Semi-Parametric Methods Applied to Modeling

Optimal Design and Model Selection

Davide Verotta, UCSF
Purposes
Describe Modern Approach to Semi-parametric Modeling
Bayesian Model Selection
Reversible Jump Markov Chain Monte Carlo
Describe Applications of Semi-Parametric Modeling to
PK
PK/PD
Population Modeling
Experimental Design
Semi-Parametric Models
Polynomial spline:

i=1
d

i
t
i1
+
i=1
m

i+d
(t
i
)
+
d1

i
B
i,
( )

1

2
t
y
Bayesian Model Selection

{M
k
}
Define a family of possible models
Prior probability distributions to the parameters of each
model
Prior probability to each model
This prior formulation induces a joint distribution

p(
k
|M
k
)

p(M
k
)

p(Y,
k
,M
k
) =p(Y|
k
,M
k
) p(
k
|M
k
) p(M
k
)
Bayesian Posteriors
Conditioning on the data Y yields the posterior model
probabilities
Comparison between a pair models is summarized by
the posterior odds

p(M
k
|Y) =
p(Y|M
k
)p(M
k
)
p(Y|M
k
)p(M
k
)
k

p(M
k
|Y)
p(M
j
|Y)
=
p(Y|M
k
)
p(Y|M
k
)

p(M
k
)
p(M
j
)
Markov Chain Monte Carlo
=(likelihood ratio) (parameters prior ratio) (proposal ratio)
=
p(y|
*
)
p(y|)

p(
*
)
p()

J( |
*
)
J(
*
| )
Metropolis-Hastings: given the current state , a new value

is
drawn from a chosen proposal density J(,). This is accepted with
probability
Reversible Jump Markov Chain Monte Carlo
=(likelihood ratio) (parameters prior ratio) (proposal ratio) (model prior ratio)
=
p(y|M*,
m*
)
p(y|M,
m
)

p(
m*
|M*)
p(
m
|M)

J
m*,m
(u*| M*, M,
m*
)
J
m,m*
(u | M, M*,
m
)

p(M*)
p(M)
Starting in state M propose a new model M* with probability J
m,m*
and determine the proposed models parameters. Accept the new
model with probability:
Reversible Jump MCMC: Flow Chart
Deconvolution
y(t ) = I (s)K(t s)ds
0
t

Disposition : y(t ) = (s)K(t s)ds

0
t

= K(t)
y(t ) =
i

i
B
i,
k
( )
0
t

K(t )d +
Kang, D. & Verotta, D. Reversible Jump Markov chain Monte Carlo
for Deconvolution. J Pharmacokin Pharmacodyn In Press (2006).
Deconvolution:
Model Selection
Problem
Deconvolution:
Confidence
Bands
Pharmacokinetics and Pharmacodynamics
Drug Specific Body Specific
PK
PD
PK/PD Models
Stimulation of production rate K
in
:

dS
dT
= K
in
( 1+ E
max
C
C + C
50
) K
out
S
Stimulation of elimination rate K
out
:

dS
dT
= K
in
K
out
( 1+ E
max
C
C + C
50
)S
Inhibition of production rate K
in
:

dS
dT
= K
in
( 1 I
max
C
C + C
50
) K
out
S
Inhibition of elimination rate K
out
:

dS
dT
= K
in
K
out
( 1 I
max
C
C + C
50
)S

Jusko, W.J. and H.C. Ko, Physiologic indirect response models characterize diverse types
of pharmacodynamic effects. Clin Pharmacol Ther, 1994. 56: p. 406-19.
PK/PD: Generalized

A:
dR
dt
= k
p
1+
E
max
Ce

EC
50

+Ce

[
\
|

)
j
k
d
R
B :
dR
dt
= k
p
1
I
max
Ce

EC
50

+Ce

[
\
|

)
j
k
d
R
|
|
|
|
|
|
|
E :
dR
dt
= h
p
Ce
( )
k
d
R
C :
dR
dt
= k
p
k
d
1+
E
max
Ce

EC
50

+Ce

[
\
|

)
j
R
D :
dR
dt
= k
p
k
d
1
I
max
Ce

EC
50

+Ce

[
\
|

)
j
R
|
|
|
|
|
|
|
F :
dR
dt
= k
p
h
d
(Ce)R
|

|
|
|
|
|
|
|
|
|
|
|
|
|
G:
dR
dt
= h R,Ce
( )
PK/PD: Semi-Parametric

M
k
:
dR
dt
= k
p
i

i
B
i ,
k
(Ce) k
d
R
M
k
:
dR
dt
= k
p
k
d
i

i
B
i ,
k
(Ce)R
M
k
:
dR
dt
= k
p
i

i
B
i ,
k
1
(Ce) k
d
i

i
B
i ,
k
2
(Ce)R
M
k
:
dR
dt
= k
p
i

i
B
i ,
k
1
(Ce
1
,Ce
2
) k
d
i

i
B
i ,
k
2
(Ce
1
,Ce
2
)R
PK/PD - Production Rate
PK/PD - Elimination Rate
PK/PD - Mixed Action
PK/PD - Interaction
Model Selection: Population Covariates

M
k
:Y = X
k

k
+
M
k
:Y = f t,
k
( )
+

k
= X
k

k
For a linear model, covariates
For a non-linear model
For example

dC
dt
V = CL C , C(0)=
Dose
V
y(t) = C(t) +
CL =
i

i
B
i ,
k
(Weight)
Model Selection: Population Covariates Example

Y = Dose / V e
Cl /Vt
Cl =
1
+
2
CrCl

5
V =
3
+
4
Weight

5
Weight, Height, Age, CrCl log normal
Corr =
1
3
0 0

3
1
2
0
0
2
1
0 0 1

Model Selection: Population Covariates Results

Experimental Design: PK
* Drover, D., et al., Input characteristics and bioavailability after administration of immediate and a new
extended relese formulation of hydromorphone in healthy volunteers,. Anesthesiology, 2002. 97(4): p. 827-
36.
Experimental Design: Optimal Sampling

q
(min)
Number of
elementary
designs (

q
)
Number of
sampling
times

n
q
for
each

q

Number of
subjects
assigned to
each

q

Parametric Semi-parametric
1: Original
design*
24 30 0 (1), 3, 5, 7, 10, 11,
12, 13, 15, 17, 20, 25,
40, 55, 70, 100, 130,
190, 250, 370, 490,
730, 970, 1450

1 6 120 0, 10,50,125,220,1440 0, 15,50,130,240,1440
6 60 0,10,30,40,60,1440 0, 15,55,130,240,1440 2
4 90 0,135,205,1440 0, 50, 240,1440
4 60 0,10,30,40,65,1440 0, 20, 240,1440
4 60 0,135,205,1440
0, 55, 240,1440
3
4 60 0,25,30,110,1440
0, 130, 240,1440

Experimental Design: Optimal Sampling
Relative efficiency (%) Standard Error for

Cl Number of
elementary
designs
(

q
)
Parametric Semi-
parametric
Parametric

Cl =V k
10
( )

Semi-parametric

Cl = Dose / AUC
( )

1: Original
design*
127 122 0.41 0.40
1 111 103 0.33 0.30
2 106 100 0.31 0.30
3 109 100 0.35 0.30

Summary
Semi-parametric modeling is intertwined with model selection
RJMCMC (or equivalent MCMC):
- Solves model selection problem
- Incorporates prior knowledge
- Provides statistics of interest
RJMCMC Semi-parametric models can be applied to
- PK
- PK/PD
- Population Modeling
- Experimental design
Software
PK. Availble in the near future from http://code.dkang.net
PK/PD: None available, in development
Population Modeling. Using BUGS http://www.mrc-
bsu.cam.ac.uk/bugs/ and in particular JUMP
Experimental design. The algorithm used to compute the
experimental design is derived from PFIM, see also Retout, S. &
Mentre, F. Optimization of Individual and Population Designs
using Splus. J Pharmacokin Pharmacodyn 30, 417-443 (2003).
Acknowledgements
Many thanks to Dongwoo Kang and Chantal Csajka, who, during
their postdoctoral fellowships in my Lab, provided much insight,
and a lot of number crunching, for the deconvolution and model
selection projects described in this presentation.