters
Learn priors from data to maximize joint posterior
2. Learn ß
Maximize P(E,ß,W|X)
1. Learn regulatory
Regulatory potentials
programs
Regulatory programs
Module i+2
Module i+1
Module
x1 xi2 …… xN
x1 x2 …
xN
x1 x2
xN
Ei
Ei
Ei
Lee et al., PLOS Genetics 2009
x1
…
x2
Ei
Maximize P(E,ß,W|X)
xN
Transfer Learning

What do regulatory potentials do?



They do not change selection of “strong” regulators –
those where prediction of targets is clear
They only help disambiguate between weak ones
Strong regulators help teach us what to look for in
other regulators
Transfer of knowledge
between different prediction tasks
Learned regulatory weights
Yeast regulatory weights
Regulatory features 0
Non-synonymous coding
Stop codon
Synonymous coding
3' UTR
500 bp upstream
5' UTR
500 bp downstream
Conservation score
Cis-regulation
Change of average mass (Da)
Change of isoelectric point (pI)
Change of pK1
Change of pK2
Change of hydro-phobicity
Change of pKa
Change of polarity
Change of pH
Change of van der Waals
Transcription regulator activity
Telomere organization and
Protein folding
Glucose metabolic process
RNA modification
Hexose metabolic process
Cell cycle checkpoint
Proteolysis
same GO process
same GO function
ChIP-chip binding
0.1
Lee et al., PLOS Genetics 2009
0.2
0.3
0.4
Human regulatory weights
0.5
0
Non-synonymous coding
Synonymous coding
Location
Intron
Locus region
Splice site
UTR region
Conservation score
Cis-regulation
Change of average mass (Da)
Change of isoelectric point
Change of pK1
Change of pKa
AA property
Change of pH
changeChange of van der Waals volume
Change of polarity
cell communication
signal transduction
transcription
Gene
function
Pairwise
feature
0.1
0.2
0.3
0.4
0.5
Statistical Evaluation
500 1000 1500 2000 2500
PGV: Percent genetic variation explained by the predicted
regulatory program for each gene
# of genes with PGV > X %

1650 genes (Lirnet)
1450 genes (Lirnet without
regulatory prior)
850 genes (Geronemo*)
250 genes (Brem & Kruglyak)
100 90 80 70 60 50 40 30 20 10
Lee et al., PLOS Genetics 2009
PGV (%)
0
* Lee et al. PNAS 2006
Biological evaluation I

How many predicted interactions have support in other data?





Deletion/ over-expression microarrays [Hughes et al. 2000; Chua et al. 2006]
ChIP-chip binding experiments [Harbison et al. 2004]
Transcription factor binding sites [Maclsaac et al. 2006]
mRNA binding pull-down experiments [Gerber et al. 2004]
Literature-curated signaling interactions
Lirnet without
regulatory features
90
% Supported interactions
80
Geronemo
Flat Lirnet
Lirnet
70
60
50
40
Reg
30
TF
20
10
0
%interactions
Module
%modules
Lee et al., PLOS Genetics 2009
%interactions
%modules
Via cascade
% supported regulatory predictions
Biological Evaluation II
100
90
80
Lirnet
Zhu et al (Nature Genet 2008)
70
Random
60
50
40
30
20
10
0
0
15
30
45
60
75
significance
of support
Significance
(%)
Lee et al., PLOS Genetics 2009
90
What Regulates the P-Bodies?

The regulatory potential over all 318 SNPs in the region
Regulatory potential
ChrXIV:415,000-495,000
MKT1
0.7
High-scoring regulatory features
Conservation
0.230
0.6
Cis-regulation 0.208
Same GO proc
0.204
0.5
Non-syn
0.158
Translation regulation
0.4
Mass (Da)
pK1
Polar
pI
0.066
0.060
0.057
0.050
0.046
ChrXIV:449639
Lee et al., PLOS Genetics 2009
Saccharomyces Genome Database (SGD)
Mkt1


Mkt1 binds to mRNAs at 3’ UTR
BY has SNP at conserved residue in nuclease domain
XPG-N
XPG-I
*
35
MKT1
RM
Pbp1 Interaction
mkt1D in RM
30
All others
P-body Module
Puf3 Module
BY
25
P-body module
20
15
Puf3 module
10
5
0
Lee et al., PLOS Genetics 2009
-1.2
-0.8
-0.4
0
0.4
0.8
1.2
8 validated regulators
in 7 regions
14 validated regulators
in 11 regions
Predicting Causal Regulators

Finding causal regulators for 13 “chromosomal hotspots”
Region
Zhu et al [Nat Genet 08]
Lirnet (top 3 are considered)
SEC18
RDH54
SPT7
1
None
AMN1
CNS1
TOS1
2
TBS1, TOS1, ARA1, CSH1, SUP45,
CNS1, AMN1
3
None
TRS20
ABD1
PRP5
4
LEU2, ILV6, NFS1, CIT2, MATALPHA1
LEU2
PGS1
ILV6
5
MATALPHA1
MATALPHA1
MATALPHA2
RBK1
6
URA3
URA3
NPP2
PAC2
7
GPA1
STP2
GPA1
NEM1
8
HAP1
HAP1
NEJ1
GSY2
9
YRF1-4, YRF1-5, YLR464W
SIR3
HMG2
ECM7
10
None
ARG81
TAF13
CAC2
11
SAL1, TOP2
MKT1
TOP2
MSK1
12
PHM7
PHM7
ATG19
BRX1
13
None
ADE2
ORT1
CAT5
Lee et al., PLOS Genetics 2009
Learning Regulatory Priors


Learns regulatory potentials that are specific to
organism and even data set
Can use any set of regulatory features:




Sequence features
Functional features for relevant gene
Features of regulator/target(s) pairs
Applicable to any organism, including ones where
functional data may not be readily available
Lee et al., PLOS Genetics 2009
Conclusion

Framework for modeling gene regulation



Use machine learning to identify regulatory program
Hierarchical Bayesian techniques to capture regularities
in effects of perturbations on network
Uncovers diverse regulatory mechanisms


Chromatin remodeling
mRNA degradation
Acknowledgements
Su-In Lee
University of
Washington


Aimee Dudley
Dana Pe’er
Institute for
System Biology
Columbia
University
Nevan Krogan, UCSF
Pam Silver, David Drubin, Harvard Medical School
National Science Foundation