From Sequence to Expression: A Probabilistic Framework Eran
Segal (Stanford) Joint work with: Yoseph Barash (Hebrew U.) Itamar
Simon (Whitehead Inst.) Nir Friedman (Hebrew U.) Daphne Koller
(Stanford)
Slide 2
Understanding Cellular Processes u Complex biological processes
(e.g. cell cycle) Coordination of multiple events Each event
requires different modules S G2 M G1 Can we recover the regulatory
circuits that control such processes?
Slide 3
Gene Structure Coding Region Promoter Region CTAGTAGATATCGATCAG
mRNA Protein
Model of Gene Regulation GeneExperiment Expression Sequence
Probabilistic Relational Models (PRMs) Pfeffer and Koller (1998)
Friedman et al (1999) Segal et al (2001) Promoter sequences
Regulation by transcription factors Expression measurements Context
Cluster
Slide 11
Regulation to Expression Level GeneExperiment Expression R(t 1
) R(t 2 ) Exp. type R(t 1 ) = yes t 1 regulates gene R(t 1 ) = no t
1 does not regulate gene Exp. cluster
Slide 12
Regulation to Expression Level GeneExperiment Expression R(t 1
) R(t 2 ) Exp. type R(t 1 ) R(t 2 ) E type 0 0 I -0.7 1.2 0 1 II
0.8 0.6 CPD P(Level) Level -0.7 0.8 P(Level) Level Exp.
cluster
Slide 13
Modeling Context Specificity Level GeneExperiment Expression
R(t 1 ) Exp. type Exp. type = G1 R(t 2 )=ye s true false true R(t 1
) = Yes false true false... 3 P(Level) Level 0 P(Level) Level 2
P(Level) Level u Gaussian decision tree u T1 only relevant in G1 u
T2 only relevant in G2 Exp. cluster R(t 2 )
Slide 14
Sequence Model Level GeneExperiment Expression R(t 1 ) R(t 2 )
Exp. type Sequence Assumptions: Binding site is of length k Binding
may occur at any k-mer TF regulates gene if binding occurs anywhere
Exp. cluster
Slide 15
From Sequence to Regulation u Assumptions: Binding site is of
length k Binding may occur at any k-mer TF regulates gene if
binding occurs anywhere u PSSM: Background distribution Motif
distribution Discriminative training where
Slide 16
From Sequence to Regulation u Model for one gene g, promoter
region of length 5 and k=2 S1S1 S3S3 S2S2 S4S4 S5S5 sequence
residues g.R(t) variable for t regulates g m[1].B m[2].B m[3].B
m[4].B k-mer binding events Logistic function motif model
DNA Localization Assay Swi5 Gene Bound Gene Not Bound TF binds
to targets u Induce TF protein level
Slide 21
Localization Assay DNA u Measure TF binding to promoter of
every gene Assign confidence for each binding Swi5 Gene Bound Gene
Not Bound TF binds to targets u Induce TF protein level
Slide 22
Localization Assay Simon et al (2001) u Localization data:
measure TF binding to promoter of each gene (assign binding
confidence)
Slide 23
Is Regulation Observed? u Not quite u Localization is measured
for specific conditions u Localization is measured for large DNA
regions u Localization is noisy
Slide 24
Incorporating Localization Level GeneExperiment Expression R(t
1 ) R(t 2 ) Exp. type Exp. Cluster L(t 1 ) L(t 2 ) Observed
localization u 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
Slide 25
Gene R(t 1 ) L(t 1 ) Localization Model u 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
Observed
Learning the Models ACGCCTAACGCCTA Experimental Details L E A R
N E R Level Gene R(t 1 ) R(t 2 ) Ehase ster Clu s1s1 sksk B(t 1
)B(t 2 ) Localization Data Exp. Phase = IV R(t 1 ) true false true
R(t 1 ) = Yes false R(t 2 ) = Yes true false truefalse R(t 1 ) R(t
2 ) E Phase 0 0 I 0.8 1.2 0 1 II -0.7 0.6
Slide 28
Learning the Models u Ndd1 activates Ace2 and Swi5 in G1, which
together activate in S u Mcm1 activates the DNA repair pathway in S
ACGCCTAACGCCTA Experimental Details L E A R N E R Level Gene R(t 1
) R(t 2 ) Ehase ster Clu s1s1 sksk B(t 1 )B(t 2 ) Localization
Data
Slide 29
Model Learning u Structure Learning: Tree structure u Missing
Data: Experiment cluster Regulation variables u Motif Model:
Parameter estimation u Expectation Maximization u Bayesian score u
Heuristic search u Discriminative training (conjugate
gradient)
Slide 30
Model Learning Gene Expression R(t 2 ) R(t 1 ) Experiment Exp.
type Level + Experimental Details Localization Data ACGCCTAACGCCTA
promoter s1s1 sksk Exp. cluster L(t 1 )
Generalization Level Gene Expression L(t 1 ) L(t 2 ) Experiment
Exp. type Gene log-likelihood u Clustering genes -112.24 u
Localization -121.48 -112.24
Slide 37
Generalization Level Gene Expression R(t 1 ) R(t 2 ) Experiment
Exp. type Exp. Cluster L(t 1 ) L(t 3 ) Gene log-likelihood u
Clustering genes -112.24 u Localization -121.48 u Localization +
exp. cluster -103.76 -112.24
Slide 38
Generalization Level Gene Expression R(t 1 ) R(t 2 ) promoter
s1s1 sksk Experiment Exp. type Exp. Cluster L(t 1 ) L(t 3 ) Gene
log-likelihood u Clustering genes -112.24 u Localization -121.48 u
Localization + exp. cluster -103.76 u + Sequence -94.59
-112.24
Slide 39
Generating Hypotheses Example: Genes regulated by Swi6, not by
Mcm1 and not by Fkh2, exhibit unique expression pattern in phase G1
in the cell cycle Gene functions: DNA repair [P 3e-09] DNA
synthesis [P 7e-05]
Slide 40
Expression vs Regulation
02142638410510701001301601902202500306090120150090180270360 alpha
cdc15cdc28elu -0.5 0 0.5 1 Phase Swi5 regulated Swi5 expression
Genes predicted to be regulated by Swi5 are probably real Swi5
targets
Induced Interaction Network u TF pairs whose regulation
predicts expression of same gene cluster Ace2 Swi5 Ndd1 Fkh2 Fkh1
Swi4 Swi6 Mcm1 Mbp1 G1 S G2 M M/G1 M G1 G2 S
Slide 47
Conclusions u Unified probabilistic model explaining gene
regulation using sequence, localization and expression data u
Models complex interactions between regulators u Discriminative
model maximizing P(Expr. | Seq.) u Sequence data helps explain
expression patterns
Slide 48
Big Picture u Goal: unified probabilistic framework Models
complex biological domains Incorporates heterogeneous data u
Framework incorporates explicitly within model basic biological
building blocks: Genes, TFs, proteins, patients, cells, species, u
Much closer connection between biology and model Can read biology
directly from model Can incorporate prior knowledge easily u Can
explicitly represent and learn biological models