Motivation
Design
Data and Trial
Simulation
Conclusion
Sub-group C luster-Based Adaptive Designs for
Precision Medicine (S C UBA )
Yuan Ji, PhD
Program for Computational Genomics & Medicine
NorthShore University HealthSystem
Department of Public Health Sciences
The University of Chicago
5.8.2015
6/4/13
Warwick Exp. Design and Big Data 2015
upload.wikimedia.org/wikipedia/en/9/93/University_of_Chicago_logo.svg
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
1 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
SUBA and SCUBA
SUBA design (Xu et al. 2014, Statistics in Biosciences)
SCUBA design (Guo, Catenacci, and Ji, 2015. To be submitted)
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
2 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
One-size Fit All Cancer Treatment
Lung
Cancer
Leukemia
Breast
Cancer
Warwick Exp. Design and Big Data 2015
Surgery
Chemo
Radiation
Surgery
Chemo
Radiation
Surgery
Chemo
Radiation
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
3 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
One-size Fit All Cancer Treatment – 2
The problem of one-size fit all:
Lung
Cancer
Leukemia
Surgery
Chemo
Radiation
Surgery
Chemo
Radiation
• Treat the “phenotypes”
with brute force
• Severe side effects and
poor quality of life
• Unpredictable prognosis
Breast
Cancer
Warwick Exp. Design and Big Data 2015
Surgery
Chemo
Radiation
• Risk of over-treatment
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
4 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Targeted Cancer Treatment
EGFR
Lung
Cancer
Erlotinib
BCR-ABL
Leukemia
Imabtinib
Breast
Cancer
Trastuzumab
HER2
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
5 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Targeted Cancer Treatment – 2
The benefits of being on target:
• Treat the “genotypes”
EGFR
Lung
Cancer
Erlotinib
that are causal of
phenotypes
• Typically mild side
BCR-ABL
Leukemia
Imabtinib
effects and high quality
of life
• Predictable prognosis
Breast
Cancer
Trastuzumab
HER2
Warwick Exp. Design and Big Data 2015
• Less chance of
over-treatment
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
6 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Precision Cancer Care
Disease agnostic; Genotype specific (e.g., NCI MATCH trial)
EGFR
Lung
Cancer
BCR-ABL
HER2
Leukemia
Breast
Cancer
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
7 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Patients with NO actionable genotypes
EGFR
BCR-ABL
Lung
Cancer
HER2
Investigational
gene X
Lung
Cancer
Warwick Exp. Design and Big Data 2015
?
gene Y
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
8 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
When to use SCUBA
Which treatment is the best depends on status of biomarkers X
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
9 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
When to use SCUBA
Which treatment is the best depends on status of biomarkers X
A hypothetical example
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
1.0
prob
0.0
0.2
0.4
0.6
0.8
1.0
0.0
0.2
0.4
prob
0.6
0.8
1.0
0.0
0.2
0.4
prob
0.6
0.8
1.0
0.0
0.2
0.4
prob
0.6
0.8
1.0
0.8
-1.0
2nd Bmkr=
-0.4
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
0.4
0.0
0.0 0.5 1.0
1.0
1st bmker
2nd Bmkr=
0.3
1.0
1st bmker
2nd Bmkr=
0.2
1.0
1st bmker
2nd Bmkr=
0.1
1.0
1st bmker
2nd Bmkr=
0
1.0
1st bmker
2nd Bmkr=
-0.1
1.0
1st bmker
-1.0
0.0 0.5 1.0
-1.0
0.0 0.5 1.0
1st bmker
2nd Bmkr=
1
-1.0
0.0 0.5 1.0
1st bmker
Warwick Exp. Design and Big Data 2015
-1.0
0.0 0.5 1.0
1st bmker
-1.0
0.0 0.5 1.0
1st bmker
-1.0
0.0 0.5 1.0
1st bmker
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
prob
0.0
0.2
0.4
0.8
0.6
0.4
0.2
0.0 0.5 1.0
1st bmker
1.0
1st bmker
2nd Bmkr=
0.9
1.0
1st bmker
2nd Bmkr=
0.8
1.0
1st bmker
2nd Bmkr=
0.7
1.0
1st bmker
2nd Bmkr=
0.6
1.0
1st bmker
2nd Bmkr=
0.5
1.0
1st bmker
2nd Bmkr=
0.4
0.0
-1.0
0.6
prob
0.2
0.0
0.0 0.5 1.0
2nd Bmkr=
-0.5
2nd Bmkr=
-0.2
1.0
-1.0
0.4
0.8
0.6
prob
0.2
0.0
-1.0
2nd Bmkr=
-0.6
1st bmker
0.2
prob
0.4
0.8
0.6
prob
0.4
0.2
0.0
0.0 0.5 1.0
2nd Bmkr=
-0.7
2nd Bmkr=
-0.3
1.0
-1.0
prob
2nd Bmkr=
-0.8
1.0
2nd Bmkr=
-0.9
1.0
2nd Bmkr=
-1
-1.0
0.0 0.5 1.0
1st bmker
-1.0
0.0 0.5 1.0
1st bmker
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
9 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Trial Setup for SCUBA
• Patients without known targeted drugs (e.g., relapsed patients out
of options)
• A set of relevant biomarkers (or PCs) X = (X1 , X2 , ...Xp ), p small
• A set of candidate drugs (t1 , t2 , ..., tT ), T ≥ 1.
Goal: find a rule that allocates patient subgroup Sk (X) to drug tk , such
that the response rate under the rule is better than standard strategy ,
such as treating ALL patients with drug tk or randomization between
different drugs (in a trial) .
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
10 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Overview of SCUBA
Run-­‐in Phase Equal Randomiza.on Treatment 1 Treatment 2 Treatment T Record (X, Y) for run-­‐in Warwick Exp. Design and Big Data 2015
E N R O L L P A T I E N T B
i
o
p
s
y G
e
n
o
t
y
i
n
g X Random Subgroups S1 Response S2 Treatment t* Y . . . B
i
o
p
s
y G
e
n
o
t
y
i
n
g X . . . E N R O L L P A T I E N T SUBA Phase Adap.ve Op.mal Treatment Alloca.on SK Adap.ve Subgroup Learning Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
11 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Overview of Probability Model and Inference
Random par//ons π1 π2 S1 π3 S2 π4 Prior Pr(π=πi) Posterior predicted for new pa/ent Pr(Ynew = 1 | Y, trt) =
Given par//on π1 = (S1, S2) ∑ Pr(Y
new
Op/mal Decision = 1 | π , trt) p(π | Y )
π
Pr(Y=1 |S1 ,tj ) Pr(Y=1|S2 ,tj ) Warwick Exp. Design and Big Data 2015
Likelihood p(Y|π,trt) Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
12 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
CART-type model for binary outcomes
Consider binary outcome yi ∈ {0, 1} where 0 and 1 denotes no response
and response.
[Π]
Let Π = (B1 , B1 , . . . , BM ) be a random partition of X = Rk ;
[θ | Π] :
iid
θj,m | Π ∼ Beta(a, b)
j = 1, 2, 3, m = 1, . . . , M
[Y |X, t, Π, θ] :
Yi | Xi , ti = j, Π, θ ∼ Bernoulli(θj,mXi ),
mXi = (m : Xi ∈ Bm )
A simple random partition P (Π)
is constructed by randomly selecting one biomarker and partition the
patient groups into half according to the median.
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
13 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Prior on Π
X2
P
Choose one Xi with probability pi ( i pi = 1), and with probability q to
split the space by I(Xi > median(Xi )). Do the same for the new subset
if the split does occur. Repeat 3 times.
X1
Prior probability p2 q × p2 qp1 q × p21 q 2 (1 − p1 − p2 )2
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
14 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Subgroup-Based Trial Design
Let N be the total sample size. For patient i, let xi be the biomarker
profile, ti be the treatment allocation, and yi be the response outcome.
1 An initial run-in with an equal randomization of n ≤ N patients.
3 Compute for patient n + 1,
qn+1 (t) = P r(yn+1 = 1 | yn , xn , tn , xn+1 , tn+1 = t) =
Z
P r(yn+1 = 1 | xn+1 , tn+1 = t, π, θ)p(π, θ | yn , xn , tn )dθ
.
4 Allocate patient n + 1 to treatment t∗ = arg maxt qn+1 (t).
5 Update the observed data as (yn+1 , xn+1 , tn+1 ), and repeat steps
2-4 for patient n + 2, n + 3, ...N .
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
15 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Subgroup-Based Trial Design
Let N be the total sample size. For patient i, let xi be the biomarker
profile, ti be the treatment allocation, and yi be the response outcome.
1 An initial run-in with an equal randomization of n ≤ N patients.
Q
2 Fit a Bayesian model i p(yi | xi = x, ti = t, π, θ) · p(π)p(θ) to
the data of n patients from step 1, denoted as (yn , xn , tn ).
3 Compute for patient n + 1,
qn+1 (t) = P r(yn+1 = 1 | yn , xn , tn , xn+1 , tn+1 = t) =
Z
P r(yn+1 = 1 | xn+1 , tn+1 = t, π, θ)p(π, θ | yn , xn , tn )dθ
.
4 Allocate patient n + 1 to treatment t∗ = arg maxt qn+1 (t).
5 Update the observed data as (yn+1 , xn+1 , tn+1 ), and repeat steps
2-4 for patient n + 2, n + 3, ...N .
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
15 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Subgroup-Based Trial Design
Let N be the total sample size. For patient i, let xi be the biomarker
profile, ti be the treatment allocation, and yi be the response outcome.
1 An initial run-in with an equal randomization of n ≤ N patients.
Q
2 Fit a Bayesian model i p(yi | xi = x, ti = t, π, θ) · p(π)p(θ) to
the data of n patients from step 1, denoted as (yn , xn , tn ).
3 Compute for patient n + 1,
qn+1 (t) = P r(yn+1 = 1 | yn , xn , tn , xn+1 , tn+1 = t) =
Z
P r(yn+1 = 1 | xn+1 , tn+1 = t, π, θ)p(π, θ | yn , xn , tn )dθ
.
4 Allocate patient n + 1 to treatment t∗ = arg maxt qn+1 (t).
5 Update the observed data as (yn+1 , xn+1 , tn+1 ), and repeat steps
2-4 for patient n + 2, n + 3, ...N .
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
15 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Subgroup-Based Trial Design
Let N be the total sample size. For patient i, let xi be the biomarker
profile, ti be the treatment allocation, and yi be the response outcome.
1 An initial run-in with an equal randomization of n ≤ N patients.
Q
2 Fit a Bayesian model i p(yi | xi = x, ti = t, π, θ) · p(π)p(θ) to
the data of n patients from step 1, denoted as (yn , xn , tn ).
3 Compute for patient n + 1,
qn+1 (t) = P r(yn+1 = 1 | yn , xn , tn , xn+1 , tn+1 = t) =
Z
P r(yn+1 = 1 | xn+1 , tn+1 = t, π, θ)p(π, θ | yn , xn , tn )dθ
.
4 Allocate patient n + 1 to treatment t∗ = arg maxt qn+1 (t).
5 Update the observed data as (yn+1 , xn+1 , tn+1 ), and repeat steps
2-4 for patient n + 2, n + 3, ...N .
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
15 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Subgroup-Based Trial Design
Let N be the total sample size. For patient i, let xi be the biomarker
profile, ti be the treatment allocation, and yi be the response outcome.
1 An initial run-in with an equal randomization of n ≤ N patients.
Q
2 Fit a Bayesian model i p(yi | xi = x, ti = t, π, θ) · p(π)p(θ) to
the data of n patients from step 1, denoted as (yn , xn , tn ).
3 Compute for patient n + 1,
qn+1 (t) = P r(yn+1 = 1 | yn , xn , tn , xn+1 , tn+1 = t) =
Z
P r(yn+1 = 1 | xn+1 , tn+1 = t, π, θ)p(π, θ | yn , xn , tn )dθ
.
4 Allocate patient n + 1 to treatment t∗ = arg maxt qn+1 (t).
5 Update the observed data as (yn+1 , xn+1 , tn+1 ), and repeat steps
2-4 for patient n + 2, n + 3, ...N .
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
15 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
A Breast Cancer Trial
• Patients eligible to the trial are
• have undergone neoadjuvant systemic therapy (NST) and surgery
• have their protein biomarkers measured (through biopsy samples) at
the end of NST but before surgery
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
16 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
A Breast Cancer Trial
• Patients eligible to the trial are
• have undergone neoadjuvant systemic therapy (NST) and surgery
• have their protein biomarkers measured (through biopsy samples) at
the end of NST but before surgery
• Three candidate treatments
• Poly (ADP-ribose) polymerase (PARP) inhibitor – DNA repair and
programmed cell death
• PI3K pathway inhibitor – cell growth, proliferation, differentiation,
motility, survival and intracellular trafficking
• Cell cycle inhibitor
• About 300 patients had expression measurements for a number of
proteins from MAPK and PI3K pathways.
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
16 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Simulation Setup
Basic setup
• Samp size N = 300, run-in phase n = 100 (equal randomization),
T = 3 treatment arms
• Six scenarios, 1,000 simulated trial per scenario.
• Compare to ER, AR (outcome adaptive), and probit-reg designs.
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
17 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Simulation Scenarios
-1.0
0.5
bmker
0.5
bmker
-1.0
0.5
-1.0
0.5
-1.0
0.5
2nd
-1.0
-1.0
0.5
-1.0
0.5
2nd
1.0
0.5
1.0
bmker
Bmkr=
0.3
0.0
1st
Bmkr=
0.9
bmker
0.0
1st
2nd
1.0
bmker
0.0
1st
0.0 0.2 0.4 0.6 0.8 1.0
prob
0.0
1st
1.0
1.0
prob
-1.0
Bmkr=
0.8
bmker
0.5
bmker
Bmkr=
0.2
0.0 0.2 0.4 0.6 0.8 1.0
1.0
bmker
0.0
1st
0.0
1st
2nd
0.0 0.2 0.4 0.6 0.8 1.0
-1.0
prob
2nd
0.5
2nd Bmkr=
-0.4
0.0 0.2 0.4 0.6 0.8 1.0
prob
0.0
1st
1.0
1.0
prob
-1.0
Bmkr=
0.7
bmker
0.5
bmker
Bmkr=
0.1
0.5
1.0
bmker
Bmkr=
1
0.0 0.2 0.4 0.6 0.8 1.0
1.0
bmker
0.0
1st
0.0
1st
2nd
prob
2nd
0.5
2nd Bmkr=
-0.5
0.0 0.2 0.4 0.6 0.8 1.0
prob
0.0
1st
1.0
-1.0
prob
-1.0
Bmkr=
0.6
bmker
1.0
0.0 0.2 0.4 0.6 0.8 1.0
1.0
bmker
0.0
1st
0.5
bmker
Bmkr=
0
prob
2nd
1.0
0.5
0.0
1st
2nd
0.0 0.2 0.4 0.6 0.8 1.0
-1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.0
1st
2nd Bmkr=
-0.6
0.0 0.2 0.4 0.6 0.8 1.0
1.0
prob
-1.0
Bmkr=
0.5
0.0
1st
0.5
bmker
0.0 0.2 0.4 0.6 0.8 1.0
1.0
bmker
prob
-1.0
0.0
1st
2nd Bmkr=
-0.1
prob
2nd
1.0
0.5
2nd Bmkr=
-0.7
prob
-1.0
prob
0.0
1st
Bmkr=
0.4
0.0
1st
1.0
0.0 0.2 0.4 0.6 0.8 1.0
-1.0
prob
prob
1.0
bmker
0.0 0.2 0.4 0.6 0.8 1.0
2nd
Sc 1-2
0.5
0.5
bmker
0.0 0.2 0.4 0.6 0.8 1.0
prob
prob
0.0
1st
0.0
1st
2nd Bmkr=
-0.2
0.0 0.2 0.4 0.6 0.8 1.0
-1.0
2nd Bmkr=
-0.8
prob
-1.0
0.0 0.2 0.4 0.6 0.8 1.0
1.0
0.0 0.2 0.4 0.6 0.8 1.0
0.5
bmker
0.0 0.2 0.4 0.6 0.8 1.0
0.0
1st
2nd Bmkr=
-0.3
0.0 0.2 0.4 0.6 0.8 1.0
2nd Bmkr=
-0.9
prob
-1.0
0.0 0.2 0.4 0.6 0.8 1.0
prob
Bmkr=
-1
0.0 0.2 0.4 0.6 0.8 1.0
2nd
-1.0
0.0
1st
0.5
1.0
bmker
Sc 3
Sc 4-5
Sc 6
all treatments are the same regardless of X.
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
18 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Sc
1
2
3
4
5
6
S*
/
S1
S3
S1
S2
S3
S1
S2
S1
S2
/
1
66.76
33.49
33.27
19.49
25.23
22.05
33.26
33.50
33.26
33.50
66.76
ER
2
66.60
33.09
33.51
19.09
25.17
22.34
33.11
33.49
33.11
33.49
66.60
3
66.64
33.24
33.40
19.29
25.35
22.00
33.44
33.20
33.44
33.20
66.64
1
83.02
33.37
33.41
22.21
21.13
24.61
43.01
42.32
39.14
38.29
66.66
AR
2
65.35
33.19
33.25
17.63
26.81
20.52
42.32
43.46
38.49
39.32
66.89
3
51.63
33.25
33.53
18.03
27.80
21.26
14.49
14.41
22.19
22.58
66.46
1
119.46
35.24
35.42
18.65
24.10
21.27
51.81
51.75
51.51
51.22
65.04
Reg
2
70.13
32.88
33.01
16.40
21.86
18.99
48.00
48.44
48.25
48.92
67.84
3
10.41
31.69
31.76
22.81
29.79
26.12
0
0
0.05
0.05
67.12
1
177.11
72.57
8.63
41.11
13.67
11.33
52.76
50.78
51.13
47.07
66.90
SUBA
2
18.67
18.37
17.79
8.94
35.91
11.54
46.96
49.29
47.05
51.53
64.20
3
4.22
8.88
73.77
7.82
26.17
43.52
0.10
0.11
1.63
1.59
68.90
*: St is the subset of of biomarker space X in which the t-th treatment
has the highest response rate.
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
19 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Define the overall response rate (ORR) as the proportion of responders among
those patients who are treated after the run-in phase
ORR =
N
X
1
I(yi = 1),
N − n i=n+1
• Plot ORR(SUBA) - ORR(design): Sc 1 Sc 2 Sc 3 Sc 4 Sc 5 Sc 6
0.0 0.2 0.4 0.6
−0.4
Difference in ORR
SUBA versus ER
200
400
600
800
1000
800
1000
800
1000
0.0 0.2 0.4 0.6
SUBA1:1000
versus AR
−0.4
Difference in ORR
0
200
400
600
0.0 0.2 0.4 0.6
SUBA 1:1000
versus Reg
−0.4
Difference in ORR
0
0
200
Warwick Exp. Design and Big Data 2015
400
600
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
20 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Define the overall response rate (ORR) as the proportion of responders among
those patients who are treated after the run-in phase
ORR =
N
X
1
I(yi = 1),
N − n i=n+1
• Plot ORR(SUBA) - ORR(design): Sc 1 Sc 2 Sc 3 Sc 4 Sc 5 Sc 6
0.0 0.1 0.2 0.3 0.4
−0.2
Difference in ORR
SUBA versus ER
200
400
600
800
1000
800
1000
800
1000
0.0 0.1 0.2 0.3 0.4
SUBA1:1000
versus AR
−0.2
Difference in ORR
0
200
400
600
0.0 0.1 0.2 0.3 0.4
SUBA 1:1000
versus Reg
−0.2
Difference in ORR
0
0
200
Warwick Exp. Design and Big Data 2015
400
600
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
20 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Define the overall response rate (ORR) as the proportion of responders among
those patients who are treated after the run-in phase
ORR =
N
X
1
I(yi = 1),
N − n i=n+1
• Plot ORR(SUBA) - ORR(design): Sc 1 Sc 2 Sc 3 Sc 4 Sc 5 Sc 6
0.0 0.1 0.2 0.3 0.4
−0.2
Difference in ORR
SUBA versus ER
200
400
600
800
1000
800
1000
800
1000
0.0 0.1 0.2 0.3 0.4
SUBA1:1000
versus AR
−0.2
Difference in ORR
0
200
400
600
0.0 0.1 0.2 0.3 0.4
SUBA 1:1000
versus Reg
−0.2
Difference in ORR
0
0
200
Warwick Exp. Design and Big Data 2015
400
600
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
20 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Define the overall response rate (ORR) as the proportion of responders among
those patients who are treated after the run-in phase
ORR =
N
X
1
I(yi = 1),
N − n i=n+1
• Plot ORR(SUBA) - ORR(design): Sc 1 Sc 2 Sc 3 Sc 4 Sc 5 Sc 6
0.0 0.1 0.2 0.3 0.4
−0.2
Difference in ORR
SUBA versus ER
200
400
600
800
1000
800
1000
800
1000
0.0 0.1 0.2 0.3 0.4
SUBA1:1000
versus AR
−0.2
Difference in ORR
0
200
400
600
0.0 0.1 0.2 0.3 0.4
SUBA 1:1000
versus Reg
−0.2
Difference in ORR
0
0
200
Warwick Exp. Design and Big Data 2015
400
600
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
20 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Define the overall response rate (ORR) as the proportion of responders among
those patients who are treated after the run-in phase
ORR =
N
X
1
I(yi = 1),
N − n i=n+1
• Plot ORR(SUBA) - ORR(design): Sc 1 Sc 2 Sc 3 Sc 4 Sc 5 Sc 6
0.0 0.1 0.2 0.3 0.4
−0.2
Difference in ORR
SUBA versus ER
200
400
600
800
1000
800
1000
800
1000
0.0 0.1 0.2 0.3 0.4
SUBA1:1000
versus AR
−0.2
Difference in ORR
0
200
400
600
0.0 0.1 0.2 0.3 0.4
SUBA 1:1000
versus Reg
−0.2
Difference in ORR
0
0
200
Warwick Exp. Design and Big Data 2015
400
600
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
20 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Operating Characteristics – All scenarios
Define the overall response rate (ORR) as the proportion of responders among
those patients who are treated after the run-in phase
ORR =
N
X
1
I(yi = 1),
N − n i=n+1
• Plot ORR(SUBA) - ORR(design): Sc 1 Sc 2 Sc 3 Sc 4 Sc 5 Sc 6
0.1
−0.2 −0.1 0.0
Difference in ORR
0.2
SUBA versus ER
0
200
800
1000
800
1000
800
1000
0.2
0.1
−0.2 −0.1 0.0
Difference in ORR
600
SUBA1:1000
versus AR
0
200
400
600
0.1
0.2
SUBA 1:1000
versus Reg
−0.2 −0.1 0.0
Difference in ORR
400
0
200
Warwick Exp. Design and Big Data 2015
400
600
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
20 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Bayesian nonparametric modeling for Clustering (SCUBA)
Model extension A nonparatric Bayesian model using Dirichlet process
priors
Flexible boundaries Allowing a varying number of boundaries
Precision medicine Report subgroup-treatment pairs for confirmatory
studies
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
21 / 23
Motivation
Design
Data and Trial
Simulation
Truth
Conclusion
SCUBA estimate
biomarker 1
Significant subgroups for scenario 7
1.0
1.0
0.5
0.5
0.0
0.0
-0.5
-0.5
-1.0
-1.0
-1.0
-0.5
0.0
0.5
1.0
biomarker 1
biomarker 2
Significant subgroups for scenario 6
1.0
1.0
0.5
0.5
0.0
0.0
-0.5
-0.5
-1.0
-1.0
-1.0
-0.5
0.0
0.5
1.0
biomarker 2
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
22 / 23
Motivation
Design
Data and Trial
Simulation
Conclusion
Conclusions
SCUBA is about precision medicine and targeted therapy.
• Precision medicine: Response to treatment (its order) is assumed to
depend on X – biomarkers.
• Adaptive learning based on Bayesian hierarchical models
• Subgroup-treatment pair report with confidence – multiple
confirmatory trials for targeted drugs/companion diagnositics
Warwick Exp. Design and Big Data 2015
Sub-group C luster-Based Adaptive Designs for Precision Medicine (S C UBA )
23 / 23