Oliver Stegle and Karsten Borgwardt - ETH Z
Transcript of Oliver Stegle and Karsten Borgwardt - ETH Z
Oliver Stegle and Karsten Borgwardt: Computational Approaches for Analysing Complex Biological Systems, Page 1
GP ApplicationsOliver Stegle and Karsten Borgwardt
Machine Learning andComputational Biology Research Group,
Max Planck Institute for Biological Cybernetics andMax Planck Institute for Developmental Biology, Tübingen
Outline
Outline
O. Stegle & K. Borgwardt GP Applications Tubingen 1
Application1: modelling physiological time series
Outline
Application1: modelling physiological time seriesOverviewGaussian process prior for heart rateResults
Application 2: differential gene expressionOverviewA Gaussian process two-sample testExperimental Results on ArabidopsisDetecting Temporal Patterns of Differential Expression
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Summary
O. Stegle & K. Borgwardt GP Applications Tubingen 2
Application1: modelling physiological time series
Motivation
I Human heart rate is an important physiological trait.
I Measurement over long periods only viable with poor sensors.
I Motivation Gaussian process model for heart rate.
O. Stegle & K. Borgwardt GP Applications Tubingen 3
Application1: modelling physiological time series
Motivation
I Human heart rate is an important physiological trait.
I Measurement over long periods only viable with poor sensors.
I Motivation Gaussian process model for heart rate.
O. Stegle & K. Borgwardt GP Applications Tubingen 3
Application1: modelling physiological time series
The problemDataset
4 days of heart data
0 1000 2000 3000 4000 5000 6000 70000
50
100
150
200
250
300
t/min
hr/
bm
p
Features
I Different noise sources
I Two time scales, 24hrhythms
I Asymmetric around themean
I Auxiliary variablesindicative of noise
O. Stegle & K. Borgwardt GP Applications Tubingen 4
Application1: modelling physiological time series
The problemDataset
4 days of heart data
50
100
150
200
HR
/ b
pm
0
1
∆HRmin
0
1
∆HRmax
1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 60000
1
t
Tlost
Features
I Different noise sources
I Two time scales, 24hrhythms
I Asymmetric around themean
I Auxiliary variablesindicative of noise
O. Stegle & K. Borgwardt GP Applications Tubingen 4
Application1: modelling physiological time series
The problemDataset
zoomed view
2100 2200 2300 2400 2500 2600
20
40
60
80
100
120
t/min
hr/
bm
p
Features
I Different noise sources
I Two time scales, 24hrhythms
I Asymmetric around themean
I Auxiliary variablesindicative of noise
O. Stegle & K. Borgwardt GP Applications Tubingen 5
Application1: modelling physiological time series Overview
modelThe Inference Model
B
C
A
z1noisyclean very noisy
noiseaux. data
y1
f1
time step 1
z2noisyclean very noisy
noiseaux. data
y2
f2
time step 2
znnoisyclean very noisy
noiseaux. data
yn
fn
time step n
...
...
...
I (a) Gaussian process prior on latent heart rate
I (b) Clustering of auxiliary data to extract noise classes
I (c) Heavy tailed noise model taking classes into account
O. Stegle & K. Borgwardt GP Applications Tubingen 6
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Covarianc function for shortrange fluctuations
I Long-range perdiodic signal
I Sum of (a) and (b)
I Log transformation (asymmetry)
0 50 100 150
−200
−100
0
100
200
t / min
HR
/ b
pm
(a)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(b)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(c)
0 1000 2000 3000 40000
50
100
150
200
250
t / min
HR
/ b
pm
(d)
O. Stegle & K. Borgwardt GP Applications Tubingen 7
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Covarianc function for shortrange fluctuations
I Long-range perdiodic signal
I Sum of (a) and (b)
I Log transformation (asymmetry)
0 50 100 150
−200
−100
0
100
200
t / min
HR
/ b
pm
(a)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(b)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(c)
0 1000 2000 3000 40000
50
100
150
200
250
t / min
HR
/ b
pm
(d)
O. Stegle & K. Borgwardt GP Applications Tubingen 7
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Covarianc function for shortrange fluctuations
I Long-range perdiodic signal
I Sum of (a) and (b)
I Log transformation (asymmetry)
0 50 100 150
−200
−100
0
100
200
t / min
HR
/ b
pm
(a)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(b)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(c)
0 1000 2000 3000 40000
50
100
150
200
250
t / min
HR
/ b
pm
(d)
O. Stegle & K. Borgwardt GP Applications Tubingen 7
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Covarianc function for shortrange fluctuations
I Long-range perdiodic signal
I Sum of (a) and (b)
I Log transformation (asymmetry)
0 50 100 150
−200
−100
0
100
200
t / min
HR
/ b
pm
(a)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(b)
0 1000 2000 3000 4000
−200
−100
0
100
200
t / min
HR
/ b
pm
(c)
0 1000 2000 3000 40000
50
100
150
200
250
t / min
HR
/ b
pm
(d)
O. Stegle & K. Borgwardt GP Applications Tubingen 7
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Short-range covariance
CS(x, x′) = C1 Matern3/2(x, x
′, δS)
I Periodic covariance
CL(x, x′) = C0 exp
[− (x− x′)2
2δ2L
]exp
[−2 ·
sin2( 2πpL · (x− x′))
A2L
]I Total covariance
C(x, x′) = CL(x, x′) + CS(x, x
′)
I Non-linear transformation, reflecting strict positivity and asymmetryof heart rate
y∗t = logβ(yt)
O. Stegle & K. Borgwardt GP Applications Tubingen 8
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Short-range covariance
CS(x, x′) = C1 Matern3/2(x, x
′, δS)
I Periodic covariance
CL(x, x′) = C0 exp
[− (x− x′)2
2δ2L
]exp
[−2 ·
sin2( 2πpL · (x− x′))
A2L
]I Total covariance
C(x, x′) = CL(x, x′) + CS(x, x
′)
I Non-linear transformation, reflecting strict positivity and asymmetryof heart rate
y∗t = logβ(yt)
O. Stegle & K. Borgwardt GP Applications Tubingen 8
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Short-range covariance
CS(x, x′) = C1 Matern3/2(x, x
′, δS)
I Periodic covariance
CL(x, x′) = C0 exp
[− (x− x′)2
2δ2L
]exp
[−2 ·
sin2( 2πpL · (x− x′))
A2L
]I Total covariance
C(x, x′) = CL(x, x′) + CS(x, x
′)
I Non-linear transformation, reflecting strict positivity and asymmetryof heart rate
y∗t = logβ(yt)
O. Stegle & K. Borgwardt GP Applications Tubingen 8
Application1: modelling physiological time series Gaussian process prior for heart rate
A GP prior for heart rate
I Short-range covariance
CS(x, x′) = C1 Matern3/2(x, x
′, δS)
I Periodic covariance
CL(x, x′) = C0 exp
[− (x− x′)2
2δ2L
]exp
[−2 ·
sin2( 2πpL · (x− x′))
A2L
]I Total covariance
C(x, x′) = CL(x, x′) + CS(x, x
′)
I Non-linear transformation, reflecting strict positivity and asymmetryof heart rate
y∗t = logβ(yt)
O. Stegle & K. Borgwardt GP Applications Tubingen 8
Application1: modelling physiological time series Gaussian process prior for heart rate
Clustering of auxiliary data
I Every datapoint can be member in one of K clusters.
I Use clustering approach to determine cluster membership
πn = {πn,1, . . . , πn,K} where
K∑k=1
πn,k = 1
O. Stegle & K. Borgwardt GP Applications Tubingen 9
Application1: modelling physiological time series Gaussian process prior for heart rate
Clustering of auxiliary data
I Every datapoint can be member in one of K clusters.
I Use clustering approach to determine cluster membership
πn = {πn,1, . . . , πn,K} where
K∑k=1
πn,k = 1
O. Stegle & K. Borgwardt GP Applications Tubingen 9
Application1: modelling physiological time series Gaussian process prior for heart rate
Robust noise model
I Remember: standard robust noise model, accounting for outliers
p(yn | fn) = πokN(yn∣∣ fn, σ2)+ (1− πok)N
(yn∣∣ fn, σ2∞)
I Here: clustering results (auxiliaries S) encode useful information.
z1
y1|f1 S1
z2
y2|f2 S2
zN
yN |fN SN
π
I Generalize likelihood K noise components, one for each cluster anduse data-specific mixture probabilities πn
p(yn | fn) =K∑k=1
πk,nN(yn∣∣ fn, σ2k)
O. Stegle & K. Borgwardt GP Applications Tubingen 10
Application1: modelling physiological time series Gaussian process prior for heart rate
Robust noise model
I Remember: standard robust noise model, accounting for outliers
p(yn | fn) = πokN(yn∣∣ fn, σ2)+ (1− πok)N
(yn∣∣ fn, σ2∞)
I Here: clustering results (auxiliaries S) encode useful information.
z1
y1|f1 S1
z2
y2|f2 S2
zN
yN |fN SN
π
I Generalize likelihood K noise components, one for each cluster anduse data-specific mixture probabilities πn
p(yn | fn) =K∑k=1
πk,nN(yn∣∣ fn, σ2k)
O. Stegle & K. Borgwardt GP Applications Tubingen 10
Application1: modelling physiological time series Gaussian process prior for heart rate
Robust noise model
I Remember: standard robust noise model, accounting for outliers
p(yn | fn) = πokN(yn∣∣ fn, σ2)+ (1− πok)N
(yn∣∣ fn, σ2∞)
I Here: clustering results (auxiliaries S) encode useful information.
z1
y1|f1 S1
z2
y2|f2 S2
zN
yN |fN SN
π
I Generalize likelihood K noise components, one for each cluster anduse data-specific mixture probabilities πn
p(yn | fn) =K∑k=1
πk,nN(yn∣∣ fn, σ2k)
O. Stegle & K. Borgwardt GP Applications Tubingen 10
Application1: modelling physiological time series Results
Regression for Heart Data
Single data set
I Clusteringcolor-coded andHinton diagrams
I Three clusters- g,b,r-valuesfor responsi-bilities
I Noise parameters{σc}3c=1
optimised alongwith other hyperparameters
O. Stegle & K. Borgwardt GP Applications Tubingen 11
Application1: modelling physiological time series Results
Regression for Heart Data
Double data-set
I Two sensors forone heart
I GPs overlap wellwithin error-bars
I Lower panel:difference plotand error bars
O. Stegle & K. Borgwardt GP Applications Tubingen 12
Application1: modelling physiological time series Results
Fill-in testMotivation
I Evaluate predictive performance to benchmark alternative modelsI Fill-in test:
I Model is trained on a subset of the data to predict the remainder.I Log probability as criteria rewards models with appropriately sized error
bars.
O. Stegle & K. Borgwardt GP Applications Tubingen 13
Application1: modelling physiological time series Results
Block size 1 minute
Block size 1 minute
0.1 0.2 0.3 0.4 0.5 0.6−1
0
1
2
3
Fraction of missing data
<ln
(P(d
ata
)>
(i)(ii)(iii)
(iv)
(v)
I 5 different models comparedI (i) baselineI (ii) GP with ksI (iii) GP with ks + klI (iv) as (iii) but with Tlost
threshold noise modelI (v) as (iv) but with full robust
noise model
O. Stegle & K. Borgwardt GP Applications Tubingen 14
Application1: modelling physiological time series Results
Block size 60 minute
Block size 60 minutes
0.1 0.2 0.3 0.4 0.5 0.6−1
−0.5
0
0.5
Fraction of missing data
<ln
(P(d
ata
)>
(i)
(ii)
(iii)
(iv)
(v)
I 5 different models comparedI (i) baselineI (ii) GP with ksI (iii) GP with ks + klI (iv) as (iii) but with Tlost
threshold noise modelI (v) as (iv) but with full robust
noise model
I Larger blocks removed rewardslong range covariance
O. Stegle & K. Borgwardt GP Applications Tubingen 15
Outline
Outline
O. Stegle & K. Borgwardt GP Applications Tubingen 16
Application 2: differential gene expression
Outline
Application1: modelling physiological time seriesOverviewGaussian process prior for heart rateResults
Application 2: differential gene expressionOverviewA Gaussian process two-sample testExperimental Results on ArabidopsisDetecting Temporal Patterns of Differential Expression
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Summary
O. Stegle & K. Borgwardt GP Applications Tubingen 17
Application 2: differential gene expression Overview
Response to External Stimuli
I Organisms such as plants respond toexternal stimuli:
I Heat/ColdI StarvationI Biotic stresses (fungus)I ...
I Changes in observed gene expressionlevels.
O. Stegle & K. Borgwardt GP Applications Tubingen 18
Application 2: differential gene expression Overview
Response to External Stimuli
I Organisms such as plants respond toexternal stimuli:
I Heat/ColdI StarvationI Biotic stresses (fungus)I ...
I Changes in observed gene expressionlevels.
O. Stegle & K. Borgwardt GP Applications Tubingen 18
Application 2: differential gene expression Overview
Response to External Stimuli
I Organisms such as plants respond toexternal stimuli:
I Heat/ColdI StarvationI Biotic stresses (fungus)I ...
I Changes in observed gene expressionlevels.
O. Stegle & K. Borgwardt GP Applications Tubingen 18
Application 2: differential gene expression Overview
Response to External Stimuli
I Organisms such as plants respond toexternal stimuli:
I Heat/ColdI StarvationI Biotic stresses (fungus)I ...
I Changes in observed gene expressionlevels.
O. Stegle & K. Borgwardt GP Applications Tubingen 18
Application 2: differential gene expression Overview
Differential Gene Expression
I An important goal is the identificationof differentially expressed genes.
I Identification of involved regulatorycomponents.
I Uncovering parts of the biologicalnetwork.
O. Stegle & K. Borgwardt GP Applications Tubingen 19
Application 2: differential gene expression Overview
Differential Gene Expression
I An important goal is the identificationof differentially expressed genes.
I Identification of involved regulatorycomponents.
I Uncovering parts of the biologicalnetwork.
O. Stegle & K. Borgwardt GP Applications Tubingen 19
Application 2: differential gene expression Overview
Differential Gene Expression in Time SeriesTime Series
I The response to external stimuli is a dynamic process.
I Hence the response should be studied as a function of time.
O. Stegle & K. Borgwardt GP Applications Tubingen 20
Application 2: differential gene expression Overview
Differential Gene Expression in Time SeriesTime Series
I The response to external stimuli is a dynamic process.
I Hence the response should be studied as a function of time.
10 20 30 40Time/h
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
Log
expr
essi
onle
vel
TreatmentControl
Non-differentially expressed
10 20 30 40Time/h
−1
0
1
2
3
4
5
6
7
Log
expr
essi
onle
vel
TreatmentControl
Differentially expressed
O. Stegle & K. Borgwardt GP Applications Tubingen 20
Application 2: differential gene expression Overview
Differential Gene Expression in Time SeriesChallenges
I Time series expression profiles vary smoothly over time.
I Noisy observations – outliers.
I Multiple replicates.
I Few observations.
I Temporal patterns (intervals) of differential gene expression.
10 20 30 40Time/h
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
5.5
Log
expr
essi
onle
vel
TreatmentControl
Non-differentially expressed
10 20 30 40Time/h
−1
0
1
2
3
4
5
6
7
Log
expr
essi
onle
vel
TreatmentControl
Differentially expressed
O. Stegle & K. Borgwardt GP Applications Tubingen 21
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process ModelModel comparison
I The basic idea – a comparison of two models:I The shared model: Expression levels are explained by a single process.I The independent model: Expression levels are explained by two
separate processes.
O. Stegle & K. Borgwardt GP Applications Tubingen 22
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process ModelModel comparison
I The basic idea – a comparison of two models:I The shared model: Expression levels are explained by a single process.I The independent model: Expression levels are explained by two
separate processes.
O. Stegle & K. Borgwardt GP Applications Tubingen 22
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process ModelModel comparison
I The basic idea – a comparison of two models:I The shared model: Expression levels are explained by a single process.I The independent model: Expression levels are explained by two
separate processes.
O. Stegle & K. Borgwardt GP Applications Tubingen 22
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process modelBayesian Network: Shared Model
I Data in conditions A and Bobserved at N time points withR replicates.
I A Gaussian process priorincorporates beliefs aboutsmoothness.
I Noise is is modeled separatelyper-replicate, σA/Br .
O. Stegle & K. Borgwardt GP Applications Tubingen 23
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process modelBayesian Network: Shared Model
I Data in conditions A and Bobserved at N time points withR replicates.
I A Gaussian process priorincorporates beliefs aboutsmoothness.
I Noise is is modeled separatelyper-replicate, σA/Br .
O. Stegle & K. Borgwardt GP Applications Tubingen 23
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process modelBayesian Network: Shared Model
I Data in conditions A and Bobserved at N time points withR replicates.
I A Gaussian process priorincorporates beliefs aboutsmoothness.
I Noise is is modeled separatelyper-replicate, σA/Br .
O. Stegle & K. Borgwardt GP Applications Tubingen 23
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian process modelBayesian Network: Both Models
I The independent model follows in an analogous manner.
Shared model HS Independent model HI
O. Stegle & K. Borgwardt GP Applications Tubingen 24
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelInference
I Models are compared using the Bayes factor
Score = log
Independent model︷ ︸︸ ︷P (DA,DB |HI)
P (DA,DB |HS)︸ ︷︷ ︸Shared model
.
I Writing out the GP models explicitly leads to
Score = logP (YA |HGP,T
A, )P (YB |HGP,TB, )
P (YA ∪YB |HGP,TA ∪TB, ).
(YA/B : expression levels in conditions A and B; TA/B : observation time points)
O. Stegle & K. Borgwardt GP Applications Tubingen 25
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelInference
I Models are compared using the Bayes factor
Score = log
Independent model︷ ︸︸ ︷P (DA,DB |HI)
P (DA,DB |HS)︸ ︷︷ ︸Shared model
.
I Writing out the GP models explicitly leads to
Score = logP (YA |HGP,T
A,θI)P (YB |HGP,T
B,θI)
P (YA ∪YB |HGP,TA ∪TB,θS).
(YA/B : expression levels in conditions A and B; TA/B : observation time points)
O. Stegle & K. Borgwardt GP Applications Tubingen 25
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelInference
I Models are compared using the Bayes factor
Score = log
Independent model︷ ︸︸ ︷P (DA,DB |HI)
P (DA,DB |HS)︸ ︷︷ ︸Shared model
.
I Writing out the GP models explicitly leads to
Score = logP (YA |HGP,T
A, θI )P (YB |HGP,T
B, θI )
P (YA ∪YB |HGP,TA ∪TB, θS ).
(YA/B : expression levels in conditions A and B; TA/B : observation time points)
O. Stegle & K. Borgwardt GP Applications Tubingen 25
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelShared Model
I Given observed data from both conditions D = {YA/B,TA/B} theposterior distribution over latent function values f is
P (f |Y,T,θK,θL) ∝N (f |0,KT(θK))
×∏
c∈{A,B}
R∏r=1
N∏n=1
pL(ycr,tn | ftn ,θL),
I Covariance funciton (kernel)
I Noise model
I Hyperparameters θS = {θK,θL} (length scale, noise levels)
I For Gaussian noise, pL(ycr,t | f cr,t,θL) = N
(ycr,t
∣∣∣ f cr,t, (σcr)2), the
model is tractable in closed form.
O. Stegle & K. Borgwardt GP Applications Tubingen 26
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelShared Model
I Given observed data from both conditions D = {YA/B,TA/B} theposterior distribution over latent function values f is
P (f |Y,T,θK,θL) ∝N(f∣∣∣0, KT(θK)
)×
∏c∈{A,B}
R∏r=1
N∏n=1
pL(ycr,tn | ftn ,θL),
I Covariance funciton (kernel)
I Noise model
I Hyperparameters θS = {θK,θL} (length scale, noise levels)
I For Gaussian noise, pL(ycr,t | f cr,t,θL) = N
(ycr,t
∣∣∣ f cr,t, (σcr)2), the
model is tractable in closed form.
O. Stegle & K. Borgwardt GP Applications Tubingen 26
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelShared Model
I Given observed data from both conditions D = {YA/B,TA/B} theposterior distribution over latent function values f is
P (f |Y,T,θK,θL) ∝N(f∣∣∣0, KT(θK)
)×
∏c∈{A,B}
R∏r=1
N∏n=1
pL(ycr,tn | ftn ,θL) ,
I Covariance funciton (kernel)
I Noise model
I Hyperparameters θS = {θK,θL} (length scale, noise levels)
I For Gaussian noise, pL(ycr,t | f cr,t,θL) = N
(ycr,t
∣∣∣ f cr,t, (σcr)2), the
model is tractable in closed form.
O. Stegle & K. Borgwardt GP Applications Tubingen 26
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelShared Model
I Given observed data from both conditions D = {YA/B,TA/B} theposterior distribution over latent function values f is
P (f |Y,T, θK,θL ) ∝N(f∣∣∣0, KT(θK)
)×
∏c∈{A,B}
R∏r=1
N∏n=1
pL(ycr,tn | ftn ,θL) ,
I Covariance funciton (kernel)
I Noise model
I Hyperparameters θS = {θK,θL} (length scale, noise levels)
I For Gaussian noise, pL(ycr,t | f cr,t,θL) = N
(ycr,t
∣∣∣ f cr,t, (σcr)2), the
model is tractable in closed form.
O. Stegle & K. Borgwardt GP Applications Tubingen 26
Application 2: differential gene expression A Gaussian process two-sample test
Gaussian Process ModelShared Model
I Given observed data from both conditions D = {YA/B,TA/B} theposterior distribution over latent function values f is
P (f |Y,T, θK,θL ) ∝N(f∣∣∣0, KT(θK)
)×
∏c∈{A,B}
R∏r=1
N∏n=1
pL(ycr,tn | ftn ,θL) ,
I Covariance funciton (kernel)
I Noise model
I Hyperparameters θS = {θK,θL} (length scale, noise levels)
I For Gaussian noise, pL(ycr,t | f cr,t,θL) = N
(ycr,t
∣∣∣ f cr,t, (σcr)2), the
model is tractable in closed form.
O. Stegle & K. Borgwardt GP Applications Tubingen 26
Application 2: differential gene expression A Gaussian process two-sample test
Robustness With Respect to Outliers
I Outliers in the expression profilecan obscure the regressionresults.
I A mixture noise-model accountsfor outliers.
I Inference in this model is doneusing Expectation Propagation.
O. Stegle & K. Borgwardt GP Applications Tubingen 27
Application 2: differential gene expression A Gaussian process two-sample test
Robustness With Respect to Outliers
I Outliers in the expression profilecan obscure the regressionresults.
I A mixture noise-model accountsfor outliers.
6 4 2 0 2 4 60.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
I Inference in this model is doneusing Expectation Propagation.
O. Stegle & K. Borgwardt GP Applications Tubingen 27
Application 2: differential gene expression A Gaussian process two-sample test
Robustness With Respect to Outliers
I Outliers in the expression profilecan obscure the regressionresults.
I A mixture noise-model accountsfor outliers.
6 4 2 0 2 4 60.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
I Inference in this model is doneusing Expectation Propagation.
O. Stegle & K. Borgwardt GP Applications Tubingen 27
Application 2: differential gene expression Experimental Results on Arabidopsis
Illustration of the Model ComparisonA Differentially Expressed Gene
I Shared model.
O. Stegle & K. Borgwardt GP Applications Tubingen 28
Application 2: differential gene expression Experimental Results on Arabidopsis
Illustration of the Model ComparisonA Differentially Expressed Gene
I Independent model.
O. Stegle & K. Borgwardt GP Applications Tubingen 28
Application 2: differential gene expression Experimental Results on Arabidopsis
Illustration of the Model ComparisonA Differentially Expressed Gene
I Model comparison.
O. Stegle & K. Borgwardt GP Applications Tubingen 28
Application 2: differential gene expression Experimental Results on Arabidopsis
Predictive Performance (RECOMB09)
I Data:I 30,336 Arabidopsis thaliana gene probesI Biotic stress: fungus infectionI 24 time points, 4 biological replicates
I Evaluation of alternative methods on 2000 randomly chosenhuman-labeled probes:
I GP no robustI GP robustI F-Test (FT) (MAANOVA package)I Timecourse (TC) (Tai and Speed)
O. Stegle & K. Borgwardt GP Applications Tubingen 29
Application 2: differential gene expression Experimental Results on Arabidopsis
Predictive Performance (RECOMB09)
I Data:I 30,336 Arabidopsis thaliana gene probesI Biotic stress: fungus infectionI 24 time points, 4 biological replicates
I Evaluation of alternative methods on 2000 randomly chosenhuman-labeled probes:
I GP no robustI GP robustI F-Test (FT) (MAANOVA package)I Timecourse (TC) (Tai and Speed)
O. Stegle & K. Borgwardt GP Applications Tubingen 29
Application 2: differential gene expression Experimental Results on Arabidopsis
Predictive Performance (RECOMB09)ROC Curves
0 0.2 0.4 0.6 0.8 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
False positive rate
Tru
e po
sitiv
e ra
te
GP robust (AUC 0.985)GP standard (AUC 0.944)FT (AUC 0.859)TC (AUC 0.869)
O. Stegle & K. Borgwardt GP Applications Tubingen 30
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleMotivation
I Differential expressionis not static over time.
I The responsedevelops over time.
I Different regulatorsmay be active atdistinct times andtrigger each other.
O. Stegle & K. Borgwardt GP Applications Tubingen 31
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleMotivation
I Differential expressionis not static over time.
I The responsedevelops over time.
I Different regulatorsmay be active atdistinct times andtrigger each other.
O. Stegle & K. Borgwardt GP Applications Tubingen 31
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleModel
I The shared andindependent model areinterleaved over time.
I Indicator variables ztnswitch between bothmodels.
I Inference is done usingGibbs sampling (later).
O. Stegle & K. Borgwardt GP Applications Tubingen 32
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleModel
I The shared andindependent model areinterleaved over time.
I Indicator variables ztnswitch between bothmodels.
I Inference is done usingGibbs sampling (later).
O. Stegle & K. Borgwardt GP Applications Tubingen 32
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleModel
I The shared andindependent model areinterleaved over time.
I Indicator variables ztnswitch between bothmodels.
I Inference is done usingGibbs sampling (later).
O. Stegle & K. Borgwardt GP Applications Tubingen 32
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleExample Results
The example from before
O. Stegle & K. Borgwardt GP Applications Tubingen 33
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Time-local GPTwoSampleExample Results
The example from before Periodic differential expression
O. Stegle & K. Borgwardt GP Applications Tubingen 33
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSample
I Transitions betweennon-differential anddifferential expressioncan occur at unobservedtime points and aresmooth.
I Extending the time-localmodel with a GP prior asgating network on theswitch variables.
I Inference using Gibbssampling.
O. Stegle & K. Borgwardt GP Applications Tubingen 34
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSample
I Transitions betweennon-differential anddifferential expressioncan occur at unobservedtime points and aresmooth.
I Extending the time-localmodel with a GP prior asgating network on theswitch variables.
I Inference using Gibbssampling.
O. Stegle & K. Borgwardt GP Applications Tubingen 34
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSampleGibbs Sampling
I Gibbs sampling exploits tractableconditional distributions.
I Individual indicators zti are resampledin turn
P(zti = s | z\ti ,T,Y,θS,θI,θG
)∼ P
(Y | zti = s, z\ti ,T,θI,θS
)× P
(zti = s | z\ti ,θG
).
I Data likelihood from GP experts
I Predictive distribution from gating network
O. Stegle & K. Borgwardt GP Applications Tubingen 35
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSampleGibbs Sampling
I Gibbs sampling exploits tractableconditional distributions.
I Individual indicators zti are resampledin turn
P(zti = s | z\ti ,T,Y,θS,θI,θG
)∼ P
(Y | zti = s, z\ti ,T,θI,θS
)× P
(zti = s | z\ti ,θG
).
I Data likelihood from GP experts
I Predictive distribution from gating network
O. Stegle & K. Borgwardt GP Applications Tubingen 35
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSampleGibbs Sampling
I Gibbs sampling exploits tractableconditional distributions.
I Individual indicators zti are resampledin turn
P(zti = s | z\ti ,T,Y,θS,θI,θG
)∼ P
(Y | zti = s, z\ti ,T,θI,θS
)× P
(zti = s | z\ti ,θG
).
I Data likelihood from GP experts
I Predictive distribution from gating network
O. Stegle & K. Borgwardt GP Applications Tubingen 35
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSampleGibbs Sampling
I Gibbs sampling exploits tractableconditional distributions.
I Individual indicators zti are resampledin turn
P(zti = s | z\ti ,T,Y,θS,θI,θG
)∼ P
(Y | zti = s, z\ti ,T,θI,θS
)× P
(zti = s | z\ti ,θG
).
I Data likelihood from GP experts
I Predictive distribution from gating network
O. Stegle & K. Borgwardt GP Applications Tubingen 35
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Smooth Time-local GPTwoSampleExample Results
Early differential expression Transient differential expression
O. Stegle & K. Borgwardt GP Applications Tubingen 36
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Detecting Transition Points in Arabidopsis Microarray Time SeriesStart/Stop Times
I Studying the temporaldistribution of differentialexpression across genes.
I Considered were the top 6000differentially expressed genes.
I Differential expression appearsto occur in two waves with starttimes at 20h an 25h afterinfection.
I Only very few genes stopdifferential behavior within themeasured time interval.
Distribution of differential start/stop time
O. Stegle & K. Borgwardt GP Applications Tubingen 37
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Detecting Transition Points in Arabidopsis Microarray Time SeriesStart/Stop Times
I Studying the temporaldistribution of differentialexpression across genes.
I Considered were the top 6000differentially expressed genes.
I Differential expression appearsto occur in two waves with starttimes at 20h an 25h afterinfection.
I Only very few genes stopdifferential behavior within themeasured time interval.
Distribution of differential start/stop time
O. Stegle & K. Borgwardt GP Applications Tubingen 37
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Detecting Transition Points in Arabidopsis Microarray Time SeriesStart/Stop Times
I Studying the temporaldistribution of differentialexpression across genes.
I Considered were the top 6000differentially expressed genes.
I Differential expression appearsto occur in two waves with starttimes at 20h an 25h afterinfection.
I Only very few genes stopdifferential behavior within themeasured time interval.
Distribution of differential start/stop time
O. Stegle & K. Borgwardt GP Applications Tubingen 37
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Detecting Transition Points in Arabidopsis Microarray Time SeriesStart Times for Gene Categories
I This distribution can bebroke down into genecategories.
I WRKY Family oftranscription factors isknown to be involved instress response.
Distribution of differential start time
O. Stegle & K. Borgwardt GP Applications Tubingen 38
Application 2: differential gene expression Detecting Temporal Patterns of Differential Expression
Detecting Transition Points in Arabidopsis Microarray Time SeriesStart Times for Gene Categories
I This distribution can bebroke down into genecategories.
I WRKY Family oftranscription factors isknown to be involved instress response.
Distribution of differential start time
O. Stegle & K. Borgwardt GP Applications Tubingen 38
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Outline
Application1: modelling physiological time seriesOverviewGaussian process prior for heart rateResults
Application 2: differential gene expressionOverviewA Gaussian process two-sample testExperimental Results on ArabidopsisDetecting Temporal Patterns of Differential Expression
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Summary
O. Stegle & K. Borgwardt GP Applications Tubingen 39
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Motivation
I This part of the course is inspired and based on results of apublication by Neil D. Lawrence et al.Modelling transcriptional regulation using Gaussian processesftp://ftp.dcs.shef.ac.uk/home/neil/gpsim.pdf.
O. Stegle & K. Borgwardt GP Applications Tubingen 40
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Motivation
I Microarray technologies allow tomeasure mRNA levels.
I The functional proteins and theirconcentration levels remain unobserved.
I Motivation: Infer the hidden proteinconcentrations?
O. Stegle & K. Borgwardt GP Applications Tubingen 41
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Motivation
I Microarray technologies allow tomeasure mRNA levels.
I The functional proteins and theirconcentration levels remain unobserved.
I Motivation: Infer the hidden proteinconcentrations?
O. Stegle & K. Borgwardt GP Applications Tubingen 41
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Motivation
I Microarray technologies allow tomeasure mRNA levels.
I The functional proteins and theirconcentration levels remain unobserved.
I Motivation: Infer the hidden proteinconcentrations?
O. Stegle & K. Borgwardt GP Applications Tubingen 41
Application 3: Modeling transcriptional regulation using Gaussianprocesses
A single gene model
I The change in gene expression abundance yi for a gene i isapproximately described by a differential equation model of the form
dyi(t)
dt= Bi + Sif(t)−Diyi(t)
I f(t) regulatory transcription factor.I Bi basal transcription rate.I Si sensitivity of the gene to the transcription factor.I Di decay rate of the mRNA.
I Goal: infer the unobserved activation f(t) from mRNA measurementsof multiple target genes.
O. Stegle & K. Borgwardt GP Applications Tubingen 42
Application 3: Modeling transcriptional regulation using Gaussianprocesses
A single gene model
I The change in gene expression abundance yi for a gene i isapproximately described by a differential equation model of the form
dyi(t)
dt= Bi + Sif(t)−Diyi(t)
I f(t) regulatory transcription factor.I Bi basal transcription rate.I Si sensitivity of the gene to the transcription factor.I Di decay rate of the mRNA.
I Goal: infer the unobserved activation f(t) from mRNA measurementsof multiple target genes.
O. Stegle & K. Borgwardt GP Applications Tubingen 42
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Derivative observations
I The key to solving this problem are derivative observations.
I Given knowledge about the derivative of a function f we would like toinfer its function values:
dy =∂f(t)
∂t
I We wish to find the joint probability of function values and functionderivatives
cov(dyi, yj) =∂
∂ticov(yi, yj)
cov(dyi, dyj) =∂2
∂ti∂tjcov(yi, yj)
I Using these covariance functions we can combine functionobservations and derivatives as training data in GP regression.
O. Stegle & K. Borgwardt GP Applications Tubingen 43
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Derivative observations
I The key to solving this problem are derivative observations.
I Given knowledge about the derivative of a function f we would like toinfer its function values:
dy =∂f(t)
∂t
I We wish to find the joint probability of function values and functionderivatives
cov(dyi, yj) =∂
∂ticov(yi, yj)
cov(dyi, dyj) =∂2
∂ti∂tjcov(yi, yj)
I Using these covariance functions we can combine functionobservations and derivatives as training data in GP regression.
O. Stegle & K. Borgwardt GP Applications Tubingen 43
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Derivative observationsSquared exponential kernel
I For the squared exponential kernel we obtain:
cov(yi, yj) = k(ti, tj) = A2e−0.5(ti−tj)
2
L2
cov(dyi, yj) = −cov(yi, yj)(ti − tj)L2
cov(dyi, dyj) = cov(yi, yj)1
L2
[δi,j −
1
L2(ti − tj)2
]
O. Stegle & K. Borgwardt GP Applications Tubingen 44
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Derivative observationsExample
(From E. Solak et al.
Derivative observations in Gaussian Process models of dynamic systems)
O. Stegle & K. Borgwardt GP Applications Tubingen 45
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Back to the ODE model for gene regulation
dyi(t)
dt= Bi + Sif(t)−Diyi(t)
I An explicit solution of the ODE system can be derived (standard ODEtechniques)
yi(t) =BiDi
+ kie−Dit + Sie
−Dit
∫ t
0
f(u) exp(Diu)du
yi(t) =BiDi
+ Li[f ](t)
I Realizing that Li is a linear operator (like taking the derivative), wecan again evaluate the covariance between f(t) and Li[f ](t).
O. Stegle & K. Borgwardt GP Applications Tubingen 46
Application 3: Modeling transcriptional regulation using Gaussianprocesses
ODE model for gene regulationInference results
(From N. D. Lawrence et al.
Modelling transcriptional regulation using Gaussian processes)
O. Stegle & K. Borgwardt GP Applications Tubingen 47
Summary
Outline
Application1: modelling physiological time seriesOverviewGaussian process prior for heart rateResults
Application 2: differential gene expressionOverviewA Gaussian process two-sample testExperimental Results on ArabidopsisDetecting Temporal Patterns of Differential Expression
Application 3: Modeling transcriptional regulation using Gaussianprocesses
Summary
O. Stegle & K. Borgwardt GP Applications Tubingen 48
Summary
Summary
I The design and choice of covariance functions allows for flexiblemodeling tasks.
I Prior on heart rateI Derivative observations, ODE systems
I Model comparison using Gaussian processes.I Testing for differential gene expression.
O. Stegle & K. Borgwardt GP Applications Tubingen 49