Title: From Sequence to Expression: A Probabilistic Framework
1From Sequence to ExpressionA Probabilistic
Framework
Joint work with
Nir Friedman (Hebrew U.) Daphne Koller (Stanford)
Yoseph Barash(Hebrew U.) Itamar Simon (Whitehead
Inst.)
2Understanding Cellular Processes
- Complex biological processes (e.g. cell cycle)
- Coordination of multiple events
- Each event requires different modules
Can we recover the regulatory circuits that
control such processes?
3Gene Structure
4Gene Regulation
A
mRNA
5Gene Regulation
A
mRNA
6Gene Regulation
Swi5
mRNA
7Gene Regulation
Swi5
mRNA
8Gene Regulation
mRNA
9Goal
ACTAGTGCTGA
CTATTATTGCA
CTGATGCTAGC
10Model of Gene Regulation
Probabilistic Relational Models (PRMs) Pfeffer
and Koller (1998)Friedman et al (1999)Segal et
al (2001)
Sequence
Promoter sequences
Gene
Experiment
Regulation by transcription factors
Expression measurements
Expression
11Regulation to Expression
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. cluster
Level
Expression
R(t1) yes ? t1 regulates geneR(t1) no ?
t1 does not regulate gene
12Regulation to Expression
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. cluster
Level
Expression
13Modeling Context Specificity
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. cluster
Level
Expression
14Sequence Model
- Assumptions
- Binding site is of length k
- Binding may occur at any k-mer
- TF regulates gene if binding occurs anywhere
Sequence
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. cluster
Level
Expression
15From Sequence to Regulation
- Assumptions
- Binding site is of length k
- Binding may occur at any k-mer
- TF regulates gene if binding occurs anywhere
16From Sequence to Regulation
- Model for one gene g, promoter region of length
5 and k2
17Joint Probabilistic Model
k-mer
s1
sk
B(t1)
B(t2)
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. Cluster
Level
Expression
18Localization Assay
19Localization Assay
Swi5
20Localization Assay
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
? TF binds to targets
21Localization Assay
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Swi5
Gene Not Bound
Gene Bound
- Measure TF binding to promoter of every gene
- Assign confidence for each binding
? TF binds to targets
22Localization Assay
- Localization data measure TF binding to promoter
of each gene (assign binding confidence)
Simon et al (2001)
23Is Regulation Observed?
- Not quite
- Localization is measured for specific conditions
- Localization is measured for large DNA regions
- Localization is noisy
24Incorporating Localization
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. Cluster
L(t2)
L(t1)
Observedlocalization
Level
Expression
- Localization p-value is noisy sensor of actual
regulation - If regulation occurs, p-value likely to be low
- If no regulation, p-value likely to be high
25Localization Model
Gene
R(t1)
L(t1)
Observed
- Localization p-value is noisy sensor of actual
regulation - If regulation occurs, p-value likely to be low
- If no regulation, p-value likely to be high
26Joint Probabilistic Model
promoter
s1
sk
Gene
Experiment
R(t2)
R(t1)
Exp. type
Exp. Cluster
L(t2)
L(t1)
Level
Expression
27Learning the Models
Experimental Details
LocalizationData
28Learning the Models
Experimental Details
LocalizationData
29Model Learning
- Structure Learning
- Tree structure
- Missing Data
- Experiment cluster
- Regulation variables
- Motif Model
- Parameter estimation
30Model Learning
promoter
s1
sk
Gene
Experiment
R(t2)
R(t1)
Exp. type
L(t1)
Exp. cluster
L(t1)
Level
Expression
31Resulting Bayesian Network
Exp. type
Exp. type2
s11
sk1
Exp. cluster
Exp. cluster
R(t2)1
Level1,2
Level1,1
L(t2)1
R(t1)1
L(t1)1
s12
sk2
R(t2)2
Level2,2
Level2,1
L(t2)2
R(t1)2
L(t1)2
s13
sk3
R(t2)3
Level3,2
Level3,1
L(t2)3
R(t1)3
L(t1)3
32Model Learning E-Step
Exp. type
Exp. type2
s11
sk1
Exp. cluster
Exp. cluster
R(t2)1
Level1,2
Level1,1
L(t2)1
R(t1)1
L(t1)1
s12
sk2
R(t2)2
Level2,2
Level2,1
L(t2)2
R(t1)2
L(t1)2
s13
sk3
R(t2)3
Level3,2
Level3,1
L(t2)3
R(t1)3
Loopy belief propagation
L(t1)3
33Model Learning M-Step
Exp. type
Exp. type2
s11
sk1
Exp. cluster
Exp. cluster
R(t2)1
Level1,2
Level1,1
L(t2)1
R(t1)1
L(t1)1
s12
sk2
R(t2)2
Level2,2
Level2,1
L(t2)2
R(t1)2
L(t1)2
s13
sk3
R(t2)3
ConjugateGradient
Level3,2
Level3,1
L(t2)3
R(t1)3
Standard ML estimation
L(t1)3
34Experimental Results
- Yeast
- Cell Cycle expression data (Spellman et al)
- Localization data for 9 TFs (Simon et al)
- Yeast genome (promoters)
35Generalization
Gene log-likelihood
-112.24
Experiment
Gene
R(t2)
R(t1)
Exp. Cluster
Level
Expression
36Generalization
Gene log-likelihood
-121.48
-112.24
Experiment
Gene
L(t2)
L(t1)
Exp. type
Level
Expression
37Generalization
Gene log-likelihood
-112.24
Experiment
Gene
R(t2)
R(t1)
Exp. type
Exp. Cluster
Level
Expression
38Generalization
Gene log-likelihood
-112.24
Experiment
Gene
R(t2)
R(t1)
Exp. type
Exp. Cluster
Level
Expression
39Generating Hypotheses
40Expression vs Regulation
1
0.5
Genes predicted to be regulated by Swi5 are
probably real Swi5 targets
0
-0.5
-1
0
21
42
63
84
105
10
70
100
130
160
190
220
250
0
30
60
90
120
150
0
90
180
270
360
cdc15
cdc28
elu
alpha
41Combinatorial Effects
1
Phase
Fkh2 Swi4
Fkh2 Ndd1
0.5
0
-0.5
-1
0
21
42
63
84
105
10
70
100
130
160
190
220
250
0
30
60
90
120
150
0
90
180
270
360
cdc15
cdc28
elu
alpha
42Combinatorial Effects
1
Phase
0.5
Mcm1 Ndd1
Mcm1 Ace2
Mcm1 Swi5
0
-0.5
-1
0
21
42
63
84
105
10
70
100
130
160
190
220
250
0
30
60
90
120
150
0
90
180
270
360
cdc15
cdc28
elu
alpha
43Localization Assignment Changes
44Motifs Found
Simonet al.
17
Expanded set identified additional genes
regulated by Ndd1
ExpandedSet
28
1
Remaining Genes
45TF Simon Expanded Rest P-Value
Ace2 10 9 1 1.4e-6
Fkh1 29 25 8 4.4e-10
Fkh2 29 29 10 5.4e-11
Mbp1 66 56 8 1.9e-45
Mcm1 28 24 2 4.2e-18
Ndd1 17 28 1 1.9e-24
Swi4 41 37 5 6.4e-26
Swi5 28 23 2 4.9e-15
Swi6 50 52 6 2.3e-48
46Induced Interaction Network
- TF pairs whose regulation predicts expression of
same gene cluster
G1
Swi5
Swi6
Ace2
M
Mcm1
Mbp1
Swi4
Fkh1
S
G2
Ndd1
Fkh2
47Conclusions
- Unified probabilistic model explaining gene
regulation using sequence, localization and
expression data - Models complex interactions between regulators
- Discriminative model maximizing P(Expr. Seq.)
- Sequence data helps explain expression patterns
48Big Picture
- Goal unified probabilistic framework
- Models complex biological domains
- Incorporates heterogeneous data
- Framework incorporates explicitly within model
basic biological building blocks - Genes, TFs, proteins, patients, cells, species,
- Much closer connection between biology and model
- Can read biology directly from model
- Can incorporate prior knowledge easily
- Can explicitly represent and learn biological
models